Double Policy Network for Aspect Sentiment Triplet Extraction (Student Abstract)

Xuting Li, Daifeng Li, Ruo Du, Dingquan Chen, Andrew D. MaddenAAAI 2023 · DOI

Abstract

Aspect Sentiment Triplet Extraction (ASTE) aims to identify aspect terms, their corresponding opinion terms, and the associated sentiment polarity. This paper introduces the Double Policy Network with Multi-Tag based Reward (DPN-MTR), which frames ASTE as a hierarchical reinforcement learning problem. By employing a trigger mechanism to link aspect detection with opinion extraction, the model effectively captures logical dependencies and improves the identification of multi-token boundaries.

Introduction

In sentiment analysis, the extraction of triplets—the aspect (what is being discussed), the opinion (how it is described), and the sentiment (the polarity)—is considered a critical task. Existing approaches often struggle with two main issues: weak interaction logic between aspect and opinion extraction, and the difficulty of accurately identifying multi-token boundaries (e.g., "could have been better"). Many models treat these tasks in parallel, which can lead to "mismatched" pairs where extracted aspects and opinions are contextually unrelated.

Methodology

The proposed DPN-MTR framework addresses these challenges through a hierarchical reinforcement learning architecture that mimics human reading strategies.

Architecture of the DPN-MTR framework showing the hierarchical relationship between ATE and TSOTE.

Double Policy Network

The architecture consists of two distinct levels:

  1. Upper-level (ATE): This network acts as the master policy, performing Aspect Term Extraction using a standard BIO (Beginning, Inside, Outside) tagging scheme.
  2. Lower-level (TSOTE): Triggered by the upper level, this subordinate network identifies specific opinion terms related to the detected aspect.

Trigger Mechanism and Reward

A "trigger mechanism" serves as the bridge, passing a context vector from the aspect detection level to the opinion extraction level. This ensures that opinion search is conditioned on the identified aspect. To improve span precision, the authors introduced a Multi-Tag based Reward (MTR). This reinforcement learning reward specifically accounts for term length, encouraging the model to capture full multi-word expressions accurately. Both levels leverage a fine-tuned BERT encoder for high-quality hidden state embeddings.

Experimental Results

The model was evaluated on the ASTE-Data-V2 benchmark, encompassing four datasets: 14lap (laptops) and 14res, 15res, 16res (restaurants).

Performance comparison of DPN-MTR against baseline models on ASTE-Data-V2 benchmarks.

DPN-MTR outperformed seven strong baselines, including state-of-the-art models like BMRC and Span-ASTE. The model achieved significant F1-score improvements, reaching 60.36 on 14lap and 73.19 on 14res. Notably, the model demonstrated strong Recall, indicating that the hierarchical trigger mechanism is superior at capturing triplets that parallel models often miss. The MTR component specifically contributed to a ~3% improvement in detecting multi-word terms compared to traditional detection strategies.

Conclusion

DPN-MTR provides a robust solution for ASTE by modeling the task as a hierarchical dependency. While the model shows state-of-the-art performance on standard benchmarks, future work will explore its generalization to other domains and its robustness against informal text. Additionally, refining the dependency between the two levels could mitigate issues where failures in the upper-level ATE propagate to the lower-level TSOTE.

Summary updated

BibTeX
@inproceedings{li2023double,
  author = {Xuting Li and Daifeng Li and Ruo Du and Dingquan Chen and Andrew D. Madden},
  title = {{Double Policy Network for Aspect Sentiment Triplet Extraction (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26988},
}
Figure