| Summary: | Traditionally, approaches based on neural networks to solve the problem of disambiguation of the meaning of words (WSD) use a set of classifiers at the end, which results in a specialization in a single set of words-those for which they were trained. This makes impossible to apply the learned models to words not previously seen in the training corpus. This paper seeks to address a generalization of the problem of WSD in order to solve it through deep neural networks without limiting the method to a fixed set of words, with a performance close to the state-of-the-art, and an acceptable computational cost. We explore different architectures based on multilayer perceptrons, recurrent cells (Long Short-Term Memory-LSTM and Gated Recurrent Units-GRU), and a classifier model. Different sources and dimensions of embeddings were tested as well. The main evaluation was performed on the Senseval 3 English Lexical Sample. To evaluate the application to an unseen set of words, learned models are evaluated in the completely unseen words of a different corpus (Senseval 2 English Lexical Sample), overcoming the random baseline.
|
|---|