{Roberto Tedesco Vincenzo Scotti Irene Morazzoni}

Abstract

Def2Vec introduces a novel paradigm for word embeddings, leveraging dictionary definitions to learn semantic representations. By constructing term-document matrices from definitions and applying Latent Semantic Analysis (LSA), Def2Vec generates embeddings that offer both strong performance and extensibility. In evaluations encompassing Part-of-Speech tagging, Named Entity Recognition, chunking, and semantic similarity, Def2Vec often matches or surpasses state-of-the-art models like Word2Vec, GloVe, and fastText. Our model’s second factorised matrix resulting from LSA enables efficient embedding extension for out-of-vocabulary words. By effectively reconciling the advantages of dictionary definitions with LSA-based embeddings, Def2Vec yields informative semantic representations, especially considering its reduced data requirements. This paper advances the understanding of word embedding generation by incorporating structured lexical information and efficient embedding extension.