Polysemy in Compositional Distributional Semantics

Research in distributional semantics has made good progress in capturing individual word meanings using contextual frequencies obtained from a large corpus. While vocabulary of a language is limited, its generative power for combinatorial expressions is nonrestrictive, and so lexical semantic methods cannot be applied directly to phrasal or sentential semantics irrespective of the corpus size. Any distributional model that aims to describe a language adequately needs to address the issue of compositionality. Very recently, a new field called Compositional Distributional Semantics (CDS) emerged, stretching the boundaries of distributional semantics from word level meaning representation to higher levels such phrasal and sentential semantic representations. CDS models deal with the task of composing the meaning of a phrase/sentence from the distributional meaning of its constituents. Polysemy of words have been a major focus in distributional semantics. The challenges posed at lexical level make a transition to phrasal and higher levels, making polysemy a major threat to CDS models. In this thesis, we aim to build better CDS models by performing sense disambiguation. We test our hypothesis, sense disambiguation benefits compositional models, on different compositionality based evaluation tasks. The evaluation of compositional models is an uncertain topic. Since we humans do not know the way we compose semantics of expressions, it is hard to prepare datasets for evaluation, thus making the evaluation of CDS models a challenging topic. In this thesis, we focus on evaluation methods for compositional models and develop a dataset with a novel annotation scheme.