Open-Domain Targeted Sentiment Analysis via Span-Based Extraction and Classification

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 537–546 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 537 Open-Domain Targeted Sentiment Analysis via Span-Based Extraction and Classification Minghao Hu†, Yuxing Peng†, Zhen Huang†, Dongsheng Li†, Yiwei Lv§ † National University of Defense Technology, Changsha, China § University of Macau, Macau, China {huminghao09,pengyuxing,huangzhen,dsli}@nudt.edu.cn Abstract Open-domain targeted sentiment analysis aims to detect opinion targets along with their senti- ment polarities from a sentence. Prior work typically formulates this task as a sequence tagging problem. However, such formulation suffers from problems such as huge search space and sentiment inconsistency. To ad- dress these problems, we propose a span-based extract-then-classify framework, where multi- ple opinion targets are directly extracted from the sentence under the supervision of target span boundaries, and corresponding polarities are then classified using their span representa- tions. We further investigate three approaches under this framework, namely the pipeline, joint, and collapsed models. Experiments on three benchmark datasets show that our ap- proach consistently outperforms the sequence tagging baseline. Moreover, we find that the pipeline model achieves the best performance compared with the other two models. 1 Introduction Open-domain targeted sentiment analysis is a fun- damental task in opinion mining and sentiment analysis (Pang et al., 2008; Liu, 2012). Com- pared to traditional sentence-level sentiment anal- ysis tasks (Lin and He, 2009; Kim, 2014), the task requires detecting target entities mentioned in the sentence along with their sentiment polarities, thus being more challenging. Taking Figure 1 as an ex- ample, the goal is to first identify “Windows 7” and “Vista” as opinion targets and then predict their corresponding sentiment classes. Sentence: I love [Windows 7]+ which is a vast improv- ment over [Vista]-. Targets: Windows 7, Vista Polarities: positive, negative Figure 1: Open-domain targeted sentiment analysis. Typically, the whole task can be decoupled into two subtasks. Since opinion targets are not given, we need to first detect the targets from the in- put text. This subtask, which is usually denoted as target extraction, can be solved by sequence tagging methods (Jakob and Gurevych, 2010; Liu et al., 2015; Wang et al., 2016a; Poria et al., 2016; Shu et al., 2017; He et al., 2017; Xu et al., 2018). Next, polarity classification aims to predict the sentiment polarities over the extracted target en- tities (Jiang et al., 2011; Dong et al., 2014; Tang et al., 2016a; Wang et al., 2016b; Chen et al., 2017; Xue and Li, 2018; Li et al., 2018; Fan et al., 2018). Although lots of efforts have been made to design sophisticated classifiers for this subtask, they all assume that the targets are already given. Rather than using separate models for each sub- task, some works attempt to solve the task in a more integrated way, by jointly extracting targets and predicting their sentiments (Mitchell et al., 2013; Zhang et al., 2015; Li et al., 2019). The key insight is to label each word with a set of tar- …

Content truncated to 500 words