Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ZZU-NLP at SIGHAN-2024 dimABSA Task: Aspect-Based Sentiment Analysis with Coarse-to-Fine In-context Learning

Senbin Zhu, Hanjie Zhao, Xingren Wang, Shanhong Liu, Yuxiang Jia, Hongying Zan

TL;DR

DimABSA requires fine-grained valence-arousal intensity predictions for restaurant aspects. The authors introduce a two-stage coarse-to-fine in-context learning (CFICL) framework based on Baichuan2-7B, where Stage 1 uses fixed in-context examples to form initial predictions and Stage 2 refines these via similarity-based retrieval of Opinion-field–aware training examples encoded with BERT, coupled with polarity filtering. This approach leverages both sentence-level context and word-level score cues to boost accuracy and reduce polarity bias, achieving notable improvements in VA quality and quadruple/triple extraction on the SIGHAN 2024 DimABSA task. The work demonstrates a practical, data-efficient pathway for fine-grained sentiment analysis, while highlighting limitations such as compute costs and dependency on effective similarity measures for the Opinion field.

Abstract

The DimABSA task requires fine-grained sentiment intensity prediction for restaurant reviews, including scores for Valence and Arousal dimensions for each Aspect Term. In this study, we propose a Coarse-to-Fine In-context Learning(CFICL) method based on the Baichuan2-7B model for the DimABSA task in the SIGHAN 2024 workshop. Our method improves prediction accuracy through a two-stage optimization process. In the first stage, we use fixed in-context examples and prompt templates to enhance the model's sentiment recognition capability and provide initial predictions for the test data. In the second stage, we encode the Opinion field using BERT and select the most similar training data as new in-context examples based on similarity. These examples include the Opinion field and its scores, as well as related opinion words and their average scores. By filtering for sentiment polarity, we ensure that the examples are consistent with the test data. Our method significantly improves prediction accuracy and consistency by effectively utilizing training data and optimizing in-context examples, as validated by experimental results.

ZZU-NLP at SIGHAN-2024 dimABSA Task: Aspect-Based Sentiment Analysis with Coarse-to-Fine In-context Learning

TL;DR

DimABSA requires fine-grained valence-arousal intensity predictions for restaurant aspects. The authors introduce a two-stage coarse-to-fine in-context learning (CFICL) framework based on Baichuan2-7B, where Stage 1 uses fixed in-context examples to form initial predictions and Stage 2 refines these via similarity-based retrieval of Opinion-field–aware training examples encoded with BERT, coupled with polarity filtering. This approach leverages both sentence-level context and word-level score cues to boost accuracy and reduce polarity bias, achieving notable improvements in VA quality and quadruple/triple extraction on the SIGHAN 2024 DimABSA task. The work demonstrates a practical, data-efficient pathway for fine-grained sentiment analysis, while highlighting limitations such as compute costs and dependency on effective similarity measures for the Opinion field.

Abstract

The DimABSA task requires fine-grained sentiment intensity prediction for restaurant reviews, including scores for Valence and Arousal dimensions for each Aspect Term. In this study, we propose a Coarse-to-Fine In-context Learning(CFICL) method based on the Baichuan2-7B model for the DimABSA task in the SIGHAN 2024 workshop. Our method improves prediction accuracy through a two-stage optimization process. In the first stage, we use fixed in-context examples and prompt templates to enhance the model's sentiment recognition capability and provide initial predictions for the test data. In the second stage, we encode the Opinion field using BERT and select the most similar training data as new in-context examples based on similarity. These examples include the Opinion field and its scores, as well as related opinion words and their average scores. By filtering for sentiment polarity, we ensure that the examples are consistent with the test data. Our method significantly improves prediction accuracy and consistency by effectively utilizing training data and optimizing in-context examples, as validated by experimental results.
Paper Structure (13 sections, 3 equations, 4 figures, 2 tables)

This paper contains 13 sections, 3 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. System Overview
  4. Fixed In-context Learning Stage
  5. Example Retrieval Enhancement Stage
  6. Experimental Setup
  7. Dataset
  8. Implementation Details
  9. Evaluation Metrics
  10. Evaluation Results
  11. Case Study
  12. Conclusion
  13. Appendix

Figures (4)

  • Figure 1: The architecture of our system. The figure illustrates a two-stage in-context learning method based on the Baichuan2-7B model to improve prediction accuracy in the DimABSA task. In the first stage, fixed in-context examples (Fixed ICL) are used to process training data. The model's sentiment recognition ability is enhanced through a prompt template, and initial predictions are made for the test data. In the second stage, the Opinion field is encoded using BERT, and the most similar training data is selected as new in-context examples based on similarity. These examples include the Opinion field and its scores, as well as related opinion words and average scores. Sentiment polarity filtering ensures that the in-context examples are consistent with the test data. Finally, these new in-context informations are input into the model along with the test data for re-prediction, yielding optimized quadruple results.
  • Figure 2: The distribution of valence and arousal scores of train dataset
  • Figure 3: The distribution of continuous real-valued scores in the valence-arousal dimensions
  • Figure 4: Case Study