Combining Seemingly Incompatible Corpora for Implicit Semantic Role Labeling

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 40–50, Denver, Colorado, June 4–5, 2015. Combining Seemingly Incompatible Corpora for Implicit Semantic Role Labeling Parvin Sadat Feizabadi and Sebastian Pad´o Institut f¨ur Maschinelle Sprachverarbeitung University of Stuttgart, Germany {parvin.feizabadi,pado}@ims.uni-stuttgart.de Abstract Implicit semantic role labeling, the task of retrieving locally unrealized arguments from wider discourse context, is a knowledge- intensive task. At the same time, the annotated corpora that exist are all small and scattered across different annotation frameworks, genres, and classes of predicates. Previous work has treated these corpora as incompatible with one another, and has concentrated on optimizing the exploitation of single corpora. In this paper, we show that corpus combination is effective after all when the differences between corpora are bridged with domain adaptation methods. When we combine the SemEval-2010 Task 10 and Gerber and Chai noun corpora, we obtain substantially improved performance on both corpora, for all roles and parts of speech. We also present new insights into the properties of the implicit semantic role labeling task. 1 Introduction Semantic role labeling (SRL) is the task of identify- ing semantic arguments of predicates in text. It is an important step in text analysis and has applications in information extraction (Christensen et al., 2010), question answering (Shen and Lapata, 2007; Moreda et al., 2011) and machine translation (Wu and Fung, 2009; Xiong et al., 2012) . A large body of work exists on algorithms for SRL (Gildea and Jurafsky, 2002; Srikumar and Roth, 2011). Their success is closely connected to the availability of two large, hand-constructed semantic role resources, FrameNet (Fillmore et al., 2003) and PropBank (Palmer et al., 2005). They used to concentrate on overt semantic roles, that is, semantic roles that are realized within the local syntactic structure of the predicate. Recent years have seen a broadening of the focus in SRL to implicit semantic roles, that is all roles that remain locally unrealized but can be retrieved in the (typically prior) context (Ruppenhofer et al., 2010). In the following example annotated with PropBank roles (cf. Section 2), the target predicate come has two roles, a locally realized one (A1, the entity in motion), it, and an implicit role mentioned in the previous sentence (A4, the goal): Well, sir, it’s [A4 this lonely, silent house] and the queer thing in the kitchen . ... I thought [A1 it] had come again. Implicit SRL is useful to complete predicates’ ar- gument structures for inference (Mirkin et al., 2010) and paraphrasing (Roth and Frank, 2013), or to assess the coherence of discourse (Burchardt et al., 2005). It however requires (even) more training data than traditional SRL. One reason is that potential argu- ments come from the whole text rather than just the sentence. Another one is that most of the power- ful syntactic features that are a staple in traditional SRL are unavailable across sentence boundaries. Un- fortunately, existing corpora for implicit SRL are quite small: The task requires full-text annotation, which is …

Content truncated to 500 words