LLM-Based Annotation and Benchmarking for ABSA in Bengali-English Code-Mixed E-Commerce Reviews
Year: 2026
Type: conference-paper
Abstract: Aspect-Based Sentiment Analysis (ABSA) in Bangladeshi e-commerce reviews is di cult because real customer feedback is noisy, code-mixed (Bengali + English), often written in Romanized Bengali (Banglish), and contains multiple opinions per sentence. This paper presents an end-to-end ABSA study on Daraz and Pickaboo reviews: (i) a scalable data preparation pipeline that reduces a public corpus of ~78k e-commerce reviews into a higher-quality set of 12k reviews; (ii) an LLM-as-annotator strategy using GPT-5 to generate structured ABSA outputs (aspect term, aspect category, opinion term, sentiment) with validation by three human validators achieving an agreement score of κ = 0.72 (substantial agreement); (iii) a hierarchical e-commerce aspect taxonomy whose category usage is checked through human annotation review; and (iv) a comparative evaluation of four open source LLMs (Gemma 2 9B, Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B) in a zero-shot setting and after lightweight LoRA-based fine-tuning. In the experiments, Gemma 2 achieves strong component performance on code-mixed reviews, reaching 89.6% F1 for aspect term extraction(ATE) and 92.4% F1 for sentiment in zero-shot inference, and further improving to 95.6% and 96.4% respectively after lightweight domain adaptation, indicating that parameter-e cient tuning substantially improves extraction reliability for Banglish-heavy e-commerce feedback.
Institution: Patuakhali Science and Technology University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204451128