Syntax-Aware Opinion Role Labeling with Dependency Graph Convolutional Networks

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3249–3258 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3249 Syntax-Aware Opinion Role Labeling with Dependency Graph Convolutional Networks Bo Zhang1, Yue Zhang2, Rui Wang2, Zhenghua Li1∗, Min Zhang1 1. Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, Suzhou, China 2. Alibaba Inc., Hangzhou, China {bzhang17}@stu.suda.edu.cn, {zhli13,minzhang}@suda.edu.cn {shiyu.zy,masi.wr}@alibaba-inc.com Abstract Opinion role labeling (ORL) is a fine-grained opinion analysis task and aims to answer “who expressed what kind of sentiment to- wards what?”. Due to the scarcity of labeled data, ORL remains challenging for data-driven methods. In this work, we try to enhance neu- ral ORL models with syntactic knowledge by comparing and integrating different represen- tations. We also propose dependency graph convolutional networks (DEPGCN) to encode parser information at different processing lev- els. In order to compensate for parser inac- curacy and reduce error propagation, we in- troduce multi-task learning (MTL) to train the parser and the ORL model simultaneously. We verify our methods on the benchmark MPQA corpus. The experimental results show that syntactic information is highly valuable for ORL, and our final MTL model effectively boosts the F1 score by 9.29 over the syntax- agnostic baseline. In addition, we find that the contributions from syntactic knowledge do not fully overlap with contextualized word repre- sentations (BERT). Our best model achieves 4.34 higher F1 score than the current state-of- the-art. 1 Introduction Opinion and sentiment analysis has a wide range of real-world applications like social media monitor- ing (Bollen et al., 2011), stock market prediction (Nguyen et al., 2015), box office prediction (Yu et al., 2010), and general e-commerce applications (Kim et al., 2013; Hu et al., 2017; Cui et al., 2017). In particular, fine-grained opinion analysis aims to identify users’ opinions in a text, including opinion expressions, holders of the opinions, targets of the opinions, target-dependent attitude, and intensity of opinions (Marasovi´c and Frank, 2018), which is very important for understanding political stance, ∗Corresponding author $ Cardoso says challenge facing Chavez is . . . Holder Expression Target Figure 1: An Example of ORL (bottom) and syntactic dependency tree (top) for “Cardoso says challenge fac- ing Chavezis is reestablishing normalcy.” customers’ reviews, marketing trends, and other subjective information (Ravi and Ravi, 2015). As a typical fine-grained opinion mining task, opinion role labeling (ORL) aims to identify different roles relevant to each opinion, i.e., who expressed what kind of sentiment towards what (Liu, 2012). Due to the lack of large-scale labeled data, ORL remains a challenging task to tackle. As a reference point, semantic role labeling (SRL) is very similar to ORL in the problem definition, but has 10 times more labeled data and thus achieves much higher performance than ORL (80∼90 vs. 60∼70 in F1 score). Motivated by the correlations between the two tasks, SRL has been utilized to help the ORL task by many previous studies (Ruppenhofer et al., 2008; Marasovi´c and Frank, 2018; Zhang et al., 2019b). …

Content truncated to 500 words