Supervised attention method with co-reference resolution for aspect extraction Authors: Ganpat Singh Chauhan, Shalini Puri, Ravi Nahta, Neeraj Garg Published: 2026 ABSTRACT AbstractAspect-based sentiment analysis (ABSA) of review texts remains a challenging problem due to the contextual interdependence of aspect evaluations across multiple sentences. Most existing approaches perform aspect extraction in isolation, limiting their ability to capture aligned semantics and multi-word (M-W) aspect terms in complex, multi-sentence (M-S) reviews. Furthermore, baseline and recent methods struggle to model long-range contextual dependencies, resulting in incomplete or fragmented aspect identifications. To address these limitations, this paper proposes a supervised hierarchical attention-based model for enhanced aspect term detection. The proposed model incorporates co-referencing resolution (CRR)-based sentence-alignment mechanism before ABSA, which resolves pronouns and referring expressions across sentences, generating semantically aligned inputs for the hierarchical attention network (HAN). This preprocessing step enhances contextual coherence, enabling the word-level and sentence-level attention mechanisms in HAN to more effectively capture aspect-related information across multi-sentence reviews. Furthermore, contextual representations are learned using word vectors enriched with semantic context, which are then processed through HAN using sequence-tagged labeled data. Experiments conducted on the SemEval-16 benchmark datasets demonstrate that the proposed model consistently outperforms strong baselines and state-of-the-art methods across the Laptop and Restaurant domains. The experimental evaluation shows that the proposed model improves F-Score by + 2.77% and + 3.52% over AspectGCN, and by + 4.16% and + 2.32% over IDGNN + BERT on the laptop and restaurant review datasets, respectively. It highlights that the inclusion of the CRR module yields an additional 3–4% gain over the HAN baseline, strengthening the effectiveness of the context-aware and alignment-driven modeling for aspect term extraction (ATE) from M-S reviews.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-52992-9.Keywords: Aspect-based sentiment analysis, Hierarchical attention, Co-reference resolution, Sustainable learning, Sustainable development, Aspect term extractionSubject terms: Engineering, Mathematics and computing Introduction The rapid growth of e-commerce platforms and online services has led to a substantial increase in user-generated reviews on the internet1,2. These reviews contain valuable opinions about various aspects of the product and services, making them an important source for understanding customer preferences and experiences. Consequently, Aspect-Based Sentiment Analysis (ABSA) has emerged as an important technique for identifying specific aspects mentioned in reviews and analyzing the sentiment associated with them3,4. Consider the review “The food is delicious, but the ambiance is uncomfortable” as shown in Fig. 1. The customer expresses confidence in the quality of the meal because the ASPect CATegory (ASP-CAT) is “FOOD#QUALITY” but disapproves of the ambiance using the “SERVICE#GENERAL” as ASP-CAT. So, the ATE identifies explicit single-word and M-W Aspect Terms (AT) in the sentence that are “food” and “ambiance”. In contrast, ACD maps each extracted and implied aspect to a predefined category, such as “FOOD#QUALITY” and “SERVICE#GENERAL”. Here, categories do not need to appear explicitly in the sentence. They represent semantic groupings defined by the task or domain, such as Restaurant domain reviews5. An example of ATE and ACD. Every textual review generally follows an underlying pattern. Figure 2 illustrates examples of background and elaboration relations in text. The background example shows an entire reference where the pronoun links back to the main subject, while the elaboration example provides additional descriptive details that expand on previously mentioned aspects6–8. This example highlights how contextual information is structured and connected in discourse. The sentiment of the present review is assessed by learning about the association between sentences and the surrounding sentences’ sentiments. It is more likely that the sentiment can remain the same throughout the phrases if the attitude towards the component is good in any statement9. To extract the valuable features and categories, it is essential to comprehend the context of the debate for the connections between and within sentences. Feedback-based aspect extraction should identify single-word and M-W aspects from large feedback sentences10. To extract the important elements, it ascertains the dependencies inside and across sentences11. Background and elaboration example. ABSA needs semantic data for the Aspect Extraction Task (AET)12,13. Online comments do not adhere to these rules when expressing opinions regarding any region of the task domain. In existing research contributions on sentiment analysis, it is found that ABSA techniques fall short in their ability to extract intra- and inter-sentence links between various review components. Since most reviews today consist of multiple sentences, the inability of discourse reviews to use aspect co-referencing leads to confusion over which ASP-CAT should be chosen. Table 1 shows the ambiguity, predicted ASP-CAT, and actual ASP-CAT for various reviews where represents the “review sentence” for i = 1 and 2. Deep Learning (DL)-based approaches articulate the linguistic and interpretative links through words to construct intra-sentence links using Bidirectional Encoder Representations from Transformers (BERT)14. However, conventional ABSA methods take isolated sentences, failing to capture the contextual interdependencies across sentences within a single review. This results in fragmented or incomplete ATE, especially when aspects are implied, referenced indirectly, or expressed as M-W phrases scattered across multiple sentences. Table 1Examples illustrating ambiguity in ACD brought up by inter-sentence relationships.ReviewPredicted ASP-CATActual ASP-CAT1S 1:Long Power backup with pleasant voice[LAPTOP#GENERAL][SOUND#QUALITY] [BATTERY#GENERAL] [SERVICE#GENERAL]S 2:I recommend it!2S 1:The burger is yummy[RESTAURANT#GENERAL][FOOD#QUALITY]S 2:Don’t ignore to buy it3S 1:I had food here few days back[RESTAURANT#GENERAL][FOOD#QUALITY]S 2:It was really tasty4S 1:The service here is very pleasant[RESTAURANT#GENERAL][SERVICE#GENERAL]S 2:I liked it and definitely revisit Examples illustrating ambiguity in ACD brought up by inter-sentence relationships. Need for sentence alignment In real-world opinion mining scenarios, user reviews frequently consist of multiple sentences, where ATs are often introduced in one sentence and further described, contrasted, or evaluated in subsequent sentences. Despite notable advancements in ABSA, several challenges remain unresolved, particularly in the context of M-S review and context-dependent reviews. Existing sentence-level and HAN-based approaches primarily process sentences independently or aggregate them sequentially, which can lead to context fragmentation and misalignment of aspect-related information. Moreover, the extraction of M-W and domain-specific aspects remains a persistent challenge due to the insufficient modeling of contextual interactions. Sentence-alignment strategies and their integration with the HAN mechanism have received limited attention in prior research. Furthermore, the role of CRR in supervised ATE has not been systematically investigated. To address these limitations, a sentence-alignment mechanism is introduced using CRR that explicitly captures cross-sentence semantic relationships, enabling more accurate ATE in the M-S reviews. The sentiment of the most recent review was evaluated by analyzing the association between individual sentences and the sentiment expressed in adjacent sentences. It is crucial to address CRR to gain an insightful analysis in ABSA. The proposed model executes this process in two steps. First, it incorporates CRR for review, which involves aligning sentences to provide context that helps clarify any ambiguity in AET. CRR connects pronouns and refers to the specific entities or aspects mentioned in the text. It identifies M-W aspects like “battery backup”, “customer service”, or “screen resolution”. In this context, CRR ensures that references to various aspects are accurately interpreted. M-W aspects are typically introduced once and subsequently referenced using pronouns or general nouns. So, CRR allows the system to recognize that “it” in the sentence “It lasts all day” refers to “battery life” and improves aspect detection continuity. The proposed model establishes connections among various elements to enhance the user’s understanding of reviews. It is trained on labeled data using Begin-Inside-Outside (BIO) encoding, which improves aspect word prediction13. Its performance is evaluated on a benchmark dataset55 (described in Sect. "Dataset description"), showing that the addition of contextual information reduces uncertainty caused by insufficient data, thereby enhancing the ASP-CAT and AET. The key contributions are given below. Proposing a supervised HAN-based framework for context-aware ATE from M-S reviews. Proposing a supervised HAN-based framework for context-aware ATE from M-S reviews. Introducing a novel sentence-alignment preprocessing strategy using CRR that aligns semantically related aspect expressions across sentences and reviews before ABSA. Introducing a novel sentence-alignment preprocessing strategy using CRR that aligns semantically related aspect expressions across sentences and reviews before ABSA. M-W and domain-specific aspect extraction enhancement by integrating contextual embeddings with sequence-tagged supervision. M-W and domain-specific aspect extraction enhancement by integrating contextual embeddings with sequence-tagged supervision. Analyze the impact of CRR semantically on supervised HAN-based ATE performance. Analyze the impact of CRR semantically on supervised HAN-based ATE performance. Validation of the proposed approach on the SemEval-16 benchmark datasets (explained in Sect. “Dataset description”), achieving consistent improvements over state-of-the-art methods. Validation of the proposed approach on the SemEval-16 benchmark datasets (explained in Sect. “Dataset description”), achieving consistent improvements over state-of-the-art methods. Novelty of the proposed work Unlike existing HAN-based ABSA approaches, which primarily operate at the sentence-level or rely solely on syntactic graph representations, the proposed HAN-based method introduces a sentence-aligned, supervised HAN framework that jointly models words, sentences, and review-level contextual semantics. The novelty of the work lies in the integration of sentence-alignment using CRR as a preprocessing step for AET, which enables consistent semantic mapping of ATs across M-S reviews. This aspect has been largely overlooked in prior studies. Furthermore, the proposed framework systematically incorporates and evaluates CRR within an HAN architecture, demonstrating its impact on supervised ATE. This combination of alignment-driven preprocessing, hierarchical contextual modeling, and co-reference-aware analysis distinguishes the proposed approach from existing attention-based, GCN-based, and hybris methods. From a sustainable perspective, the proposed CRR-based framework promotes model reuse across both domains by effectively leveraging contextual relationships, thereby reducing the need for extensive retraining and the scalability of annotated data. This is how the paper is structured: Sect. “Related work” discusses the related contributions on the ATE. The suggested approach for the subtasks of category detection and aspect extraction is then shown. This discusses sentence alignment for ACD using CRR and the ASP-CAT detection process. Subsequently, the article covers the hierarchical review-level model stages, which include BIO labeling, Part-Of-Speech (POS) tagging, and WE as input functionalities for the network. The proposed network architecture presented in Sect. “The proposed approach of aspect terms extraction” utilizes sentence-alignment-based HAN for AET. Section “Analysis of the findings” analyzes the findings of the proposed model, detailing the dataset used and the evaluation metrics employed. It evaluates the proposed work’s performance compared to other baselines and state-of-the-art approaches. Next, the article presents ABSA applications in Sect. “Applications of ABSA” and conducts significance testing for the proposed model in Sect. “Significance testing of the proposed model”. A case study is then provided to discuss the PREcision (PRE), RECall (REC), and F-score achieved using the proposed model in Sect. “Case study: discussion on precision, recall, and F-score”. The final Sect. “Conclusions and future recommendations” concludes the article and offers future recommendations. Related work ABSA has been extensively studied over the past decade with particular emphasis on ATE and ACD as fundamental subtasks2,15,61. Early approaches primarily relied on rule-based techniques, frequency statistics, and topic modeling to identify ATs from review texts16. The only aspect phrases that could be extracted using these rule-based methods are explicit ones. While ASP-CAT recognition addresses implicit and