Inferences on regions in semantic space

Katrin Erk

Eighth International Conference on Computational Semantics (IWCS-8 2009)
Tilburg University, Netherlands, January 7-9, 2009

---

Summary

Semantic space models represent the meaning of a word as a vector in high-dimensional space. This framework facilitates inferences involving paraphrases, but it is not clear how it could support inferences involving hyponymy, like "horse ran -> animal moved". In this paper, we propose that by extending the semantic space representation of a word from a point to a region, we can represent hyponymy as the subregion relation.

We present two models for inducing a region representation for word meaning in semantic space, based on fact that points at close distance tend to represent similar meanings. Both models perform over 95% F-score on an occurrence classification task.

---

START Conference Manager (V2.56.8 - Rev. 414)