Teaching: Distributional Models for Lexical Semantics
A short course at the International Summer School on Typology and Lexicon (TyLex)
Annotation
The course discusses vector-based distributional semantic models and their application to linguistic problems.
Distributional semantic models, also known as word embeddings, are an increasingly popular tool in computational semantics. In a distributional model, each word (or other linguistic expression of interest) is represented as a multidimensional numeric vector. Vector operations such as addition, pointwise multiplication, or linear maps can serve as analogs of semantic composition operations for vector based representations.
Comparisons between vectors are used to predict various semantic relations between words, such as semantic relatedness or hyperonymy, various psycholinguistic measures such as the degree of semantic priming between words, and, lately, diachronic lexical semantic development and lexical semantic typology.
Demo
If you want to do vectors yourself, try the Dissect toolkit which is a set of pre-built tools for working with distributional vector models and vector composition. You can try seeing how it works using demo materials.
List of references (suggested reading)
Slides
Assignments