ProdRev: A DNN framework for empowering customers using generative pre-trained transformers
Aakash Gupta, Nataraj Das · 2025
Read ProdRev: A DNN framework for empowering customers using generative pre-trained transformers by Aakash Gupta, Nataraj Das (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-13491.txt.
Introduction
Aakash Gupta and Nataraj Das propose in 2025 a framework that reads the mountain of reviews under an online product so shoppers do not have to. E-commerce boomed after the pandemic, and a single listing can carry thousands of reviews. Facing that pile, buyers freeze, and star scores can be gamed by manipulation.
Their system, ProdRev, fine tunes a large generative transformer, in the GPT-3 family, on review text. Instead of snipping sentences, it writes an abstractive summary in its own words, weighing the reviews with something like common sense. Shoppers get the distilled pros and cons and decide for themselves. The authors pitch the shift as empowerment: the machine does the reading, the human keeps the judgment. Guarding against fake reviews is part of the motivation too, since a model that reads everything can spot patterns a star average hides.
Catalog record
| CREATOR | Aakash Gupta, Nataraj Das |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 16 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-13491
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ProdRev: A DNN framework for empowering customers using generative pre-trained transformers Aakash Gupta, Nataraj Das Abstract Following the pandemic, customers, preference for using e-commerce has accelerated. Since much information is available in multiple reviews (sometimes running in thousands) for a single product, it can create decision paralysis for the buyer. This scenario disempowers the consumer, who cannot be expected to go over so many reviews since its time consuming and can confuse them. Various commercial tools are available, that use a scoring mechanism to arrive at an adjusted score. It can alert the user to potential review manipulations. This paper proposes a framework that fine-tunes a generative pre-trained transformer to understand these reviews better. Furthermore, using ”common-sense” to make better decisions. These models have more than 13 billion parameters. To fine-tune the model for our requirement, we use the curie engine from generative pre-trained transformer (GPT3). By using generative models, we are introducing abstractive summarization. Instead of using a simple extractive method of summarizing the reviews. This brings out the true relationship between the reviews and not simply copy-paste. This introduces an element of ”common sense” for the user and helps them to quickly make the right decisions. The user is provided the pros and cons of the processed reviews. Thus the user/customer can take their own decisions. I Introduction The paper illustrates a methodology for empowering customers through a deliberate decision-making process through fine-tuning a generative pre-trained transformer. There has been a global shift of users towards e-commerce platforms. This trend was accelerated during the covid-19 induced lock-downs. These platforms also provide extensive information for the concerned product in user reviews. By analyzing them, users can anticipate the quality of the product. Reviews provide essential details regarding the product and its performance in real life. However, the user can easily get overwhelmed with too much information. Such an overload can cause decision paralysis and lead to users making sub-optimal decisions. We propose an architecture based on deep learning techniques to mitigate this scenario and empower the customers. • We fine-tune a GPT3 model from OpenAI. • This architecture helps the user to get a clear understanding of the product reviews by summarizing and clustering the reviews into pros and cons. • The generative transformer model also provides a final verdict about the concerned product. • Our proposed framework aims at minimizing misleading machine verdicts. By providing transparent summarization of the reviews as pros and cons, the user can make their own decisions. For ease of understanding, we have structured the paper into the following sections. In Section II we provide a literature review on the research conducted in this space. Section III provides a broad overview of our working methodology, which includes data set preparation. In section IV we discuss the results of our experiments and section V provides a summary and future scope for research. Fig. 1: An overview of the working framework II Literature review [ 1 ] introduces generative pre-trained transformer(GPT). These models can capture semantic and syntactic features through an attention selective mechanism. By extending the standard transformer architecture with an attentive selection mechanism for the multi-instance learning scenario – we can achieve state-of-art performance on predicting distinct relationships with high confidence. In [ 2 ] we runs three tests, to check how well GPT3 performs with logico-mathematical, semantic, and ethical requests. GPT3, is a third-generation language model, that is auto-regressive. This transformer uses deep learning techniques, that are capable of producing texts, at par with humans. The paper also discusses the various limits and consequences of using the pre-trained transformer. The language model presented in this paper [ 3 ] , has been trained on 175 billion parameters, which performs outstandingly well on various natural language processing tasks and various other benchmarks. The paper also introduces the concept of zero-shot, one-shot, and few-shot learning, which gives an example to the model about what completion part needs to create and what should be the structure of the output. It also provides various ways to manipulate the output, produced by the language model. The paper also discusses the various ways, where the model can be turned into a zero-shot model, for a specific purpose, so that one can achieve better performance. [ 4 ] talks about the art of fine-tuning the GPT3 model, for a particular task, that outperforms other state-of-art in that field. Here the fine-tuning job was summarising Russian text, which was tested upon BERT-based metrics. III Approach Humans tend to use short-term memory for reviewing information and making a decision on that basis. It often turns out to be a cumbersome task for the human brain to review something heterogeneous in nature, especially when the item to be reviewed is large. Such situations lead to poor decision-making, not only in terms of conclusion but also the time taken to come up with one. Here comes the product of our research, which makes the review items homogeneous and provides a final verdict depending upon the pros and cons, thus assisting the human decision-making process. Making the reviews [ 5 ] , tend to make better neural connections in the human brain, providing the user, the opportunity to analyze the pros and cons more efficiently, and thus making the user independent in terms of final decision making. As the model also provides a final verdict depending upon its understanding of the reviews, one can also compare how well he or she is using his/her decision making abilities. III-A Synopsis of the working methodology The figure 1 depicts the data flow of the stated methodology. We start with data preparation. Since the dataset can contain unsafe words, we need to filter them out. The next step is prompt creation. The reviews are structured such that the model performs better during the fine-tuning process. The OpenAI fine-tuning endpoint accepts the training data in a specific file format. The dataset is then converted from CSV file format to a jsonl file format. Once converted, the jsonl file is uploaded to the fine-tuning API endpoint. This process generates a file-id that can track the fine-tuning process. We optimize the model by experimenting with different hyper-parameters I . We also train with multiple (train) dataset sizes 6 5 . The final generative fine-tuned transformer is then available for deployment. III-B Data set preparation The data set chosen [ 6 ] was of reviews about various products that were fetched from the AWS S3 bucket, after configuring the concerned AWS account. After pushing all the S3 bucket contents from the cloud to the local drive, based on product category, the files were sorted out. Algorithm 1 Preparation of data set 1: Storing review files in terms of product category. 2: path_1 : The path to local drive 3: for Files in path_1 : do 4: if review\_len\geq 120: then 5: {extract,toarray(),reshape,tocsv(),Save} 6: end if 7: end for 8: path_2 :The path to saved reshaped files 9: Initialising instance of KMeans cluster 10: k=90 11: for File in path_2 : do 12: Empty List declaration 13: toarray(), reshape, vectorize 14: KMeans : fit_transform() 15: for cluster in clusters(k=90): do 16: Reshaping 17: Appending the list 18: end for 19: Concatenating the list