General estimation and evaluation of compositional distributional semantic models

摘要

In recent years, there has been widespread interest in compositional distributional semantic models (cDSMs), that derive meaning representations for phrases from their parts. We present an evaluation of alternative cDSMs under truly comparable conditions. In particular, we extend the idea of Baroni and Zamparelli (2010) and Guevara (2010) to use corpus-extracted examples of the target phrases for parameter estimation to the other models proposed in the literature, so that all models can be tested under the same training conditions. The linguistically motivated functional model of Baroni and Zamparelli (2010) and Coecke et al. (2010) emerges as the winner in all our tests.