Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Aspect Extraction from E-Commerce Product and Service Reviews

Valiant Lance D. Dionela, Fatima Kriselle S. Dy, Robin James M. Hombrebueno, Aaron Rae M. Nicolas, Charibeth K. Cheng, Raphael W. Gonda

TL;DR

This study tackles aspect extraction in Taglish e-commerce reviews, addressing the challenge of low-resource, code-switched language by building a hierarchical, data-driven taxonomy (HAF) and a multi-method AE pipeline. It combines rule-based heuristics, a Generative LLM (Gemini 2.0 Flash), and two fine-tuned Gemma 3 1B models, coupled with a two-stage general-to-specific training regime to identify and extract both explicit and implicit aspects. Empirical results show that the Generative LLM achieves the strongest performance (e.g., Macro F1 ≈ 0.91 and EM ≈ 0.78 for general aspects), while fine-tuned models suffer from data imbalance and capacity constraints; rule-based methods offer a computationally efficient baseline but struggle with implicit cues. The work provides a scalable, linguistically adaptive framework for ABSA in Taglish and points to hybrid pipelines that leverage the precision of rules with the reasoning capabilities of LLMs to balance accuracy and cost in real-world deployment.

Abstract

Aspect Extraction (AE) is a key task in Aspect-Based Sentiment Analysis (ABSA), yet it remains difficult to apply in low-resource and code-switched contexts like Taglish, a mix of Tagalog and English commonly used in Filipino e-commerce reviews. This paper introduces a comprehensive AE pipeline designed for Taglish, combining rule-based, large language model (LLM)-based, and fine-tuning techniques to address both aspect identification and extraction. A Hierarchical Aspect Framework (HAF) is developed through multi-method topic modeling, along with a dual-mode tagging scheme for explicit and implicit aspects. For aspect identification, four distinct models are evaluated: a Rule-Based system, a Generative LLM (Gemini 2.0 Flash), and two Fine-Tuned Gemma-3 1B models trained on different datasets (Rule-Based vs. LLM-Annotated). Results indicate that the Generative LLM achieved the highest performance across all tasks (Macro F1 0.91), demonstrating superior capability in handling implicit aspects. In contrast, the fine-tuned models exhibited limited performance due to dataset imbalance and architectural capacity constraints. This work contributes a scalable and linguistically adaptive framework for enhancing ABSA in diverse, code-switched environments.

Aspect Extraction from E-Commerce Product and Service Reviews

TL;DR

This study tackles aspect extraction in Taglish e-commerce reviews, addressing the challenge of low-resource, code-switched language by building a hierarchical, data-driven taxonomy (HAF) and a multi-method AE pipeline. It combines rule-based heuristics, a Generative LLM (Gemini 2.0 Flash), and two fine-tuned Gemma 3 1B models, coupled with a two-stage general-to-specific training regime to identify and extract both explicit and implicit aspects. Empirical results show that the Generative LLM achieves the strongest performance (e.g., Macro F1 ≈ 0.91 and EM ≈ 0.78 for general aspects), while fine-tuned models suffer from data imbalance and capacity constraints; rule-based methods offer a computationally efficient baseline but struggle with implicit cues. The work provides a scalable, linguistically adaptive framework for ABSA in Taglish and points to hybrid pipelines that leverage the precision of rules with the reasoning capabilities of LLMs to balance accuracy and cost in real-world deployment.

Abstract

Aspect Extraction (AE) is a key task in Aspect-Based Sentiment Analysis (ABSA), yet it remains difficult to apply in low-resource and code-switched contexts like Taglish, a mix of Tagalog and English commonly used in Filipino e-commerce reviews. This paper introduces a comprehensive AE pipeline designed for Taglish, combining rule-based, large language model (LLM)-based, and fine-tuning techniques to address both aspect identification and extraction. A Hierarchical Aspect Framework (HAF) is developed through multi-method topic modeling, along with a dual-mode tagging scheme for explicit and implicit aspects. For aspect identification, four distinct models are evaluated: a Rule-Based system, a Generative LLM (Gemini 2.0 Flash), and two Fine-Tuned Gemma-3 1B models trained on different datasets (Rule-Based vs. LLM-Annotated). Results indicate that the Generative LLM achieved the highest performance across all tasks (Macro F1 0.91), demonstrating superior capability in handling implicit aspects. In contrast, the fine-tuned models exhibited limited performance due to dataset imbalance and architectural capacity constraints. This work contributes a scalable and linguistically adaptive framework for enhancing ABSA in diverse, code-switched environments.
Paper Structure (18 sections, 1 figure, 4 tables)

This paper contains 18 sections, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. Topic Modeling and Framework Development
  5. Data Annotation Pipeline
  6. Aspect Identification Models
  7. Rule-Based System
  8. Large Language Model (LLM)
  9. Fine-Tuned Gemma Models
  10. Results
  11. General Aspect Identification
  12. Specific Aspect Identification
  13. Aspect Extraction Results
  14. Discussion and Error Analysis
  15. Implicit vs. Explicit Aspects
  16. ...and 3 more sections

Figures (1)

  • Figure 1: Fine-Tuning Hierarchical Architecture