Exploiting Unlabeled Data for Target-Oriented Opinion Words Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 7075–7085 October 12–17, 2022. 7075 Exploiting Unlabeled Data for Target-Oriented Opinion Words Extraction Yidong Wang1 ∗ Hao Wu1 ∗ Ao Liu1 Wenxin Hou2 Zhen Wu3 † Jindong Wang4 Takahiro Shinozaki1 Manabu Okumura1 Yue Zhang5 1Tokyo Institute of Technology 2Microsoft STCA 3Nanjing University 4Microsoft Research Asia 5Westlake University {yidongwang37, wu.364371691}@gmail.com, wuz@nju.edu.cn Abstract Target-oriented Opinion Words Extraction (TOWE) is a fine-grained sentiment analysis task that aims to extract the corresponding opin- ion words of a given opinion target from the sentence. Recently, deep learning approaches have made remarkable progress on this task. Nevertheless, the TOWE task still suffers from the scarcity of training data due to the expen- sive data annotation process. Limited labeled data increase the risk of distribution shift be- tween test data and training data. In this pa- per, we propose exploiting massive unlabeled data to reduce the risk by increasing the ex- posure of the model to varying distribution shifts. Specifically, we propose a novel Multi- Grained Consistency Regularization (MGCR) method to make use of unlabeled data and de- sign two filters specifically for TOWE to filter noisy data at different granularity. Extensive experimental results on four TOWE benchmark datasets indicate the superiority of MGCR compared with current state-of-the-art meth- ods. The in-depth analysis also demonstrates the effectiveness of the different-granularity filters. Our codes are available at https: //github.com/TOWESSL/TOWESSL. 1 Introduction Target-oriented Opinion Words Extraction (TOWE) (Fan et al., 2019) is an important subtask of aspect-based sentiment analysis (ABSA) (Pon- tiki et al., 2014), which aims to extract the corresponding opinion words for a given opinion target from the sentence. For the TOWE task, opinion targets, also called aspect terms, are the entities or objects in the sentence toward which users show attitudes; opinion words, sometimes known as opinion expressions, are those words explicitly mentioned in the sentence and used to express attitudes or opinions. Figure 1 shows an ∗ Equal Contribution. † Corresponding author. The dishes are amazingly delicious but the waiter is so rude . Given opinion target: dishes Corresponding opinion words: amazingly delicious Given opinion target: waiter Corresponding opinion words: rude Figure 1: Example of TOWE. Words in red are opinion targets and words in blue are corresponding opinion words. TOWE extracts corresponding opinion words when given opinion targets. example of the TOWE task. For the sentence “The dishes are amazingly delicious but the waiter is so rude.”, the terms “dishes” and “waiter” are two opinion targets. The goal of TOWE is to extract “amazingly delicious” as the opinion words for the opinion target “dishes” and opinion word “rude” when given the opinion target “waiter”. While seminal work casts TOWE as a sequence labeling problem, using fully supervised learning methods to identify opinion words and phrases from sentences, recent work shows that external sources of information can be highly useful for improving the performance. In particular, both syn- tactic knowledge (Dai et al., 2022; Veyseh et al., 2020; Jiang et al., 2021; Zhang et al., …

Content truncated to 500 words