explicit aspects, the ATE task concentrates on explicit aspects. However, these methods were limited in handling linguistic variability, contextual ambiguity, and M-W aspect expressions. Next, CRR between ATs is not extracted using rule-based approaches17,18. Therefore, rule-based approaches disregard contextual information while identifying aspect categories from the reviews. Subsequent research introduced using supervised learning techniques and DL models has improved performance19,20,52; however, since a review having numerous phrases must be evaluated, they produce subpar aspect extraction results22. Through the capturing of the review’s deeper significance, the DL-based approaches presented in21 have enhanced ATE performance. A convolutional network introduced to unexplored languages significantly improved the performance of aspect detection. The accuracy of identifying aspect categories has been greatly enhanced by incorporating POS and WE into the network model14,23. This approach utilized a hybrid strategy to extract the most salient contextual features, extending aspect extraction capabilities12. A DL network for ACD is trained using a sentence vector representation approach for word vectors14. Over the past decade, multiple LSTM-based methods27,28 have been proposed for ATE, particularly Laptop and Restaurant samples of the SemEval-16 reviews29–35. LSTM networks are widely used in ABSA as they effectively capture long-range dependencies and contextual information in sequential text data. Recurrent architectures, such as Bi-LSTM and GRU, utilize contextual dependencies among sentences by modeling sequential information57,58. However, LSTMs face limitations, such as challenges in capturing global text context, a lack of parallelization, difficulty handling multiple aspects simultaneously, and sensitivity to input length. Bi-LSTM with attention suffers from the exponential memory delay and discrete attention weighting, which limits its ability to capture long-range and cross-sentence semantic dependencies, leading to over-focusing on high-frequency sentiment cues while missing implicit or low-occurrence aspects27. These approaches leverage sequence labeling and contextual representation to capture local semantic dependencies within sentences. However, the majority of existing models operate at a single-sentence level, implicitly assuming that aspect evaluations are independent across sentences, thereby overlooking inter-sentence contextual dependencies present in real-world reviews. Recent studies have increasingly adopted transformer-based architectures for ABSA due to their ability to capture long-range contextual dependencies. This section discusses the recent models, specifically published in 2025, based on BERT, RoBERTa, and DeBERTa, which have demonstrated strong performance across multiple ABSA benchmarks. These models utilize transformer encoders to obtain contextual word representations, which enhance aspect detection and sentiment classification61. More recent studies have explored BERT-based2,10,43, hierarchical22,34,61, and attention-based mechanisms32,40 to model sentence-level importance and contextual relevance. A hybrid ABSA model, SVCCL-GCN50, is a strong recent baseline because it combines BERT + Bi-LSTM to obtain contextual embedding, GCN built on syntactic dependency trees to capture relationships between words, a self-attention based variational encoder to capture global and local semantic information, and a capsule network with contrastive learning to improve feature discrimination between sentiment classes. This architecture enables the model to capture long-distance dependencies and both syntactic and semantic information for better AT extraction73,75. 58 proposed an adaptive contextual memory graph transformer-based contextual embedding with a domain-adaptive knowledge graph. This model utilizes graph transformer layers with memory modules to capture semantic relations between aspects and opinion words, thereby enhancing contextual interaction modeling for ABSA2,45,57. Although these models enhance representation learning, they largely focus on document-level sentiment aggregation rather than explicit aspect alignment across sentences or reviews24–26,51. BERT and its variants introduce contextual embeddings, address these limitations by generating dynamic word representations conditioned on surrounding context, and improve token-level semantics59,60,63. Despite these improvements, transformer-based models primarily focus on linear contextual dependencies and do not explicitly model syntactic and semantic relational structures among words across multiple sentences64. 65,74 proposed an LLM-augmented syntax-aware domain adaptation framework for aspect extraction that integrates contextual knowledge from large language models with dependency-based syntactic features. The model utilizes domain adaptation mechanisms to align representations across domains, thereby enabling improved aspect extraction in cross-domain settings.72 proposes a self-attention-driven transformer-based sentiment classification with speaker intent (TASCI) framework that integrates aspect prediction with intent understanding61,62. In recent years, to better exploit syntactic and relational information, graph-based neural networks, particularly Graph Convolutional Networks (GCN), have been incorporated into ABSA frameworks. These approaches model sentences as dependency graphs, enabling the capture of long-range syntactic dependencies between words and aspects61,66,67. GCN-based models effectively capture non-local syntactic relations and structural dependencies, leading to improved aspect boundary detection68. The combination of BERT with GCN further enhanced performance by jointly modeling contextual semantics and structural information69. GCNs are robust to sentence complexity and naturally support cross-sentence aspect referencing, enabling improved contextual understanding39. GCN models that leverage BERT along with many weight mechanisms36–38, BERT-based maximal activated weighted memory39, and components focusing on multi-head co-attention40, have significantly benefited ABSA. These methods facilitate the extraction of four aspects using a two-stage neural network41,44,53,54. GCN-based models have demonstrated improvements in ATE by leveraging structural linguistics information; however, they typically rely on sentence-level dependency parsing and remain limited in modeling cross-sentence semantic alignment and review-level contextual dependencies61. In parallel, unsupervised and hybrid approaches have been proposed to reduce dependency on labeled data by exploiting dependency relations, topic coherence, and graph-based representations. While such methods demonstrate reasonable performance, they face limited generalization and reduced accuracy when handling complete contextual interactions. Additionally, sentence-alignment and preprocessing strategies aimed at eliminating contextual discontinuities are seldom integrated with GCN-based or attention-driven ABSA pipelines61,70,71. Many research works on ABSA and AET methods have been proposed to date. Table 2 compares various ABSAs from 2015 to 2025 based on the problem-focused approach and techniques applied. They have used several methods for sentiment analysis, such as BERT, LSTM, Bi-Directional LSTM (Bi-LSTM), Recurrent Neural Network (RNN), DL, Convolutional Neural Networks (CNN), Deep CNN (DCNN), Global Vectors (GloVe) for word representation, GCN, dual Graph Neural Networks (GNN), Artificial Neural Network (ANN), RCut algorithm, Linear Conditional Random Fields (CRF), Natural Language Processing (NLP)-based sentiment analysis, dyadic memory networks, deep Gated Recurrent Unit (GRU), Multi-Layer Perceptron (MLP), association rule mining, attention-based DL, regularized neural attention model, aligning aspect embeddings, multiple Word Embedding (WE) method, Large Language Model (LLM), memory network, transformation network, multi-task learning, and two-stage neural model. Table 2Comparison among various existing ABSA methods.Ref. NoYearProblem focusedTechniques applied172015Combining web-based similarity metrics with frequency-based techniques to extract aspectsRCut Algorithm202015Aspect-based opinion mining from product reviewsLinear CRF232015Fine-grained opinion mining with WEsRNN122016Multilingual ABSADL142016Aspect extraction in opinion miningDCNN222016ABSA hierarchical modelBi-LSTM242016ABSAANN282016ABSA through ANN feature integrationANN292016Sentiment analysisDL12017Cross-domain opinion target extraction using auxiliary labelsRNN32017YouTube video retrieval via sentiment analysis of user commentsNLP-based ABSA252017Classifying aspects of restaurant evaluations and extracting termsLSTM262017Sentiment analysis using classification of sentence typesBi-LSTM-CRF and CNN272017ABSADyadic memory networks132018ATE with opinion targetsDeep GRU442018Supervised and unsupervised detection of aspect categories using co-occurrence dataAssociation rule mining312019ABSA with alternating co-attention networksLSTM102020AETAttention-based DL model352020Sentence-level ABSA using lexicalized domain ontologyRegularized neural attention model362020ABSAAligning aspect embeddings422020Multi-embedding attention-based model for cross-domain ATEMultiple WE methods302021Self-supervised attention learning for enhanced ABSAMemory network, BERT, transformation network322021Aspect and context for ABSA using co-attention networksGloVe and BERT332021ABSA via a position-aware approachCNN342021End-to-end frameworks for ABSAMulti-task learning372021Sentiment-aware personality detection from textDL382022ABSA via multiple weight mechanismsGCN392022ABSA with maximal activation weighted memoryBERT402022ABSA via multi-head co-attention networks for component-focusedBERT412022Full ABSA quadruple extractionTwo-stage neural model432022Unsupervised aspect extractionBERT112023Implicit AET using a weakly supervised learningBERT452024Hierarchical attention for fine-grained AETLSTM462024ABSA through aspect–sentence relationshipsGCN472024ABSA with syntactic and semantic informationDual GNN482024Multimodal ABSAGCN602024ABSA with BERT-driven context generation and quality filteringBERT-based662024TCKGCN: GCN for ABSA with three-channel knowledge fusionGCN with knowledge fusion672024A knowledge-augmented heterogeneous GCN for ABSAHeterogeneous GCN492025Information-assisted, sentiment relation-driven approach for ABSAMulti-task learning502025ABSAGCN512025Boosting cross-document event CRR using discourse and structural cuesMLP572025ABSA based on deep implicit feature extractionCABiLSTM-BERT582025Adaptive contextual memory graph transformer with domain-adaptive knowledge graph for ABSAGraph-Based592025Hierarchical and position-aware GCN with external knowledge and prompt learning for ABSAHierarchical GCN632025Multi-Feature fusion graph attention network for ABSAGraph Attention642025Adaptive contextual memory graph transformer with domain-adaptive knowledge graph for ABSAMemory and Knowledge Graph652025LLM augmented syntax-aware domain adaptation ABSA methodLLM-based682025Syntactic dependency enhanced and integrated GCN for ABSASDE-GCN692025Semantic enhanced bi-syntactic GCN for ABSABi-syntactic GCN702025Augmented dynamic GCNs by semantic and syntactic dependenciesDSS-GCN722025Transformers for ABSA with contextual intent integrationIntent-based Transformer732025GCN model based on transformer and frequency domain for ABSAGCN-based Transformer742025ABSA used a permutation language-based transformer modelLanguage Transformer752025Target-oriented Cross-Modal Transformer for Multimodal ABSATarget-oriented Transformer Comparison among various existing ABSA methods. It was observed that most researchers utilized DL, hybrid, BERT, attention, HAN, and GCN-based techniques for ABSA and ATE. Most current ATE and ACD methods perform well on the SemEval-14, but their effectiveness declines for SemEval-16 samples (described in Sect. “Dataset description”). They primarily applied the SemEval-14 and SemEval-16 datasets, where the Restaurant data yielded better aspect extraction accuracy than the Laptop domain. Furthermore, it was determined that extracting the context of the reviews was essential for a complete understanding of the dataset reviews46–50. This required replacing all personal pronouns with reference nouns to clarify the context of the previous reviews. By simulating the CRR with the ATs, a hierarchical model improved the results of the ATE and ACD subtasks22. Still, this hierarchical approach struggled to record the relationships between sentences in long evaluations. While intra-sentence relationships are no longer problematic, capturing inter-sentence linkages that connect various sentence aspects remains an unsyntactic challenge45. Finally, CRR has been acknowledged as a useful linguistic feature for improving sentiment interpretation; however, its role in supervised ATE has not been systematically analyzed, particularly in conjunction with HAN and context-aware modeling frameworks. In contrast to existing attention-based and GCN-driven approaches that emphasize sentence-level syntactic and semantic dependencies, this work focuses on pre-alignment of contextual information across sentences and reviews, enabling HAN to operate on semantically consistent input representations. Problem statement and contributions Existing aspect extraction approaches, such as Bi-LSTM with attention and standard