Improving Aspect Sentiment Quad Prediction by Relational Mask Multi-Head Attention and Template-Order Grouping

ACL ARR 2025 May Submission284 Authors

Abstract: Aspect sentiment quad prediction (ASQP) has become a popular task in the field of aspect-based sentiment analysis, which aims to predict four sentiment elements: aspect category, aspect term, opinion term, sentiment polarity. Although its great success, existing methods still have shortcomings. First, the sentiment element is only related to the specific words in the input sentence. The existing works predict quads based on the whole input, which adds redundant information. Second, recent methods convert quad prediction into a generative task through a pre-defined templates. Constructing different template orders can improve the performance of the model. However, most methods simply utilize pre-trained language models to select template order groupings without deeply analyzing the relationships between template orders. In this paper, we propose a relational mask multi-head attention and template-order grouping method, which not only reduces the redundant information in the input but also select appropriate template order groupings. Specifically, we construct a trainable relation mask matrix and fuse the multi-head attention of the T5 decoder. Then we introduce relation constraint loss to reduce redundant information in the input. In addition, we quantify the effect of one template order's gradient on another template order's loss to determine the template order groupings. Experiments on multiple public datasets demonstrate that our method outperforms state-of-the-art methods.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Sentiment Analysis

Contribution Types: Theory

Languages Studied: English

Keywords: Aspect Sentiment Quad Prediction, Relational Mask Multi-Head Attention, Template-Order Grouping

Submission Number: 284