Spanish Word Vectors from Wikipedia

Spanish word vectors from Wikipedia Mathias Etcheverry, Dina Wonsever mathiase@fing.edu.uy, wonsever@fing.edu.uy Universidad de la Republica Uruguay Abstract Contents analisys from text data requires semantic representations that are difficult to obtain automatically, as they may require large handcrafted knowledge bases or manually annotated examples. Unsupervised autonomous methods for generating semantic representa- tions are of greatest interest in face of huge volumes of text to be exploited in all kinds of applications. In this work we describe the generation and validation of semantic representations in the vector space paradigm for Spanish. The method used is GloVe (Pennington et al., 2014), one of the best performing reported methods , and vectors were trained over Spanish Wikipedia. The learned vectors eval- uation is done in terms of word analogy and similarity tasks (Pennington et al., 2014; Baroni et al., 2014; Mikolov et al., 2013a). The vector set and a Spanish version for some widely used semantic relatedness tests are made publicly available. Keywords: Distributional Semantics, Word Embeddings, Word Analogies 1. Introduction Learning representations for words from their contexts pro- vides us with new instruments to tackle natural language processing (NLP) tasks. In many supervised NLP tasks the training data is lim- ited and expensive but plain text is easily accessible thanks to the web. In opposition to NLP systems that considers words as atomic units, representing words as non-supervised learned vectors is a way to use unstructured data to improve results. These methods have a long history in Natural Language Processing, starting from Salton’s proposal in the field of Information Retrieval, where a document collection is represented as a two-dimensional array, with the element ai,j indicating the number of times that the word i occurs in document j. The word-document model gave a bag-of-words characterization to the subject contents of documents, in a completely unsupervised way. As words that co-occur in the same document tend to belong to a same subject, this methodology led to the implementation of the distributional hypothesis of lexical similarity: words that occur in the same contexts have similar meanings. From the array with all words i.e., the vocabulary, as rows and documents as columns we now have to consider a vocabulary x vocabulary array. And from the idea that similar contexts imply similar meanings we are now facing the fact that from this word-word co-occurrence array we can compute a new representation for words (using a huge volume of text) that ”summarizes” in some way its contexts (syntax and semantics included) and that has proved to be very effective in different NLP tasks. In this paper we present the generation and evaluation of a Spanish word-vector resource, obtained from Spanish Wikipedia text, by means of the method employed in GloVe (Pennington et al., 2014). In section 2 we present a general background, in sec- tion 3 we describe the GloVe method while in section 4 we explain the generation process. We have conducted a first evaluation of the generated resources (section 5) using some standard data sets for word …

Content truncated to 500 words