Improving Implicit Semantic Role Labeling by Predicting Semantic Frame Arguments

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 90–99, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Improving Implicit Semantic Role Labeling by Predicting Semantic Frame Arguments Quynh Ngoc Thi Do1, Steven Bethard2, Marie-Francine Moens1 1Katholieke Universiteit Leuven, Belgium 2University of Arizona, United States quynhngocthi.do@cs.kuleuven.be bethard@email.arizona.edu sien.moens@cs.kuleuven.be Abstract Implicit semantic role labeling (iSRL) is the task of predicting the semantic roles of a predicate that do not appear as explicit ar- guments, but rather regard common sense knowledge or are mentioned earlier in the discourse. We introduce an approach to iSRL based on a predictive recurrent neu- ral semantic frame model (PRNSFM) that uses a large unannotated corpus to learn the probability of a sequence of semantic arguments given a predicate. We lever- age the sequence probabilities predicted by the PRNSFM to estimate selectional preferences for predicates and their argu- ments. On the NomBank iSRL test set, our approach improves state-of-the-art perfor- mance on implicit semantic role labeling with less reliance than prior work on manu- ally constructed language resources. 1 Introduction Semantic role labeling (SRL) has traditionally fo- cused on semantic frames consisting of verbal or nominal predicates and explicit arguments that oc- cur within the clause or sentence that contains the predicate. However, many predicates, especially nominal ones, may bear arguments that are left implicit because they regard common sense knowl- edge or because they are mentioned earlier in a discourse (Ruppenhofer et al., 2010; Gerber et al., 2009). These arguments, called implicit arguments, are resolved by another semantic task, implicit se- mantic role labeling (iSRL). Consider a NomBank (Meyers et al., 2004) annotation example: [A0 The network] had been expected to have [NP losses] [A1 of $20 million] . . . Those [NP losses] may widen because of the short Series. The predicate loss in the first sentence has two ar- guments annotated explicitly: A0, the entity losing something, and A1, the thing lost. Meanwhile, the other instance of the same predicate in the second sentence has no associated arguments. However, for a good reader, a reasonable interpretation of the second loss should be that it receives the same A0 and A1 as the first instance. These arguments are implicit to the second loss. As an emerging task, implicit semantic role la- beling faces a lack of resources. First, hand-crafted implicit role annotations for use as training data are seriously limited: SemEval 2010 Task 10 (Baker et al., 1998) provided FrameNet-style (Baker et al., 1998) annotations for a fairly large number of pred- icates but with few annotations per predicate, while Gerber and Chai (2010) provided PropBank-style (Palmer et al., 2005) data with many more anno- tations per predicate but covering just 10 predi- cates. Second, most existing iSRL systems depend on other systems (explicit semantic role labelers, named entity taggers, lexical resources, etc.), and as a result not only need iSRL annotations to train the iSRL system, but annotations or manually built resources for all of their sub-systems as …

Content truncated to 500 words