We investigate the relationship between lexical spaces and contextually-defined conceptual spaces, offering applications to creative concept discovery. We define a computational method for discovering members of concepts based on semantic spaces: starting with a standard distributional model derived from corpus co-occurrence statistics, we dynamically select characteristic dimensions associated with seed terms, and thus a subspace of terms defining the related concept. This approach performs as well as, and in some cases better than, leading distributional semantic models on a WordNet-based concept discovery task, while also providing a model of concepts as convex regions within a space with interpretable dimensions. In particular, it performs well on more specific, contextualized concepts; to investigate this we therefore move beyond WordNet to a set of human empirical studies, in which we compare output against human responses on a membership task for novel concepts. Finally, a separate panel of judges rate both model output and human responses, showing similar ratings in many cases, and some commonalities and divergences which reveal interesting issues for computational concept discovery.
translated by 谷歌翻译