Una aproximación al uso de word embeddings en una tarea de similitud de textos en español

In this paper we show how a vector representation of words based on word embeddings can help to improve the results in tasks focused on the semantic similarity of texts. Thus we have experimented with two methods that rely on the vector representation of words to calculate the degree of similarity of two texts, one based on the aggregation of vectors and the other one based on the calculation of alignments. The alignment method relies on the similarity of word vectors to determine the semantic link between them. The aggregation method allows us to construct vector representations of the texts from the individual vectors of each word. These representations are compared by means of two classic distance measures: Euclidean distance and cosine similarity. We have evaluated our systems with the corpus based on Wikipedia distributed in the competition of similarity of texts in Spanish of SemEval-2015. Our experiments show that the method based on the alignment of words performs much better, obtaining results that are very close to the best system at SemEval. The method based on vector representations of texts behaves substantially worse. However, this second approach seems to capture aspects of similarity not detected by the first one, as when the outputs of both systems are combined the results of the alignment method are surpassed, even exceeding the results of the best system at SemEval.