Chinese Opinion Role Labeling with Corpus Translation: A Pivot Study

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10139–10149 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10139 Chinese Opinion Role Labeling with Corpus Translation: A Pivot Study Ranran Zhen1, Rui Wang2, Guohong Fu3∗, Chengguo Lv1, Meishan Zhang4 1School of Computer Science and Technology, Heilongjiang University, China 2Vipshop (China) Co., Ltd. 3Institute of Artificial Intelligence, Soochow University, China 4School of New Media and Communication, Tianjin University, China {zenrran, mason.zms}@gmail.com {mars198356, ghfu}@hotmail.com 2004085@hlju.edu.cn Abstract Opinion Role Labeling (ORL), aiming to iden- tify the key roles of opinion, has received in- creasing interest. Unlike most of the previ- ous works focusing on the English language, in this paper, we present the first work of Chi- nese ORL. We construct a Chinese dataset by manually translating and projecting an- notations from a standard English MPQA dataset. Then, we investigate the effective- ness of cross-lingual transfer methods, includ- ing model transfer and corpus translation. We exploit multilingual BERT with Contextual Pa- rameter Generator and Adapter methods to examine the potentials of unsupervised cross- lingual learning and our experiments and anal- yses for both bilingual and multilingual trans- fers establish a foundation for the future re- search of this task1. 1 Introduction Fine-grained opinion mining has been a crucial task in natural language processing (NLP) for a long time (Kim and Hovy, 2006a; Breck et al., 2007; Wilson et al., 2009; Qiu et al., 2011; Irsoy and Cardie, 2014; Liu et al., 2015; Wiegand et al., 2016) and it aims to discover useful structural informa- tion of user opinions from unstructured text, which is the relation between expression and entities such as Who expressed what kind of sentiment towards what?. The EXPRESSION conveys attitudes includ- ing sentiments, agreements, beliefs, or intentions (e.g., voiced his condolences in Figure 1); the enti- ties consist of the HOLDER who expresses the opin- ion (e.g., Chen.) and the TARGET which the opin- ion is expressed to (e.g., the families) (Breck et al., 2007; Yang and Cardie, 2012; Katiyar and Cardie, 2016a). Here we focus on the opinion role labeling (ORL) task which is to identify opinion holders and ∗Corresponding author. nese1We release the code and way of obtaining Chi- dataset at https://github.com/zenRRan/ ChineseORL-with-Corpus-Translation. [Chen]holder [voiced his condolences]expression to [the families]target. Figure 1: An example of fine-grained opinion mining. targets (Marasovi´c and Frank, 2017; Zhang et al., 2019a) when the expressions are given. Most of the previous researches focus on the En- glish ORL, benefiting from the benchmark MPQA dataset (Wiebe et al., 2005) which includes span- based annotations of opinion expressions, holders and targets. The task is commonly solved by se- quence labeling models with the BIO conversion scheme (Kim and Hovy, 2006b; Choi et al., 2006; Yang and Cardie, 2013; Johansson and Moschitti, 2013). Recently, neural BiLSTM-CRF models have achieved state-of-the-art performance on this task (Katiyar and Cardie, 2016b; Marasovi´c and Frank, 2017; Zhang et al., 2019a). However, the studies on other languages are relatively rare due to the scarcity of annotated datasets. …

Content truncated to 500 words