GEmFakeABSA: graph-embedded fake review detection using aspect-based sentiment analysis Authors: Ganpat Singh Chauhan, Atul Kumar Verma, Shikha Chaudhary, Geetanjali Maurya Published: 2026 ABSTRACT The bulk of customers and manufacturers who use e-business platforms depend largely on user comments when making purchase decisions and improving product quality. Unfortunately, this reliance also creates an opening for dishonest reviewers to deceive both consumers and producers by posting fake reviews. Addressing this problem directly supports the United Nations Sustainable Development Goal 12: ’Responsible Consumption and Production’ because reliable review ecosystems foster transparency and informed decision making in online markets. To mitigate this threat, the present article proposes GEmFakeABSA: Graph-Embedded Fake Review Detection using Aspect-Based Sentiment Analysis (ABSA). We further enhance fake review detection (FRD) by integrating ABSA with a Graph Convolutional Network (GCN) that leverages dependency graphs to capture contextual relationships among words, aspects, and sentiments, thereby exposing unnatural aspect-sentiment patterns in fabricated reviews. Because a GCN can also encode intricate relationships among users, reviews, and products, the resulting GCN-based ABSA model is highly effective when sentiment is not logically associated with specific aspects. Our proposed model combines linguistic information, part-of-speech (POS) features and ABSA-derived embeddings to improve every standard FRD metric in Amazon reviews spanning 30 product categories, even when each set of features is evaluated in isolation. Evaluated on 21,034 Amazon reviews across 30 product categories, GEmFakeABSA achieves 92.4% accuracy and surpasses the strongest BiLSTM-ABSA baseline by +2.3% precision, and +2.46% recall. All results are averaged over five random seeds with 95% confidence intervals; significance is assessed via two-sided paired t-tests across the 30 categories, and splits are stratified by product and reviewer to preclude train–test leakage. These results highlight the model’s potential for practical, real-time content moderation in responsible e-commerce. Introduction The Internet has grown immensely, and as a result, people commonly shop online. Thus, it encourages everyone to buy and sell via online marketplaces and to write product reviews. Online textual comments and ratings are essential in determining the accomplishments of e-businesses across various industries, including movies, hotels, restaurants, products, and YouTube contents, etc. People trust user-generated content more than advertisements, and positive reviews and comments serve as social proof, making a brand or person appear more trustworthy. These comments draw attention to the positive or negative aspects of online products, which influence opinions on what to buy next and how to enhance the product1. Ensuring the authenticity and reliability of online reviews directly contributes to achieving Sustainable Development Goal (SDG) 12: Responsible Consumption and Production. This SDG emphasizes the importance of transparency and informed consumer decision making to foster sustainable market practices. By effectively identifying and mitigating fraudulent online reviews, businesses and consumers can make better-informed decisions, ultimately promoting more responsible and sustainable patterns of consumption and production. However, because there are so many reviews, reading them all one at a time to comprehend the views expressed by consumers and the salient features of the product can be tedious and time-consuming. People rarely read the full review before assessing a product’s feature level descriptions, overall rating, or individual rating. Customers are therefore highly interested in the routine review of evaluations by sentiment analysis rather than manual reading. Emotion detection, another name of opinion mining, is a branch of natural language processing, computer linguistics, and text mining2. Sentiment analysis evaluates people’s views, assessments, feelings, emotions, and attitudes towards various topics3. Determining the polarity of the evaluation’s text is the main goal of the sentiment monitoring on consumer feedback4. Sentiment analysis is crucial for e-businesses because it helps glean insightful information from a vast amount of consumer reviews. According to5,6, it establishes the subjectivity, polarity, and degree of polarity in the sentimental article. Three methods exist for conducting evaluation of emotion at the phrase, documenting, or characteristic level7. Sentence-level opinion evaluation categorises a phrase as neutral, positive, or negative. Sentiment analysis is used to classify the complete record, which is made up of numerous neutral, positive, and negative expressions, on a document basis. Opinion Analysis, which determines the emotional tone behind text, has the limitation of a lack of context awareness, in which sentiment analysis often fails to understand the context in which the sentiment word is used. For example, in the review “The movie is insane”, the sentiment might be favorable or negative based on the situation. Traditional sentiment analysis gives an overall sentiment score but does not tell which feature of a good or service those feelings are about. Considered as the reviewer’s assessment of the product from the standpoint of its aspects, the ABSA computes the complete polarity belonging to the review by considering the pertinent product aspects of each line. Manufacturers can enhance the quality of their products by tracking client feedback on various product parameters through aspect-level sentiment analysis. ABSA is used to obtain the object aspects out of evaluations, subsequently ascertaining related thoughts of every aspect8. Compared to the previous phases of opinion analysis, ABSA can provide accurate sentiment on certain opinion goals9. ABSA requires a pair of aspect-oriented opinion determination and aspect extraction10. Various methodologies have been described in aspect-level sentiment analysis to address feature extraction and opinion identification; however, these approaches either fail to extract aspects in different domains or rely on labelled datasets. Depending on the domain, different features may potentially be retrieved. A GCN embedding-based method was used in this article to perform ABSA, in which features are extracted, and opinions are determined. Importance of implicit sentiment and figurative language In online reviews, users often express opinions through implicit sentiment and figurative language such as sarcasm, satire, irony, and self-deprecating sarcasm. These forms of expression create significant challenges for sentiment analysis because the intended meaning may differ from the literal meaning of the text. For instance, statements such as “Great battery life, it died within one hour” or “This phone is unbelievably good at hanging all the time” contain hidden negative sentiment despite the presence of positive words. Traditional sentiment analysis approaches frequently fail to capture these contextual and implicit emotional expressions accurately. Modern literature has revealed the significance of figurative language understanding and contextual semantic analysis for detecting implicit sentiments expressed through sarcasm and irony. Although11 reviewed computational methods for figurative language, whereas12,13 evaluated deep learning techniques for sarcasm and satire detection, still handling figurative language remains a significant challenge in fake review analysis. Therefore, understanding figurative language is essential for reliable aspect-based sentiment analysis and fake review detection because deceptive reviews may intentionally use exaggerated, sarcastic, or contradictory language patterns to manipulate customer opinions. Research gap While aspect-based and behavior/metadata driven approaches have each shown promise for online review analysis, prior work seldom unifies fine-grained ABSA signals with platform/behavioral cues under rigorous, leakage safe evaluation or demonstrates robustness across many product categories. This limits reliability when deploying models in real marketplaces. The benefits of using aspect-level sentiment analysis by analyzing sentiment at a more granularity level include identifying specific aspects, improving decision making, better handling of mixed sentiments, and enhancing consumer experience analysis. Instead of just saying a review is positive or negative, ABSA identifies what a sentiment is about the mentioned aspect. For instance, in the comment, “The camera is amazing, nonetheless battery life is terrible”, opinion about “Camera” is positive whereas about “battery life” is negative. Business can get incredible insights by knowing which aspects of their product or service need improvement. ABSA can analyze reviews that contain both positive and negative opinions about different aspects. It helps brands understand customer opinion in more detail, leading to more targeted improvements. Additionally, recent research further examined innovative transformer-based, multimodal, and contextual semantic modelling methods for opinions and information extraction applications, covering conversational ABSA, multimodal opinion extraction, and affective analysis of social media posts14–16. Document embeddings capture global semantic information for classification tasks17, while mixture-of-experts models improve performance by combining multiple specialized experts18. Moreover, ensemble-based fake news detection methods have incorporated information diffusion and network propagation embeddings to capture how fake news spreads across social networks19, while other studies leverage social network structure and user interaction patterns within ensemble frameworks to improve detection performance20. The above study shows the impact of ABSA on online reviews. The online consumer reviews offer customers valuable product insights, having a significant influence on buyer behavior and influencing retail sales. Thus, in fiercely competitive online marketplaces, there is a strong temptation to generate positive reviews for one’s own product or negative reviews for a competitor’s product. Recently, advanced language models are being used to create phony evaluations21. Fake reviews are misleading or deceptive customer reviews that are created to manipulate public perception of a product, service or business. These reviews can be positive (to boost a brand reputation) or negative (to harm a competitor). Fake reviews are integrated for several reasons. Reviewers frequently deceive readers by using polarity and enthusiasm5. The fake reviews can significantly distort sentiment analysis in several ways. Reviewers are considered dishonest if they often submit reviews with the intention of misleading readers. The overall impression that reviews create is impacted by these fraudulent reviews. These reviewers could fabricate positive or negative evaluations to damage a product’s reputation22. Nearly half of the buyers place an equal amount of faith in online reviews and recommendations from friends and family. Consumers who fall for promotional emails could have a negative opinion of the products. ABSA is thus provides information about the evaluation’s information about truthfulness or untruth5. Hence, ABSA is an effective and practical instrument that might be used to identify fraudulent review articles. Impact of aspect-based sentiment analysis on fake reviews Companies post fake positive reviews to attract more customers and improve their ratings, whereas some businesses create fake negative reviews to lower their rivals’ ratings and drive customers away. Many online platforms rank businesses based on reviews, and fake reviews can help manipulate algorithms to get better visibility. Some third-party agencies also sell fake reviews to businesses looking for an artificial boost. People often trust highly rated products and services, and a high number of positive reviews can create a false sense of credibility. The online platforms like Google, Amazon, Yelp, and many more utilize various AI- driven technologies to filter the fake reviews out. Removal of fake negative reviews from these platforms is a serious issue, as it can hamper the reputation of platform and can break the trust of customers. Sentiment Analysis models trained on manipulated data may develop biases because of over representation of certain words/phrases, which can mislead sentiment classification. Dominated fake negative reviews in the dataset may lead to inaccurate prediction of genuine customer sentiments. Also, fake reviews containing repetitive phrases/words or also interpretation makes complicated for models to distinguish between false and true data. Apart from this the marketing and product improvement decisions, which relies totally on sentiments analysis got affected. Further, consumers misled by the manipulated sentiment analysis may lose trust in brands and platforms. Fake review are serious issues, and platforms like Amazon, Google and Yelp have implemented AI driven tools to detect and remove them. However, they still exist and can mislead customers. Therefore, fake review detection with sentiment analysis plays an important part by guaranteeing and accuracy of sentiment classification. Sentiment analysis relies on genuine user opinions to access public opinions on products, services, or brands. However, the occurrence of fradulent comments even it is positive or negative, can distort sentiment analysis results, leading to misleading insights. In FRD the sentiment detection is crucial in the improvement of the above discussed issues as it improves accuracy, prevents biased sentiments, improves decision making, and boosting trust on online platforms. By integrating fake review detection with sentiment analysis pipelines, businesses and researchers can derive more reliable insights and make informed decisions. As ABSA goes deeper by analyzing specific aspects of a product, fake review poses an even greater challenge. Fake reviews make aspect extraction less reliable as it introduce unnatural and exaggerated language patterns which makes ABSA model fail to extract meaningful aspects properly. Lack of real user experiences in fake reviews makes the aspect-sentiment pair errorful leading to reduce the models ability to generalize. ABSA also considers multiple characteristics in a single review, the presence of false information about multiple features can cause more complicated erroneous results. ABSA’s reliance on proceedings and words makes it too exploitable for misleading speech while sentiment in SA is often each otherless informative, with only minor manipulations. It also becomes more challenging when the number of reviews on each aspect is small because it is easy for a fake review to be associated with that dataset, meaning that such conditions are limited with ABSA. False reviews also add noise, bias and manipulability which greatly affect the reliability of aspect extraction. Therefore, accurate aspect extraction ensures correct fake review detection, prevents aspect-sentiment manipulations by businesses or competitors, and enhances consumer trust and decision making. To better define the objectives and scope of the proposed framework, the following research questions are addressed in this study: RQ1: Does the integration of linguistic features, dependency parsing, word embeddings, POS taggings, and attention mechanisms improve contextual understanding and fake review detection performance? RQ2: How effectively can the proposed dependency-aware GCN and Bi-LSTM with hierarchical attention framework capture aspect-opinion relationships from textual reviews for fake review detection? RQ3: Can the proposed aspect-based sentiment analysis integrated for fake review detection outperform existing baseline approaches in identifying genuine and deceptive reviews on benchmark datasets? Our contributions The proposed GEmFakeABSA architecture incorporates linguistic features, dependency parsing, POS tagging, and word embeddings to enhance semantic understanding and contextual representation.A novel dependency-aware aspect-based sentiment analysis for a fake review detection framework integrating GCN, Bi-LSTM, and hierarchical attention mechanisms is proposed to effectively capture contextual and aspect-opinion relationships from textual reviews.Extensive quantitative and qualitative analyses are conducted to evaluate the effectiveness of the proposed framework.Experimental results demonstrate that the proposed approach achieves improved performance compared to baseline and