JN-NLP at SIGHAN-2024 dimABSA Task: Extraction of Sentiment Intensity Quadruples Based on Paraphrase Generation

Yunfan Jiang, Tianci Liu, Hengyang LuACL 2024

TL;DR

This paper tackles the challenge of extracting fine-grained sentiment intensity (valence and arousal) from Chinese text, going beyond simple positive/negative labels. They propose a novel paraphrase generation approach using the T5 model to directly predict sentiment intensity quadruples, achieving competitive results on the dimABSA task.

What Problem Does This Paper Address?

Traditional Aspect-Based Sentiment Analysis (ABSA) typically extracts four elements: aspect term, aspect category, opinion term, and sentiment polarity (positive/negative/neutral). However, this simplified polarity classification fails to capture the nuanced emotional intensity expressed in reviews. The SIGHAN-2024 dimABSA task introduces a more challenging goal: predicting sentiment intensity as continuous valence and arousal scores. This allows for a richer understanding of user emotions. The paper addresses the difficulty of directly predicting these continuous scores and the potential for error propagation in traditional multi-stage ABSA pipelines.

Methods

The core innovation of this paper is framing the dimABSA task as a paraphrase generation problem. Instead of directly predicting sentiment intensity scores, they transform the task into generating a natural language sentence that encapsulates the sentiment quadruple (aspect term, aspect category, opinion term, and sentiment intensity).

Specifically, they use the T5 pre-trained language model (specifically, t5-base-chinese). The process involves:

  1. Input Transformation: The sentiment quadruple is linearized into a structured format.
  2. Paraphrase Generation: The T5 model is fine-tuned to generate a target sentence that expresses the sentiment quadruple. This generated sentence acts as a paraphrase of the original review, but explicitly encodes the sentiment information.
  3. Output Extraction: The generated target sentence is then processed to extract the sentiment intensity quadruple.

This end-to-end approach aims to leverage the full semantic understanding of the T5 model and reduce errors that can accumulate in traditional pipeline methods where each element is predicted sequentially.

Datasets

  • dimABSA dataset: This is the primary dataset used for training and evaluation. It's a collection of Chinese catering industry social media comments, manually annotated with aspect term, aspect category, opinion term, and sentiment intensity (valence and arousal). The sentiment intensity is represented by continuous scores from the "Chinese EmoBank."

Key Insights and Ideas

  • Paraphrase Generation as a Unified Framework: Reframing sentiment intensity extraction as a paraphrase generation task is a clever way to leverage powerful text-to-text models like T5. It allows the model to learn complex mappings between raw text and structured sentiment information.
  • End-to-End Learning for Reduced Errors: By avoiding a multi-stage pipeline, the proposed method aims to mitigate error propagation, which is a common issue in traditional ABSA systems.
  • Capturing Nuanced Sentiment: The focus on continuous valence and arousal scores, rather than discrete polarity, allows for a much deeper understanding of the emotional tone of reviews.

Key Results

  • The proposed model, JN-NLP, achieved competitive results on the dimABSA task, ranking fifth among 7 participating teams.
  • The paper reports F1-scores for valence, arousal, and combined valence-arousal. For instance, their model achieved an F1-score of 0.482 for valence and 0.439 for arousal.
  • A comparison with a test model that did not use the paraphrase generation paradigm showed that their approach performed better in extracting sentiment triplets (aspect term, aspect category, opinion term).

Limitations and Future Work

  • Handling Multiple Aspect Terms: The model struggles when a sentence contains multiple aspect terms and opinion terms without clear conjunctions, sometimes leading to the repetition of the same sentiment score for different aspects.
  • Chinese Language Complexity: The diversity and complexity of the Chinese language can lead to the generation of incorrect target sentences that do not conform to the paraphrase generation rules.
  • Sentiment Intensity Prediction Nuances: The paper acknowledges limitations in predicting sentiment intensity, particularly when multiple aspect terms are present. The model may predict the same intensity score for different aspects.

Future work could focus on improving the model's ability to disambiguate sentiment for multiple aspects within a single sentence and further enhancing its robustness to the complexities of Chinese language expression.

Summary updated

BibTeX
@inproceedings{jiang2024jn,
  author = {Yunfan Jiang and Tianci Liu and Hengyang Lu},
  title = {{JN-NLP at SIGHAN-2024 dimABSA Task: Extraction of Sentiment Intensity Quadruples Based on Paraphrase Generation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure