Comparing Different Word Embeddings for Multiword Expression Identification

Abstract

The identification of Multi-Word Expressions (MWEs) is central to resolving ambiguity of phrases. Recent works show that deep learning methods outperform statistical and lexical based approaches. The deep learning approaches mostly use word2vec embedding; our paper aims at comparing the use of word2vec, GloVe, and a combination of the two word embeddings in identifying MWEs. GloVe, and the combination of word2vec and GloVe were marginally better in terms of F-score, identifying more unique words, and identifying words not seen in the train data. GloVe was marginally better at identifying Verbal Multi-Word Expressions (VMWEs) which tend to be the hardest group of MWEs because they can be gappy, which is caused by interleaving of words that are part of the MWE and words that are not part of the MWE. The major purpose of the paper is to compare the use of different word embeddings in identifying MWEs and not to suggest improvements to the state-of-the-art. Future work using different dimensions of word embedding vectors and use of fasttext are suggested.