state-of-the-art methods on the Amazon Reviews dataset.Novelty of the Proposed Work The proposed GEmFakeABSA model proposes a unified architecture that incorporates aspect-based sentiment analysis-based fake review detection. In contrast to the prior approaches that work in a standalone manner using deep learning or ABSA models, the proposed methodology first concatenated WE with POS tagging, which further integrates dependency parser-based GCN-embedding to capture structural contextual relationships across review words, followed by Bi-LSTM and attention mechanisms for contextual representation for complex review structure. Moreover, the model deliberately exploits the aspect–opinion pair retrieval to collectively model aspect terms and fake review characteristics, confirming effective fine-grained examination of reviews. This incorporation of ABSA with a graph-based and sequential neural framework presents an extended architecture for fake review detection. The proposed GEmFakeABSA architecture incorporates linguistic features, dependency parsing, POS tagging, and word embeddings to enhance semantic understanding and contextual representation. The proposed GEmFakeABSA architecture incorporates linguistic features, dependency parsing, POS tagging, and word embeddings to enhance semantic understanding and contextual representation. A novel dependency-aware aspect-based sentiment analysis for a fake review detection framework integrating GCN, Bi-LSTM, and hierarchical attention mechanisms is proposed to effectively capture contextual and aspect-opinion relationships from textual reviews. A novel dependency-aware aspect-based sentiment analysis for a fake review detection framework integrating GCN, Bi-LSTM, and hierarchical attention mechanisms is proposed to effectively capture contextual and aspect-opinion relationships from textual reviews. Extensive quantitative and qualitative analyses are conducted to evaluate the effectiveness of the proposed framework. Extensive quantitative and qualitative analyses are conducted to evaluate the effectiveness of the proposed framework. Experimental results demonstrate that the proposed approach achieves improved performance compared to baseline and state-of-the-art methods on the Amazon Reviews dataset. Experimental results demonstrate that the proposed approach achieves improved performance compared to baseline and state-of-the-art methods on the Amazon Reviews dataset. Paper roadmap Section Related work reviews related work; Section Proposed methodology details the methodology of the proposed GEmFakeABSA framework. Section Experimental setup and results reports quantitative and qualitative result analysis. The section also discusses strengths, limitations, and future scope. Section Conclusion presents the conclusion. Related work Although anyone can write a product review, some contributors post fake ones to mislead consumers or influence business decisions for financial gain23. Previous studies generally classified spam into three categories24,25: (1) class-1 deceptive reviews, that include overly-promotional and unfair negative content; (2)class-2 brand-only posts containing information about a brand reference but no specific mention of the target product; and (3) class-3 non-opinionated messages such as advertisement, Q&A type question or other irrelevant comments. Class-1 is typically the most harmful and hardest to detect because it mimics genuine opinion while subtly shifting sentiment. ML based approaches Surveys identify machine learning methods using content and/or author signals as one of the major families for detecting fake reviews26,27. ML settings include unsupervised, semi-supervised and supervised approaches which often utilize author behavior together with the textual indicators. For example, Bayesian clustering with a Latent Spam Model to measure author or review spam intensity28, thresholded scoring on weighted features29 and KNN-style unsupervised filtering based on review–reviewer–product triplets30. Another set of popular methods are temporal and singleton anomaly detection techniques31,32, as well as outlier detection in a more general sense33–35; trends on the metadata aspect (time or space) have also shown to be effective36. Pattern mining techniques find common or abnormal textual and temporal patterns, while pipelines built on rules depend on a fixed set of heuristics for spam detection. While rule-based heuristics are interpretable, they adapt poorly to evolving tactics; pattern mining scales better but can struggle with context and class imbalance. The choice of features strongly affects accuracy37. Content signals (lexical, stylistic), author behaviour, product context, and spam grouping are widely used22,25,38–40. Linguistic markers that differentiate genuine from deceptive include parts of speech distributions, readability and subjectivity profiles, and specific function/content word patterns36,39,41–43. A particularly robust cue is the rating–text inconsistency: sentiment expressed in text need not align with the star rating44–46. Prior work operationalizes this through rating deviation, content similarity, and reviewer activity to generate spam scores47. Although traditional ML-based methods achieved promising results in fake review detection, their reliance on handcrafted features and limited contextual understanding restricts their effectiveness. DL based approaches To overcome the limitations of traditional ML-based approaches, deep learning methods have gained significant attention in fake review detection due to their capability for automatic feature extraction and contextual representation learning. Conventional sentiment analysis yields an overall polarity but overlooks the attributes that drive the polarity. ABSA extracts aspects and their associated sentiments, enabling fine-grained opinion analysis that is highly useful for identifying deceptive patterns in fake review detection8,9. Early aspect extraction used association rules, PMI, language models, and pattern/bootstrapping strategies9,48–50; later, CRF/ME/DP/ontology driven methods improved precision but still faced domain dependence and context gaps46,51,52. Supervised and neural approaches increased accuracy53–56, yet long-range aspect–opinion dependencies remain challenging for sequence models. Deep learning based models improved sentiment classification and aspect extraction performance through neural architecture and contextual representations. Recent studies have explored ensemble learning, contextual sentiment analysis, topic modeling, and embedding-based approaches for analyzing opinions and textual semantics in social media and review datasets57–61. These methods can capture contextual sentiment and semantic features for fake review detection. However, they continue to struggle with implicit sentiment analysis, figurative language/semantic ambiguity, and sarcasm/irony/nice fake review written by a powerful language model. Recent work has shown large language model output to be fluent, templated and effective at obfuscating stylistic signals from fake reviews62. For such cases, both aspect-sentiment inconsistency and rating-text mismatch are still important signals to detect deceitful reviews. Though they have pushed the boundaries of fake review detection and sentimental classification, artificial deception generated by powerful LMs still present major challenges to conventional deep learning techniques. Sequential models like LSTM and CNN have shown effectiveness in many natural language processing subtasks, but they face several limitations when applied to FRD, particularly on complex datasets like those from Amazon products. These models primarily focus on local textual features and sequential dependencies with individual reviews. These models struggle with long-range dependencies, as their context is inherently limited by their sequential nature (even with attention mechanisms). However, FRD often requires understanding interconnected linguistic patterns within a review, patterns embedded in its syntactic structure, which traditional models fail to capture effectively. They also struggle with sparse metadata, reviewer behaviour, and contextual relationships between entities. Moreover, fake reviews can be linguistically similar to genuine ones, making it difficult for text-only models to distinguish them. Despite significant advancements, existing deep learning and ABSA-based methods still face challenges in capturing complex semantic dependencies and detecting sophisticated AI-generated deceptive reviews. Graph based approaches To further enhance contextual and relational representation learning in fake review detection, graph-based approaches have recently gained significant attention. These methods model relationships among words, reviews, users, and aspects as graph structures, enabling more effective capture of semantic dependencies and contextual interactions within review data. Network based methods exploit user–review–product relations and capture coordinated behaviour through graph structures63,64. Graph encoders (e.g., GCNs) represent tokens as nodes and syntactic dependencies as edges, allowing models to bridge long range ties between aspect terms and sentiment expressions65,66. Compared with purely sequential LSTM/CNN models, dependency-aware encoders better handle multi-clause sentences, multiple aspects per sentence, and non local cues. When combined with attention and pretrained embeddings, graph based ABSA achieves strong results across datasets67,68. These properties are attractive for fraud settings, where fabricated reviews often include exaggerated or templated sentiment that is weakly grounded in aspect details. Current deceptive reviews identification techniques maintain context-aware embeddings, transformer architectures, and attention-based networks to strengthen contextual semantic learning and accurate fake review classification69–72. In addition, some latest approaches evaluated on modern deep learning and representation learning methodologies for complex semantic analysis and classification tasks73–75. Recent studies have explored various transformer variants for sentiment and fake review detection, including standard transformers76, fine-tuned transformers for domain adaptation77, architectures with disentangled attention and enhanced mask decoders for improved contextual modeling78, sentence transformer models for semantic embeddings79, as well as adaptive and language-based mixture-of-transformer frameworks for improved robustness and flexibility [8,9]. Aspect–sentiment grounding and rating–text inconsistencies remain informative in such settings; our ABSA-graph + behaviour fusion explicitly captures both. Such research efforts provide further evidence for the incorporation of a contextual and aspect-aware contextual semantic framework in the proposed ABSA-based GEmFakeABSA model. Role of ABSA in fake review detection and motivation of the proposed work Although existing fake review detection approaches have shown promising performance, many methods primarily rely on overall sentiment and general linguistic patterns. However, fake reviews often contain deceptive aspect-level opinion patterns, making Aspect-Based Sentiment Analysis (ABSA) important for fine-grained sentiment analysis and improved fake review detection. Fake reviews often have generic sentiments without specific details about product aspects. ABSA can analyze sentiment for different aspects and flag reviews where sentiments are overly broad or do not match specific product attributes. Fake comments are generally extremely positive (for promotion) or extremely negative (for defamation), and ABSA can detect unnatural sentiment distributions across different aspects. Real reviews typically contain balanced positive and negative opinions, whereas fake reviews may contain contradictory or repetitive sentiments across multiple aspects. Fake reviews often copy-paste phrases across multiple aspects, lacking variety in expression. Genuine reviews provide a natural distribution of the sentiment across aspect, while fake ones may overemphasize a single aspect. Fake reviews often repeat phrases across several reviews, lack variability in expression, and provide unnatural sentiment distributions. ABSA enables comparison of sentiment trends for individual aspects across reviews, thereby supporting the identification of suspicious opinion patterns and semantic inconsistencies frequently observed in deceptive reviews. Despite recent advances in deep learning and graph-based fake review detection, existing approaches still face challenges in capturing fine-grained semantic dependencies, contextual inconsistencies, reviewer behavior patterns, and sophisticated AI-generated deceptive reviews. To address these limitations, the proposed GEmFakeABSA framework integrates ABSA, Graph Convolutional Networks (GCN), POS-based linguistic feature extraction, contextual semantic embeddings, and hierarchical attention mechanisms within a unified fake review detection architecture80–82. In the framework we propose, ABSA performs fine-grained extraction of aspects and sentiments to identify deceptive opinion patterns with respect to these product attributes. The POS based Linguistic Identity Extraction identifies important features related to grammar and opinion-related content, such as nouns, adjectives, verbs and sentiment-carrying expressions that tend to be frequently used in deceptive reviews. Contextual semantic embeddings improve the learning of semantic representations and help to better capture the contextual relationship between words in reviews. Furthermore, GCN is utilized by the framework to represent complicated relational dependencies among reviews, aspects, contextual semantic features and sentiment patterns. GCN is not often designed for sequential models (such as LSTM), but rather it capture latent semantic interactions and graph based contextual relationships among review data, which would induce stronger discriminative power of deception detection. Combined with ABSA, GCN can jointly learn aspect-level sentiment representations and contextual relational dependencies so that the framework further improves fake review detection. Therefore, the proposed GEmFakeABSA framework is designed to address the limitations of existing fake review detection methods by combining fine-grained aspect-aware sentiment analysis, linguistic feature extraction, contextual semantic representation learning, and graph-based dependency modeling within a unified architecture. This integration enhances the capability of the framework to identify sophisticated, context-aware, and AI-generated deceptive reviews more effectively than traditional fake review detection approaches. Current status and limitations of existing studies This technology has achieved remarkable results, but many challenges remain83: (1) label quality–manual annotation is expensive and often subjective; “fake” datasets from crowdsourcing may not generalize well to live platforms29; (2) data imbalance–fake reviews are rare vs. genuine proposals leading to biased learning; (3) domain transfer–models trained on one vertical or platform perform poorly elsewhere, (4) feature access–much behavioral metadata such as user history and timing information will be unavailable, affecting model performance; (5) explainability–neural networks provide higher accuracy than traditional models but lower transparency; finally, (6) due to root-level restrictions of real-time streaming data across multiple platforms the computational cost remains high when scaling automated moderation for marketplaces. Recent research cautions that random splits at the review level may lead to overestimation of performance as signals from users or products may leak into the test set. Several methods depend on either content or behavioral signals, or utilize ABSA without including platform-level indications. Certain evaluation methods contain entity leakage since they use the same entity (a user or a product) in both the training and test sets. Moreover, current work still faces scalability issues, adversarial fake reviews, how to effectively model reviewer–product interactions as well as detecting implicit sentiment indicators (e.g., sarcasm, satire, irony and self-deprecating humor). Objective of the work: To develop an effective fake review detection framework by integrating Aspect-Based Sentiment Analysis (ABSA), Graph Convolutional Networks (GCN), contextual embeddings, and POS-based linguistic feature extraction within a unified architecture.To perform fine-grained aspect-level sentiment analysis for identifying deceptive opinion patterns, sentiment inconsistencies, and