SPT-ABSA

We continue to pre-train BERT-base via Sentiment-enhance pre-training (SPT).

• Title: An Empirical Study of Sentiment-Enhanced Pre-Training for Aspect-Based Sentiment Analysis
• Author: Yice Zhang, Yifan Yang, Bin Liang, Shiwei Chen, Bing Qin, and Ruifeng Xu
• Conference: ACL-2023 Finding (Long)

GitHub Repository: https://github.com/HITSZ-HLT/SPT-ABSA

What Did We Do?

Aspect-Based Sentiment Analysis (ABSA) is an important problem in sentiment analysis. Its goal is to recognize opinions and sentiments towards specific aspects from user-generated content. Many research efforts leverage pre-training techniques to learn sentiment-aware representations and achieve significant gains in various ABSA tasks. We conduct an empirical study of SPT-ABSA to systematically investigate and analyze the effectiveness of the existing approaches.

We mainly concentrate on the following questions:

• (a) what impact do different types of sentiment knowledge have on downstream ABSA tasks?;
• (b) which knowledge integration method is most effective?; and
• (c) does injecting non-sentiment-specific linguistic knowledge (e.g., part-of-speech tags and syntactic relations) into pre-training have positive impacts?

Based on the experimental investigation of these questions, we eventually obtain a powerful sentiment-enhanced pre-trained model. The powerful sentiment-enhanced pre-trained model has two versions, namely zhang-yice/spt-absa-bert-400k and zhang-yice/spt-absa-bert-10k, which integrates three types of knowledge:

• aspect words: masking aspects' context and predicting them.
• review's rating score: rating prediction.
• syntax knowledge:
  • part-of-speech,
  • dependency direction,
  • dependency distance.

Experimental Results

<img width="75%" alt="image" src="https://github.com/HITSZ-HLT/SPT-ABSA/assets/9134454/38fc2db0-6ccf-47a7-a93c-cf54667e1a23">

<img width="75%" alt="image" src="https://github.com/HITSZ-HLT/SPT-ABSA/assets/9134454/20c5a976-014e-433f-a2ec-4bb259e5a382">