COM-MRC: A COntext-Masked Machine Reading Comprehension Framework for Aspect Sentiment Triplet Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3230 - 3241 December 7-11, 2022 ©2022 Association for Computational Linguistics COM-MRC: A COntext-Masked Machine Reading Comprehension Framework for Aspect Sentiment Triplet Extraction Zepeng Zhai1, Hao Chen1, Fangxiang Feng1,2, Ruifan Li1,2∗, Xiaojie Wang1,2 1School of Artificial Intelligence, Beijing University of Posts and Telecommunications, China 2Engineering Research Center of Information Networks, Ministry of Education, China {zepeng, ccchenhao997, fxfeng, rfli, xjwang}@bupt.edu.cn Abstract Aspect Sentiment Triplet Extraction (ASTE) aims to extract sentiment triplets from sen- tences, which was recently formalized as an ef- fective machine reading comprehension (MRC) based framework. However, when facing mul- tiple aspect terms, the MRC-based methods could fail due to the interference from other aspect terms. In this paper, we propose a novel COntext-Masked MRC (COM-MRC) frame- work for ASTE. Our COM-MRC framework comprises three closely-related components: a context augmentation strategy, a discrimina- tive model, and an inference method. Specif- ically, a context augmentation strategy is de- signed by enumerating all masked contexts for each aspect term. The discriminative model comprises four modules, i.e., aspect and opin- ion extraction modules, sentiment classification and aspect detection modules. In addition, a two-stage inference method first extracts all aspects and then identifies their opinions and sentiment through iteratively masking the as- pects. Extensive experimental results on bench- mark datasets show the effectiveness of our pro- posed COM-MRC framework, which outper- forms state-of-the-art methods consistently1. 1 Introduction Aspect Sentiment Triplet Extraction (ASTE) has re- cently been proposed, which is a variant of the fine- grained Aspect-based Sentiment Analysis (ABSA) task. For a given sentence, ASTE aims to extract the sentiment triplets, including aspect term, opin- ion term and the corresponding sentiment polarity. As shown in Figure 1, ASTE could produce two triplets from the given sentence. For ASTE task, early methods adopt a two-stage pipeline framework that first identifies aspects with sentiment and opinions then pairs them, producing *Corresponding author. 1Code and datasets are available at https://github.com/ zzp-seeker/COM-MRC. Nice ambience , but highly overrated place . positive negative { ambience, Nice, POS } ambience Nice, POS Triplets： { place, overrated, NEG } place overrated, NEG query about ambience context query context masked place query context masked ambience query about place context Aspect Accessory Inference Stage Traditional MRC COM-MRC (Ours) Figure 1: A sentence and its sentiment triplets are shown at the top half. Moreover, the primary difference be- tween the traditional MRC and our COM-MRC is high- lighted at the bottom. the sentiment triplets (Peng et al., 2020). However, these pipeline-based methods ignore the interac- tion among triplets, which could result in the error propagation. To alleviate this problem, some re- cent studies jointly extract the sentiment triplets in an end-to-end framework (Xu et al., 2020; Wu et al., 2020a; Zhang et al., 2020; Chen et al., 2021b; Yan et al., 2021), which is constructed mainly by designing a tagging scheme. Very recently, Mao et al. (2021) and Chen et al. (2021a) formalized ASTE by using a machine read- ing …

Content truncated to 500 words