rating–text mismatches commonly observed in fake reviews.To model complex semantic and contextual dependencies among reviews, aspects, and sentiment features using graph-based relational learning for improved fake review detection performance.To enhance the detection of sophisticated, context-aware, and AI-generated deceptive reviews through contextual semantic representation learning and hierarchical attention mechanisms.To evaluate the effectiveness and robustness of the proposed GEmFakeABSA framework through comprehensive quantitative, qualitative, comparative, and ablation-based performance analysis against baseline and state-of-the-art methods. To develop an effective fake review detection framework by integrating Aspect-Based Sentiment Analysis (ABSA), Graph Convolutional Networks (GCN), contextual embeddings, and POS-based linguistic feature extraction within a unified architecture. To develop an effective fake review detection framework by integrating Aspect-Based Sentiment Analysis (ABSA), Graph Convolutional Networks (GCN), contextual embeddings, and POS-based linguistic feature extraction within a unified architecture. To perform fine-grained aspect-level sentiment analysis for identifying deceptive opinion patterns, sentiment inconsistencies, and rating–text mismatches commonly observed in fake reviews. To perform fine-grained aspect-level sentiment analysis for identifying deceptive opinion patterns, sentiment inconsistencies, and rating–text mismatches commonly observed in fake reviews. To model complex semantic and contextual dependencies among reviews, aspects, and sentiment features using graph-based relational learning for improved fake review detection performance. To model complex semantic and contextual dependencies among reviews, aspects, and sentiment features using graph-based relational learning for improved fake review detection performance. To enhance the detection of sophisticated, context-aware, and AI-generated deceptive reviews through contextual semantic representation learning and hierarchical attention mechanisms. To enhance the detection of sophisticated, context-aware, and AI-generated deceptive reviews through contextual semantic representation learning and hierarchical attention mechanisms. To evaluate the effectiveness and robustness of the proposed GEmFakeABSA framework through comprehensive quantitative, qualitative, comparative, and ablation-based performance analysis against baseline and state-of-the-art methods. To evaluate the effectiveness and robustness of the proposed GEmFakeABSA framework through comprehensive quantitative, qualitative, comparative, and ablation-based performance analysis against baseline and state-of-the-art methods. Summary In summary, literature converges on three pillars: (i) ABSA improves fine grained evidence; (ii) graph encoders capture non local aspect–opinion ties; and (iii) behaviour and metadata add orthogonal signals. The gap lies in principled fusion under leakage safe evaluation with robust reporting. Table 1 shows that the existing studies either focus mainly on content cues, behavioural cues, or graph-based sentiment representations independently. In the table “R” (Robust reporting) column, △ indicates partial support for robust reporting practices, whereas \checkmark represents comprehensive reporting support. Very limited studies integrate all these components under leakage-safe evaluation settings. Our work addresses this gap and demonstrates gains across diverse categories.Table 1Compact taxonomy of prior approaches. C = Content cues; B = Behavior/metadata; G = Graph/ABSA; L = Leakage safe splits; R = Robust reporting.Method familyCBGLRContent only FRD\checkmarkBehavior only FRD\checkmarkHybrid (Content + Behavior)\checkmark\checkmark△Graph/Dependency ABSA in SA\checkmark\checkmark△Ours (GEmFakeABSA)\checkmark\checkmark\checkmark\checkmark\checkmark Table 1Compact taxonomy of prior approaches. C = Content cues; B = Behavior/metadata; G = Graph/ABSA; L = Leakage safe splits; R = Robust reporting.Method familyCBGLRContent only FRD\checkmarkBehavior only FRD\checkmarkHybrid (Content + Behavior)\checkmark\checkmark△Graph/Dependency ABSA in SA\checkmark\checkmark△Ours (GEmFakeABSA)\checkmark\checkmark\checkmark\checkmark\checkmark Compact taxonomy of prior approaches. C = Content cues; B = Behavior/metadata; G = Graph/ABSA; L = Leakage safe splits; R = Robust reporting. Proposed methodology This section details the architecture, training strategy, and feature design of the GEmFakeABSA framework. In order to uncover features unique to every individual product category, this is a crucial benefit because the well-known baseline datasets that are currently available mainly concentrate on small number of products, like laptop and restaurants, and do not adequately represent the wide variety that can be encountered on online shopping sites. Fig. 1 shows the proposed fake review detection framework GEmFakeABSA architecture. The entire system procedure is divided into more manageable segments because of the intricacy of the fake review detection challenge. The input reviews are first pre processed in order to be cleaned along with subsequently examined. Afterwards, GCN is used to extract product features and opinion extraction with ABSA. Lastly, the retrieved features and aspects are applied to identify fake reviews. The four primary components of the proposed approach for identifying fake reviews are displayed in Fig. 1: (I) Text preparation, (II) Extraction of features, (III) GCN-embedded attention-driven ABSA, and (IV) fake review detection. Any website that deals with online retail should be able to use the framework to identify fake reviews. Each component is briefly illustrated in the subsequent sections.Fig. 1Fake Review Detection using ABSA. The pipeline comprises: (I) text preparation, (II) feature extraction, (III) ABSA encoding with a GCN over dependency graphs, and (IV) fake review classification. Feature fusion occurs directly before the final classifier. Fake Review Detection using ABSA. The pipeline comprises: (I) text preparation, (II) feature extraction, (III) ABSA encoding with a GCN over dependency graphs, and (IV) fake review classification. Feature fusion occurs directly before the final classifier. Data Collection and Preprocessing Dataset: Training and assessment of the technique use a corpus of 21,034 Amazon reviews spanning 30 product categories. Unless otherwise noted, all analyses refer to this multi category dataset; category-specific findings are explicitly marked. The system takes user reviews, verified purchases (VP), product category, product ID, product title, star rating, and labels. The features include behavioural features (e.g., rating score features) and Linguistic features (e.g., textual content of the reviews)Text Preprocessing: Prior to feature extraction and GCN-embedded ABSA, the preprocessing pipeline consists of sentence splitting, text cleaning, tokenization, lowercasing, stop word removal, lemmatization, POS tagging, and dependency parsing, as illustrated in Fig. 1. These steps were employed to reduce textual noise, normalize linguistic variations, and preserve syntactic dependency required for aspect-opinion extraction and further fake and genuine review classification. The adopted preprocessing strategy follows commonly used NLP preprocessing practices reported in the literature84. Furthermore, 300-dimensional pre-trained Word2Vec embedding concatenated with 6-dimensional POS-tag representations were utilized for semantic feature representation.POS tagging: POS tagging was performed to extract syntactic features, aiding in the identification of linguistic patterns commonly found in fake reviews. Feature Extraction: Linguistic Features: Extracted features include POS tags, syntactical structures, and writing style cues that differentiate genuine from fake reviews.Behavioural features: extraction: Overall rating score features are used as behavioural features.POS concatenated with Word Embedding:POS concatenated with Word Embedding: Part-of-Speech (POS) features are concatenated with word embedding vectors to incorporate both semantic information and syntactic structure, thereby enhancing contextual feature representation for accurate classification.Dependency Graph Construction: A dependency parser (SpaCy V3.7) generates a direct dependency tree for each review to construct a graph representation of syntactic relationships among words.Aspect Term Extraction: An ABSA module was used to identify aspect terms (e.g., “battery”, “screen”) and their associated sentiments from the reviews. ABSA Framework: Traditional ABSA baseline: A traditional ABSA pipeline was implemented using attention-based mechanisms to extract aspect–sentiment pairs.GCN embedded ABSA model: The proposed model integrates GCN-based contextual embedding into the ABSA pipeline. The dependency graph is formally defined as G = (V, E), where:Nodes (V): Each node represents a token (word) in the review.Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser.Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. Fake Review Detection GEmFakeABSA Module: Classification Objective: Using the output from the GCN-enhanced ABSA model, a binary classifier (e.g., a feedforward deep learning model) was learned to differentiate across authentic and fraudulent comments.Feature Fusion (clarified): Outputs from linguistic analysis, POS tagging, and ABSA are concatenated with the ABSA/GCN representation and fed to a lightweight two layer MLP with layer normalization and dropout; the final sigmoid/softmax layer produces the fake/genuine probability. Training uses class weighted cross entropy with early stopping on validation macro-F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}. Data Collection and Preprocessing Dataset: Training and assessment of the technique use a corpus of 21,034 Amazon reviews spanning 30 product categories. Unless otherwise noted, all analyses refer to this multi category dataset; category-specific findings are explicitly marked. The system takes user reviews, verified purchases (VP), product category, product ID, product title, star rating, and labels. The features include behavioural features (e.g., rating score features) and Linguistic features (e.g., textual content of the reviews)Text Preprocessing: Prior to feature extraction and GCN-embedded ABSA, the preprocessing pipeline consists of sentence splitting, text cleaning, tokenization, lowercasing, stop word removal, lemmatization, POS tagging, and dependency parsing, as illustrated in Fig. 1. These steps were employed to reduce textual noise, normalize linguistic variations, and preserve syntactic dependency required for aspect-opinion extraction and further fake and genuine review classification. The adopted preprocessing strategy follows commonly used NLP preprocessing practices reported in the literature84. Furthermore, 300-dimensional pre-trained Word2Vec embedding concatenated with 6-dimensional POS-tag representations were utilized for semantic feature representation.POS tagging: POS tagging was performed to extract syntactic features, aiding in the identification of linguistic patterns commonly found in fake reviews. Data Collection and Preprocessing Dataset: Training and assessment of the technique use a corpus of 21,034 Amazon reviews spanning 30 product categories. Unless otherwise noted, all analyses refer to this multi category dataset; category-specific findings are explicitly marked. The system takes user reviews, verified purchases (VP), product category, product ID, product title, star rating, and labels. The features include behavioural features (e.g., rating score features) and Linguistic features (e.g., textual content of the reviews)Text Preprocessing: Prior to feature extraction and GCN-embedded ABSA, the preprocessing pipeline consists of sentence splitting, text cleaning, tokenization, lowercasing, stop word removal, lemmatization, POS tagging, and dependency parsing, as illustrated in Fig. 1. These steps were employed to reduce textual noise, normalize linguistic variations, and preserve syntactic dependency required for aspect-opinion extraction and further fake and genuine review classification. The adopted preprocessing strategy follows commonly used NLP preprocessing practices reported in the literature84. Furthermore, 300-dimensional pre-trained Word2Vec embedding concatenated with 6-dimensional POS-tag representations were utilized for semantic feature representation.POS tagging: POS tagging was performed to extract syntactic features, aiding in the identification of linguistic patterns commonly found in fake reviews. Dataset: Training and assessment of the technique use a corpus of 21,034 Amazon reviews spanning 30 product categories. Unless otherwise noted, all analyses refer to this multi category dataset; category-specific findings are explicitly marked. The system takes user reviews, verified purchases (VP), product category, product ID, product title, star rating, and labels. The features include behavioural features (e.g., rating score features) and Linguistic features (e.g., textual content of the reviews) Dataset: Training and assessment of the technique use a corpus of 21,034 Amazon reviews spanning 30 product categories. Unless otherwise noted, all analyses refer to this multi category dataset; category-specific findings are explicitly marked. The system takes user reviews, verified purchases (VP), product category, product ID, product title, star rating, and labels. The features include behavioural features (e.g., rating score features) and Linguistic features (e.g., textual content of the reviews) Text Preprocessing: Prior to feature extraction and GCN-embedded ABSA, the preprocessing pipeline consists of sentence splitting, text cleaning, tokenization, lowercasing, stop word removal, lemmatization, POS tagging, and dependency parsing, as illustrated in Fig. 1. These steps were employed to reduce textual noise, normalize linguistic variations, and preserve syntactic dependency required for aspect-opinion extraction and further fake and genuine review classification. The adopted preprocessing strategy follows commonly used NLP preprocessing practices reported in the literature84. Furthermore, 300-dimensional pre-trained Word2Vec embedding concatenated with 6-dimensional POS-tag representations were utilized for semantic feature representation. Text Preprocessing: Prior to feature extraction and GCN-embedded ABSA, the preprocessing pipeline consists of sentence splitting, text cleaning, tokenization, lowercasing, stop word removal, lemmatization, POS tagging, and dependency parsing, as illustrated in Fig. 1. These steps were employed to reduce textual noise, normalize linguistic variations, and preserve syntactic dependency required for aspect-opinion extraction and further fake and genuine review classification. The adopted preprocessing strategy follows commonly used NLP preprocessing practices reported in the literature84. Furthermore, 300-dimensional pre-trained Word2Vec embedding concatenated with 6-dimensional POS-tag representations were utilized for semantic feature representation. POS tagging: POS tagging was performed to extract syntactic features, aiding in the identification of linguistic patterns commonly found in fake reviews. POS tagging: POS tagging was performed to extract syntactic features, aiding in the identification of linguistic patterns commonly found in fake reviews. Feature Extraction: Linguistic Features: Extracted features include POS tags, syntactical structures, and writing style cues that differentiate genuine from fake reviews.Behavioural features: extraction: Overall rating score features are used as behavioural features.POS concatenated with Word Embedding:POS concatenated with Word Embedding: Part-of-Speech (POS) features are concatenated with word embedding vectors to incorporate both semantic information and syntactic structure, thereby enhancing contextual feature representation for accurate classification.Dependency Graph Construction: A dependency parser (SpaCy V3.7) generates a direct dependency tree for each review to construct a graph representation of syntactic relationships among words.Aspect Term Extraction: An ABSA module was used to identify aspect terms (e.g., “battery”, “screen”) and their associated sentiments from the reviews. Feature Extraction: Linguistic Features: Extracted features include POS tags, syntactical structures, and writing style cues that differentiate genuine from fake reviews.Behavioural features: extraction: Overall rating score features are used as behavioural features.POS concatenated with Word Embedding:POS concatenated with Word