A character-based part-of-speech tagger with feedforward neural networks

Abstract

This article presents a simple method to perform part-of-speech (POS) tagging with feedforward neural networks applied to learnable character embeddings. The motivation of the research is based on the fact that for some languages a human can find out the part of speech for a word just by its spelling even without knowing the meaning of the word (see C.Fries’s example “woggles ugged diggles”). One of the goals was to achieve high accuracy tagging without using semantic information (e.g. without word embeddings). This allows performing tagging for out of vocabulary words. Also, the dependency of the performance from context size was studied. The plausibility of the method was proved by building a POStagger with the accuracy comparable to state-of-the-art results.