Workshop on Semantic Spaces at the Intersection of NLP, Physics, and Cognitive Science
Martha Lewis (University of Amsterdam, Netherlands), Dan Marsden (University of Oxford, United Kingdom) and Mehrnoosh Sadrzadeh (Queen Mary University of London, United Kingdom)
Vector embeddings of word meanings have become a mainstream tool in large scale natural lan- guage processing tools through implementations publicly available by Google and Facebook (Word2Vec and FastText). These models rely on a distributional hypothesis, put forward by Harris and Firth, that words that often occur in the same contexts have similar meanings. They have been applied to many language tasks, such as summarisation, disambiguation, name entity recognition, all through their key characteristic, which is the ability to reason about degrees of similarity between word meanings via the distances between the embeddings. The notion of vectors and their distances have also been employed in cognitive science, where we have the vector symbolic architectures of Smolensky  and the conceptual spaces of Gärdenfors . Unrelated to natural language and cognitive science, vectors and vector spaces have been extensively used as models of physical theories and especially the theory of quantum mechanics. These were first put forward by Hilbert and led to one of the first axiomatisations of quantum theory. They were successively used by Birkhoff and von Neumann  to develop the first logic to reason about properties of quantum processes. The similarities between the vector representations of quantum mechanics and those of natural language were first discovered by Van Rijsbergen  in the context of vector models of documents in information retrieval. Recently this connection was rediscovered through the work of Clark and Pulman  and successively of Clark, Coecke, and Sadrzadeh  also independently through the work of Lambek  on bicompact linear logic and compact closed structures in natural language.
Based on these, we propose the presentation of a workshop on the related areas of natural language processing (NLP), physics, and cognitive science. The interplay between the three disciplines will foster theoretically motivated approaches to understanding how meanings of words are formed and how they interact to form meanings of sentences and other discourse units, how concepts form and develop in the mind, and how the meanings of these words and concepts get exchanged via utterances, such as those in a dialogue. Commonalities between these different levels of compositional mechanisms will be extracted, and applications and phenomena traditionally thought of as ‘non-compositional’ will be examined. Experimental evaluations of the models using corpora of text, dialogue, and interactive data will form part of the body of submissions.
This workshop is the third in the series of workshops run previously in 2016 collocated with the 13th International Conference on Quantum Physics and Logic, and in 2018 with the International Symposium on Quantum Interactions. Keynote speakers so far have included Paul Smolensky (physics and cognitive science), Peter Gärdenfors (cognitive science and linguistics), and Dominic Widdows (vector space models of natural language), each covering a related strand of the workshop themes.