Neural word embeddings with multiplicative feature interactions for tensor-based compositions

Categorical compositional distributional models unify compositional formal semantic models and distributional models by composing phrases with tensor-based methods from vector representations. For the tensor-based compositions, Milajevs et al. (2014) showed that word vectors obtained from the continuous bag-of-words (CBOW) model are competitive with those from co-occurrence based models. However, because word vectors from the CBOW model are trained assuming additive interactions between context words, the word composition used for the training mismatches to the tensor-based methods used for evaluating the actual compositions including pointwise multiplication and tensor product of context vectors. In this work, we show whether the word embeddings from extended CBOW models using multiplication or tensor product between context words, reflecting the actual composition methods, can show better performance than those from the baseline CBOW model in actual tasks of compositions with multiplication or tensor-based methods.