Corpus-based Induction of a Frame Semantics Projection for LFG

In computational linguistics there is growing insight that high-quality NLP applications for information access (question anwering, etc.) are in need of deeper linguistic analysis, in particular, semantic analysis. A bottleneck for semantic processing is the lack of large-scale domain-independent lexical semantic resources. While WordNets for several languages are important lexical resources for paratactic semantic relations, we are lacking large-scale lexical resources for the semantic description of relational lexical items, in particular predicate-argument structure. There are now efforts for the creation of large lexical semantic resources that provide information on predicateargument structure. The Berkeley FrameNet project [2], building on Fillmore’s theory of frame semantics [5], has a strong lexicographic character, concentrating on the definition of an inventory of frames and their semantic roles, and manual annotation of selected example sentences. A strictly corpus-based approach is carried out at the UPenn, with the creation of “PropBank” [8] – a manual predicate-argument annotation on top of the PennTreebank. Similar to corpus-based induction of stochastic grammars from syntactically annotated corpora, there are first approaches for learning models for semantic role assignment from semantically annotated corpora (e.g. [7]). However, since semantic role assignment is based on a syntactic annotation layer, fully automatic processing for semantic role assignment requires an interface between a syntactic analyser and the targeted semantic representation. In [1] we presented a design study for an LFG syntax-semantics projection architecture for frame semantics. In this paper we report on the corpus-based induction of an operational computational LFG interface for frame semantic annotation. We describe the methods we apply in order to derive generic frame assignment rules from annotation instances of a manually annotated corpus (see example in Fig. 1(a)). We further present first experiments to assess the potential of frame semantic processing for semantics-driven syntactic disambiguation.