Embedding: Part-of-Speech (POS) features are concatenated with word embedding vectors to incorporate both semantic information and syntactic structure, thereby enhancing contextual feature representation for accurate classification.Dependency Graph Construction: A dependency parser (SpaCy V3.7) generates a direct dependency tree for each review to construct a graph representation of syntactic relationships among words.Aspect Term Extraction: An ABSA module was used to identify aspect terms (e.g., “battery”, “screen”) and their associated sentiments from the reviews. Linguistic Features: Extracted features include POS tags, syntactical structures, and writing style cues that differentiate genuine from fake reviews. Linguistic Features: Extracted features include POS tags, syntactical structures, and writing style cues that differentiate genuine from fake reviews. Behavioural features: extraction: Overall rating score features are used as behavioural features. Behavioural features: extraction: Overall rating score features are used as behavioural features. POS concatenated with Word Embedding:POS concatenated with Word Embedding: Part-of-Speech (POS) features are concatenated with word embedding vectors to incorporate both semantic information and syntactic structure, thereby enhancing contextual feature representation for accurate classification. POS concatenated with Word Embedding:POS concatenated with Word Embedding: Part-of-Speech (POS) features are concatenated with word embedding vectors to incorporate both semantic information and syntactic structure, thereby enhancing contextual feature representation for accurate classification. Dependency Graph Construction: A dependency parser (SpaCy V3.7) generates a direct dependency tree for each review to construct a graph representation of syntactic relationships among words. Dependency Graph Construction: A dependency parser (SpaCy V3.7) generates a direct dependency tree for each review to construct a graph representation of syntactic relationships among words. Aspect Term Extraction: An ABSA module was used to identify aspect terms (e.g., “battery”, “screen”) and their associated sentiments from the reviews. Aspect Term Extraction: An ABSA module was used to identify aspect terms (e.g., “battery”, “screen”) and their associated sentiments from the reviews. ABSA Framework: Traditional ABSA baseline: A traditional ABSA pipeline was implemented using attention-based mechanisms to extract aspect–sentiment pairs.GCN embedded ABSA model: The proposed model integrates GCN-based contextual embedding into the ABSA pipeline. The dependency graph is formally defined as G = (V, E), where:Nodes (V): Each node represents a token (word) in the review.Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser.Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. ABSA Framework: Traditional ABSA baseline: A traditional ABSA pipeline was implemented using attention-based mechanisms to extract aspect–sentiment pairs.GCN embedded ABSA model: The proposed model integrates GCN-based contextual embedding into the ABSA pipeline. The dependency graph is formally defined as G = (V, E), where:Nodes (V): Each node represents a token (word) in the review.Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser.Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. Traditional ABSA baseline: A traditional ABSA pipeline was implemented using attention-based mechanisms to extract aspect–sentiment pairs. Traditional ABSA baseline: A traditional ABSA pipeline was implemented using attention-based mechanisms to extract aspect–sentiment pairs. GCN embedded ABSA model: The proposed model integrates GCN-based contextual embedding into the ABSA pipeline. The dependency graph is formally defined as G = (V, E), where:Nodes (V): Each node represents a token (word) in the review.Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser.Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. GCN embedded ABSA model: The proposed model integrates GCN-based contextual embedding into the ABSA pipeline. The dependency graph is formally defined as G = (V, E), where:Nodes (V): Each node represents a token (word) in the review.Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser.Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. Nodes (V): Each node represents a token (word) in the review. Nodes (V): Each node represents a token (word) in the review. Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser. Edges (E): An undirected edge is created between two nodes if a syntactic dependency relation exists between two corresponding words as determined connected by the dependency parser. Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. Graph Type: The graph is constructed at the individual review level and captures intra-sentence dependencies. Fake Review Detection GEmFakeABSA Module: Classification Objective: Using the output from the GCN-enhanced ABSA model, a binary classifier (e.g., a feedforward deep learning model) was learned to differentiate across authentic and fraudulent comments.Feature Fusion (clarified): Outputs from linguistic analysis, POS tagging, and ABSA are concatenated with the ABSA/GCN representation and fed to a lightweight two layer MLP with layer normalization and dropout; the final sigmoid/softmax layer produces the fake/genuine probability. Training uses class weighted cross entropy with early stopping on validation macro-F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}. Fake Review Detection GEmFakeABSA Module: Classification Objective: Using the output from the GCN-enhanced ABSA model, a binary classifier (e.g., a feedforward deep learning model) was learned to differentiate across authentic and fraudulent comments.Feature Fusion (clarified): Outputs from linguistic analysis, POS tagging, and ABSA are concatenated with the ABSA/GCN representation and fed to a lightweight two layer MLP with layer normalization and dropout; the final sigmoid/softmax layer produces the fake/genuine probability. Training uses class weighted cross entropy with early stopping on validation macro-F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}. Classification Objective: Using the output from the GCN-enhanced ABSA model, a binary classifier (e.g., a feedforward deep learning model) was learned to differentiate across authentic and fraudulent comments. Classification Objective: Using the output from the GCN-enhanced ABSA model, a binary classifier (e.g., a feedforward deep learning model) was learned to differentiate across authentic and fraudulent comments. Feature Fusion (clarified): Outputs from linguistic analysis, POS tagging, and ABSA are concatenated with the ABSA/GCN representation and fed to a lightweight two layer MLP with layer normalization and dropout; the final sigmoid/softmax layer produces the fake/genuine probability. Training uses class weighted cross entropy with early stopping on validation macro-F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}. Feature Fusion (clarified): Outputs from linguistic analysis, POS tagging, and ABSA are concatenated with the ABSA/GCN representation and fed to a lightweight two layer MLP with layer normalization and dropout; the final sigmoid/softmax layer produces the fake/genuine probability. Training uses class weighted cross entropy with early stopping on validation macro-F\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_1$$\end{document}. Constructing a model of GCN-embedded GEmFakeABSA aspect term extraction The GCN-embedded deep attention model presented in this section is capable of capturing advanced semantics for improved aspect term prediction. Additionally, sentence-level attention aggregates GCN embeddings across tokens so that aspect cues dispersed within the sentence are emphasized during decoding.Table 2Aspect encoding samples (IOB2). Example illustrates a multi-word aspect (Camera Quality) with tags B-I, and a single word aspect (Performance) with tag B.SLongCameraQualityWithFastPerformance\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{Y}$$\end{document}OBIOOBType:OutsideAspect BeginAspect InsideOutsideOutsideAspect BeginY[0 0 1][1 0 0][0 1 0][0 0 1][0 0 1][1 0 0]Aspect:Multi wordSingle word Table 2Aspect encoding samples (IOB2). Example illustrates a multi-word aspect (Camera Quality) with tags B-I, and a single word aspect (Performance) with tag B.SLongCameraQualityWithFastPerformance\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{Y}$$\end{document}OBIOOBType:OutsideAspect BeginAspect InsideOutsideOutsideAspect BeginY[0 0 1][1 0 0][0 1 0][0 0 1][0 0 1][1 0 0]Aspect:Multi wordSingle word Aspect encoding samples (IOB2). Example illustrates a multi-word aspect (Camera Quality) with tags B-I, and a single word aspect (Performance) with tag B. IOB2 labelling In order to solve a sequential labelling issue that is then used for the ATE subtask, the model leverages the review labels. Given an input sentence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R=\{S_1,S_2,\ldots ,S_N\}$$\end{document} of N words, the system generates results as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y=\{Y_1,Y_2,\ldots ,Y_N\}$$\end{document} and the labels are given as IOB2 encoding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{Y}=\{\bar{Y}_1,\bar{Y}_2,\ldots ,\bar{Y}_N\}$$\end{document} to train the GCN-embedded ABSA model. In our method, the tagging scheme used for generating pseudo labels is shown in Table 2. IOB2 tagging details. We use the standard IOB2 scheme with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{B}$$\end{document} (aspect-begin), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{I}$$\end{document} (aspect-inside), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{O}$$\end{document} (outside). For a multi-word aspect such as camera quality, the first token is tagged B and subsequent tokens are tagged I; single-word aspects (e.g., performance) receive B. The corresponding one hot mapping used in the row Y is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[1,0,0]=\textrm{B}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[0,1,0]=\textrm{I}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[0,0,1]=\textrm{O}$$\end{document}. The table is illustrative; the Type row descriptions align directly with the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{Y}$$\end{document} tags and the one hot vectors shown (B→Aspect Begin, I→Aspect Inside, O→Outside). POS and word embedding for graph convolutional embedding The concatenated embedding from Word embedding with POS tagging is converted into GCN-embedding are utilised as the deep neural network’s input characteristics in our proposed work. Firstly, we used the supplied sentence as our model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R = \{S_1, S_2, S_3, \ldots , S_N\}$$\end{document} as a sequence vector for feature embedding that corresponds to the specific word as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W = \{we_1, we_2, we_3, \ldots , we_N\}$$\end{document}, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$we_i \in \mathbb {R}^{300}$$\end{document}. Additionally, we displayed the POS tag sequence embeddings for each word in the provided review phrase as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = \{pos_1, pos_2, pos_3, \ldots , pos_N\}$$\end{document}, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i \in \mathbb {R}^{6}$$\end{document}. After that, W and the matching P were concatenated to create the input sequence for GCN as shown in Eq. 1:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \begin{aligned} X^{\text {Emb}d} = \{x_1, x_2, \ldots , x_N\} = \{(we_1, pos_1)^\top , (we_2, pos_2)^\top , \ldots , (we_N, pos_N)^\top \} \end{aligned} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\quad x_i \in \mathbb {R}^{300 + 6}$$\end{document}Algorithm 1Dependency graph construction for a review. Dependency graph construction for a review. The dependency graph is constructed as presented in Algorithm 1. The embedding generated through Algorithm 1 serves as initial feature matrix to a two-layer GCN implementation using PyTorch Geometric (PyG) library. The layer dimensions are set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{in}->64$$\end{document} for the first layer and 64 -> 64 for the second layer. As in the standard GCN, the mean aggregation method is used. In order to determine the syntactic ties across words and create an interaction network among them, the dependency parser next employs the embedding. Edges are represented by associations that the parser has learned to represent the adjacency matrix (A). Consequently, both adjacency matrix (A) and feature matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X^{Embd}$$\end{document} presented below serve as the GCN’s input:Feature Matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X^{\text {Emb}d} \in \mathbb {R}^{N \times 306}$$\end{document}): each node in the graph has an associated feature vector.Adjacency Matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A \in \mathbb {R}^{N \times N}$$\end{document}): represents connections between nodes from the dependency graph.Adjacency Normalization is formulated as in Eq. 2-3 :2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{A}= & A + I, \qquad D_{ii} = \sum _{j} \hat{A}_{ij}; \end{aligned}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{A}= & D^{-\frac{1}{2}} \hat{A} D^{-\frac{1}{2}} \end{aligned}$$\end{document}Here I is the identity matrix (self loops) and D is the degree matrix computed from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{A}$$\end{document}. The normalised matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A}$$\end{document} improves numerical stability and enables smooth feature propagation. Feature Matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X^{\text {Emb}d} \in \mathbb {R}^{N \times 306}$$\end{document}): each node in the graph has an associated feature vector. Feature Matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X^{\text {Emb}d} \in \mathbb {R}^{N \times 306}$$\end{document}): each node in the graph has an associated feature vector. Adjacency Matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A \in \mathbb {R}^{N \times N}$$\end{document}): represents connections between nodes from the dependency graph. Adjacency Matrix (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A \in \mathbb {R}^{N \times N}$$\end{document}): represents connections between nodes from the dependency graph. In GCN, convolution is adapted and the main concept is to improve the vector representation by combining data from the node’s neighbours. This operation is similar to the convolution process in the CNN, where neighbouring pixels’ information is aggregated. Each GCN layer takes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X^{\text {Emb}d}$$\end{document} (feature matrix) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A}$$\end{document} (normalised adjacency) as input and captures the relationships between aspect terms. Each GCN layer applies the following propagation rule:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} H^{(l+1)} = \sigma \!