[Research Contribution] Enhancing E-Commerce Review Data Using Large Language Models
21 October, 2026
Keywords: Large language models; Sentiment analysis; Data augmentation; E-commerce.
In e-commerce, customer reviews provide valuable insights into product quality, services, and shopping experiences. However, when positive and negative reviews are unevenly distributed, artificial intelligence (AI) models may struggle to accurately identify dissatisfied customers’ feedback. To address this issue, a research team from UEH Mekong, University of Economics Ho Chi Minh City (UEH), investigated the use of the GPT-3.5 Turbo large language model to generate additional Vietnamese review data and evaluate its potential to improve sentiment analysis performance in e-commerce.
Customer reviews and the challenge of imbalanced data
Customer reviews not only reflect satisfaction after a purchase but also help businesses identify issues related to products, delivery, and customer service. Effectively analyzing this data can help businesses better understand consumer needs and make more informed decisions.
However, real-world review datasets often contain uneven distributions of sentiment classes. When positive reviews predominate, machine learning models may become biased toward the majority class and become less effective at identifying negative feedback. Yet negative reviews provide particularly valuable information for businesses seeking to identify shortcomings in their products and shopping experiences.
One way to address this issue is data augmentation, which involves generating additional samples based on existing data. However, traditional techniques, such as synonym replacement and random word insertion or deletion, may alter the meaning or emotional tone of a sentence.
In this context, Large Language Models (LLMs) offer a new approach through their ability to rephrase sentences while considering context. This study leverages that capability to generate additional negative reviews that preserve the original meaning but use different wording, thereby expanding the training data available for sentiment analysis models.
Using GPT-3.5 turbo to augment Vietnamese review data
The study uses the AIVIVN 2019 dataset, which contains 16,087 Vietnamese e-commerce reviews, including 9,280 positive reviews (57.69%) and 6,807 negative reviews (42.31%). To ensure a consistent basis for comparison, the researchers split the dataset into training and testing sets using an 80:20 ratio while maintaining the corresponding proportions of the sentiment classes.
Rather than simply replacing a few words in each sentence, the researchers used GPT-3.5 Turbo to paraphrase negative reviews. The model was instructed to preserve the core meaning and negative sentiment of each original review while changing its vocabulary and sentence structure.
As a result, 1,979 new negative reviews were generated to balance the number of samples from the two sentiment classes in the training set. These reviews were added to the training data, while the test set remained unchanged to enable a fair comparison between the model trained on the original data and the model trained on the augmented data.
This approach makes it possible to leverage LLMs’ language-processing capabilities to expand datasets without having to collect an entirely new set of user reviews. However, the quality and relevance of AI-generated data still need to be assessed before practical use.
Analysis Results: Different wording, but meaning preserved
The analysis shows that GPT-3.5 Turbo substantially changed the wording of the reviews it paraphrased. The total number of words decreased by approximately 15.7%, from more than 50,000 to over 42,000. The number of unique words also fell from 3,813 to 2,207.
By contrast, sentence length increased slightly. For example, a short review such as “Not good” could be rephrased as “Does not meet expectations or standards.” Although the revised expression is longer, its negative sentiment is preserved.
These findings suggest that GPT-3.5 Turbo tends to normalize wording and reduce the use of less common words. However, according to the vocabulary-based measures examined, the generated data did not become more diverse in terms of vocabulary. Instead, it primarily introduced alternative ways of expressing the same meaning. This is an important consideration when using AI-generated data.
Data Augmentation Improves Sentiment Analysis Performance
To evaluate the effectiveness of the approach, the researchers developed two sentiment analysis models using the Random Forest algorithm. One model was trained on the original dataset, while the other used the AI-augmented data. Both models were then evaluated on the same test set.
The results show that the model trained on augmented data performed better on most evaluation metrics. Overall accuracy increased from 88.56% to 89.34%, while the macro-averaged F1-score across the sentiment classes rose from 88.40% to 89.22%.
Notably, the most pronounced improvement was in identifying negative reviews. The recall for the negative class increased from 90.68% to 92.88%, while its F1-score rose from 87.03% to 88.06%. These results indicate that adding augmented data helped the model identify negative feedback more effectively.
