New Directions in Vector Space Models of Meaning
Edward Grefenstette‚ Karl Moritz Hermann‚ Georgiana Dinu and Phil Blunsom
Symbolic approaches have dominated NLP as a means to model syntactic and semantic aspects of natural language. While powerful inferential tools exist for such models, they suffer from an inability to capture correlation between words and to provide a continuous model for word, phrase, and document similarity. Distributed representations are one mechanism to overcome these constraints. This tutorial will supply NLP researchers with the mathematical and conceptual background to make use of vector-based models of meaning in their own research. We will begin by motivating the need for a transition from symbolic representations to distributed ones. We will briefly cover how collocational (distributional) vectors can be used and manipulated to model word meaning. We will discuss the progress from distributional to distributed representations, and how neural networks allow us to learn word vectors and condition them on metadata such as parallel texts, topic labels, or sentiment labels. Finally, we will present various forms of semantic vector composition, and discuss their relative strengths and weaknesses, and their application to problems such as language modelling, paraphrasing, machine translation and document classification. This tutorial aims to bring researchers up to speed with recent developments in this fast-moving field. It aims to strike a balance between providing a general introduction to vector-based models of meaning, an analysis of diverging strands of research in the field, and also being a hands-on tutorial to equip NLP researchers with the necessary tools and background knowledge to start working on such models. Attendees should be comfortable with basic probability, linear algebra, and continuous mathematics. No substantial knowledge of machine learning is required.