Boundary-Driven Table-Filling for Aspect Sentiment Triplet Extraction

Boundary-Driven Table-Filling for Aspect Sentiment Triplet Extraction Yice Zhang1,2∗, Yifan Yang1,2∗, Yihui Li1,2, Bin Liang1,2† , Shiwei Chen1,3, Yixue Dang4, Ming Yang5, and Ruifeng Xu1,2,3† 1 Harbin Insitute of Technology, Shenzhen, China 2 Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies 3 Peng Cheng Laboratory, Shenzhen, China 4 Joint Lab of HITSZ and China Merchants Securities 5 SIAT, Chinese Academy of Sciences, Shenzhen, China {zhangyc_hit,evanyfyang,liyihui0413}@163.com bin.liang@stu.hit.edu.cn, chenshw@pcl.ac.cn dangyixue@cmschina.com.cn, min.yang@siat.ac.cn xuruifeng@hit.edu.cn Abstract Aspect Sentiment Triplet Extraction (ASTE) aims to extract the aspect terms along with the corresponding opinion terms and the expressed sentiments in the review, which is an important task in sentiment analysis. Previous research ef- forts generally address the ASTE task in an end- to-end fashion through the table-filling formal- ization, in which the triplets are represented by a two-dimensional (2D) table of word-pair rela- tions. Under this formalization, a term-level re- lation is decomposed into multiple independent word-level relations, which leads to relation in- consistency and boundary insensitivity in the face of multi-word aspect terms and opinion terms. To overcome these issues, we propose Boundary-Driven Table-Filling (BDTF), which represents each triplet as a relation region in the 2D table and transforms the ASTE task into detection and classification of relation regions. We also notice that the quality of the table rep- resentation greatly affects the performance of BDTF. Therefore, we develop an effective rela- tion representation learning approach to learn the table representation, which can fully exploit both word-to-word interactions and relation-to- relation interactions. Experiments on several public benchmarks show that the proposed ap- proach achieves state-of-the-art performances1. 1 Introduction As a fine-grained task, Aspect-Based Sentiment Analysis (ABSA) focuses on the attitudes ex- pressed on specific aspect terms (Pontiki et al., 2014). Opinion terms refer to words or phrases ∗ The first two authors contribute equally to this work. † Corresponding Authors 1We make our code publicly available at https:// github.com/HITSZ-HLT/BDTF-ABSA . Figure 1: An example of the ASTE task. The aspect terms and the opinion terms are marked with orange and blue, respectively. expressing subjective sentiments. Intuitively, opin- ion terms are important clues when determining the sentiment polarity and could provide a more detailed sentiment description for the aspect terms. Previous work has generally concentrated on ex- tracting aspect terms and opinion terms and classi- fying the sentiment expressed on the aspect term without explicitly considering the relations between aspect terms and opinion terms (He et al., 2019; Li et al., 2019; Chen and Qian, 2020). Therefore, Peng et al. (2020) propose the Aspect Sentiment Triplet Extraction (ASTE) task, which is exempli- fied in Figure 1. In the ASTE task, a triplet consists of an aspect term, the corresponding opinion term, and the expressed sentiment. Peng et al. (2020) adopt the pipeline approach to address the ASTE task. They first decompose the ASTE task into several subtasks and then learn models separately for each subtask. A more ad- vanced alternative is learning a joint model to ex- ploit the interactions between different subtasks (Xu et al., …

Content truncated to 500 words