\left( \tilde{A}\, H^{(l)} W^{(l)} \right) \end{aligned}$$\end{document}which includes:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(l)}$$\end{document} : node feature matrix at layer l (initially \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(0)} = X^{\text {Emb}d}$$\end{document});\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A}$$\end{document} : normalised adjacency matrix defined above;\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^{(l)}$$\end{document} : layer level learnable weight matrix;σ : non linearity (ReLU).The key operations in a GCN layer are as: Neighbourhood aggregation (message passing): each node gathers feature information from its neighbours.Linear transformation: the weight matrix translates the aggregated features.Non linearity: activation functions are applied to introduce non linearity.The proposed GCN embedded ABSA model will receive the input as the sequence learned based on the GCN embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(l+1)}$$\end{document} by Eq. 4 as the hidden layer vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{h}_t$$\end{document}. Unless otherwise specified, token embeddings are 300-d and POS embeddings are 6-d; concatenation yields 306-d node features, used consistently in all figures and equations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(l)}$$\end{document} : node feature matrix at layer l (initially \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(0)} = X^{\text {Emb}d}$$\end{document}); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(l)}$$\end{document} : node feature matrix at layer l (initially \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(0)} = X^{\text {Emb}d}$$\end{document}); \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A}$$\end{document} : normalised adjacency matrix defined above; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A}$$\end{document} : normalised adjacency matrix defined above; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^{(l)}$$\end{document} : layer level learnable weight matrix; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^{(l)}$$\end{document} : layer level learnable weight matrix; σ : non linearity (ReLU). σ : non linearity (ReLU). Neighbourhood aggregation (message passing): each node gathers feature information from its neighbours. Neighbourhood aggregation (message passing): each node gathers feature information from its neighbours. Linear transformation: the weight matrix translates the aggregated features. Linear transformation: the weight matrix translates the aggregated features. Non linearity: activation functions are applied to introduce non linearity. Non linearity: activation functions are applied to introduce non linearity. Network proposed with hierarchical attention The ABSA module in Fig. 1 presents the GCN-embedded hierarchical attention aspect extraction network. The architecture includes multiple components (GCN +Bi-LSTM + Hierarchical Attention), each of which operates on relatively small units. Here, the baseline dependency-aware HAN model already incorporates dependency-driven syntactic features as part of its input representation but processes them sequentially without explicit graph-based message passing. As a result, the computational load is significantly lower than Transformer-based models. The training was conducted on an NVIDIA RTX 2080Ti GPU; thus, the model achieves an average training cost 18 seconds per epoch. For inference latency of 4.3 ms per review on GPU and 11.8 ms per review on CPU. For inference, batch size of 32 is used to measure the GPU latency using batched processing, and the reported per-review time was computed by dividing the total batch processing time by the batch size which is 32. On the other hand, CPU latency was measured using single-sample inference with batch size of 1. To ensure the stable timing, all the measurements were averaged over multiple runs after a warm-up phase. Inference was evaluated on both CPU and GPU environments. The hidden vector at time step t is defined as5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_t = B_t(\textbf{h}_{t-1}, \textbf{x}_t; \Theta _{\text {enc}}) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{h}_{t-1}$$\end{document} is the previous hidden state, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_t$$\end{document} the current input, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Theta _{\text {enc}}$$\end{document} the encoder parameters. For a Bi-LSTM encoder, we use the standard concatenation:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_t = \textbf{h}_t^{\rightarrow } \oplus \textbf{h}_t^{\leftarrow } = B_t^{\rightarrow }(\textbf{h}_{t-1}, \textbf{x}_t; \Theta _{\text {enc}}) \oplus B_t^{\leftarrow }(\textbf{h}_{t-1} \textbf{x}_t; \Theta _{\text {enc}}) \end{aligned}$$\end{document}The gate updates follow:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{i}_t&= \sigma (W_i \cdot \textbf{h}_{t-1} + W_i \cdot \textbf{x}_t + \textit{bias}_i) \end{aligned}$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{f}_t&= \sigma (W_f \cdot \textbf{h}_{t-1} + W_f \cdot \textbf{x}_t + \textit{bias}_f) \end{aligned}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{o}_t&= \sigma (W_o \cdot \textbf{h}_{t-1} + W_o \cdot \textbf{x}_t + \textit{bias}_o) \end{aligned}$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{C}_t&= \tanh \!\big (W_c[\textbf{h}_{t-1} \textbf{x}_t] + \textit{bias}_c\big ) \end{aligned}$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{c}_t&= \textbf{i}_t \odot \tilde{C}_t + \textbf{f}_t \odot \textbf{c}_{t-1} \end{aligned}$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{h}_t&= \textbf{o}_t \odot \tanh (\textbf{c}_t) \end{aligned}$$\end{document}Notation note. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_\bullet$$\end{document} compactly denotes the collection of input and recurrent weight matrices; we use this shorthand while retaining standard LSTM semantics. Given the input sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X=\{x_1,\ldots ,x_N\}$$\end{document}, the decoder predicts labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_t$$\end{document} conditioned on the decoder state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_t$$\end{document}, the encoder states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H=\{h_1,\ldots ,h_N\}$$\end{document}, and a sentence level context \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{ti}$$\end{document}:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} S_t = f(S_{t-1}, y_{t-1}, C_{ti}) \end{aligned}$$\end{document}The context vector uses attention weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha _{ti}=\{\alpha _{t1},\ldots ,\alpha _{tN}\}$$\end{document}:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C_t = \sum _{i=1}^{N} \alpha _{ti} \, h_i \end{aligned}$$\end{document}Attention energies are computed by a learnable scoring function (implemented with a Bi-LSTM-based scorer):15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} e_{ti} = \textrm{BiLSTM}\big (S_{t-1}, h_i\big ) \end{aligned}$$\end{document}and normalised via softmax,16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha _{ti} = \textrm{softmax}(e_{ti}) = \frac{\exp (e_{ti})}{\sum _{j=1}^{N} \exp (e_{tj})} \end{aligned}$$\end{document}The sequence labelling probability factorises as17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} P(Y|X) = \prod _{t=1}^{N} P\big (y_t \mid y[1{:}t{-}1], C_{ti}\big ) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y[1{:}t{-}1]=[y_1,\ldots ,y_{t-1}]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P(y_t \mid \cdot )$$\end{document} is parameterised by the attention-weighted context. The cross-entropy over the IOB2 tag set is18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} H(e_t, p_t) = - \sum _{k \in K} e_t(k) \log p_t(k), \quad \text {with } K=\{\textrm{B},\textrm{I},\textrm{O}\} \end{aligned}$$\end{document}consistent with the one-hot mapping used earlier: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[1,0,0]{=}\textrm{B}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[0,1,0]{=}\textrm{I}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$[0,0,1]{=}\textrm{O}$$\end{document}. Next, the extracted aspect-opinion representation generated from the ABSA module be denoted as:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} z_i = \left[ a_i \oplus o_i \oplus s_i \right] \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_i$$\end{document} indicates the aspect embedding, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$o_i$$\end{document} represents the opinion representation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_i$$\end{document} denotes the sentiment-aware contextual feature, and ⊕ represents vector concatenation. The extracted aspect-opinion representations are next forwarded to the dense neural layers for fake review classification:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h_i^l = \sigma \left( W_f z_i + b_f \right) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_f$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b_f$$\end{document} are the trainable weight matrix and bias vector of the dense layer, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma \left( \cdot \right)$$\end{document} denotes the nonlinear activation function. Finally, the probability of a review being fake or genuine is computed using the Softmax classifier:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{y}_i = \text {Softmax} \left( W_o h_i^1 + b_o \right) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_i = \left[ P(\textrm{Genuine} \mid r_i), \; P(\textrm{Fake} \mid r_i) \right]$$\end{document} and the predicted class label is obtained as:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} y_i^{*} = \arg \max \left( \hat{y}_i \right) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_i^*\in \{\textrm{Fake}, \textrm{Genuine}\}$$\end{document} Behavioural and linguistic features Because it takes lots of mental effort to create a phoney review that does not reflect experience, previous research has shown a negative correlation between phrase complexity and faking. Consequently, compared to people who speak the truth, dishonest people typically have a weaker mental capacity to produce complicated and cohesive reviews. Similar to this, research indicates that dishonest people purposefully use harder-to-read and more complicated language to make their opinions appear more thorough and thus more trustworthy among readers. However, as genuine evaluations are created primarily to express the customer’s thoughts about the good or service rather than to trick others into believing they are reliable, they might be simpler to read. Additionally, recent data indicates that the semantic content of reviews may be important in identifying fraudulent reviews. Linguistic visualization of the comments is not practicable due to the sparseness of typical phrase arrangements based on BOW. Review context needs to be converted into context based dense word vectors in order to solve this problem. The challenge with this method is that word embeddings need a lot of data to be trained well85. Nonetheless, it is possible to produce a more comprehensive picture of reviews. Consequently, an extensive dataset of around 84 million Amazon reviews was used to train the Word2Vec model. Consistent with the model inputs used throughout this paper, we employ 300-dimensional Word2Vec embeddings with a context window of 5 (subsampling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-3}$$\end{document}) (This paragraph refers to embedding pretraining. These embeddings are then concatenated with 6-d POS vectors to yield 306-d node features (Sections 3.1 and Constructing a model of GCN-embedded GEmFakeABSA aspect term extraction).). Moreover, prior empirical research indicates that non-linguistic and behavioural variables, i.e. nonverbal characteristics of the commenter, have strong importance for FRD and can be more pertinent than linguistic characteristics in some settings. In order to determine whether a review score is anomalous and deviates beyond the standard score, the median score of a particular product and the review rating have been incorporated in accordance with previous research. The deceiver’s intention to either encourage or disparage the goods may be communicated through an excessive rating. In our feature set, we explicitly model this as a rating text gap feature, defined as the discrepancy between the star rating and the sentiment polarity inferred from the review text (via ABSA). Verified Purchase (VP) and rating information are used strictly as features, never as labels. Table 3 summarizes the linguistic and behavioural characteristics examined in proposed model.Table 3Linguistic and behavioural properties used in GEmFakeABSA. VP and rating are features only, not labels.Aspect categoryTypeName of featurePart-of-SpeechLinguistic FeaturesCounts of: sentiment tokens, nouns, verbs, adverbs, pronouns, adjectivesUnderstandabilityLinguistic FeaturesEase of reading, ease of interpretation, adequacy/accuracy of meaningReview compositionLinguistic FeaturesTotals: sentences, words, characters (per review)Word VectorLinguistic FeaturesWord2Vec embeddings (300-d, window=5, subsampling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-3}$$\end{document})Verified Purchase (VP)Behavioural FeaturesBinary indicator: whether the purchase is verifiedRating StatisticsBehavioural FeaturesHighest, lowest, average rating (per product); rating text gap \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta _r$$\end{document} (rating vs. ABSA inferred polarity) Table 3Linguistic and behavioural properties used in GEmFakeABSA. VP and rating are features only, not labels.Aspect categoryTypeName of featurePart-of-SpeechLinguistic FeaturesCounts of: sentiment tokens, nouns, verbs, adverbs, pronouns, adjectivesUnderstandabilityLinguistic FeaturesEase of reading, ease of interpretation, adequacy/accuracy of meaningReview compositionLinguistic FeaturesTotals: sentences, words, characters (per review)Word VectorLinguistic FeaturesWord2Vec embeddings (300-d, window=5, subsampling \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10^{-3}$$\end{document})Verified Purchase (VP)Behavioural FeaturesBinary indicator: whether the purchase is verifiedRating StatisticsBehavioural FeaturesHighest, lowest, average rating (per product); rating text gap \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta _r$$\end{document} (rating vs. ABSA inferred polarity) Linguistic and behavioural properties used in GEmFakeABSA. VP and rating are features only, not labels. Experimental setup and results In this section, the experimental setup and result analysis were conducted to evaluate the effectiveness and robustness of the proposed model GEmFakeABSA on the benchmark Amazon Reviews dataset. Quantitative and qualitative evaluations, including performance metrics, comparative analysis, and ablation studies, statistically significant testing, were performed to validate the capability of the proposed model in accurately detecting fake reviews using aspect-level contextual information Dataset for fake review detection In our investigation, we have utilized the Amazon dataset1, which contains a total of 21,000 English-language reviews distributed equally between fake and genuine reviews, with 10,500 reviews in each class, thereby maintaining a balanced class distribution. In addition, the dataset offers a number of other crucial details for every review as shown in Table 4, including Product_Category, Product_Title, Product_ID, Rating, Verified_Purchase, along with Review_Title and Review_Text. The average review length is approximately 85 words, with review lengths varying from a minimum of 5 words to a maximum of 450 words. The vocabulary size of the corpus is approximately 18,000 unique terms. The average review rating across the dataset is 4.22 on a 5-point scale, where 1 indicates highly negative sentiment and 5 indicates highly positive sentiment. The average rating for verified reviews is 4.24, whereas non-verified reviews exhibit a comparatively lower average rating of 3.98. In addition, the dataset contains both verified and non-verified purchase reviews, where verified reviews constitute approximately 55.7% (11,668 reviews) and non-verified reviews account for 44.3% (9,302 reviews). These dataset characteristics provide sufficient diversity in linguistic patterns, sentiment expressions, and review authenticity for evaluating the effectiveness of the proposed fake review detection framework. Since the utilized Kaggle Amazon review dataset does not contain predefined aspect annotations, aspect terms and aspect-opinion pairs were automatically extracted using the proposed BIO labelling-based ABSA framework combined with dependency parsing. The comments uniformly spanned 30 distinct types of products, with every group having approximately 700 comments, among them 350 were fake (counts vary slightly by category; exact per category tallies appear in the Appendix). Labels for fraudulent/genuine reviews are provided by the source and are retained without modification; only de-identified text and metadata are used. All statistics refer to the full corpus before modeling; train/validation/test partitions are product/user disjoint (80/10/10) as detailed in Section Proposed methodology. Data Availability Note. The dataset used and analyzed in this study is publicly available at the Kaggle repository: https://www.kaggle.com/datasets/lievgarcia/amazon-reviews. No new data were generated or collected specifically for this research. All analyses were conducted using this publicly accessible, de-identified dataset. Next, the product categories are divided into the following classification as search, experience, and confidence (also known as credence) goods. This classification depends on how challenging it is for customers to find and evaluate details concerning a product or service. If a product category is search it refers to goods and services whose key attributes (like price, quality, or performance) can be easily evaluated before purchase through available information. The more accessible this information is, the