HAN, primarily operate at the sentence-level or perform sequential aggregation of sentence representations. As a result, they struggle to accurately extract ATs from M-S reviews, where aspect information is often distributed across multiple sentences and requires alignment of semantically related contexts. The core problem addressed in this work is how to effectively model and align cross-sentence dependencies to improve aspect term detection in M-S reviews. To the best of our knowledge, this is one of the first studies to explicitly integrate sentence-alignment and CRR within a supervised HAN for ATE. Motivated by the gaps identified in prior studies, this work aims to advance ATE in ABSA through the following objectives and contributions. To propose a novel supervised HAN-based framework that captures word, sentence, and review-level contextual semantics for robust aspect term detection in M-S reviews. To propose a novel supervised HAN-based framework that captures word, sentence, and review-level contextual semantics for robust aspect term detection in M-S reviews. To introduce sentence-alignment using CRR as a preprocessing mechanism for ABSA, enabling the alignment of semantically related aspects across sentences and reviews, which is largely unexplored in existing AET literature. To introduce sentence-alignment using CRR as a preprocessing mechanism for ABSA, enabling the alignment of semantically related aspects across sentences and reviews, which is largely unexplored in existing AET literature. To enhance the extraction of M-W and domain-specific ATs by leveraging contextual embedding and sequence-tagged supervision within an HAN architecture. To enhance the extraction of M-W and domain-specific ATs by leveraging contextual embedding and sequence-tagged supervision within an HAN architecture. To systematically investigate the impact of CRR as ATE, demonstrating its effectiveness when integrated with alignment-aware and context-sensitive models. To systematically investigate the impact of CRR as ATE, demonstrating its effectiveness when integrated with alignment-aware and context-sensitive models. To empirically validate the proposed approach on benchmark SemEval-16 datasets (described in Sect. “Dataset description”), achieving consistent and significant F-score improvements over state-of-the-art and baselines supervised as well as hybrid unsupervised methods across multiple domains. To empirically validate the proposed approach on benchmark SemEval-16 datasets (described in Sect. “Dataset description”), achieving consistent and significant F-score improvements over state-of-the-art and baselines supervised as well as hybrid unsupervised methods across multiple domains. The proposed approach of aspect terms extraction This section introduces the proposed HAN model, which extends conventional sentence-level and HAN approaches by explicitly modeling cross-sentence dependencies. Unlike Bi-LSTM with attention, which applies attention over a single sentence sequence, the proposed HAN-based deep network explicitly captures inter-sentence dependencies through sentence-alignment. In contrast to standard HAN, which hierarchically aggregates sentence representations with explicit semantic alignment, our approach introduces a sentence-alignment mechanism. This design enables iterative refinement of aspect representation using aligned cross-sentence context, making the model particularly effective for M-S aspect extraction. The system develops contextual embeddings using BERT to represent the semantics of words across various task domains. Extracting domain-relevant ATs requires an understanding of the overall context of the reviews. The methodology is illustrated in Fig. 3, which shows a HAN integrated with CRR for aspect extraction in customer reviews. The following is a high-level step-by-step description of each component of the methodology as shown in Fig. 3. The subsequent subsections explain the methodology, roles, and interaction within the overall framework. The input consists of a multi-sentence review. For instance:Sentence 1: “Long power backup with pleasant voice.”Sentence 2: “I recommend it.” The input consists of a multi-sentence review. For instance:Sentence 1: “Long power backup with pleasant voice.”Sentence 2: “I recommend it.” Sentence 1: “Long power backup with pleasant voice.” Sentence 1: “Long power backup with pleasant voice.” Sentence 2: “I recommend it.” Sentence 2: “I recommend it.” CRR is applied to resolve pronouns or references, such as “it.” Consequently, the second sentence becomes: “I recommend the laptop.” The pronouns are not directly substituted instead, a mapping is used between pronouns and their respective antecedents. This step enhances clarity and context for downstream tasks by explicitly replacing ambiguous terms. CRR is applied to resolve pronouns or references, such as “it.” Consequently, the second sentence becomes: “I recommend the laptop.” The pronouns are not directly substituted instead, a mapping is used between pronouns and their respective antecedents. This step enhances clarity and context for downstream tasks by explicitly replacing ambiguous terms. The ATs in the co-referenced sentences are annotated using BIO tagging: “Long power backup” is tagged as B I O, where each token is labeled as B-ASP, I-ASP, or O, enabling sequence labeling for ATE. The ATs in the co-referenced sentences are annotated using BIO tagging: “Long power backup” is tagged as B I O, where each token is labeled as B-ASP, I-ASP, or O, enabling sequence labeling for ATE. The co-referenced reviews are then passed through a BERT encoder to generate contextual embeddings for each token. The co-referenced reviews are then passed through a BERT encoder to generate contextual embeddings for each token. The model employs two levels of attention based on a HAN model:Sentence-Level Attention: Captures the importance of words within each sentence to create a sentence-level context vector.Review-Level Attention: Aggregates sentence-level representations to model the overall review, incorporating cross-sentence information, which is essential in M-S reviews. The model employs two levels of attention based on a HAN model:Sentence-Level Attention: Captures the importance of words within each sentence to create a sentence-level context vector.Review-Level Attention: Aggregates sentence-level representations to model the overall review, incorporating cross-sentence information, which is essential in M-S reviews. Sentence-Level Attention: Captures the importance of words within each sentence to create a sentence-level context vector. Sentence-Level Attention: Captures the importance of words within each sentence to create a sentence-level context vector. Review-Level Attention: Aggregates sentence-level representations to model the overall review, incorporating cross-sentence information, which is essential in M-S reviews. Review-Level Attention: Aggregates sentence-level representations to model the overall review, incorporating cross-sentence information, which is essential in M-S reviews. The final representation is processed through a softmax classifier to predict BIO labels for each token. These labels correspond to the identified ATs in the review. The final representation is processed through a softmax classifier to predict BIO labels for each token. These labels correspond to the identified ATs in the review. The model outputs the ATs recognized from the review, such as “power backup,” “voice,” and “laptop.” The model outputs the ATs recognized from the review, such as “power backup,” “voice,” and “laptop.” The proposed CRR-based Hierarchical Attention method for ATE. The algorithm and its step-by-step mathematical formulation are available in the supplementary document for the readers. Sentence alignment using CRR for ATE and ACD Although CRR is emphasized in many NLP-related domains, it has not been extensively explored for ATE and ACD in ABSA. For a better understanding, placeholder symbols like Noun Phrase as (NP) or a set of Noun Phrases as (NPs) are defined and replaced. CRR identifies all nouns and NPs related to the same entity within the textual review. It captures the context of the entire review to enable accurate aspect and category detection. In the proposed framework, co-reference resolution is applied during the preprocessing stage to resolve pronouns and referring expressions across sentences. The pre-trained language model en_core_web_sm-2.3 is used for tokenization, part-of-speech tagging, and dependency parsing. The CRR process is implemented using the neural coreference resolution module available in the spaCy NLP library (version 2.3.5) with the NeuralCoref (version 4.0) extension, which identifies mentions referring to the same entity and establishes mappings between pronouns and their corresponding antecedents based on coreference clusters generated by NeuralCoref. The NeuralCoref module is integrated into the spaCy pipeline using its default configuration for CRR. The proposed framework relies on the default predictions generated by the NeuralCoref module. To ensure consistency with the pre-trained model and avoid introducing additional hyperparameters, no additional filtering or explicit confidence threshold is applied to the resolved mentions. Applying a strict confidence threshold may discard moderately scored yet contextually relevant coreference links, potentially affecting the quality of downstream representation. Every cluster maintains all the mentions referring to the same aspects. In cases where multiple antecedents exist for a pronoun, the NeuralCoref module resolves the ambiguity (no manual disambiguation performed in the proposed model) using its internal scoring mechanism based on contextual and syntactic features. Specifically, the model assigns a compatibility score to each candidate antecedent within the cluster and selects the antecedent with the highest score for resolution. This best-fit selection strategy ensures a single coherent linkage, but may overlook secondary plausible antecedents in complex contexts. The coreference model performed as-is without any fine-tuning on the domain-specific dataset. This step generates semantically aligned sentences that are subsequently provided as input to the HAN for aspect term extraction. However, general-purpose coreference models may not fully capture domain-specific linguistic patterns such as implicit references or context-dependent entity relations, which may affect resolution quality in specialized review datasets. Incorporating domain-adaptive fine-tuning could further improve performance and is considered future work. Figure 4 illustrates an example of a multi-sentence CRR review, utilizing the “NeuralCorefV4.0” library alongside spaCy’sV2.3.5 NLP tools. Table 1 indicates that the expected category for Sentence 2 (S2) of Feedback 1 was [RESTAURANT#GENERAL]. CRR is essential for identifying all references, including nouns and NPs, related to the same aspect in the review. Table 3 provides examples of how applying CRR can reduce ambiguity in ACD caused by inter-sentence connections. For example, the phrase “Long Power Backup with pleasant voice” appears in the assessment. Example for CRR in a multi-sentence review. Table 3Qualitative examples demonstrating the effectiveness of CRR in mitigating inter-sentence ambiguity for ACD.Review/FeedbackSentence 2 (S2) after CRR1S 1:Long Power backup with pleasant voiceI recommend the laptop for battery and soundS 2:I recommend it!2S 1:The burger is really yummyDon’t forget to buy anything without BurgerS 2:Don’t forget to buy anything without it3S 1:I had food here few days backThe food was really tastyS 2:It was really tasty4S 1:The service here is very pleasantI liked the service and definitely revisitS 2:I liked it and definitely revisit Qualitative examples demonstrating the effectiveness of CRR in mitigating inter-sentence ambiguity for ACD. Co-referential expressions may affect how ATE in ABSA is carried out at the review level. The proposed method addresses the co-reference while enhancing the existing annotations to utilize ATE and ACD. However, the identical sentence will be changed to read as follows when CRR is executed: “Good meal, outstanding look, and significant assistant. The restaurant’s food, atmosphere, and service are all excellent”. The anticipated feature categories for S2 are now [FOOD#QUALITY], [AMBIENCE#GENERAL], and [SERVICE#GENERAL]. Currently, the primary sub-element of ABSA is identifying aspects from the text. Each word in the NP of a multi-review comment may not necessarily belong to just one feedback sentence56. Instead, these sentences could include NPs that are dependent upon the context of the preceding or following sentences in the review. Furthermore, each POS-tagged single word or M-W component can be present within the phrase. The accuracy of the Stanford Parser tool is crucial for completing ATE tasks. In recent years, online users have expressed their thoughts without adhering to grammatical or linguistic standards; as a result, AETs may extract incorrect nouns and NPs or fail to recognize valid aspects. For instance, in the evaluation “Long power backup,” the word “Long” occurs as part of the M-W entity “power backup” still, the output from the Stanford Parser56 might not accurately reflect this connection, where ‘JJ’ represents the tag for adjectives. Long/JJ power/JJ backup/NN. The word “power” is one of the words in