Probabilistic models of human-annotated semantic proximity graphs
Polysemy is everywhere in human language. Yet, the phenomenon is still very poorly understood and modeled. A central question is whether sense assignments in context are discrete or graded (discreteness). Another question is if and how senses are geometrically represented in the mental lexicon (spatial organization). Equally important is the question to what extent word meanings are shared between speakers (individuality). With this study I aim to introduce a new type of evidence and the corresponding modeling perspective into the linguistic debate around polysemy. The evidence comes from human-annotated "Word Usage Graphs", representing word uses in a weighted undirected graph with edge weights reflecting how similar in meaning two uses are (semantic proximity). These graphs can be seen as rich semantic measurements of latent structures underlying the cognitive organization of word meaning. The most adequate explanatory model for these measurements can be rigorously chosen through Bayesian model comparison. Hypotheses about discreteness, spatial organization and individuality of word meaning can be encoded into such graph models, e.g. in the form of latent blocks or spatial regions of word uses. In the talk, I will introduce you to the central data and the models used in the study, and present first model comparison results. I will also describe the main modeling challenges resulting e.g. from data sparsity.