less effort consumers need to spend searching. The examples include books, grocery, toys, watches, etc. Experience products are those whose key attributes can only be assessed after use, as complete information is not available beforehand. Therefore, searching for information is less effective and more time consuming than actually experiencing the product or service. Confidence (credence) products are those whose attributes cannot be accurately evaluated even after purchase and use. Consumers often struggle to judge these products due to a lack of technical knowledge or expertise. Table 5 displays the grouping of the product types21.Table 4Dataset statistics and characteristics.StatisticsValueTotal Reviews21,000Genuine (Non-Fake) Reviews10,500Fake Reviews10,500Class DistributionBalanced (Fake-50% / Non-Fake-50%)Average Review Length (words)85 approx.Verified Purchase Reviews55.70% approx. (11,668)Non-Verified Reviews44.30% approx. (9,302)Number of Product Categories30Reviews Per Categories700 approx.Fake Reviews per Category350 approx.Average Rating4.22 (Maximum-5 (very positive), Minimum-1 (very negative))Average Rating (verified)4.24Average Rating (non-verified)3.98Minimum Review Length (words)5Maximum Review Length (words)450Vocabulary Size18,000 approx.LanguageEnglishDataset ResourceKaggle (Amazon Reviews)Table 5Amazon product categories21.SearchExperienceConfidence (Credence)Books, grocery, kitchen, luggage, office products, pet products, tools, toys, video DVD, watchesApparel, automotive, baby, beauty, camera, electronics, furniture, home, home entertainment, home improvement, jewellery, lawn and garden, musical instruments, outdoors, PC, shoes, sports, video games, wirelessHealth & personal care Table 4Dataset statistics and characteristics.StatisticsValueTotal Reviews21,000Genuine (Non-Fake) Reviews10,500Fake Reviews10,500Class DistributionBalanced (Fake-50% / Non-Fake-50%)Average Review Length (words)85 approx.Verified Purchase Reviews55.70% approx. (11,668)Non-Verified Reviews44.30% approx. (9,302)Number of Product Categories30Reviews Per Categories700 approx.Fake Reviews per Category350 approx.Average Rating4.22 (Maximum-5 (very positive), Minimum-1 (very negative))Average Rating (verified)4.24Average Rating (non-verified)3.98Minimum Review Length (words)5Maximum Review Length (words)450Vocabulary Size18,000 approx.LanguageEnglishDataset ResourceKaggle (Amazon Reviews) Dataset statistics and characteristics. Table 5Amazon product categories21.SearchExperienceConfidence (Credence)Books, grocery, kitchen, luggage, office products, pet products, tools, toys, video DVD, watchesApparel, automotive, baby, beauty, camera, electronics, furniture, home, home entertainment, home improvement, jewellery, lawn and garden, musical instruments, outdoors, PC, shoes, sports, video games, wirelessHealth & personal care Amazon product categories21. Experiment setup and parameter settings All experiments were conducted using Python on a system equipped with Intel Core i7 processor, 16 GB RAM, and NVIDIA RTX 2080Ti GPU support. Pre-trained Word2Vec embeddings of dimension 300 were employed for semantic representation and concatenated with 6-dimension POS tagging features to enhance syntactic contextual learning. As mentioned in Section 3.2, two-layer GCN implementation using the PyTorch Geometric (PyG) library. The layer dimensions are set to din->64 for the first layer and 64->64 for the second layer. As in the standard GCN, the mean aggregation method is used. The Bi-LSTM hidden size is 64 per direction, with a dropout rate of 0.5 being applied to reduce overfitting. After concatenation, the output dimensionality is 128, while the attention layer uses a dense representation for contextual weighting at both the word and sentence levels. A dropout rate of 0.5 is applied between GCN layers, and ReLU is used as the activation function. The model is trained with a batch size of 32 and 20 epochs using the Adam optimizer (learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times 10^{-3}$$\end{document}) and categorical cross-entropy loss. These hyperparameters were selected based on empirical evaluation on the validation set through controlled experiments to achieve optimal performance. The proposed model was trained and evaluated using train-test splitting along to maintain class balance across fake and genuine reviews. Early stopping was applied based on validation loss to prevent overtraining. Performance evaluation was carried out using accuracy, precision, recall, and F1-score metrics. Evaluation metrics For the performance evaluation of the proposed fake review detection model GEmFakeABSA, the most widely used standard classification metrics such as precision, recall, F1-score and accuracy are considered for the benchmark Amazon Review dataset. These metrics offer an unbiased evaluation of the model performance by examining both false positives and false negatives. Additionally, latest state-of-the-art and baseline approaches on Amazon review datasets for fake review detection generally examined on these metrics for performance evaluation, ensuring fair and consistent comparison across models. The computation of these metrics is based on the confusion matrix which includes true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values. The accuracy measures the overall correctness of the proposed fake review detection model GEmFakeABSA.23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Acurracy=\frac{TP+TN}{TP+TN+FP+FN} \end{aligned}$$\end{document}24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Precision = \frac{TP}{TP + FP} \end{aligned}$$\end{document}25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Recall = \frac{TP}{TP + FN} \end{aligned}$$\end{document}26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} F1-Score = 2 \times \frac{Precision \times Recall}{Precision + Recall} \end{aligned}$$\end{document} where TP represents the number of fake reviews correctly identified as fake, TN denotes the number of genuine reviews correctly identified as genuine, FP indicates the number of genuine reviews incorrectly classified as fake, and FN represents the number of fake reviews incorrectly classified as genuine. Quantitative analysis In this section, the quantitative analysis evaluates the performance of the proposed framework using standard metrics such as accuracy, precision, recall, and F1-score. Comparative results with baseline and existing methods demonstrate the effectiveness and robustness of the proposed model. Effect of ABSA on FRD Using ABSA as a preprocessing or feature extraction step significantly enhances the semantic structure of raw review text. The primary aim of the preliminary investigations was to investigate the effects of ABSA on fake review detection. To perform ABSA on the Amazon dataset, the GCN embedded Aspect Term Extraction methodology presented in Section Constructing a model of GCN-embedded GEmFakeABSA aspect term extraction was initially employed to determine salient features for every product type. The aspects retrieved applying the presented framework are illustrated in Table 6. Although the relationship between ABSA and fake review detection is indirect, the use of GCN-based ABSA offers notable advantages by structuring review content into meaningful aspect, sentiment pairs. The contributions of GCN-based ABSA to fake review detection are discussed below. Unless otherwise noted, results referenced in this subsection use product /user disjoint splits and are averaged over multiple random seeds (mean±std). Improved aspect-sentiment modellingThe GCN component of the ABSA framework models the syntactic structure of each review by constructing dependency graphs.This facilitates fine grained sentiment detection at the aspect level, enabling the identification of inconsistencies in sentiment expressions.Fake reviews often exhibit overly generic or unnatural sentiments, and ABSA effectively exposes such patterns through structured analysis.Graph Structure Reveals Linguistic Anomalies GCN models can capture syntactic anomalies in fabricated text, such as repetitive structures, excessive use of positive sentiment, and the absence of concrete aspect terms. These anomalies serve as important discriminative features for fake review classification. In our pipeline, such graph derived cues are used strictly as features within the fusion head. The GCN component of the ABSA framework models the syntactic structure of each review by constructing dependency graphs. The GCN component of the ABSA framework models the syntactic structure of each review by constructing dependency graphs. This facilitates fine grained sentiment detection at the aspect level, enabling the identification of inconsistencies in sentiment expressions. This facilitates fine grained sentiment detection at the aspect level, enabling the identification of inconsistencies in sentiment expressions. Fake reviews often exhibit overly generic or unnatural sentiments, and ABSA effectively exposes such patterns through structured analysis. Fake reviews often exhibit overly generic or unnatural sentiments, and ABSA effectively exposes such patterns through structured analysis. Explainable Patterns for Detection GCN based ABSA generates interpretable graphs that illustrate how sentiment expressions are linked to specific aspects. In fake reviews, these graphs may:Exhibit disconnected sentiment nodes, where sentiment words are not logically attached to any aspect.Reveal shallow or uninformative dependency structures, lacking linguistic richness.Enhancement to Fake Review Detection Pipelines By integrating GCN-ABSA with downstream fake review classifiers:The outputs of ABSA–such as the number of detected aspects, sentiment consistency, and syntactic richness–can be used as engineered features.A secondary classifier can be trained on these features to further improve detection performance.Table 6Aspect-terms extracted for each product category.Product categoryAspect-terms extractedHome improvementShower Head, Packaging, screws, drawer, design, appearance, hose, installShoesfoot, loose, smell, earphones, glasses, glasses, durable, comfortable, lookingToolsappears, laundry room, power failure, design, price, speed, advertised, shipment, deliveryCameraprice, battery, lens, video, time, picture, qualityWirelessiPhone, phone, screen, quality, case, battery, priceElectronicsprice, battery, music, speaker, headphones, quality, soundVideo gamescharging station, AC adapter, headset, speakers, sound quality, price, look and feel, subscriptionBooksguide, life, author, time, reading, story, peopleApparelsize, price, quality, colour, look, time, beltOutdoorsbag, design, brand, carry, transporting, qualityHealthpersonal care & price, weight, exercise, quality, health, energy, tasteHomecost, distances, comfortable, cushion, durability, bedsheets, weight, flowersKitchendesign, color, cup, food, tea, coffee, dinnerware, crackers, sugar, videosToysgame, quality, price, son, year, daughter, carLuggageprice, returns policy, size, durable, traveling, quality, bag, colorLawn and gardendurable, weight, Setup, location, appearance, touch, size, pool, seatSportstraining, foot, price, optics, weight, monitor, lifestyle, healthy, qualityBeautyteeth, heal, color, price, quality, appearance, look and feel, makeupHome entertainmentwarranty, sound quality, voice quality, price, picture, designMusical instrumentsprice, cord, durability, sound quality, sizeOffice productsColors, predictability, price, quality, performance, print quality, speedVideo DVDfamily, action, people, series, video, story, movieWatchesadjustable, price, color, display, looks, fitting, batteryAutomotiverefund, polity, servicing, speed, price, maintenance, battery life, compressorFurnituresize, price, color, advertising, design, chair, manufacturing, assemblyPCcase, screen, tablet, price, mouse, laptop, keyboardPet productsweight, size, quality, consistency, priceBabyprotection, ease of use, design, price, fitting, adjustmentGroceryflavor, taste, price, size, healthy Exhibit disconnected sentiment nodes, where sentiment words are not logically attached to any aspect. Exhibit disconnected sentiment nodes, where sentiment words are not logically attached to any aspect. Reveal shallow or uninformative dependency structures, lacking linguistic richness. Reveal shallow or uninformative dependency structures, lacking linguistic richness. The outputs of ABSA–such as the number of detected aspects, sentiment consistency, and syntactic richness–can be used as engineered features. The outputs of ABSA–such as the number of detected aspects, sentiment consistency, and syntactic richness–can be used as engineered features. A secondary classifier can be trained on these features to further improve detection performance. A secondary classifier can be trained on these features to further improve detection performance. Table 6Aspect-terms extracted for each product category.Product categoryAspect-terms extractedHome improvementShower Head, Packaging, screws, drawer, design, appearance, hose, installShoesfoot, loose, smell, earphones, glasses, glasses, durable, comfortable, lookingToolsappears, laundry room, power failure, design, price, speed, advertised, shipment, deliveryCameraprice, battery, lens, video, time, picture, qualityWirelessiPhone, phone, screen, quality, case, battery, priceElectronicsprice, battery, music, speaker, headphones, quality, soundVideo gamescharging station, AC adapter, headset, speakers, sound quality, price, look and feel, subscriptionBooksguide, life, author, time, reading, story, peopleApparelsize, price, quality, colour, look, time, beltOutdoorsbag, design, brand, carry, transporting, qualityHealthpersonal care & price, weight, exercise, quality, health, energy, tasteHomecost, distances, comfortable, cushion, durability, bedsheets, weight, flowersKitchendesign, color, cup, food, tea, coffee, dinnerware, crackers, sugar, videosToysgame, quality, price, son, year, daughter, carLuggageprice, returns policy, size, durable, traveling, quality, bag, colorLawn and gardendurable, weight, Setup, location, appearance, touch, size, pool, seatSportstraining, foot, price, optics, weight, monitor, lifestyle, healthy, qualityBeautyteeth, heal, color, price, quality, appearance, look and feel, makeupHome entertainmentwarranty, sound quality, voice quality, price, picture, designMusical instrumentsprice, cord, durability, sound quality, sizeOffice productsColors, predictability, price, quality, performance, print quality, speedVideo DVDfamily, action, people, series, video, story, movieWatchesadjustable, price, color, display, looks, fitting, batteryAutomotiverefund, polity, servicing, speed, price, maintenance, battery life, compressorFurnituresize, price, color, advertising, design, chair, manufacturing, assemblyPCcase, screen, tablet, price, mouse, laptop, keyboardPet productsweight, size, quality, consistency, priceBabyprotection, ease of use, design, price, fitting, adjustmentGroceryflavor, taste, price, size, healthy Aspect-terms extracted for each product category. Feature engineering for detection ABSA enables the extraction of highly informative features before fake review detection, which are structured, interpretable opinion units that help expose sentiment-aspect inconsistencies. It also enables rich, discriminative feature engineering and improves fake review detection classifier robustness and performance. In all experiments that follow, ABSA-derived signals are concatenated with linguistic, POS, and behavioural features within the fusion head; Verified Purchase and rating are used only as features. For pairs of research were conducted to demonstrate the reliability of the findings. Only behavioral and ABSA features were merged in the initial trial; linguistic features (understandability, review convolution, and part of speech) were incorporated in the next test; and the verified purchase factor was introduced in the third analysis. In the final run, star ratings were included. Fig. 2 presents the proposed and the21 achievement for each of the four parameters differs significantly, as the figure illustrates. Particularly, when the behavioral characteristics VP was removed, the detection models failed to perform well, which is in line with another research. Put differently, the rating based behavioral indicators and the linguistics markers by themselves were insufficient to attain a satisfactory level of model precision. Subsequently needs to be mentioned that earlier research on online retailer phony review identification came to comparable findings. In contrast to the results presented in Fig 2, Fig. 3 clearly demonstrate the impact of aspect-opinion pair extraction