SRL4ORL: Improving Opinion Role Labeling Using Multi-Task Learning with Semantic Role Labeling

Proceedings of NAACL-HLT 2018, pages 583–594 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics SRL4ORL: Improving Opinion Role Labeling Using Multi-Task Learning With Semantic Role Labeling Ana Marasovi´c and Anette Frank Research Training Group AIPHES Department of Computational Linguistics Heidelberg University {marasovic,frank}@cl.uni-heidelberg.de Abstract For over a decade, machine learning has been used to extract opinion-holder-target struc- tures from text to answer the question Who ex- pressed what kind of sentiment towards what?. Recent neural approaches do not outperform the state-of-the-art feature-based models for Opinion Role Labeling (ORL). We suspect this is due to the scarcity of labeled train- ing data and address this issue using differ- ent multi-task learning (MTL) techniques with a related task which has substantially more data, i.e. Semantic Role Labeling (SRL). We show that two MTL models improve signif- icantly over the single-task model for label- ing of both holders and targets, on the devel- opment and the test sets. We found that the vanilla MTL model which makes predictions using only shared ORL and SRL features, per- forms the best. With deeper analysis we deter- mine what works and what might be done to make further improvements for ORL. 1 Introduction Fine-Grained Opinion Analysis (FGOA) aims to: (i) detect opinion expressions (O) that convey at- titudes such as sentiments, agreements, beliefs or intentions (like feared in example (1)), (ii) mea- sure their intensity (e.g. strong), (iii) identify their holders (H), i.e. entities that express an attitude (e.g. it), (iv) identify their targets (T), i.e. enti- ties or propositions at which the attitude is di- rected (e.g. violence) and (v) classify their target- dependent attitude (e.g. negative sentiment)1. (1) Australia said [it]H [feared]Oneg [violence]T if voters thought the election had been stolen. As the commonly accepted benchmark corpus MPQA (Wiebe et al., 2005) uses span-based an- notations to represent opinion entities (opinions, 1Examples are drawn from MPQA (Wiebe et al., 2005). holders and targets), the task is usually approached with sequence labeling techniques and the BIO encoding scheme (Choi et al., 2006; Yang and Cardie, 2013; Katiyar and Cardie, 2016). Initially pipeline models were proposed which first pre- dict opinion expressions and then, given an opin- ion, label its opinion roles, i.e. holders and tar- gets (Kim and Hovy, 2006; Johansson and Mos- chitti, 2013). Pipeline models have been sub- stituted with so-called joint models that simulta- neously identify all opinion entities, and predict which opinion role is related to which opinion (Choi et al., 2006; Yang and Cardie, 2013; Katiyar and Cardie, 2016). Recently an LSTM-based joint model was proposed (Katiyar and Cardie, 2016) that unlike the prior work (Choi et al., 2006; Yang and Cardie, 2013) does not depend on external re- sources (such as syntactic parsers or named en- tity recognizers). The neural variant does not out- perform the feature-based CRF model (Yang and Cardie, 2013) in Opinion Role Labeling (ORL). Both the neural and the CRF joint models achieve about 55% F1 score for predicting which targets …

Content truncated to 500 words