UFSAC: Unification of Sense Annotated Corpora and Tools

UFSAC: Unification of Sense Annotated Corpora and Tools Lo¨ıc Vial, Benjamin Lecouteux, Didier Schwab LIG – GETALP – Univ. Grenoble Alpes – France {loic.vial, benjamin.lecouteux, didier.schwab}@imag.fr Abstract In Word Sense Disambiguation, sense annotated corpora are often essential for evaluating a system and also valuable in order to reach a good efficiency. Always created for a specific purpose, there are today a dozen of sense annotated English corpora, in various formats and using different versions of WordNet. The main hypothesis of this work is that it should be possible to build a disambiguation system by using any of these corpora during the training phase or during the testing phase regardless of their original purpose. In this article, we present UFSAC: a format of corpus that can be used for either training or testing a disambiguation system, and the process we followed for constructing this format. We give to the community the whole set of sense annotated English corpora that we know, in this unified format, when the copyright allows it, with sense keys converted to the last version of WordNet. We also provide the source code for building these corpora from their original data, and a complete Java API for manipulating corpora in this format. The whole resource is available at the following URL: https://github.com/getalp/UFSAC. Keywords: Word Sense Disambiguation, sense annotated corpora, unified resource, tools 1. Introduction Whether they are used for the evaluation or for the learn- ing process of a Word Sense Disambiguation (WSD) sys- tem, the importance of sense annotated corpora in Natural Language Processing (NLP) is considerable. On one hand, the evaluation in vivo, i.e. the evaluation of a WSD sys- tem as part of a larger task, has never been really exploited. On the other hand, the evaluation in vitro, which uses di- rectly sense annotated corpora by comparing the output of a system to manual annotations, is predominant. Moreover, WSD systems exploiting examples from sense annotated corpora are generally far better than those which do not (Navigli et al., 2007; Moro and Navigli, 2015). At the time of its creation, WordNet (Miller, 1995) was un- doubtedly the only lexical database freely available for En- glish. Since the beginning of the 2000s, it has become the de facto standard for WSD in this language. Indeed, most of sense annotated corpora are either directly annotated with WordNet sense keys or they are annotated with a sense in- ventory linked to the senses of WordNet, such as BabelNet (Navigli and Ponzetto, 2010). However, it is not trivial to use these corpora, because most of them differ in their format and on the version of Word- Net they use. As a consequence, very few works in the lit- erature of WSD are trained or evaluated on more than two annotated corpora. Also, WSD systems are systematically evaluated on cor- pora that have been initially created for the purpose of eval- uation, and never on corpora that have been created for an- other purpose, such as training or for sense …

Content truncated to 500 words