module in improving fake review detection performance. The integration of aspect extraction with linguistic features enables the model to capture aspect-opinion relationships and semantic dependencies within reviews, resulting in improved precision, recall, and F1-score for distinguishing fake and genuine reviews. Moreover, considering behavior features, verified purchase information, and star ratings gradually improve the discriminative capabilities of the model by integrating semantic, syntactic, and reviewer-behaviour characteristics. As examined, the highest precision, recall, and F1-score values are achieved when aspect extraction is combined with VP and star-rating information, which also refers to the maximum classification accuracy shown in Fig. 2. The sustained progress across all evaluation measures confirms that the proposed ABSA module effectively increases contextual feature learning and significantly improves the overall competence of classifying reviews as fake or genuine.Fig. 2Comparison of accuracy using ABSA for fake review detection GEmFakeABSA (proposed) and Base21.Fig. 3Impact of ABSA on precision, recall, and F1-score. Comparison of accuracy using ABSA for fake review detection GEmFakeABSA (proposed) and Base21. Impact of ABSA on precision, recall, and F1-score. Impact of features on precision and recall (feature-level ablation) This section shows a feature-level ablation analysis to examine the role of different feature combinations towards fake review detection performance. The analysis presented in this section evaluates how semantic, syntactic, behavioural, and metadata-based features affect the precision and recall power of ABSA module of the proposed framework. Specifically, the impact of POS features, behaviour features, verified purchase information, and ABSA is systematically investigated using the results obtained in Fig. 4 and 5. The feature-level ablation analysis in Fig. 4 and Fig. 5, altogether with the respective numerical results summarized in Table 7 and Table 8, highlights the role of individual feature combinations towards improving the precision and recall performance of the propose fake review detection architecture. The key reason of this analysis is to examine how different semantic, syntactic, behavioural, and metadata-based features influence the capability of the model to accurately classify reviews as fake or genuine. As examined and obtained from Fig. 4 and Fig. 5, the baseline configuration using on WE produce lower precision and recall, respectively, this clarify that semantic embedding vector alone is not capable of capturing fake review patterns. The incorporation of ABSA with WE enhance both precision and recall by facilitating aspect-opinion driven contextual understanding, which enable the model extracting sentiment inconsistencies and implicit fake reviews characteristics hidden in the text. Moreover, linguistic features in advance increases precision and recall by integrating syntactic and grammatical information, thereby enhancing contextual dependency training and overcome ambiguous fake review categorization.Table 7Precision comparison with and without verified purchase (VP) features.FeaturesPrecision (Without VP)Precision (With VP)WE72.2677.68WE + ABSA74.4281.1WE + Linguistic Features79.3483.78WE + Linguistic Features + ABSA82.586.22WE + Linguistics and Behavioural Features84.7887.88WE + Linguistics and Behavioural Features + ABSA86.4689.98WE + POS72.9878.06WE + POS + ABSA75.1882.82WE + POS + Linguistic features81.3684.12WE + POS + Linguistic features + ABSA83.0486.92WE + POS + Linguistics and Behavioural Features85.188.2WE + POS + Linguistics and Behavioural Features + ABSA87.5291.22 Table 7Precision comparison with and without verified purchase (VP) features.FeaturesPrecision (Without VP)Precision (With VP)WE72.2677.68WE + ABSA74.4281.1WE + Linguistic Features79.3483.78WE + Linguistic Features + ABSA82.586.22WE + Linguistics and Behavioural Features84.7887.88WE + Linguistics and Behavioural Features + ABSA86.4689.98WE + POS72.9878.06WE + POS + ABSA75.1882.82WE + POS + Linguistic features81.3684.12WE + POS + Linguistic features + ABSA83.0486.92WE + POS + Linguistics and Behavioural Features85.188.2WE + POS + Linguistics and Behavioural Features + ABSA87.5291.22 Precision comparison with and without verified purchase (VP) features. Fig. 4Impact of features on precision.. Table 8Recall comparison with and without verified purchase (VP) features.FeaturesRecall (Without VP)Recall (With VP)WE75.3277.14WE + ABSA77.177.88WE + Linguistic Features79.2282.34WE + Linguistic Features + ABSA84.786.64WE + Linguistics and Behavioural Features86.4487.96WE + Linguistics and Behavioural Features + ABSA86.7888.94WE + POS77.1678.92WE + POS + ABSA82.5684.38WE + POS + Linguistic features83.285.3WE + POS + Linguistic features + ABSA85.2687.44WE + POS + Linguistics and Behavioural Features86.8888.52WE + POS + Linguistics and Behavioural Features + ABSA87.3289.64 Impact of features on precision.. Table 8Recall comparison with and without verified purchase (VP) features.FeaturesRecall (Without VP)Recall (With VP)WE75.3277.14WE + ABSA77.177.88WE + Linguistic Features79.2282.34WE + Linguistic Features + ABSA84.786.64WE + Linguistics and Behavioural Features86.4487.96WE + Linguistics and Behavioural Features + ABSA86.7888.94WE + POS77.1678.92WE + POS + ABSA82.5684.38WE + POS + Linguistic features83.285.3WE + POS + Linguistic features + ABSA85.2687.44WE + POS + Linguistics and Behavioural Features86.8888.52WE + POS + Linguistics and Behavioural Features + ABSA87.3289.64 Recall comparison with and without verified purchase (VP) features. Fig. 5Impact of features on recall.. Impact of features on recall.. Impact of using verified purchase on performance The subsequent round of tests investigated the premise that Verified Purchase (VP) increases detection efficiency in experience and confidence categories, and examined outcomes when VP is unavailable. Experiments were carried out with and without the VP attribute. To ensure a strong configuration for the detection framework, all engineered cues listed in Table 9 were included (aspect/sentiment features from ABSA, consistency signals, and linguistic statistics). Fig. 2 and Fig. 7 visualise the method’s accuracy across Amazon product categories without and with VP, respectively. Unless noted otherwise, results use product/user disjoint splits and are averaged over multiple seeds (mean±std).Table 9Informative features derived from ABSA for fake review detection.FeatureWhy it HelpsNumber of aspects detectedFake reviews may mention fewer or too manySentiment consistency across aspectsReal reviews show nuanced opinionsFrequency of frequent opinion wordsExaggerated sentiments are common in fakeAspect sentiment densityRatio of aspects to the total text can be unnatural in fake reviews Table 9Informative features derived from ABSA for fake review detection.FeatureWhy it HelpsNumber of aspects detectedFake reviews may mention fewer or too manySentiment consistency across aspectsReal reviews show nuanced opinionsFrequency of frequent opinion wordsExaggerated sentiments are common in fakeAspect sentiment densityRatio of aspects to the total text can be unnatural in fake reviews Informative features derived from ABSA for fake review detection. Figure 6 shows that, without VP, the approach achieved its highest accuracies for Home Improvement (86.12%), Shoes (85.21%), Tools (84.06%), Camera (83.13%), and Wireless (83.13%). Lower accuracies were observed for Grocery (66.23%), Baby (67.87%), PC (68.52%), and Pet Products (68.96%), while Jewellery (76.38%), Lawn and Garden (76.44%), and Sports (76.37%) were near the middle of the range. These variations are consistent with the differing observability of product attributes across categories.Fig. 6Accuracy of fake review detection across product categories without the VP feature.Fig. 7Accuracy of fake review detection across product categories with the VP feature. Accuracy of fake review detection across product categories without the VP feature. Accuracy of fake review detection across product categories with the VP feature. When including VP, Fig. 7 indicates notable changes across categories. The highest accuracies were obtained for Wireless (92.37%), Tools (91.69%), Shoes (91.54%), Furniture (91.09%), and Watches (90.12%). Lower accuracies appeared for Pet Products (75.13%), Grocery (76.32%), Office Products (80.27%), PC (80.59%), and Video DVD (80.91%). Toys, Home, and Automotive averaged around 85%. Thus, even with VP available, effectiveness varied by product type–consistent with the search/experience/credence taxonomy. Throughout, VP is used strictly as a feature (never as a label).Fig. 8Precision of fake review detection across product categories without and with VP features. Precision of fake review detection across product categories without and with VP features. Fig. 8 and Fig. 9 jointly emphasize the effectiveness of VP features in enhancing fake review detection over multiple product categories. The integration of VP features generally improves both precision and recall, demonstrating greater consistency in detecting fake reviews. From the precision analysis (Fig. 8), considerable improvements are obtained for the product category such as Sports, Office Products, Furniture, and Music Instruments, illustrating that VP features help in overcoming false positive predictions. Likewise, the recall analysis (Fig. 9) shows significant improvements for categories consisting of Automative, Video Games, Luggage, and Watches, denoting refined competence in identifying fake reviews that remained genuine review otherwise. A key finding is that categories such as Home Improvement, Shoes, Electronics, and Toy observed considerable improvement in both precision and recall, indicating balanced performance improvement in fake review detection. On the other side, a few categories exhibit minimal trade-offs, where increase in recall is obtained by slight degradation in precision, illustrating that the influence of VP metadata is product category dependent. In general, the performance achieved confirm that VP features provides additional behavioral and contextual metadata besides textual semantics strengthen the generalizability and robustness potential of the proposed fake review detection model.Fig. 9Recall of fake review detection across product categories without and with VP features.Fig. 10Accuracy with VP: GEmFakeABSA vs. baseline21 across product categories. Recall of fake review detection across product categories without and with VP features. Accuracy with VP: GEmFakeABSA vs. baseline21 across product categories. Figure 10 compares GEmFakeABSA against the baseline21 when VP is present. The proposed approach improved accuracy for most categories; the largest gains were observed for Electronics (+13%), Wireless (+10%), and Camera (+8%). For Musical Instruments and PC, accuracy was lower than21, suggesting category specific sensitivity. These outcomes align with prior observations that VP and rating related signals are beneficial but not uniformly predictive across all categories. Comparison of proposed GEmFakeABSA model with baseline and recent approaches This section presents the comparative analysis of the proposed GEmFakeABSA model with baseline and recent state-of-the-art approaches in terms of precision, recall, and F1-score. The comparative evaluation demonstrates the effectiveness and robustness of the proposed framework in achieving improved fake review detection performance.Table 10Performance comparison of the proposed GEmFakeABSA model with baseline approaches.ArticlesMethodologyPrecisionRecallF1-Score86Topic Modelling-based Model80.478.779.5487DNN Model + Word Embedding and Emotion Mining81.9881.381.6388Syntax-aware Graph Neural Network78.7283.7881.1789RHGNN: Reinforced Heterogeneous GNNs81.482.3681.8721ABSA-based DNN Model81.3382.982.10ProposedNon-ABSA Hierarchical Attention (with VP)79.8280.2480.029ABSA-based Hierarchical Attention (without VP)82.0383.7682.88ABSA-based Hierarchical Attention (with VP)83.785.3584.51 Table 10Performance comparison of the proposed GEmFakeABSA model with baseline approaches.ArticlesMethodologyPrecisionRecallF1-Score86Topic Modelling-based Model80.478.779.5487DNN Model + Word Embedding and Emotion Mining81.9881.381.6388Syntax-aware Graph Neural Network78.7283.7881.1789RHGNN: Reinforced Heterogeneous GNNs81.482.3681.8721ABSA-based DNN Model81.3382.982.10ProposedNon-ABSA Hierarchical Attention (with VP)79.8280.2480.029ABSA-based Hierarchical Attention (without VP)82.0383.7682.88ABSA-based Hierarchical Attention (with VP)83.785.3584.51 Performance comparison of the proposed GEmFakeABSA model with baseline approaches. Table 10 exhibits comparative experimental evaluation of the proposed GEmFakeABSA model related to the presented baseline, state-of-the-art, and recent fake reviews classification approaches. The proposed GEmFakeABSA framework ABSA-based attention with VP metadata observed the optimal overall outcomes with precision 83.70%, recall 85.35%, and F1-score 84.52%, outperforming all benchmark approaches presented in Fig. 11. Relative to the baseline topic modelling-based approach86, GEmFakeABSA obtained +3.30% and + 6.65 % higher precision and recall, respectively. Likewise, as compared to the DNN-based model87, GEmFakeABSA achieves higher precision and recall as 1.72% and 4.05%, respectively. These improvements reflect aspect-based attention modelling identifies deceptive review patterns more effectively than standard semantic-driven approaches. The proposed GEmFakeABSA further outclasses recent graph-based models. As compared to Syntax-aware GNN88, our approach achieves +4.98% and +3.35% higher precision and F1-score, respectively, while sustaining higher recall values. Similarly, against RHGNN89, the proposed model achieves +2.30% and 2.99% higher precision and recall, respectively. These findings demonstrate that incorporating ABSA with behavioral VP metadata extracted richer classification-relevant information than only using graph-based structural modelling. The key findings from Fig. 11 are the role of VP features when integrating with the proposed GEmFakeABSA. The ABSA-based Attention model without features obtained 82.89% F1-score, whereas VP features raised the F1-score to +1.63%. Similarly, precision and recall are increased to +1.67% and +1.59%, respectively. These results highlight that incorporating VP features-based metadata significantly enhances the fake review detection performance.Fig. 11Comparison of proposed ABSA-based Hierarchical Attention model GEmFakeABSA.. Comparison of proposed ABSA-based Hierarchical Attention model GEmFakeABSA.. Ablation study Table 11 and Fig. 12 display the ablation study of the proposed GEmFakeABSA model to examine the functional role of each component on fake review classification performance. The full model observed the highest overall performance with precision 83.70%, recall 85.35%, and F1-score 84.52%, ensuring the impact of incorporating semantic, contextual, linguistic, and behavioural representation. A valuable performance degradation is noted when aspect-level features are not considered, reducing F1-Score by ↓4.49%, which indicates that Aspect-opinion information remains the key component in the proposed model to identify deceptive review patterns. Likewise, without attention mechanism F1-score is reduced to 80.80%, indicating the relevance of contextual attention in capturing meaning semantic features. The importance of VP features is also strong, where without VP decreases the F1-score by ↓1.63%. This ensures that VP features contribute to additional behavioural characteristics besides textual semantics. Similarly, deleting behavioural features lowers down the F1-score to 82.05%, signify their role in extracting suspicious reviewer activity patterns. Next, without GCN-based contextual embedding decreases the F1-score to 81.63%, demonstrating the graph-embedded contextual dependency enhances semantic representation. Moreover, removing linguistic features and POS tagging declined the F1-score to 81.96% and 82.50%, respectively, highlighting the contribution of linguistic and syntactic information. The highest decrease in performance observed was when POS, linguistic, and behavioural features are collectively excluded, where F1-score reduced by (↓5.62%), this reflecting that these features contribute additional insights and in-group improve the robustness of the proposed GEmFakeABSA.Table 11Ablation Study of the proposed model GEmFakeABSA.Variant of proposed model GEmFakeABSAPrecisionRecallF1-scoreWithout POS Features80.8884.1882.49Without Aspect-level Features79.8280.2480.02Without