the phrase “power backup” for this sentence, but the parser marks it as an adjective. The sentence co-reference partially addresses the previously mentioned problems, but can be further enhanced if NPs are recognized at the chunk level56. Figure 5 shows an example with the “Long power backup” and “pleasent voice” NP chunks. The aspect extraction process utilizes Linguistic Rules (LR) to identify ATs. In the algorithm referenced by16, the inputs R, L, and O stand for reviews, LRs, and opinion words, respectively. The algorithm captures the relationship between nouns and adjectives, producing N ATs. Within each NP chunk, the algorithm searches for the word, , which is a personal pronoun. If is found as a personal pronoun, CRR is applied to the sentence, as described in Sect. “Sentence alignment using CRR for ATE and ACD”. Next, the algorithm checks if the chunk contains any nouns and whether its preceding word is an adjective (e.g., “yummy Burger”). If present in the set of opinion words then (yummy) will be removed. Conversely, if is an adjective, but not found in (e.g., “pleasent evening ambiance”) Then the NPs’ “evening ambiance” will be extracted as ATs. Accordingly, if is a verb (e.g., “hot paw”), will be treated as a single-word aspect. The next step states that if is an article (e.g., “the tablet”) or a pronoun (e.g., “the screen of my tablet”), it will be discarded. Sentence-alignment on multi-sentence review using CRR. Furthermore, if there is an association between every noun and any word (such as in the example “The speaker has a high sound”), the NP “speaker sound” will be extracted as NPs. In this approach, denotes an extracted noun or NP from the sentence. Next, the extracted aspects for all reviews may include irrelevant terms. For example, in the feedback “Excellent food with pleasent evening ambiance”, the word “evening” is part of the extracted aspect “evening ambiance,” but it has very low relevance to the Restaurant domain. Therefore, the system employs word vector representations to filter out irrelevant ATs. Aspect category detection Assigning an aspect, such as “battery,” to the category “laptop” and an attribute, like “backup,” to the ASP-CAT ENTITY#ATTRIBUTE (LAPTOP#BATTERY) is the task of ACD. For example, in the English language review, “The Burger is yummy and at nice location,” the aspect “Burger” is related to the ASP-CAT “FOOD#QUALITY.” Similarly, a single feedback may include several aspect categories such as FOOD#QUALITY and AMBIENCE#GENERAL. Table 4 provides example sentences for various ASP-CATs and ATs. It explains that certain phrases may present contrasting viewpoints on different topics and identify multiple classifications for the same topic. For instance, in the sentence “It is quite unsure about the food quality and ambiance of the restaurant,” the default ASP-CAT assigned is [RESTAURANT#GENERAL]. The categories of the supplied aspects include ‘service,’ ‘food,’ ‘price,’ ‘ambiance,’ and ‘general.’ Table 4Example reviews illustrating aspect categories and corresponding ATs.ReviewsASP-CATATsS 1:Great food, pleasent look, and remarkable service[FOOD#QUALITY][AMBIENCE#GENERAL][SERVICE#GENERAL]fooddecorserviceS 2:It is quite a dissatisfaction with the restaurant[RESTAURANT#GENERAL]RestaurantS 3:The service was quite presentable[SERVICE#GENERAL]Service Example reviews illustrating aspect categories and corresponding ATs. [FOOD#QUALITY] [AMBIENCE#GENERAL] [SERVICE#GENERAL] food decor service The experiments assumed that each sentence contained at least one aspect type. Users can also include a wide range of other aspects in the evaluation; however, it can be challenging to determine which category to apply to a given statement due to the limited domain coverage of the aspect categories. This program generates a set of lexical representations for each ASP-CAT as category labels using an unsupervised technique42 to identify aspect groups based on word co-occurrence rates. For each ASP-CAT, a set of Seed Words (SW) is created to represent and explain the category. The semantic lexicon utilizes WordNet to extract synonyms for the terms in the phrases. Table 5 illustrates various aspect-based categories for the ‘Restaurant (RES)’ and ‘Laptop (LAP)’ domains. It presents nine categories within the RES domain and five in the LAP domain. For each ASP-CAT in these domains, corresponding SW are identified and labeled as CATiSWj, where ‘CAT’ represents the category, ‘i’ represents the category number, and ‘j’ represents the specific SW. So, it identifies a set of SWs CATiSWj against every specified ‘CAT’. The CATiSWj includes the terms utilizing pre-trained Amazon embedding. The algorithm categorizes phrases based on individual words, such as “burger,” “pasta,” or “service,” to align them with example categories like “ambiance” and “service”45. Table 5Sample seedwords (SWs) for aspect categories detection.ASP-CAT numberASP-CAT nameSWRESRestaurant domain (from 12 categories)RES_SWRES_C1AMBIENCE#GENERALAmbiance (CAT1SW1) , Cosette (CAT1SW2), Restaurant (CAT1SW3), atmosphere (CAT1SW4)RES_C2DRINKS#PRICESPrice (CAT2SW1), priced (CAT2SW2), overpriced (CAT2SW3), prices (CAT2SW4), pay (CAT2SW5), cheap (CAT2SW6)RES_C3FOOD#PRICESInexpensive (CAT3SW1), price (CAT3SW2), cheap (CAT3SW3), overpriced (CAT3SW4), prices (CAT3SW5)RES_C4FOOD#QUALITYSushi (CAT4SW1), food (CAT4SW2), dishes (CAT4SW3), samosas (CAT4SW4), cheese (CAT4SW5), pizza (CAT4SW6)RES_C5LOCATION#GENERALView (CAT5SW1), place (CAT5SW2), location (CAT5SW3), views (CAT5SW4), located (CAT5SW5), neighborhood (CAT5SW6)RES_C6RESTAURANT#GENERALRestaurant (CAT6SW1), pace (CAT6SW2)RES_C7RESTAURANT#PRICESExpensive (CAT7SW1), price (CAT7SW2)RES_C8SERVICE#GENERALService (CAT8SW1), kitchen (CAT8SW2), staff (CAT8SW3), waitstaff (CAT8SW4), hostess (CAT8SW5), people (CAT8SW6)RES_C9DRINKS#QUALITYWine (CAT9SW1), tea (CAT9SW2), chai (CAT9SW3)LAPLaptop samples (frequently occurring samples, 81 categories)LAP_SWLAP_C1LAPTOP#GENERALLaptop (CAT1SW1), computer (CAT1SW2), HP (CAT1SW3), Toshiba (CAT1SW4), netbook (CAT1SW5), machine (CAT1SW6)LAP_C2LAPTOP#OPERATION_PERFORMANCEwireless network (CAT2SW1), speed (CAT2SW2), capacity (CAT2SW3), battery life (CAT2SW4), performance (CAT2SW5)LAP_C3LAPTOP#QUALITYLaptop (CAT3SW1), price (CAT3SW2), quality (CAT3SW3), light (CAT3SW4), durable (CAT3SW5), hot (CAT3SW6)LAP_C4SUPPORT#QUALITYwarranty (CAT4SW1), shipping (CAT4SW2), customer support (CAT4SW3)LAP_C5LAPTOP#DESIGN_FEATURESsize (CAT5SW1), looks (CAT5SW2), design (CAT5SW3), designs (CAT5SW4), designed (CAT5SW5), lightweight (CAT5SW6) Sample seedwords (SWs) for aspect categories detection. Design of the HAN for review-level aspect extraction using CRR The proposed method utilizes an HAN-based model that captures good linguistic features, enhancing aspect and ASP-CAT prediction. Additionally, review-level attention is layered on top of sentence-level attention to identify co-referential relationships among components across all customer textual comments. All mathematical symbols and equations are fully defined at their first occurrence for clarity and reproducibility. Figure 6 illustrates the end-to-end architecture of the proposed HAN for ATE, where each architectural component corresponds directly to the mathematical formulations as presented in Sect. “Labeling text sequences using BIO”. Architecture of the proposed CRR-based HAN for ATE. Labeling text sequences using BIO To overcome the sequential labeling difficulty of ATE and ACD, the model uses review labels. When an input sentence {,, …, } containing words where (with representing the jth word in the sentence ), the model generates an output The labels are formatted using. to train the model. Table 6 illustrates the tagging scheme employed for generating pseudo labels. The objective is to identify ATs A⊆ R, where each AT is an element of A and may consist of either a single word or an M-W span within the review. Table 6Example illustrating aspect term encoding using the BIO tagging scheme.SExcellentMealbutaverageroomserviceOBOOBIType:Outside Aspect BeginOutsideOutsideAspect BeginAspect InsideY[0 0 1][1 0 0][0 0 1][0 0 1][1 0 0][0 1 0]Aspect:Single-wordMulti-word Example illustrating aspect term encoding using the BIO tagging scheme. POS and WE as network input functionalities WE and POS tags are applied as input features for the DNN in the proposed study. Initially, the input sentence is represented as a feature mapping of a sequence vector, denoted by {,, , …, } corresponding to the individual words represented as {,, , …, }, where . Additionally, the POS tag sequence embeddings for each word in the provided review phrase are represented as {,, ,…, }, with . The WEs and the corresponding POS tags P are then concatenated to create the input sequence, as expressed in Eq. (1), which forms the initial token representation (X) “Word Embedding + POS Tagging” block in Fig. 6. The input representation handles the entire multi-sentence review as a single token sequence while preserving their original order. is further aligned through the CRR module NeuralCoref, which provides cross-sentence contextual consistency before contextual encoding. After applying CRR, the input sequence is transformed into a coreference-aware representation {}, where each token representation is augmented such that , and denotes the average embedding of all tokens belonging to the coreference cluster ck associated with the token . Subsequently, the conference-aware sequence is provided as input to a pre-trained BERT model to obtain contextual token representation H = BERT( where . Specifically, to maintain full reproducibility of the implemented model, the BERT-base (uncased) configuration, is employed, which consists of 12 transformer layers, 12 attention layers, and a hidden embedding dimension of 768. Each input review sentence is tokenized using the BERT tokenizer and converted into contextual embeddings. The embedding corresponding to the [CLS] token is used as the overall sentence representation, while token-level embeddings are used for aspect term detection. The BERT model is fine-tuned during training with the downstream HAN to adapt the contextual representations (H) to the ABSA task, where H serves as the input embeddings to the subsequent hierarchical attention layers of the proposed HAN-based DNN. Proposed HAN architecture's mathematical formulation for ATE In this architecture, the hidden state comprises 256 units in each direction, resulting in a total of 512 units. A dropout rate of 0.3 is applied to the input and output of the LSTM to prevent overfitting during training. This proposed network utilizes the contextualized input representation H. Equation (2) through Eq. (9) to formulate the problem within the context of the Bi-LSTM using the contextual representation H. In the proposed architecture, Bi-LSTM layers are used as the sequence encoders within the HAN framework. At the word level, the Bi-LSTM processes contextual word embeddings and captures both past and future dependencies in the sentence. The resulting hidden representations are then passed to the sentence-level attention mechanism, which identifies the most informative words for aspect extraction. Similarly, at the review-level, another Bi-LSTM encodes the sequence representations before applying review-level attention to determine the most important sentences in the review. Thus, the Bi-LSTM serves as the contextual feature extractor, while the hierarchical attention layers assign weights of importance to words and sentences. The HAN incorporated in the Bi-LSTM is defined from Eq. (10) to Eq. (15). Equation (15) calculates the attention for each word in the sentence, which is further utilized to determine the sentence-level context . These sentence-level contexts for each sentence are then aggregated to formulate the review-level context . Figure 6 also depicts attention computation at both the sentence and review levels. Due to the small training dataset and the requirement for an embedded vector, the model implemented a minimum word count of 1. The most crucial parameter for H is domain knowledge, alongside a context size of 5 words. Repeated words in the training samples are managed by downsampling with a factor of (10–3). The training and testing samples are split in an 80:20 ratio, respectively. The optimizer used is AdamW, with a learning rate of 2e−5 for fine-tuning BERT parameters and 1e-3 for other aspect extraction task-specific layers. The training set includes a batch size of 32 instances, and the model was trained for 180 epochs until convergence. Accuracy and loss are assessed every 5 epochs, and 128 tokens is the maximum sequence length. A random seed is set to 42 to maintain the reproducibility of results. The cross-entropy loss function incorporates class weighting to address label imbalance. All experiments are performed using Python 3.9 on a system equipped with an NVIDIA GeForce RTX 2080 GPU, 8 GB of memory, 16 GB of RAM, and an Intel Core i7 processor. SpaCy V2.3.5, integrated with NeuralCoref V4.0 and model en_core_web_sm-2.3, is used for natural language processing and coreference resolution. The performance of the proposed approach is evaluated using matrices, including precision, recall, and F1-score. The HAN mechanism operates at two levels: the sentence-level, which learns attention weights αij for the importance of tokens within a sentence, and the review-level attention, which aggregates context vectors from multiple sentences (Cti → Ct). Here, Ct denotes the sentence-level context vector, while represents the review-level context vector associated with the i-th sentence of the review, capturing inter-sentence dependencies. Both levels utilize the tanh activation function and softmax normalization over the attention weights. The mathematical formulation of the work presented in Fig. 6 is as follows. The contextual embedding H obtained from BERT in Sect. 3.2.2 is used as input to the HAN, where at the time step , the input corresponds to the contextual embedding . The hidden vector is expressed in Eq. (2), where represents a non-linear transition function, is activation from the previous state, is the current input, and represents the parameter set. The contextual encoding function defined in Eq. (3), corresponds to the bidirectional encoder of in the hierarchical attention network, which processes BERT-based contextual embeddings as shown in Fig. 6, where each token is enriched with inter-sentence semantic context obtained through co-referenced reviews. Next, the forward and backward hidden states updates defined in Eq. (4) to Eq. (9) map directly to the sequential modeling performed using a Bi-LSTM layer, as shown in Fig. 6, producing a contextual hidden representation for each token. where , , , denote the input gate, forget gate, output gate, candidate cell state, and updated cell state at timestamp t, respectively. The symbol represents a sigmoid activation function, , , , and represent weight matrices corresponding to the input gate, forget gate, output gate, and candidate cell state, respectively. Additionally, , , , denote the bias terms associated with the input gate, forget gate, output gate, and cell state, respectively. Further, for the encoded sequence (, , the prediction of the label depends on the decoder state , the sentence-level context (Ct), and the review-level context (. The decoder state is computed as given in Eq. (10). Subsequently, the decoder generates using and to predict the next label along with all previously generated labels ,, , …, . Equation (11) represents the sentence-level attention mechanism illustrated by directed connections from Bi-LSTM outputs to the attention nodes in Fig. 6. Here, the encoder hidden states are aggregated using attention weight over all sentences of the review. The sentence-level context vector is computed as a weighted sum of the encoder hidden states. A learnable parameter, the weight vector (v) denotes the representation of the attention energy, as defined in Eq. (12), and is subsequently used to compute the attention weights , where the relevance between the encoder hidden state and the decoder state is measured using a feedforward network as: The hierarchical aggregation defined in Eq. (11) and Eq. (12) corresponds to the review-level attention block in Fig. 6, where multiple sentence representations are combined to form a global representation. The review-level context vector is computed by aggregating the sentence-level context vectors across the review as where represents the review-level attention weight, and N denotes the number of sentences in a review. By utilizing theeview-level context vector , the proposed model formulates the overall sequence prediction as shown in Eq. (13), where the joint probability of the output label sequence Y is decomposed into a product of the conditional probability of each label , given the previously generated and the context vector . Subsequently, the model is optimized using the cross-entropy loss defined in Eq. (14), which measures the discrepancy between the predicted label distribution and the ground truth BIO tag sequence. where = ,, , … , where denote the BIO tag set, which is a one-hot ground truth label, and is the predicted probability obtained from the softmax layer. Finally, the decoder output is passed through a dense layer (D1-Dn) followed by a softmax activation to obtain the probability distribution over the BIO tag set, as formulated in Eq. (15). The logits , computed from the concatenation of the decoder state and the context vectors and , are transformed into class probabilities for final label prediction. To further demonstrate the mathematical formulation and processing flow of the proposed model described in Eq. (1)- Eq. (15), a step-by-step worked example based on a multi-sentence review, as illustrated in Fig. 6, is presented here. Consider a multi-sentence review: “ Long power backup with pleasant voice. I recommend it.”. The input sentence is first tokenized as: Sentence1 → [Long, power, backup, with, pleasant, voice], Sentence2 → [I, recommend, it], and transformed into token-level representation X using word embeddings and POS tags, such as such as and (Eq. (1)). Next, coreference resolution links the pronoun ‘it’ in Sentence2 to the entities ‘power backup’ and ‘voice’ , described in Sentence1, yielding the enhanced sequence This sequence is then passed through BERT to obtain contextual embedding H = {hLong​, hpower​, hbackup, hwith​, hpleasant​, hvoice​, hI​, hrecommend​, hit​}. Notably, due to coreference resolution, the embedding corresponds to the pronoun ‘it’ encodes semantic information of the entities ‘power backup’ and ‘voice’ described in Sentence1 such as hit​ → inherits context from ‘power backup’ and ‘voice’. The contextual embeddings H are subsequently fed into the hierarchical attention network (HAN), where a Bi-LSTM encoder captures sequential dependencies and generates hidden states {s1​, s2​,…, sM​} (Eqs. 2–9) corresponding to the tokens [Long, power, backup,…, it]. For instance, and jointly encode the aspect ‘power backup’, while sit​ captures contextual information propagated from its co-referent entity in the previous sentence. Here, represents forward and backward states which capture future and past contextual dependencies. Next, sentence-level attention computes context vector by assigning higher attention weights to relevant tokens such as ‘power’, ‘backup’, and ‘voice’, while comparatively lower weights are assigned to less relevant tokens like ‘with’ and ‘I’. At the review-level, attention further aggregates contextual information across sentences to form , enabling the model to capture cross-sentence dependencies, particularly linking the pronoun ‘it’ to the entities ‘power backup’ and ‘voice’ described in Sentence1 (Eqs. 11–12). Further, the decoder utilizes the states , along with , and , to sequentially predict labels for each token (Eqs. (10), (13)). Finally, the softmax layer produces probability over the BIO tag set, resulting in token-level labels such as [Long(O), power(B–Aspect), backup(I–Aspect), with(O), pleasant(O), voice(B–Aspect), I(O), recommend(O), it(O)], which represented in [0,1] vector form as [(0,0,1), (1,0,0), (0,1,0), (0,0,1), (0,0,1), (1,0,0), (0,0,1), (0,0,1), (0,0,1)], where each vector corresponds to the one-hot encoding of the BIO tag set (B, I, O). This correctly identifies aspect terms ‘power backup’ and ‘voice’, while assigning the label O to non-aspect tokens (Eqs. (14–15)). In summary, the proposed model extends the conventional HAN by explicitly incorporating coreference-aware alignment and contextual embedding to effectively capture cross-sentence dependencies, thereby overcoming the limitations of the standard hierarchical aggregation. Analysis of the findings This section analyzes the experimental results to evaluate the effectiveness of the proposed approach for ATE. The findings are discussed in comparison with baseline and state-of-the-art methods to highlight the impact of HAN and sentence-alignment using CRR. Dataset description We evaluate the proposed model using the SemEval-16 Task 5 ABSA datasets. Restaurant and Laptop review datasets, which are widely adopted benchmark datasets for ATE. The datasets are publicly available and were released as part of the SemEval-16 shared task. The experiment is conducted on the Restaurant and Laptop review domains. As presented in Table 7, the Laptop dataset includes 451 training reviews and 91 testing reviews, while the Restaurant dataset provides 395 training reviews and 80 testing reviews. Each review may consist of multiple sentences; therefore, during preprocessing, reviews are segmented into multiple sentences for model processing. After sentence segmentation, the Restaurant datasets contain 2,000 training sentences and 676 testing sentences, while the Laptop dataset includes 2,500 training sentences and 800 testing sentences. Since a single review can contain multiple sentences, the total number of sentences in the dataset reported in Table 7 is higher than the number of reviews. The ATs are annotated at the token level within these sentences. Compared to the Laptop dataset, which has 88 aspect categories, the Restaurant review dataset features a significantly lower range of 12 aspect categories. Both datasets comprise multiple-sentence reviews55,56. Table 8 summarizes the aspect-related statistics of the SemEval-16 Restaurant reviews used in this study, including the distribution of ATs, sentence structures, and review-level characteristics across the training and testing splits. Notably, ATs are missing from the training set in 35% of the dataset, and the testing set lacks ATs in 39% of cases55,56. Table 7Statistics of restaurant and laptop samples from SemEval-16 for training and testing.Aspect numberAspect criteriaRestaurant domainLaptop domainTraining dataTesting dataGross samplesTraining dataTesting dataGross samples1Reviews39591486451805312Sentences20006762676250080033003ASP-CAT121212816888 Statistics of restaurant and laptop samples from SemEval-16 for training and testing. Table 8Aspect criteria of Restaurant reviews from the SemEval-16.Parameter numberAspect criteriaTraining dataTesting data1Gross reviews395912Gross sentences20006763Number of AT reviews12544144No. of non-aspect sentences7462645# of sentences single word aspects8983316# of Sentences with NPs356837# of Sentences with multiple categories9013478# of sentences with multiple aspects353679Total # of aspects250765110# of unique aspects63526511Total NP aspects37913512Total single-word aspects256130 Aspect criteria of Restaurant reviews from the SemEval-16. As shown in Table 8, the Restaurant datasets exhibit a diverse and challenging aspect distribution. A substantial number of sentences contain single-word aspects, while a large proportion also involve multiple aspects and multiple aspect categories within the same sentence. Additionally, the presence of non-aspect sentences and NP-based aspects highlights the need for robust contextual modeling. These characteristics motivate the use of CRR, and feature-enriched representations show inter-sentence dependencies and complex aspect structures. Table 9, adapted from43, displays statistics for the dataset, highlighting a few representative aspect categories and their frequency of occurrences in the SemEval-16 standard domains. The Laptop dataset does not include aspect words, and for the Restaurant and Laptop domains, 14% and 19% of statements, respectively, lack any defined aspect categories. Among the two analyzed domains, 26% of sentences fall into the single category, while 60% belong to multiple categories. Some reviews have a similar category applied to several elements within a single review. For instance, inside the phrase, “I like the pizza, coffee, and the rawa, but the burger and rassam were not up to the mark”, the ATs “pizza,” “coffee,” “burger,” “rawa,” and “rassam” all fall under the same ASP-CAT, “FOOD#QUALITY”. In addition, Table 9 provides a dataset-level analysis of review characteristics, including the presence of background and elaborative sentences, which are known to influence the difficulty of ATE and ACD, particularly in M-S reviews. The statistics reported in Table 9 highlight several phenomena that directly impact ATE and ACD performance. Reviews containing a higher proportion of background sentences often introduce contextual information without explicit aspect mentions, increasing the risk of false positives. Elaboration sentences expand aspect-related information across multiple sentences and require effective cross-sentence context modeling to identify ATs and categories. These characteristics highlight the need for sentence-alignment and contextual refinement in the proposed approach. Table 9Dataset characteristics illustrating the distribution of Background and Elaborative sentences for ATE.Category numberAT categoryTraining samplesTesting samplesMeasure (in sentences)RestaurantLaptop1AMBIENCE#GENERAL25566# with multiple categories5406882DRINKS#QUALITY4722# with a single category118913773FOOD#PRICES9023# dissimilar categories for a similar aspect12204FOOD#QUALITY849313# similar categories for dissimilar aspects19005FOOD#STYLE_OPTIONS13755# without category2714356RESTAURANT#GENERAL422142# sentences with background3154547SERVICE#GENERAL449155# of reviews with Elaboration57498LAPTOP#GENERAL634158 Dataset characteristics illustrating the distribution of Background and Elaborative sentences for ATE. The dataset follows a BIO tagging scheme, where each token is labeled as B-ASP, I-ASP, or O, enabling sequence labeling to ATE. In total, the datasets contain several thousand annotated ATs, covering a wide range of explicit aspect expressions. As mentioned in Tables 8 and 9, a key characteristic of the SemEval-16 dataset is the presence of M-S reviews with varying sentence lengths and long-range contextual dependencies, which align well with the