Manifold Learning for the Semi-Supervised Induction of FrameNet Predicates: An Empirical Investigation

This work focuses on the empirical investigation of distributional models for the automatic acquisition of frame inspired predicate words. While several semantic spaces, both word-based and syntax-based, are employed, the impact of geometric representation based on dimensionality reduction techniques is investigated. Data statistics are accordingly studied along two orthogonal perspectives: Latent Semantic Analysis exploits global properties while Locality Preserving Projection emphasizes the role of local regularities. This latter is employed by embedding prior FrameNet-derived knowledge in the corresponding non-euclidean transformation. The empirical investigation here reported sheds some light on the role played by these spaces as complex kernels for supervised (i.e. Support Vector Machine) algorithms: their use configures, as a novel way to semi-supervised lexical learning, a highly appealing research direction for knowledge rich scenarios like FrameNet-based semantic parsing.