GCN-based Context Embedding80.3482.9681.62Without Linguistic Features82.281.7281.95Without Behaviour Features81.6682.4482.048Without Linguistics and Behavioural Features79.2481.1680.18Without POS and Linguistics and Behavioural Features77.8479.9878.89Without Hierarchical Attention79.5882.0680.80Without VP82.0383.7682.88Full Model83.785.3584.51 Table 11Ablation Study of the proposed model GEmFakeABSA.Variant of proposed model GEmFakeABSAPrecisionRecallF1-scoreWithout POS Features80.8884.1882.49Without Aspect-level Features79.8280.2480.02Without GCN-based Context Embedding80.3482.9681.62Without Linguistic Features82.281.7281.95Without Behaviour Features81.6682.4482.048Without Linguistics and Behavioural Features79.2481.1680.18Without POS and Linguistics and Behavioural Features77.8479.9878.89Without Hierarchical Attention79.5882.0680.80Without VP82.0383.7682.88Full Model83.785.3584.51 Ablation Study of the proposed model GEmFakeABSA. Fig. 12Model-level Ablation Study. Model-level Ablation Study. Collectively, the ablation study ensures that each component of the proposed framework plays an important role in improving performance, while their fusion enables the proposed GEmFakeABSA model to obtain optimal fake review classification performance. Significant testing To examine the consistency of the results achieved two-sided paired t-test significance testing was conducted against the proposed GEmFakeABSA model and the benchmark baseline models using F1-score. The analysis illustrates that the proposed approach achieved statistically significant performance enhancements\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} (\alpha < 0.05) \end{aligned}$$\end{document}over recent and state-of-the-art approaches, specifically, compared with ABSA-based DNN21 and RHGNN89 models. Persistent performance gains in precision, recall, and F1-score emphasizes the importance of incorporating GCN-embedded ABSA modelling with VP characteristics, which contributes effectively to the fake reviews classification model. Overall, the significance testing validates the generalization and robustness of the proposed GEmFakeABSA model. Effect of verified purchase on product classification accuracy Types of products were divided among search, confidence, and experience types based on the product category for further evaluation. Once more, we looked at how VP affected all three product kind’s effectiveness in detection. Conversely, the success rate of identifying the kind of search product for the proposed approach GEmFakeABSA secure was much lower, at just 74.32%, as contrasted with confidence 78.13% and experience product 76.20%. According to the Fig. 13, the GEmFakeABSA model’s accuracy improvement over the baseline technique21 is approximately 4%, 12% and 8% for confidence, search and experience, respectively. That being said, Fig. 14 illustrates how the results changed entirely when the VP features was encompasses. While the accuracy of identifying fake reviews improved by just 9.05% and 9.20%for confidence and search products, respectively, it climbed by 10.32% to 86.31% for experience product types. The accuracy increased by 2.3%, 1.21% and 2.92% for search, confidence and experience, respectively when VP features was introduced in comparison to the baseline21 strategy. Our findings also show that the VP characteristics considerably raises the fake review detection accuracy of every product category, particularly the confidence product groups showing the most substantial gains.Fig. 13Product classification accuracy without Verified Purchase (GEmFakeABSA and Base Model21.Fig. 14Product classification accuracy with Verified Purchase (GEmFakeABSA and Base Model21. Product classification accuracy without Verified Purchase (GEmFakeABSA and Base Model21. Product classification accuracy with Verified Purchase (GEmFakeABSA and Base Model21. Qualitative analysis In this section, the qualitative analysis was conducted to examine the contextual understanding and prediction capability of the proposed GEmFakeABSA model. In order to do further analysis to examine the impact of the proposed Graph embedded aspect-based fake review classification model GEmFakeABSA as architecture diagram illustrated in Fig. 1, a qualitative analysis is performed taking some review samples from the dataset. Distinct from quantitative analysis, qualitative analysis helps in interpreting how each architecture components like linguistic processing, POS concatenated enhanced WE, dependency-aware representation, contextual embeddings, hierarchical attention, and aspect-based features influence the fake review decision-making process. The proposed model GEmFakeABSA’s pipeline starts with sentence splitting, tokenization, WE with POS tagging, and dependency parsing for the preprocessing steps of the reviews. This pipeline phases help in converting review text into a structural form and extracting syntactic characteristics. The fusion of Word2Vec-based static embedding with POS characteristics empowers the model to combinedly extract the semantic understanding and grammatical meaning. This concatenation is especially useful in identifying genuine reviews, which particularly include aspect-level characteristics, from fake reviews that generally depend on overstated sentiment patterns. Further, the GCN-embedded component improves semantic dependency between aspect terms and opinion expressions. This enables an accurate association of sentiment with respective aspects, particularly when a review contains multiple aspects on different products. Furthermore, the attention mechanism incorporated in the proposed model significantly contributes to extracting informative patterns. The attention assigns higher importance to reviews holding well-defined aspect-opinion dependencies while filtering redundant and unimportant contents. This enhances robustness and interpretability in fake reviews classification predictions. Collectively, as presented in Table 12, the qualitative analysis ensures that the proposed model GEmFakeABSA’s accurately identify linguistic, syntactic, and contextual patterns for correct fake review classification. These insights are consistent in the qualitative findings observed from ablation study, significant testing and comparative analysis with baseline, state-of-the-art and recent fake review detection methodologies. A qualitative inspection of the dataset shows that reviews are generally related to product quality, pricing, usability, and delivery experience.Table 12Qualitative analysis of representative reviews.Review exampleClassificationQualitative observationPhone looks good but wouldn’t stay charged long.FakeRepetitive negative emphasis without detailed aspect justification identified as deceptive.Battery life is excellent for daily use but drains quickly during gaming.GenuineCorrectly captures different sentiments associated with battery usage aspects.Amazing product, best ever, highly recommended to everyone!!!FakeGeneric exaggerated sentiment without aspect-specific evidence detected as fake.Camera performs well in daylight, but low-light images are slightly noisy.GenuineAspect-level sentiment understanding improves contextual interpretation.Worst experience ever, I love this product so much!!!FakeContradictory sentiment expressions successfully identified as deceptive behavior. Table 12Qualitative analysis of representative reviews.Review exampleClassificationQualitative observationPhone looks good but wouldn’t stay charged long.FakeRepetitive negative emphasis without detailed aspect justification identified as deceptive.Battery life is excellent for daily use but drains quickly during gaming.GenuineCorrectly captures different sentiments associated with battery usage aspects.Amazing product, best ever, highly recommended to everyone!!!FakeGeneric exaggerated sentiment without aspect-specific evidence detected as fake.Camera performs well in daylight, but low-light images are slightly noisy.GenuineAspect-level sentiment understanding improves contextual interpretation.Worst experience ever, I love this product so much!!!FakeContradictory sentiment expressions successfully identified as deceptive behavior. Qualitative analysis of representative reviews. Strength, limitation and future work of the proposed model This section discusses the strengths, limitations, and future scope of the proposed GEmFakeABSA framework to provide a comprehensive understanding of its effectiveness, existing challenges, and potential research directions. The proposed GEmFakeABSA model efficiently incorporates semantic, syntactic, and behaviour features within the unified framework enables a semantic enriched representation learning for fake review classification. The attention mechanism enhances contextual understanding of multi-sentence reviews, whereas the fusion of aspect-opinion pair information improves classification performance. The POS+WE integration improved, embedding semantic representations improves features categorization effectiveness. Further, the fusion of the linguistic and reviewer behavioral characteristics enhance the model to extract additional information besides textual semantics individually. Examination results and ablation study highlights that in the proposed feature fusion mechanism each feature group contributes meaningfully thus strengthen positively towards precision, recall, and F1-score. However, the proposed work also has certain drawbacks. The proposed model GEmFakeABSA examined basically on benchmark datasets, and its efficiency can change and may not be consistent across heterogeneous real-world context and multilingual scenarios. The proposed model integrates behaviour and linguistic features; the performance still depends on labeled training dataset. In real-world scenarios, extracting large-scale correctly annotated fake-review datasets will be a real challenge. Further, the presented model largely depends on textual reviewer behaviour characteristics and ignores multimodal sources such as images and user communication patterns. Furthermore, the use of pre-trained Word2Vec embedding integrated with POS-based linguistic representation might struggle in comprehensively representing contextual semantic fine-grained relative to more recent transformer-based large language models90. Although propose model obtained comparable results with reduced computing complexity, incorporating enriched contextual embeddings still obtains improved performance. In addition, the GCN-embedded attention-based aspect-opinion pair extraction with feature fusion introduced additional computing overhead while utilizing it on large-scale, heavy real-time applications. Additionally, in online real-life user-generated reviews on product/service, opinions may be conveyed indirectly via figurative language such as sarcasm, irony, satire, and self-deprecating sarcasm instead of clear, explicit sentiment terms. These figurative patterns limit the effectiveness of the fake review detection using conventional sentiment extraction methods. However, the proposed GEmFakeABSA model by incorporating ABSA and contextual understanding better extract the important implicit sentiments expressed in the text for fake review detection. Still, explicitly modeling these figurative expressions is an open research problem for future studies. Moreover, the proposed model observed improved performance in fake review identification; still, it is not capable of resolving all the issues, some of the existing constraints of fake review classification persist. Progressive model tuning and task-specific training refinement may therefore be needed for long-term performance stability. Future scope As presented in the Experimental Setup and Results section, the proposed framework demonstrates promising classification accuracy through extensive quantitative and qualitative analysis, detailed dataset statistics, feature-level and model-level ablation studies, and statistically significant testing. Further, qualitative interpretability analysis and detailed critical insight discussion related to robustness, scalability, and generalization capability have been incorporated in the result discussion section. Furthermore, the research presented is restricted to the single-dataset evaluation whereas examination on different large-scale and domain-diverse datasets is a potential area of continued research for comprehensive verification of the proposed model’s generalizability. Moreover, future research may concentrate on extending the proposed GEmFakeABSA model to multilingual and out-of-domain fake review classification to improve adaptability in real-world environments. The incorporation of multimodal characteristics including visual data, associated metadata, and user interaction-driven information further improves the stability and reliability of fake review classification. Moreover, substituting static embedding (Word2Vec) with transformer-based BERT-driven contextual embedding contributes to deeper semantic understanding. Subsequent research may extend to exploring lightweight and scalable frameworks to encourage real-domain examination to enable continues adaptation to change the fake manipulation patterns. Finally, although the proposed ABSA-driven fake review detection model GEmFakeABSA improves performance in handling implicit sentiments, still explicit modeling of figurative sentiment expressions such as sarcasm and irony is an open research problem for for future enhancement of the proposed model. Future research may explore specialized mechanisms to better capture such implicit sentiment cues. Although GEmFakeABSA delivers strong in-domain accuracy, its evaluation is confined to a single Amazon category. Ongoing work will (i) benchmark cross domain transfer to YelpCHI and TripAdvisor, (ii) extend the graph encoder with transformer layers to capture longer reviews, and (iii) explore adversarial training to harden the model against AI-generated spam. Conclusion In the presented work, the proposed GEmFakeABSA model demonstrates a competent aspect-aware fake review detection technique by incorporating aspect-based sentiment analysis, GCN-based contextual semantic embeddings, and hierarchical attention mechanisms through a combined framework. The proposed work resolved the captured the research questions by evaluating the contribution of aspect-level opinion mining, semantic dependency learning, and contextual attention for fake review classification. The GEmFakeABSA model integrates POS with WE, contextual semantic extraction, graph-based dependency modelling, and hierarchical attention to identify fine-grained semantic and contextual dependencies. The fusion of ABSA and graph-based semantic learning improved the inter-aspect contextual relationships and deceptive opinion patterns effectively. Experimental analysis conducted on the Amazon Reviews dataset using benchmark performance metrics such as precision, recall, F1-score, and accuracy. Validated the performance of the proposed model over baseline, recent, and state-of-the-art approaches, which shows that the model outperforms the existing models. The findings demonstrate the substantial role of ABSA, GCN-based semantic embeddings, hierarchical attention, linguistic feature extraction, behavioral features, and verified purchase (VP) information for achieving better fake-review identification. The proposed GEmFakeABSA model, a hierarchical attention model with VP, obtained the optimal performance with 83.7% precision, 85.35% recall, and 84.51% F1-score. The implementation details, experimental setup and hyperparameter settings, and experimental configuration have been adequately presented and explained comprehensively to ensure reproducibility of the proposed GEmFakeABSA model. Although the GEmFakeABSA model presented significant improvement in the performance, it also increases computational complexity and training cost for large-scale datasets due to extracting aspect-opinion pairs using GCN-embedded attention-based aspect-level sentiment analysis. However, the proposed framework is currently evaluated on a single dataset, and future research can focus on large-scale and domain-diverse datasets to further validate its generalizability and robustness. Although the proposed GEmFakeABSA model improves implicit sentiment handling, effective detection of sarcasm and irony remains a challenging area for future enhancement. Future work can focus on developing lightweight graph-based architectures, multimodal fake review analysis incorporating textual and behavioral features, and domain adaptation strategies to further improve scalability, computational efficiency, and cross-domain generalization capability of fake review detection systems.