proposed HAN mechanism and sentence-alignment preprocessing. Evaluation metrics To evaluate the effectiveness of the proposed ATE framework, the PRE, REC, and F-score metrics are employed for sequence labeling and ATE. PRE measures the proportion of the correctly identified ATs among all predicted ATs. Besides, REC quantifies the proportion of the ground-truth ATs that are correctly detected, which is critical in M-S reviews where aspects may be sparsely distributed. The F-score, defined as the harmonic mean of PRE and REC, provides a balanced measure of overall performance, making it particularly suitable for datasets with class imbalance between aspect and non-aspect tokens. Accuracy is not reported as it is less informative for sequence labeling tasks with highly imbalanced class distributions. It reports micro-averaged PRE, REC, and F-score, following the standard evaluation protocol of SemEval-16. The micro-average F-score is particularly suitable for SemEval-16, as it accounts for class imbalance by weighting each instance equally rather than each class, thereby providing a more reliable measure of overall extraction performance. Moreover, ATE requires exact boundary matching, especially for the M-W aspect expressions; therefore, under the SemEval-16 Task 5 evaluation protocol (explained in Sect. “Dataset Description”), the extracted ATs are considered correct only if their boundary exactly matches the annotated gold standard. These metrics are widely adopted in prior ABSA studies, which compare them with state-of-the-art approaches and assess the improvements introduced by sentence-alignment and hierarchical contextual modeling. Evaluation and performance evaluation Baseline models: To highlight the effectiveness of the proposed approach, it is compared with a diverse set of well-established baseline methods for ATE, including traditional, neural, attention-based, and graph-based methods. DP(M1) and CRF(M2) represent traditional feature-based sequence labeling baselines. Bi-GRU + WE + POS (M3), Bi-LSTM + Google WE + CRF (M4) are widely used neural baselines that capture contextual information from sentences. AspectGCN (M5) and IDGNN + BERT (M6) are state-of-the-art syntax-aware and graph-based baselines that model dependency structures and contextual semantics. In addition, HAN without CRR (P1) is included to isolate the contribution of the proposed HAN with the CRR module. All models are evaluated under identical experimental settings for fair comparison. Performance comparison of the proposed model with baseline models The Dependency Parser (DP)14 and CRF were used to analyze the suggested technique’s findings with and without CRR. Furthermore, the proposed model was benchmarked against contemporary techniques3,12,24,42 for AET and12,26,28 ACD. Traditional baselines such as DP and CRF exhibit limited performance due to their reliance on handcrafted features for local context. Bi-GRU and Bi-LSTM-CRF models enhance performance by capturing sequential dependencies; however, they primarily operate at the sentence level. In contrast, the proposed model explicitly models cross-sentence dependencies, enabling more accurate extraction of ATs that depend on broader contextual information. Component-wise contribution analysis (feature ablation) Table 10 isolates the contribution of POS tagging and CRR under identical feature settings. While POS tagging improves performance by providing syntactic cues (i.e., WE → WE + POS), the addition of CRR consistently yields further F-score improvement across the Restaurant and Laptop domains. For example, on the Laptop domain dataset, CRR improves F-score from 74.33 to 77.04% (+ 2.71%) under WE + POS features, demonstrating its complementary contribution beyond POS information as shown in Fig. 7. Table 10 illustrates the results of applying CRR to a common dataset, highlighting its effects on the ATE task. Integrating concatenated features improved PRE and REC by capturing inter-sentence dependencies relevant to aspect extraction. The SemEval-16 dataset (explained in Sect. “Dataset description”) includes numerous sentences where a single aspect encompasses multiple aspect categories. Additionally, longer sentences often contain various aspects. The traditional Bi-LSTM-based approach overlooks legitimate aspects, extracts invalid ones, and misclassifies aspect categories. A review of several sentences reveals the presence of inter- and intra-sentence dependencies, potentially leading to the misidentification of aspects and the erroneous tagging of nouns and NPs as aspect words. CRR tends to improve PRE consistently for the Restaurant and Laptop datasets and feature sets. REC shows mixed effects; it slightly decreases with WE only in the Restaurant, but generally increases with WE + POS, especially for the Laptop domain. Overall, F-score benefited from CRR, with the largest gains in the Laptop domain with (WE + POS). Including the POS feature improves model performance, especially in terms of PRE and F-score, indicating that syntactic information helps the model better identify correct labels. Table 10Impact of features incorporating CRR in the proposed model.Domain nameFeatures usedWithout CRRWith CRRPRE (%)REC (%)F-score (%)PRE (%)REC (%)F-score (%)RestaurantWE82.7980.7281.7483.8579.6681.70RestaurantWE + POS83.4680.3681.8885.6680.4482.97LaptopWE65.5578.8471.5968.2278.3472.93LaptopWE + POS77.2271.6674.3378.6775.4877.04 Impact of features incorporating CRR in the proposed model. Impact of CRR on features (WE, WE + POS). Next, ATE must perform CRR for sentence-alignment to determine the correct ASP-CAT. Unlike the Restaurant samples, which have 12 aspect categories, the Laptop samples feature 81 aspect categories. Due to the greater number of NP characteristics in the Laptop domain, the proposed technique faces challenges in ATE and ACD tasks when evaluated on Laptop samples. Table 11 highlights the impact of sentence-alignment on the proposed model for AET at various levels, particularly results from longer sentences in M-S evaluations. In this context, the proposed CRR-based strategy enhances the Laptop domain’s ability to identify the correct ASP-CAT, even without labeled data. The PRE of the Bi-LSTM model is increased by 5% using a CRR technique that recovers pertinent M-W properties and captures long-term inter-sentence interactions. Compared to the LR baseline as shown in Table 11, when combined with a HAN-based DNN, CRR greatly improves PRE and REC, with REC gains of 13% (67.55% to 81.04%) for the Restaurant domain and 9% (62.96% to 71.98%) for the Laptop domain, and PRE increases of 23% (62.44% to 81.99%) and 25% (53.04% to 77.55% ) for the Restaurant and Laptop domains, respectively. Although hierarchical attention alone provides substantial gain over shallow baselines, incorporating CRR further improves performance. The proposed model additionally gains REC 2% (79.06% to 81.04%) for the Restaurant domain and 1.90% (69.88% to 71.98%) for the Laptop domain, and PRE increases of 2% (79.86% to 81.99%) and 2.41% (75.14% to 77.55%) for the Restaurant and Laptop domains, respectively. This indicates CRR refines contextual representations beyond the HAN mechanism. Table 11Effect of incorporating CRR on baseline and proposed hierarchical models.Domain nameMethodology appliedPRE (%)REC (%)F-score (%)RestaurantLR62.4467.5564.90RestaurantLR + CRR69.5470.8870.20RestaurantProposed model (without CRR)79.8679.0679.46RestaurantProposed model (with CRR)81.9981.0481.51LaptopLR53.0462.9657.58LaptopLR + CRR62.0465.1063.54LaptopProposed HAN model (without CRR)75.1469.8872.41LaptopProposed HAN model (with CRR)77.5571.9874.66 Effect of incorporating CRR on baseline and proposed hierarchical models. Since the similarity measure remains constant across the utilized Restaurant and Laptop datasets, the average F-score was determined to evaluate the effectiveness of the proposed strategy. It was observed that performance did not significantly change in the Restaurant and Laptop domains when the threshold value was less than five. An issue arose concerning inter-sentence dependency in the language. The suggested hierarchical DL model addressed this concern by combining CRR for sentence-alignment and capturing NPs as features, significantly improving performance. Results show that the REC value was higher than the PRE for instances containing multiple aspect categories, as the presence of diverse and repetitive aspect cues across sentences increases the likelihood of detecting at least one valid aspect term, thereby improving REC. This trend is evident in Tables 10 and 11, where reviews with richer aspect diversity exhibit higher REC but occasionally lower PRE due to increased candidate span. This behavior is now supported by references to the experimental results (Tables 10 and 11) and dataset characteristics (Tables 7, 8 and 9). For instance, the Restaurant dataset, which contains a higher average number of aspect categories per review, consistently shows higher REC values compared to the Laptop datasets, as observed in Tables 10 and 11. Furthermore, Table 10 empirically demonstrates the consistent improvement in F-score in the CRR across Restaurant and Laptop datasets and feature settings, + 1.09% for the Restaurant, and + 2.71% for the Laptop domain. Hence, the CRR is stronger for Laptops than Restaurants, which are more sparse, context-dependent, and have multiple aspect categories. CRR improves contextual filtering, increasing PRE while maintaining competitive REC. These results confirm that CRR contributes independently of model complexity by refining contextual representations and reducing noisy aspect predictions. Additionally, sentences often include related aspect categories for various ATs in the Restaurant industry. The reduced PRE in hierarchical review-level aspect extraction indicates the capture of several minor features. The parser’s inability to recognize intra- and inter-sentence dependencies in reviews with grammatical errors is reflected in its low REC value, demonstrating that some significant components are overlooked. Overall, HAN contributes the largest absolute performance gain, while CRR provides consistent and statistically meaningful refinements by reducing contextual noise and improving cross-sentence coherence. Because there are many domain-relevant items in the Restaurant domain, the Laptop PRE is lower. Additionally, the Restaurant dataset includes reviews with distinct aspect categories for related aspects or comparable categories for different aspect phrases. Furthermore, the Restaurant dataset has a higher REC for the Restaurant domain since it contains fewer phrases with multiple aspect categories than the Laptop dataset. As shown in Table 12, the proposed method43 enhanced REC by 7% and 9% for the Laptop domain and 10% and 13% for the Restaurant domain, respectively, while increasing PRE by 11% and 14% for each. The aspect reduction procedure, which considers frequency and similarity, improves PRE results. Additionally, the proposed hierarchical deep network, incorporating the sentence-alignment step, significantly enhanced category recognition performance in both domains. The GCN structure models the non-directional connections among phrases and extracts relevant sentiment information and grammatical conventions within sentences46. When applied, BERT-based GNNs yield better and more promising results47. Table 12Performance comparison of the proposed model with baseline and state-of-the-art methods for aspect extraction with and without the CRR module.ModelsLaptopRestaurantPRERECF-scorePRERECF-scoreDP (M1)43.1356.4548.9054.558.7856.56CRF (M2)54.3272.4562.0960.6765.5263.00Bi-GRU + WE + POS R9(M3)58.9470.5964.2472.8179.8176.15Bi-LSTM + Google WE + CRF (M4)61.3471.8266.1768.3276.5672.21AspectGCN (M546)66.4072.0569.1177.7082.3579.96IDGNN + BERT (M647)65.4570.1567.7285.6877.1081.16BERT + Bi-LSTM + GCN + Self-Attention (M750)--75.17--77.67Proposed HAN Model (without CRR)43 (P1)63.4474.6368.5881.4677.7479.56Proposed HAN Model (with CRR) (P2)67.6876.6471.8885.1881.8583.48 Performance comparison of the proposed model with baseline and state-of-the-art methods for aspect extraction with and without the CRR module. Comparison with baselines and state-of-the-art methods for aspect extraction As shown in Table 12, the proposed model with CRR (P2) consistently outperforms all baseline models across both Laptop and Restaurant domains in terms of F-score. The improvement over strong baselines, such as + 3.52% F-score improvement over AspectGCN46, + 2.32% F-score improvement over IDGNN + BERT)47, and a degradation of F-score around 3.29% over (BERT + Bi-LSTM + GCN + Self Attention50 on the Restaurant dataset, as well as + 2.77%, + 4.16%, and + 5.81%, respectively, on the Laptop dataset. While these models leverage syntactic graphs and contextual embeddings, they primarily focus on intra-sentence dependencies. These gains demonstrate the effectiveness of the proposed HAN and CRR mechanisms in capturing long-range inter-sentence relationships, which are critical for aspect extraction in M-S reviews. Furthermore, the performance gain over HAN model P1 (without CRR) is + 3.30% for Laptop and + 3.92% for Restaurant, highlighting the specific contribution of the CRR component. The results show that CRR significantly enhances performance, especially in natural, M-S reviews. Differences in the performance across datasets can be partially attributed to dataset-specific