Perhaps a tangent: If there is a dataset of word connotations vs word meaning closeness, this can be extended to find other language inconsistencies.