Résumé
Current NLP systems heavily rely on embedding techniques that are used to automatically encode relevant information about linguistic entities of interest (e.g., words, sentences) into latent spaces. These embeddings are currently the cornerstone of the best machine learning systems used in a large variety of problems such as text classification. Interestingly, state-of-the-art embeddings are commonly only computed using large corpora, and generally do not use additional knowledge expressed into established knowledge resources (e.g. WordNet). In this paper, we empirically study if retrofitting, a class of techniques used to update word vectors in a way that takes into account knowledge expressed in knowledge resources, is beneficial for short text classification. To this aim, we compared the performances of several state-of-the-art classification techniques with or without retrofitting on a selection of benchmarks. Our results show that the retrofitting approach is beneficial for some classifiers settings and only for datasets that share a similar domain to the semantic lexicon used for the retrofitting.