characteristics. The Restaurant datasets consistently achieve higher PRE across baselines and proposed models (Tables 10, 11 and 12). This indicates relatively lower contextual variability, suggesting comparatively lower contextual noise than the Laptop domain. In contrast, as shown in Tables 7, 8, and 9, the Laptop dataset has a higher prevalence of complex, M-W aspect expressions (e.g., battery life, power management), which raise additional challenges for aspect extraction and impact overall performance. Transformers analyze all words in a sentence simultaneously using attention, which allows them to capture long-range dependencies and contextual relationships more effectively. The proposed approach already uses BERT and a hierarchical attention network, which provides contextual embedding using the transformer architecture to some extent. Therefore, the proposed approach is compared with state-of-the-art methods for aspect extraction, including HAN-based (P1)43, AspectGCN46, IDGNN + BERT47, and more recent transformer architecture BERT + Bi-LSTM + GCN + Self Attention Encoder50, which represent strong syntax-aware and transformer-based models reported in the literature. These methods have demonstrated competitive performance on the SemEval-16 benchmark (explained in 4.1) and serve as representative state-of-the-art baselines. The performance of the proposed coreference resolution-based approach has also been considered in the analysis with several recent transformer-based ABSA models to ensure a comprehensive evaluation, reported in the literature58,65,72–75.72–75 mainly focus on aspect-level sentiment classification or primarily perform aspect category detection rather than aspect extraction; still, analyzing these approaches provides a broader perspective of including transformer-based aspect extraction models. The results clearly state that, unlike these approaches72–75, the proposed model effectively captures contextual relationships across sentences by using coreference resolution, leading to improved ATE performance. This improvement highlights the improvement of incorporating discourse-level context in ABSA. Figures 8 and 9 demonstrate superior performance in both the Laptop and Restaurant domains compared to baseline and recent techniques. Figure 8 displays REC values, while Fig. 9 presents PRE values for various state-of-the-art models, including supervised, hybrid unsupervised approaches, and recent transformer-based approaches. As50 have provided F-score only, the F-score is considered only for PRE as well as the REC value to plot the results of M7 in Figs. 8 and 9. The models range from traditional methods such as DP and CRF to more advanced DL-based methods. Initially, the REC and PRE values for Model M1, which uses a conventional linguistic approach, are lower and may not capture sufficient information compared to modern machine learning techniques. Model M2 improves REC and PRE by capturing dependencies between labels, but lacks deep contextual understanding. Model M3 demonstrates a significant increase in REC and PRE, while Model M4 offers a richer representation but does not outperform the Bi-GRU. Moreover, preprocessing steps related to CRR for sentence-alignment in the proposed supervised HAN-based model (P2) notably enhance AT recognition in ABSA. P2 likely introduces CRR, which results in consistent improvements in PRE and F-score across both Restaurant and Laptop datasets, indicating that CRR enhances the alignment and refinement of contextual representations across sentences. This method enhances context-aware aspect extraction, maintains correct aspect alignment across multiple sentence evaluations, and eliminates incomplete information. The network derives contextual meaning from M-S reviews by identifying long-term relationships between specific feature words. Comparison of REC for the proposed model versus baseline and state-of-the-art approaches in ATE. Comparison of PRE for the proposed model versus baseline and state-of-the-art approaches in ATE. In the Laptop domain, terms may overlap with more general language, leading the model to classify a higher number of words as relevant. While this increases the REC, it may also result in more false positives, reducing PRE. Conversely, restaurant-related entities are more domain-specific and less ambiguous. Restaurant reviews often follow predictable patterns, making it easier for the model to accurately identify key aspects. In contrast, the ATE yielded greater REC than PRE for the laptops, with no reviews involving varied aspect categories and only reviews featuring the same categories lacking diverse aspects. Restaurant domain reviews achieve a higher PRE than Laptop reviews, while Laptop reviews exhibit a lower PRE overall. Before executing the ATE, phrase alignment notably enhances the PRE and REC values. When sentence-alignment is implemented, the final F-score for the ATE task in Laptop reviews is 13% lower than in Restaurant reviews, as shown in Table 12. This discrepancy arises because Restaurant reviews typically include more M-W elements and fewer nouns. Similarly, when sentence-alignment is used, the final F-score for the ACD task in Laptop reviews is 11% lower than in Restaurant reviews due to the diverse aspect categories and several intra-sentence dependencies in Laptop reviews. The findings of the proposed work indicate that performing sentence-alignment before the ATE subtask and its categories more effectively captures the inter-sentence dependencies within the reviews. The effectiveness of the suggested method is validated across two well-known domains: Restaurant reviews and Laptop reviews. The results demonstrate a similarity in findings for both sectors, considering that the proposed method is resilient and practical when applied to various datasets from different fields. For example, when tested in the Restaurant domain, the model trained on Laptop data produces significant results, and vice versa, when the model trained on Restaurant data is applied to the Laptop domain. It is worth considering how much these results generalize to other datasets. Although the SemEval-16 datasets (described in Sect. “Dataset description”) for Laptop and Restaurant reviews are widely recognized in ABSA research, future work should evaluate the generalizability of the proposed approach across other datasets and domains, particularly those with diverse linguistic styles or varying annotation schemes. The model demonstrates consistent performance trends across both the Restaurant and Laptop domains when trained and evaluated within domains. The proposed model (P2) demonstrates superior performance, attaining the highest F-score across both domains. Laptop: 71.88 F-score (↑ 3.30 compared to P1, ↑ 2.77 compared to M6) Laptop: 71.88 F-score (↑ 3.30 compared to P1, ↑ 2.77 compared to M6) Restaurant: 83.48 F-score (↑ 3.92 compared to P1, ↑ 2.32 compared to M6) Restaurant: 83.48 F-score (↑ 3.92 compared to P1, ↑ 2.32 compared to M6) These improvements are consistent across PRE and REC, demonstrating that incorporating CRR plays a crucial role in identifying more ATs (resulting in higher REC) and enhances the accuracy of aspect prediction (leading to higher PRE). In Table 12, the hybrid attention model (P1) employs HAN and displays strong performance in the Restaurant domain. It outperforms several robust baselines (M1-M5) despite not utilizing CRR. This suggests that HAN alone significantly boosts performance by effectively modeling structured context. However, combining HAN with CRR (as in P2) yields even better results. The integration of these techniques leads to optimal overall performance, owing to improved contextual understanding and referential coherence. Therefore, the current evaluation focuses on within-domain performance, and explicit cross-domain training is left for future work. Qualitative error analysis and common failure modes Despite achieving strong quantitative performance on the SemEval-16 Restaurant and Laptop datasets (explained in Sect. “Dataset description”), the proposed model still exhibits certain limitations. To provide deeper insights beyond aggregate metrics, it presents representative failure cases for both domains. Cases on Restaurant Domain: Implicit ATs: “The place is always crowded, but I still keep coming back.” For this review, the Gold aspect is “ambience,” but the model failed to predict it. Here, the aspect is implicitly expressed without an explicit noun phrase. Although CRR helps capture contextual relationships across sentences, implicit aspects remain challenging due to the lack of explicit lexical cues. Multiple Aspects with Contrast: “The food was excellent, but the service was painful.” For this review, the Gold aspects are “food” and “service,” but the model only predicted “food” and failed to predict “service” as an aspect term. Here, when multiple aspects occur in a single sentence with contrasting sentiment, the model may prioritize the dominant opinion-bearing aspect, leading to partial extraction. Cases on Laptop Domain: Cross-Sentence Aspect Reference: “I brought this laptop for work. The battery barely lasts three hours.” For this review, the Gold aspect is “battery,” but the model failed to predict it correctly; it predicts “power” as ATs. Here, although CRR improves inter-sentence modeling, indirect references and paraphrased aspect mentions can still result in coarse-grained or partially incorrect predictions. Aspect Boundary Ambiguity: “The keyboard layout and key travel feel comfortable.” For this review, the Gold aspects are “keyboard layout” and “key travel,” but the model only predicted “keyboard” and failed to predict the correct aspect term. Here, closely related ATs are sometimes merged into a single generalized aspect, especially when they share overlapping syntactic and semantic structures. These failure cases indicate that while CRR enhances contextual awareness and feature interactions, challenges remain in handling implicit aspects, fine-grained aspect boundaries, and contrastive structures. Addressing these limitations may require incorporating external knowledge or discourse-aware representations in future work. The failure cases discussed here are consistent with the dataset characteristics reported in Table 8, particularly the high frequency of multi-aspect, cross-sentence aspect reference and implicit ATs. Applications of ABSA Since the advent of online social platforms, consumer-generated data on the internet has become highly valuable. ABSA has become essential for companies in the digital era, enabling them to observe consumer behavior, understand public sentiment, and evaluate feedback to elevate consumer expectations regarding brands and services6. Over the past 20 years, ABSA has been applied in various fields, including commodity trading forecasting, product reviews, government issues, education, regulatory compliance, and multimedia content platforms such as YouTube. Public opinions on popular social networking sites provide rich resources for ABSA analysis. In one study33, a vector space model was applied, which required an aspect-polarity word vector to analyze recent news and events for intraday stock forecasting. This method linked intraday returns to specific headlines. Additionally, DL techniques can classify news articles into three categories for aspect extraction: word, category, and sentiment. Correlating business information from public reviews for the extraction of significant elements by establishing elaboration and background dependencies for stock recommendation decisions. Traditionally, news reports have been viewed as the best sources of information. Since the pharmaceutical industry is sensitive to news and media commentary, ABSA can be effectively utilized to understand, identify, and extract the impacts of public sentiment across various media aspects in the industry. One promising application of ABSA is in stock price forecasting. The primary goal of stock market forecasting is to predict when the value of a stock or the dividend from an investment will rise. Stock market values naturally fluctuate due to numerous factors, including politics, economic conditions, social trends, epidemics, and natural disasters29. Additionally, news stories and business releases directly affect a company’s stock value and the overall trading market. Social media platforms, where users express and discuss their opinions and feelings about financial institutions, generate significant amounts of commercial text related to trading15. ABSA helps forecast prices by assessing public sentiment on specific topics that influence stock values. Research30 has explored the relationship between public sentiment and market forecasting. For example, some studies31 have combined sentiment analysis from Twitter with financial data to improve predictions about trading markets. Others32 have used crowdsourcing techniques to enhance market forecasts. The impact of social media on stock price predictions has been recognized, noting that social media often has a closer connection to stock market performance than traditional media. In recent years, YouTube has become one of the most popular platforms for learning about a variety of topics, including investigations, e-learning, the voice industry, education, food, games, and more3,34,35. The platform includes a dedicated section where users can share comments on the videos they watch. The volume of user comments can