Seq2Path: Generating Sentiment Tuples as Paths of a Tree

Findings of the Association for Computational Linguistics: ACL 2022, pages 2215 - 2225 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Seq2Path: Generating Sentiment Tuples as Paths of a Tree Yue Mao, Yi Shen, Jingchao Yang, Xiaoying Zhu, Longjun Cai Alibaba Group {maoyue.my, sy133447, yangjingchao.yjc, candy.zxy, longjun.clj}@alibaba-inc.com Abstract Aspect-based sentiment analysis (ABSA) tasks aim to extract sentiment tuples from a sentence. Recent generative methods such as Seq2Seq models have achieved good performance by formulating the output as a sequence of sentiment tuples. However, the orders between the sentiment tuples do not naturally exist and the generation of the cur- rent tuple should not condition on the previous ones. In this paper, we propose Seq2Path to generate sentiment tuples as paths of a tree. A tree can represent “1-to-n” relations (e.g., an aspect term may correspond to multiple opinion terms) and the paths of a tree are independent and do not have orders. For training, we treat each path as an independent target, and we calculate the average loss of the ordinary Seq2Seq model over paths. For inference, we apply beam search with constrained decoding. By introducing an additional discriminative token and applying a data augmentation technique, valid paths can be automatically selected. We conduct experiments on five tasks including AOPE, ASTE, TASD, UABSA, ACOS. We evaluate our method on four common benchmark datasets including Laptop14, Rest14, Rest15, Rest16. Our proposed method achieves state-of-the-art results in almost all cases. 1 Introduction ABSA tasks. Aspect-based sentiment analysis (ABSA) is a classic research topic and has received continuous attention. The ABSA tasks aim to ex- tract sentiment tuples of elements such as the aspect term (a), opinion term (o), aspect category (c), and sentiment polarity (s), respectively. Following the tasks definitions in (Zhang et al., 2021b), we con- sider various ABSA tasks including aspect opinion pair extraction (AOPE), aspect sentiment triplet ex- traction (ASTE), target aspect sentiment detection (TASD), unified aspect-based sentiment analysis (UABSA) and aspect category opinion sentiment (ACOS). The output formats are shown in Table 1. Throughout this paper, we assume the ASTE task is our default task to illustrate our ideas. ABSA Task Abbr Output Aspect Opinion Pair Extraction AOPE (a, o) Aspect Sentiment Triplet Extraction ASTE (a, o, s) Target Aspect Sentiment Detection TASD (c, a, s) Unified Aspect-Based Sentiment Analysis UABSA (a, s) Aspect Category Opinion Sentiment ACOS (c, a, o, s) Table 1: The ABSA tasks with their output formats: AOPE (Zhao et al., 2020; Chen et al., 2020), ASTE (Peng et al., 2020), TASD (Wan et al., 2020), UABSA (Li et al., 2019; Chen et al., 2020), ACOS (Cai et al., 2021). Throughout this paper, the ASTE task is as- sumed to be our default task to illustrate our ideas. Seq2Seq for ABSA. Instead of using separate models for each ABSA task, the recent trend is to design a unified framework to handle multi- ple ABSA tasks at the same time. Recently, the Seq2Seq models have been applied to the ABSA tasks (Yan et al., 2021; …

Content truncated to 500 words