Huge Automatically Extracted Training-Sets for Multilingual Word SenseDisambiguation

Huge Automatically Extracted Training Sets for Multilingual Word Sense Disambiguation Tommaso Pasini, Francesco Elia, Roberto Navigli Sapienza University of Rome {pasini, elia, navigli}@di.uniroma1.it Abstract We release to the community six large-scale sense-annotated datasets in multiple language to pave the way for supervised multilingual Word Sense Disambiguation. Our datasets cover all the nouns in the English WordNet and their translations in other languages for a total of millions of sense-tagged sentences . Experiments prove that these corpora can be effectively used as training sets for supervised WSD systems, surpassing the state of the art for low-resourced languages and providing competitive results for English, where manually annotated training sets are accessible. The data is available at trainomatic.org. Keywords: Multilingual Word Sense Disambiguation, Resource, Dataset 1. Introduction Word Sense Disambiguation is a crucial task in Natural Language Processing as it can be beneficial to several downstream applications, i.e., natural language understand- ing, semantic parsing and question answering. Despite the task has been around for a long time, it is far from being solved as it presents several challenges that have not fully been addressed yet, starting from the theoretical difficulty of formally establishing what a ”word sense” is and choos- ing a corresponding sense inventory to the more pragmatic problems of finding large-scale sense-annotated corpora to train supervised systems on. Although WordNet (Fellbaum, 1998) virtually solved the first problem at least for English, a wide range of other issues still remain open. In fact, since supervised WSD systems need to be trained on a word- by-word basis, creating effective datasets requires a huge effort, which is beyond reach even for resource-rich lan- guages like English. Clearly, this issue is even more severe for systems that need both lexicographic and encyclope- dic knowledge (Schubert, 2006) and/or need to work in a multilingual or domain-specific setting. Knowledge-based WSD, on the other hand, exploits the knowledge contained in resources like WordNet to build algorithms (e.g. dens- est subgraph (Moro et al., 2014) or personalized page rank (Agirre and Soroa, 2009)) that can choose the sense of a word in context, thus not requiring training data but usually adopting bag-of-words approaches that neglect the lexical and syntactic context of the word (information that is more easily exploited by supervised systems), which may be es- sential in some scenarios. Furthermore, performances of both types of systems are highly affected by distribution of word senses that are usually different for each domain of application (Pasini and Navigli, 2018). In order to address these issues different solutions have been proposed in the past years, ranging from manually an- notated resources that can be used to train WSD systems to automatic or semi-automatic approaches that aim at ex- ploiting parallel corpora or partially annotated data in or- der to produce training corpora. One of the first attempt to produce a sense annotated corpus is SemCor (Miller et al., 1993), a collection of thousand sentences manually tagged with WordNet senses. While its quality is very high thanks to the effort of specialized …

Content truncated to 500 words