significantly influence the relevance, quality, and sorting of social media. Unstudied user opinions have attracted the interest of researchers focused on sentiment analysis. By examining multiple aspects of a video’s subject through user comments, ABSA can evaluate its relevance and effectiveness for ranking YouTube videos36. Governments often introduce new programs to promote economic development, enhance human well-being, and address other concerns with sustainable development and learning. The effectiveness of these public initiatives in enhancing the quality of life for citizens is crucial for gauging their success. ABSA could serve as a valuable tool for evaluating citizen reviews and feedback, helping to gauge public sentiment. Public opinions on various policy areas can inform updates and improvements to governmental policies. Moreover, outcome-based teaching has encouraged academics and educators to enhance the educational environment by evaluating faculty performance, training, and learning through participants’ feedback. Analyzing these inputs and stakeholder feedback can provide insights into various education-related topics, including placements, research, academics, and personality development9. Despite its potential for capturing feedback on diverse topics from user-generated reviews, research indicates that ABSA is underutilized in the context of academic improvement3,37. Beyond the previously mentioned areas, other practical applications include the analysis of restaurants38, hotels9,10, film reviews39, trade forecasting, health issue predictions in the medical field40–43, fake news analysis, news reporting, movie success predictions, and business intelligence. Another interesting application utilizes various characteristics, such as social indexing, online community engagement, and digital participation, to propose potential customers. ABSA builds satisfaction across multiple sectors, including goods and services, retail, finance, restaurants, and hotels. The application of ABSA for analyzing digital textual samples is becoming increasingly significant in e-business decision-making across various domains. Some additional real-world applications of ABSA beyond current datasets include customer feedback analysis in e-commerce and retail, hospitality and travel, social media listening, news and public relations analysis, healthcare and patient experience, stock market insights, banking and insurance, election monitoring, and AI-powered chatbots and virtual assistants. Moreover, integrating CRR enhances the model’s performance, underscoring the importance of capturing semantic dependencies across discourse. These findings confirm the potential of the proposed methodology to advance domain-specific aspect extraction for sentiment analysis, particularly in technology and society, by incorporating elements such as infrastructure systems, the industrial internet, web access, and sustainable infrastructure. Significance testing of the proposed model Paired t-tests were conducted on the F-score values over 10 independent training runs with different random SWs for the top three models, M6, P1, and P2, to ensure the reported improvements are not due to chance. The results presented in Table 12 showed that the proposed model’s performance gain is statistically significant at 95% confidence level. Thus, the improvements are empirical and statistically robust, reinforcing the effectiveness of incorporating CRR and HAN in aspect extraction. Using a range of data samples, the F-score is computed to demonstrate the validity of the proposed HAN model. The achievements of the proposed approach on a variety of data examples are nearly equal to those of the latest landmark designs, as shown in Fig. 9. When the samples comprise around 2000 review sentences for the Restaurant domain and approximately 1800 review sentences for the Laptop domain, Fig. 9 illustrates that the success of the comprehensive effort described in this research is comparable to earlier work. To further clarify that the presented HAN model correctly recovers the domain-dependent characteristics, the significance test used in15 is employed. Therefore, the HAN model (P2) outperformed the prior model (P1), hybrid unsupervised + attention10, by validating its relevance, benefits, and integrity. According to the null hypothesis, there is no apparent distinction regarding the F-scores of the aspects recovered using HAN-based DNNs compared to those obtained through hybrid approaches for various data sources. The obtained p-value (0.035) for the model (P1), a hierarchical attention without CRR43, is generated by comparing the F-score of the aspects extracted using the proposed HAN- based model incorporating CRR (P2) with the F-score of the aspects extracted using hybrid approaches. At the significance threshold level α = 0.05, this p-value is below the significance threshold. For all data samples, the p-value of the framework (P1)43 is greater than the p-value of the proposed HAN model (P2) (0.030). Therefore, the null hypothesis holds that the proposed HAN model with CRR and the earlier model (P1), i.e., HAN without CRR43, do not significantly differ. These statistically significant findings indicate that the proposed methodology has reduced errors and improved the aspect retrieval PRE. A limiting factor in the proposed model is that it relies on extensive labeled training data, which may not be readily available across all domains, limiting its generalizability. Another limitation is that it only works for the English reviews. Third, fine-tuning pre-trained transformer models for ABSA tasks could improve ATE and sentiment classification. Table 13 shows the abbreviations and their descriptions. Table 13Abbreviation terms and symbols.AbbreviationDescriptionAbbreviationDescriptionABSAAspect-based sentiment analysisGloVeGlobal vectorsACDAspect category detectionGNNGraph neural networkAETAspect extraction taskGRUGated recurrent unitANNArtificial neural networkHANHierarchical attention networkASP-CATASP-CATJJTag for adjectiveATsAspect termsLLMLarge language modelATEAspect term extractionLRLinguistic rulesBERTBidirectional encoder representations from transformersLSTMLong short-term memoryBi-LSTMBidirectional long short-term memoryM-SMulti-sentenceBIOBegin, inside, outside taggingM-WMulti-wordCATcategoriesMLPMulti-layer perceptronCNNConvolutional neural networksNLPNatural language processingCRRCo-reference resolutionNPNoun phraseCRFConditional random fieldsPOSPart-of-speech taggingDCNNDeep convolutional neural networksPREPrecisionDLDeep learningRECRecallDNNDeep neural networkRNNRecurrent neural networksDPDependency parserSemEval-16Semantic evaluation 2016FNFalse negativeSWSeed wordsFPFalse positiveTPTrue positiveGCNGraph convolutional networkWEWord embeddings Abbreviation terms and symbols. Case study: discussion on precision, recall, and F-score This section explains how the PRE, REC, and F-score values are calculated for the proposed model. Tables 7, 8 and 9 provide insight into the dataset’s statistics. The data indicate that reviews utilizing nouns as aspects tend to rank higher than those that utilize NPs. There is also an imbalance in aspect reviews, aspect categorization, and other classifications. Consequently, PRE, REC, and F-score were selected as performance metrics for the HAN model. The primary goal is to increase the True Positive (TP) value while reducing the False Positive (FP) and False Negative (FN) values. A higher TP indicates that more single-word and M-W aspects are correctly identified; conversely, FP represents incorrect words or phrases identified as aspects, and FN signifies important missing characteristics in the retrieved aspects. This section focuses solely on performance metrics, as Sect. “The proposed approach of aspect terms extraction” details the specifics and workings of each phase. Initially, the numbers are evaluated: TP (275), FP (165), and FN (132) when NP chunks from the review sentences undergo a new logistic regression method for potential aspect extraction. These values indicate that 275 viable aspects were accurately extracted, but 165 nouns and NPs were incorrectly identified as valid aspects. Additionally, 132 legitimate aspects were not extracted. This means that while invalid aspects were extracted, some valid aspects were overlooked, evidenced by the higher FP and FN values, respectively. Incorporating CRR with LR resulted in improved values: FP (133) and FN (125) while increasing TP (303). Furthermore, using Bi-LSTM with CRR, TP rose to 338 while FP decreased to 85 and FN to 89. Finally, the proposed model improved the TP to 342, reducing FP to 75 and FN to 80. These results highlight the significance of each stage in the proposed hybrid unsupervised approach. Table 13 provides a list of abbreviations and symbols used in this context. Conclusions and future recommendations The proposed model introduces several novel contributions that distinguish it from existing ABSA research. The proposed model remarkably improved AT detection in ABSA by incorporating CRR for sentence alignment. This step removes partial understanding, establishes proper aspect alignment across multiple sentence reviews, and enhances context-aware aspect detection. The approach successfully represents long-term dependencies among precise features in M-S reviews and detects contextual understanding from those dependencies. Once incorporating word embeddings with linguistic features, contextualized representations, HAN, and CRR, the proposed model constructively detects both local and global contextual cues from M-S reviews. Evaluations conducted on SemEval-16 Restaurant and Laptop datasets show that the proposed approach improved baselines and state-of-the-art methods in terms of PRE, REC, and F-score. Extensive feature-wise and ablation analysis authenticate that the POS features combined with CRR improve extraction efficiency, specifically in REC, and underline the significance of inter-sentence contextual modeling. Qualitative error analysis, backed by dataset statistics, shows that the proposed technique is robust over varying review constructions; however tackles problems while dealing with inbuilt aspects, and multi-sentence reviews. In terms of computational complexity, the model presents subsidiary processing because of hierarchical attention and sequence modeling layers; however, the training remains feasible on standard-GPU-enabled systems. In spite of optimistic outcomes, the work introduces restrictions, like dependence on annotated datasets and increased computational demands for large-scale corpora. Although the proposed model improves results in aspect detection, the CRR step might generate errors if cross-references are inaccurately linked. Consequently, various future research directions open, comprising domains namely healthcare, finance, education, and e-commerce to evaluate generalizability by integrating multilingual capabilities with pre-trained language models. Integrating transformer architectures will strengthen contextual embedding by employing bidirectional attention mechanisms. The proposed model can be expanded to address the above limitations and enhance accuracy, efficiency, and applicability in various real-world problems. Future research can focus on improving computational efficiency and exploring lightweight architectures. This will extend the model to multilingual datasets, which focus on addressing cross-domain variability, such as training on one domain and testing on another, by enhancing the proposed framework’s ability to generalize across heterogeneous domains with minimal domain-specific adaptation. Overall, the proposed model presents an effective and scalable solution for aspect extraction in real-world review analysis, providing a strong balance between contextual richness and model interpretability. By improving fine-grained analysis from large-scale user reviews, the proposed approach supports the sustainable development of intelligent decision support systems in domains such as e-commerce and service analytics. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Author contributions Ganpat Singh Chauhan conceived the idea, curated the data, and conducted the formal analysis. He also contributed to performing the experiments. Shalini Puri contributed to methodology design and formal analysis and handled software and visualization. Ravi Nahta provided resources and supervision throughout the research process. Neeraj Garg validated the findings and assisted in reviewing and editing the manuscript. All authors contributed equally to the writing and revision of the manuscript. Funding Open access funding provided by Manipal University Jaipur. Manipal University Jaipur, India, provides open access funding. Data availability The datasets generated and analysed during the current study are available in the SemEval-16 repository, [http://alt.qcri.org/semeval2016/task5/index.php?id=data-and-tools]. Declarations Competing interests The authors declare no competing interests. Research involving human and/or animals This study did not involve any human participants or animals. Informed consent Not applicable, as the study did not involve human participants or biological material. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets generated and analysed during the current study are available in the SemEval-16 repository, [http://alt.qcri.org/semeval2016/task5/index.php?id=data-and-tools].