Unsupervised Adversarial Domain Adaptation for Implicit Discourse Relation Classification

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 686–695 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 686 Unsupervised Adversarial Domain Adaptation for Implicit Discourse Relation Classification Hsin-Ping Huang Department of Computer Science The University of Texas at Austin hsinping@cs.utexas.edu Junyi Jessy Li Department of Linguistics The University of Texas at Austin jessy@austin.utexas.edu Abstract Implicit discourse relations are not only more challenging to classify, but also to annotate, than their explicit counterparts. We tackle sit- uations where training data for implicit rela- tions are lacking, and exploit domain adapta- tion from explicit relations (Ji et al., 2015). We present an unsupervised adversarial do- main adaptive network equipped with a recon- struction component. Our system outperforms prior works and other adversarial benchmarks for unsupervised domain adaptation. Addi- tionally, we extend our system to take advan- tage of labeled data if some are available. 1 Introduction Discourse relations capture the relationship be- tween units of text—e.g., sentences and clauses— and are an important aspect of text coherence. While some relations are expressed explicitly with a discourse connective (e.g., “for example”, “how- ever”), relations are equally often expressed im- plicitly without an explicit connective (Prasad et al., 2008); in these cases, the relation needs to be inferred. Resources for implicit discourse relations are scarce compared to the explicit ones, since they are harder to annotate (Miltsakaki et al., 2004). For example, among corpora annotated with dis- course relations such as Arabic (Al-Saif and Mark- ert, 2010), Czech (Pol´akov´a et al., 2013), Chi- nese (Zhou and Xue, 2015), English (Prasad et al., 2008), Hindi (Oza et al., 2009), and Turk- ish (Zeyrek et al., 2013), only the Chinese, English and Hindi corpora include implicit discourse rela- tions (Prasad et al., 2014). In this low-resource scenario, Ji et al. (2015) proposed training with explicit relations via unsupervised domain adapta- tion, viewing explicit relations as a source domain with labeled training data, and implicit relations as a target domain with no labeled data. The do- main gap between explicit and implicit relations is acknowledged by prior observations that the two types of discourse relations are linguistically dis- similar (Sporleder and Lascarides, 2008; Ruther- ford and Xue, 2015). We present a new system for the unsuper- vised domain adaptation setup on the Penn Dis- course Treebank (Prasad et al., 2008). Our sys- tem is based on Adversarial Discriminative Do- main Adaptation (Tzeng et al., 2017), which de- couples source domain training and representa- tion mapping between source and target. We im- prove this framework by proposing a reconstruc- tion component to preserve the discriminability of target features, and incorporating techniques for stabler training on textual data. Experimental results show that even with a simple architecture for representation learn- ing, our unsupervised domain adaptation sys- tem outperforms prior work by 1.4-2.3 macro F1, with substantial improvements on Temporal and Contingency relations. It is also superior to DANN (Ganin et al., 2016), an adversarial frame- work widely used in NLP (Chen et al., …

Content truncated to 500 words