Type-theoretic frameworks for compositional semantics are aimed at producing structured meaning representations of natural language utterances. Using elements of lexical semantics, these frameworks are able to represent many difficult phenomena related to the polysemy of words and their context-dependent meanings. However, they are just as powerful as their built-in lexical resources. This paper explores ways to create and enrich wide-coverage, weighted lexical resources from crowd-sourced data, specifically, investigating how existing rich lexical networks-created and validated by serious games-can be used to infer linguistic coercions with ranking.