For positive reviews, precision increased from 92.71% to 94.32%, and the F1-score rose from 89.77% to 90.37%, although recall for the positive class decreased slightly.
Overall, using AI to generate additional training data shows potential for improving sentiment analysis, particularly in identifying negative reviews. The findings also highlight the potential of LLMs as a tool for addressing class imbalance in datasets.
However, the effectiveness of this approach requires further validation across different datasets and models to determine its applicability in real-world settings.
Business Implications
Based on the findings, e-commerce businesses may consider several practical applications.
First, augment data for underrepresented feedback categories. When developing sentiment analysis systems, businesses can consider using LLMs to generate additional samples for less frequently represented categories, potentially improving the system’s ability to identify negative feedback.
Second, connect sentiment analysis results to service improvement. Once identified, negative reviews can be categorized and monitored to help detect issues related to product quality, delivery, and customer service. This provides a basis for prioritizing corrective actions and improving the shopping experience.
Third, ensure data quality before deployment. Reviews generated by LLMs should be checked for semantic consistency, sentiment preservation, and relevance to real-world data. Businesses should also evaluate model performance on actual customer reviews before deployment, rather than relying entirely on synthetic data.
The value of LLMs therefore lies not only in their ability to generate additional data but also in their potential to help businesses extract greater value from customer feedback. However, their effectiveness in practical applications still needs to be validated.
Overall, the study demonstrates that using GPT-3.5 Turbo to paraphrase negative reviews is a promising approach to addressing class imbalance in Vietnamese-language datasets. In this experiment, data augmentation improved sentiment analysis performance, particularly in identifying negative reviews.
However, the study was limited to GPT-3.5 Turbo, the Random Forest algorithm, and the AIVIVN 2019 dataset. Future research should therefore test the approach using a wider range of large language models, algorithms, and datasets to provide a more comprehensive assessment of its effectiveness and practical applicability.
View the full research paper “Enhancing E-Commerce Review Data Using Large Language Models” HERE.
Authors: Tran Son Nam, Nguyen Nha Yen – University of Economics Ho Chi Minh City (UEH)
This article is part of the series disseminating research and applied knowledge under the message “For a More Sustainable Mekong,” within the “Research Contribution For All” program conducted by UEH. UEH respectfully invites readers to stay tuned for the next edition of the UEH Research Insights newsletter.
News, photos: Authors, Department of Admissions – Communications UEH Mekong, Department of Communications and Partnerships UEH
Chia sẻ:
[Research Contribution] Co-innovating for a Sustainable Ocean Economy
12 December, 2025
[Podcast] Recommendations for University education development
14 February, 2025
[Podcast] Developing Vung Tau into a World-Class Tourism City
16 January, 2025
[Podcast] Postdigital Design Strategies for Media Art
6 January, 2025
[Podcast] NFTs – Artistic Innovation or Just a New Hype?
27 December, 2024
[Podcast] Boosting Employee Creativity with Constructive Feedback
23 November, 2024
[Podcast] “Dutch Disease” in Remittances and the Case of Vietnam
4 November, 2024
[Podcast] Latest approaches for sustainable universities
11 July, 2024
Data Law – Part 1: Necessity for a New Approach
18 May, 2024
Advertising Evaluation on Tiktok Platform
14 May, 2024
[Podcast] Advertising Evaluation on Tiktok Platform
13 May, 2024
Promoting Learner Autonomy in English Language Learning (Part 1)
24 November, 2023
ArtTech And Sustainable Development
27 October, 2023
Motivation of EFL Vietnamese Students in Economics-related Majors
12 October, 2023
Climate Change And Our Society: The Urgency Of Gender Inequality
12 October, 2023
People Analytics in Vietnam
10 March, 2022
Revolution in Experimental Economics
30 January, 2022
The Sharing Economy: Governance Issues in Vietnam
24 January, 2022
Employment Policy For Ho Chi Minh City in Post-social distancing Period
28 December, 2021
Lifelong learning at UEH: Towards a Sustainable University
28 December, 2021
Building a Decentralized Stock Market based on Blockchain Technology
24 December, 2021
Consumer price index from big data mining perspective (Big data)
17 November, 2021
Overview Of Digital Currency – Part 5: DIEM Private Stabilized Currency
11 November, 2021
