KMITL Expo 2026 LogoKMITL 66th Anniversary Logo

Aspect-Based Sentiment Analysis for E-Commerce Product Reviews

Abstract

In today’s rapidly expanding e-commerce environment, the massive volume of product reviews makes it crucial to summarize user opinions in a way that is both comprehensible and practically applicable. This research presents a system for analyzing product reviews using Aspect-Based Sentiment Analysis (ABSA), a Natural Language Processing (NLP) technique that identifies key aspects of a review (such as shipping, product quality, and packaging) and evaluates the sentiment (positive, negative, or neutral) associated with each aspect, allowing both consumers and merchants to gain more efficient access to in-depth insights. This project focuses on developing AI for Thai-language ABSA by utilizing WangchanBERTa, a model trained on Thai data, and comparing it with various standard approaches such as TF-IDF + Logistic Regression, Word2Vec + BiLSTM, and Multilingual BERT (mBERT/XLM-R) to assess their performance in terms of accuracy, speed, and resource usage. Additionally, a dashboard visualization is provided to help users quickly grasp review trends. The expected outcome is to create an AI tool that can be practically employed in the e-commerce industry, enabling consumers to make easier purchasing decisions and assisting merchants in effectively improving their products and services.

Objective

ปัจจุบันการซื้อขายสินค้าออนไลน์เติบโตขึ้นอย่างรวดเร็ว ทำให้ข้อมูลรีวิวสินค้าจากผู้บริโภคมีปริมาณเพิ่มขึ้นเป็นจำนวนมาก รีวิวเหล่านี้มีบทบาทสำคัญในการตัดสินใจซื้อของลูกค้าและการปรับปรุงคุณภาพสินค้าของร้านค้า อย่างไรก็ตาม ปริมาณข้อมูลที่มากเกินไปและความหลากหลายของรูปแบบการแสดงความคิดเห็นทำให้การสรุปและวิเคราะห์ข้อมูลเหล่านี้เป็นไปได้ยาก ผู้บริโภคต้องใช้เวลามากในการอ่านรีวิวจำนวนมากเพื่อสรุปแนวโน้มความคิดเห็น ในขณะที่ร้านค้าประสบปัญหาในการวิเคราะห์ข้อมูลรีวิวเพื่อปรับปรุงผลิตภัณฑ์และบริการ เพื่อแก้ไขปัญหานี้ งานวิจัยนี้นำเสนอการประยุกต์ใช้ Aspect-Based Sentiment Analysis (ABSA) ซึ่งเป็นเทคนิคใน Natural Language Processing (NLP) ที่สามารถแยกแยะ แง่มุมสำคัญของรีวิวสินค้า (Aspects) และวิเคราะห์ อารมณ์ของแต่ละแง่มุม (Sentiments) โดยอัตโนมัติ การนำเทคนิคนี้มาใช้จะช่วยให้ผู้บริโภคสามารถรับข้อมูลเชิงลึกจากรีวิวได้ง่ายขึ้น และช่วยให้ร้านค้าสามารถใช้ข้อมูลรีวิวเพื่อปรับปรุงสินค้าและบริการอย่างมีประสิทธิภาพ โครงงานนี้ยังมุ่งเน้นการศึกษาว่า แนวทาง AI แบบใดมีประสิทธิภาพสูงสุดในการทำ ABSA สำหรับภาษาไทย โดยเปรียบเทียบ วิธีการประมวลผลภาษาธรรมชาติแบบดั้งเดิม กับเทคนิคการเรียนรู้เชิงลึกที่ทันสมัย เพื่อให้ได้แนวทางที่เหมาะสมที่สุด พร้อมทั้งพัฒนา Dashboard Visualization ที่ช่วยให้ข้อมูลรีวิวถูกนำเสนอในรูปแบบที่เข้าใจง่ายและสามารถนำไปใช้งานจริงในอุตสาหกรรมอีคอมเมิร์ซ

Other Innovations

In Silico Drug Discovery of Emerging Immune Checkpoint TIGIT-Binding Compounds for Cancer Immunotherapy: Computational Screening, Docking Studies, and Molecular Dynamics Analysis

คณะวิทยาศาสตร์

In Silico Drug Discovery of Emerging Immune Checkpoint TIGIT-Binding Compounds for Cancer Immunotherapy: Computational Screening, Docking Studies, and Molecular Dynamics Analysis

