Deeper syntax for better semantic parsing

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 409–420, Osaka, Japan, December 11-17 2016. Deeper syntax for better semantic parsing Olivier Michalon⋆ Corentin Ribeyre†∗ Marie Candito⋄ Alexis Nasr⋆ ⋆ Aix Marseille Univ, CNRS, Centrale Marseille, LIF, Marseille, France † Faculty of Humanities, University of Geneva, Switzerland ⋄ Alpage, Universit´e Paris Diderot, Sorbonne Paris Cit´e, INRIA, Paris, France ⋆ firstname.lastname@lif.univ-mrs.fr † firstname.lastname@unige.ch ⋄ firstname.lastname@inria.fr Abstract Syntax plays an important role in the task of predicting the semantic structure of a sentence. But syntactic phenomena such as alternations, control and raising tend to obfuscate the relation between syntax and semantics. In this paper we predict the semantic structure of a sentence using a deeper syntax than what is usually done. This deep syntactic representation abstracts away from purely syntactic phenomena and proposes a structural organization of the sentence that is closer to the semantic representation. Experiments conducted on a French corpus annotated with semantic frames showed that a semantic parser reaches better performances with such a deep syntactic input. 1 Introduction FrameNet (Baker et al., 1998) is an English resource containing a set of inter-related semantic frames, each frame containing a set of semantic roles (frame elements in FrameNet’s terminology). Frames offer semantic generalizations over individual predicates, since different lexical units can evoke the same frame, and semantic roles offer generalizations over syntactic arguments. Hence FrameNet parsing can be viewed as mixing predicate disambiguation and semantic role labelling.1 Although FrameNet is more semantically-oriented than other semantic role labeling resources such as PropBank (Palmer et al., 2005), syntactic information has been shown to be decisive for predicting (FrameNet) semantic roles since the early days of the task (Gildea and Jurafsky, 2002). Linking regular- ities provide the theoretical justification of this result: there exist regularities in how semantic arguments are realized in syntax. Yet it is well known that the mapping from syntactic arguments to semantic ones is not straightforward. First, lexical idiosyncrasies can come into play, for instance the Addressee of communication verbs may correspond to the indirect object for verbs like to say and to the direct object for a verb like to inform. Second, it is also well known that surface syntax exhibits variation that can obfuscate regularities. For instance though the Speaker is generally the subject of communication verbs, this does not hold when the verb is passivized. This difference disappears if syntactic alternations are neutralized, and the “canonical” diathesis of a verb is made explicit: the Speaker is the canonical subject in both active and passive voices. In this paper, we investigate the syntax-semantic interface in FrameNet annotated data, and study the impact of using “deeper” syntactic features to predict semantic frames and roles. More precisely, we take advantage of a deep syntactic dependency graphbank for French (Candito et al., 2014b; Ribeyre et al., 2014), which provides a level of representation that abstracts away from purely syntactic variation. The main contributions of the paper are (i) a comparison of the syntax/semantic regularities observed …

Content truncated to 500 words