Distributional semantics has contributed much to the recent advances in NLP, but it still has many theoretical challenges - and one of the biggest challenges is compositionality. This tutorial will cover the current proposals for representation and interpretation of semantic features in word-level vector space models (VSMs), representation of morphological information, building sentence representations and encoding abstract linguistic structures that are necessary for grammar but hard to capture distributionally. For each problem we will discuss the existing evaluation datasets and ways to improve them. The tutorial will conclude with a hands-on section covering how some of these issues could be addressed with Vecto, a new open-source Python library for conducting large-scale, reproducible experiments with VSMs.
This page contains slides and materials for the hands-on session.