Cancer remains a major global health challenge as the second-leading cause of human death worldwide. The traditional treatments for cancer beyond surgical resection include radiation and chemotherapy; however, these therapies can cause serious adverse side effects due to their high killing potency but low tumor selectivity. The FDA approved monoclonal antibodies (mAbs) that target TIGIT/PVR (T-cell immunoglobulin and ITIM domain/poliovirus receptor) which is an emerging immune checkpoint molecules has been developed; however, the clinical translation of immune checkpoint inhibitors based on antibodies is hampered due to immunogenicity, immunological-related side effects, and high costs, even though these mAbs show promising therapeutic efficacy in clinical trials. To overcome these bottlenecks, small-molecule inhibitors may offer advantages such as better oral bioavailability and tumor penetration compared to mAbs due to their smaller size. Here, we performed structure-based virtual screening of FDA-approved drug repertoires. The 100 screened candidates were further narrowed down to 10 compounds using molecular docking, with binding affinities ranging from -9.152 to -7.643 kcal/mol. These compounds were subsequently evaluated for their pharmacokinetic properties using ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) analysis, which demonstrated favorable drug-like characteristics. The lead compounds will be further analyzed for conformational changes and binding stability against TIGIT through molecular dynamics (MD) simulations to ensure that no significant conformational changes occur in the protein structure. Collectively, this study represents the potential of computational methods and drug repurposing as effective strategies for drug discovery, facilitating the accelerated development of novel cancer treatments.

Read more
Bacteriocinogenomic analysis and anti-pathogenic activity of potential Lactococcus lactis TKP1-5 isolated from the feces of Anas platyrhynchos

คณะวิทยาศาสตร์

Bacteriocinogenomic analysis and anti-pathogenic activity of potential Lactococcus lactis TKP1-5 isolated from the feces of Anas platyrhynchos

Bacteriocins are microbial peptides that demonstrate potency against pathogens. This study evaluated the inhibitory effects on pathogens and characterized the bacteriogenomic profile of strain TKP1-5, isolated from the feces of Anas platyrhynchos domesticus. Strain TKP1-5 was characterized using phenotypic traits, 16S rRNA sequencing, and Whole-Genome Sequencing (WGS). It exhibited growth in the presence of 2-6% NaCl, temperatures of 25-45°C, and pH levels ranging from 3 to 9. Based on ANIb, ANIm, and dDDH values, strain TKP1-5 was identified as Lactococcus lactis. Whole genome analysis revealed that strain TKP1-5 harbors the Nisin Z peptide gene cluster with a bit-score of 114.775. The antimicrobial spectrum of bacteriocin TKP1-5 showed inhibitory effects against pathogenic bacteria including Pediococcus pentosaceus JCM5885, Listeria monocytogenes ATCC 19115, Enterococcus faecalis JCM 5803T, Salmonella Typhimurium ATCC 13311ᵀ, Aeromonas hydrophila B1 AhB1, Streptococcus agalactiae 1611 and Streptococcus cowan I. Genomic analysis confirmed L. lactis TKP1-5 as a non-human pathogen without antibiotic resistance genes or plasmids. Furthermore, L. lactis TKP1-5 contains potential genes associated with various probiotic properties and health benefits. This suggests that L. lactis TKP1-5, with its antibacterial activity and probiotic potential, could be a promising candidate for further research and application in the food industry.

Read more
Application of Machine Learning, Stochastic Process, and Game Theory in Short-Term Financial Asset Investment Strategies

คณะวิทยาศาสตร์

Application of Machine Learning, Stochastic Process, and Game Theory in Short-Term Financial Asset Investment Strategies

This project focuses on the study and development of a short-term investment framework via gold trading in the foreign exchange market. Machine learning techniques are applied to analyze and forecast pricing trends. Moreover, we develop the system using a stochastic process to determine optimal stop-loss points, with the aim of maximizing expected returns. Additionally, we apply game theory to guide the decision-making process regarding order holding or closure. The system is implemented and tested on the MetaTrader 5 (MT5) platform. This project outlined the clear process that includes data preparation, machine learning model training, probabilistic modeling of gold price movements, stop-loss strategy formulation, strategic decision modeling based on game theory, the development of an automated trading program, and backtesting to evaluate system performance.

Read more