Corpus-based Induction of a Frame Semantics Projection for LFG

We present a method for the corpus-based induction of an LFG syntax-semantics interface for frame semantic processing within a computational LFG parsing architecture.

Building on (Frank and Erk, 2004) we show how to model frame semantic annotations in an LFG projection architecture. We follow a modular DBA approach, by projecting frame semantic projections from completed f-structure output of LFG parsing, using the XLE term rewriting (transfer) system as processing machinery. We discuss the treatment of special phenomena that involve non-isomorphic mappings between levels, such as uncertain embedding and multiword expressions, as well as coordination and underspecification phenomena.

For the induction of LFG-based frame semantic projection rules we follow a corpus-based approach. Frame semantic annotations from the manually annotated SALSA corpus (Erk et al. 2003) are ported to a ``parallel'' LFG corpus (Forst 2003) by use of sentence-specific transfer rules. From this frame-enriched LFG corpus we extract functional path descriptions, to derive general LFG frame assignment rules that can be applied to new sentences.

We apply the induced frame assignment rules to the output of free LFG parsing to control the results and give first evaluation figures.