Using Explicit Discourse Connectives in Translation for Implicit Discourse Relation Classification

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 484–495, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Using Explicit Discourse Connectives in Translation for Implicit Discourse Relation Classification Wei Shi1, Frances Yung1, Raphael Rubino1,3 and Vera Demberg1,2 1Dept. of Language Science and Technology 2Dept. of Mathematics and Computer Science, Saarland University 3German Research Center for Artificial Intelligence (DFKI) Saarland Informatic Campus, 66123 Saarbr¨ucken, Germany {w.shi, frances, vera}@coli.uni-saarland.de raphael.rubino@dfki.de Abstract Implicit discourse relation recognition is an extremely challenging task due to the lack of indicative connectives. Various neural network architectures have been proposed for this task recently, but most of them suffer from the shortage of la- beled data. In this paper, we address this problem by procuring additional training data from parallel corpora: When humans translate a text, they sometimes add con- nectives (a process known as explicita- tion). We automatically back-translate it into an English connective, and use it to in- fer a label with high confidence. We show that a training set several times larger than the original training set can be generated this way. With the extra labeled instances, we show that even a simple bidirectional Long Short-Term Memory Network can outperform the current state-of-the-art. 1 Introduction When humans comprehend language, their inter- pretation consists of more than just the sum of the content of the sentences. Additional seman- tic relations (known as coherence relations or dis- course relations) are inferred between sentences in the text. Identification of discourse relations is useful for various NLP applications such as ques- tion answering (Jansen et al., 2014; Liakata et al., 2013), summarization (Maskey and Hirschberg, 2005; Yoshida et al., 2014; Gerani et al., 2014), machine translation (Guzm´an et al., 2014; Meyer et al., 2015) and information extraction (Cimiano et al., 2005). Recently, the task has drawn increas- ing attention, including two CoNLL shared tasks (Xue et al., 2015, 2016). Discourse relations are sometimes expressed with an explicit discourse connective (DC), such as “because”, “but”, “if”. Example 1 shows an ex- plicit discourse relation marked by “because”; the text spans between which the relation holds are marked as Arg1 and Arg2. DCs serve as strong cues and allow us to classify discourse relations with high accuracy (Pitler et al., 2008, 2009; Lin et al., 2014). However, more than half of the discourse rela- tions in a text are not signalled by a connective. See for example 2: a contrastive relation can be in- ferred between the text spans marked as Arg1 and Arg2. Implicit relation classification is very chal- lenging and represents a bottleneck of the entire discourse parsing system. 1. [The city’s Campaign Finance Board has re- fused to pay Mr Dinkins $95,142 in matching funds]Arg1 because [his campaign records are incomplete.]Arg2 — Explicit, Contingency.Cause 2. [They desperately needed somebody who showed they cared for them, who loved them.]Arg1 [The last thing they needed was another drag-down blow.]Arg2 —Implicit, Comparison.Contrast In order to classify an implicit discourse rela- tion, it is necessary …

Content truncated to 500 words