A Multi-Task Dual-Tree Network for Aspect Sentiment Triplet Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 7065–7074 October 12–17, 2022. 7065 A Multi-Task Dual-Tree Network for Aspect Sentiment Triplet Extraction Yichun Zhao1, Kui Meng1, Gongshen Liu1∗ Jintao Du2, Huijia Zhu2 1Shanghai Jiao Tong University 2Ant Group Co., Ltd. 1{zhaoyichun, mengkui, lgshen}@sjtu.edu.cn 2{lingke.djt, huijia.zhj}@antgroup.com Abstract Aspect Sentiment Triplet Extraction (ASTE) aims at extracting triplets from a given sen- tence, where each triplet includes an aspect, its sentiment polarity, and a corresponding opin- ion explaining the polarity. Existing methods are poor at detecting complicated relations be- tween aspects and opinions as well as classify- ing multiple sentiment polarities in a sentence. Detecting unclear boundaries of multi-word as- pects and opinions is also a challenge. In this paper, we propose a Multi-Task Dual-Tree Net- work (MTDTN) to address these issues. We employ a constituency tree and a modified de- pendency tree in two sub-tasks of Aspect Opin- ion Co-Extraction (AOCE) and ASTE, respec- tively. To enhance the information interaction between the two sub-tasks, we further design a Transition-Based Inference Strategy (TBIS) that transfers the boundary information from tags of AOCE to ASTE through a transition matrix. Extensive experiments are conducted on four popular datasets, and the results show the effectiveness of our model. 1 Introduction Aspect Based Sentiment Analysis (ABSA), also known as Target Based Sentiment Analysis, has received widespread attention in both academia and industry in recent years. ABSA allows a sen- timent analysis of different aspects in a given sen- tence, which can be applied in many areas, such as social media and E-commerce reviews. Com- pared with sentence sentiment analysis, ABSA is more fine-grained and more in line with reality. ABSA contains many sub-tasks, such as Aspect Term Extraction (ATE) (Xu et al., 2018; Yang et al., 2020), Opinion Term Extraction (OTE) (Fan et al., 2019; Wu et al., 2020b), Aspect Level Sentiment Classification (ALSC) (Xiao et al., 2021; Li et al., 2021), Aspect Sentiment Pair Extraction (ASPE) (Li et al., 2019; Ji et al., 2020; Chen and Qian, 2020; *Corresponding author. Luo et al., 2020), Aspect Opinion Co-Extraction (AOCE) (Dai and Song, 2019), Aspect Opinion Pair Extraction (AOPE) (Chen et al., 2020; Zhao et al., 2020) and Aspect Sentiment Triplet Extrac- tion (ASTE) (Wu et al., 2020a; Chen et al., 2021a). Table 1: An example of different ABSA sub-tasks. As- pects, opinions and sentiment polarities are in blue, red and green respectively. Sentence: Good service but poor taste Aspect Term Extration: {service, taste} Opinion Term Extraction: {Good, poor} Aspect Sentiment Pair Extraction: {(service, pos), (taste, neg)} Aspect Opinion Co-Extraction: {service, Good, taste, poor} Aspect Opinion Pair Extraction: {(service, Good), (taste, poor)} Aspect Sentiment Triplet Extraction: {(service, Good, pos), (taste, poor, neg)} Table 1 gives an example of different ABSA sub- tasks for the sentence ’Good service but poor taste’. This paper mainly concentrate on ASTE, which extracts triplets of all aspects in a sentence with the corresponding opinion and the sentiment polarity for each aspect simultaneously. Although researches have been conducted in the area, ASTE …

Content truncated to 500 words