Semantic and Syntactic Enhanced Aspect Sentiment Triplet Extraction

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1474–1483 August 1–6, 2021. ©2021 Association for Computational Linguistics 1474 Semantic and Syntactic Enhanced Aspect Sentiment Triplet Extraction Zhexue Chen1,2,3, Hong Huang1,2,3∗, Bang Liu4∗, Xuanhua Shi1,2,3, Hai Jin1,2,3 1National Engineering Research Center for Big Data Technology and System 2Service Computing Technology and System Lab 3Huazhong University of Science and Technology, China 4RALI & Mila, University of Montreal {chenzhexue,honghuang,xhshi,hjin}@hust.edu.cn bang.liu@umontreal.ca Abstract Aspect Sentiment Triplet Extraction (ASTE) aims to extract triplets from sentences, where each triplet includes an entity, its associated sentiment, and the opinion span explaining the reason for the sentiment. Most existing re- search addresses this problem in a multi-stage pipeline manner, which neglects the mutual information between such three elements and has the problem of error propagation. In this paper, we propose a Semantic and Syntactic Enhanced aspect Sentiment triplet Extraction model (S3E2) to fully exploit the syntactic and semantic relationships between the triplet ele- ments and jointly extract them. Specifically, we design a Graph-Sequence duel representa- tion and modeling paradigm for the task of ASTE: we represent the semantic and syntac- tic relationships between word pairs in a sen- tence by graph and encode it by Graph Neu- ral Networks (GNNs), as well as modeling the original sentence by LSTM to preserve the se- quential information. Under this setting, we further apply a more efficient inference strat- egy for the extraction of triplets. Extensive evaluations on four benchmark datasets show that S3E2 significantly outperforms existing approaches, which proves our S3E2’s superi- ority and flexibility in an end-to-end fashion. 1 Introduction Aspect-based Sentiment Analysis (ABSA) usually requires to extract comment targets in a review and judge corresponding sentiment polarities (Liu, 2012; Pontiki et al., 2014). Such a research field has received widespread attention (Zhang et al., 2015; Li and Lu, 2017, 2019; Li et al., 2019a). In this paper, we concentrate on a more relatively fine-grained task - Aspect Sentiment Triplet Extrac- tion (ASTE) (Peng et al., 2020), which aims to extract triplets, including aspects (e.g., entities), *Corresponding Author the corresponding sentiment for each aspect, and the opinion spans explaining the reason for the sentiment. An example is shown in Fig. 1. It contains two triplets, (Waiters, friendly, +) and (fruit salad, so so, 0) where we use +, -, and 0 to represent positive, negative, and neutral senti- ment. Unlike the ABSA task that extracts two tuples, (Waiters, +) and (fruit salad, 0) in this sentence, such triplets extracted by ASTE task can better reflect multiple emotional factors (aspect, opinion, sentiment) from the user reviews and are more suitable for practical application scenarios. The ASTE task is extremely challenging because it requires extracting these three elements in one shot. Straightforwardly, one naive solution is to split the ASTE task into two stages in a pipeline manner using a unified tagging schema 1 (Peng et al., 2020). Such a pipeline approach lacks an effective mechanism to capture the three elements’ relationship and suffers from error propagation. Another solution for …

Content truncated to 500 words