Date of Award


Document Type

Thesis (Undergraduate)

Department or Program

Cognitive Science

First Advisor

Jonathan Phillips

Second Advisor

Soroush Vosoughi


In this paper I evaluate the ability of different Natural Language Processing (NLP)
techniques to make human-like word relatedness judgements in a variant of the word-based board game Codenames. I analyze a variety of statistical and knowledge based approaches, combinations of these, and techniques for incorporating the wider game context into relatedness judgements. While no approach explored here reaches human performance, simple word embedding based approaches incorporate a surprising amount of the useful information captured by other techniques. I attempt to characterize the limitations of these approaches in relation to human game play, although differences are largely not systematic. Finally, I discuss these results in terms of future directions for the field of NLP.