Enhancing Modern Supervised Word Sense Disambiguation Models by Semantic Lexical Resources

Enhancing Modern Supervised Word Sense Disambiguation Models by Semantic Lexical Resources Stefano Melacci1,2, Achille Globo2, Leonardo Rigutini2 1DIISM - University of Siena, Via Roma 56 – 53100 Siena (Italy) 2QuestIT s.r.l., Via Leonida Cialfi 23 – 53100 Siena (Italy) mela@diism.unisi.it, {globo, rigutini}@quest-it.com Abstract Supervised models for Word Sense Disambiguation (WSD) currently yield to state-of-the-art results in the most popular benchmarks. Despite the recent introduction of Word Embeddings and Recurrent Neural Networks to design powerful context-related features, the interest in improving WSD models using Semantic Lexical Resources (SLRs) is mostly restricted to knowledge-based approaches. In this paper, we enhance “modern” supervised WSD models exploiting two popular SLRs: WordNet and WordNet Domains. We propose an effective way to introduce semantic features into the classifiers, and we consider using the SLR structure to augment the training data. We study the effect of different types of semantic features, investigating their interaction with local contexts encoded by means of mixtures of Word Embeddings or Recurrent Neural Networks, and we extend the proposed model into a novel multi-layer architecture for WSD. A detailed experimental comparison in the recent Unified Evaluation Framework (Raganato et al., 2017) shows that the proposed approach leads to supervised models that compare favourably with the state-of-the art. Keywords: Word Sense Disambiguation, Supervised Models, Lexical Resources 1. Introduction Determining the correct meaning of a target word in a given context is a problem that is commonly referred to as Word Sense Disambiguation (WSD). WSD has a long tradition in the Natural Language Processing (NLP) community (Nav- igli, 2009), and it is still a very challenging task, subject to recent studies (Camacho-Collados et al., 2016; Raganato et al., 2017). The universe of WSD approaches is usually divided into the two main categories of “supervised” and “knowledge-based” methods (Raganato et al., 2017). The first category includes those algorithms that exploit manu- ally annotated corpora (Zhong and Ng, 2010), and, recently, we also observe the growth of techniques that benefit from additional (semi) automatically annotated data (Taghipour and Ng, 2015a; Bovi et al., 2017) or that cast the learning problem into the semi-supervised setting (Taghipour and Ng, 2015b; Yuan et al., 2016). The second category con- siders those approaches that disambiguate words only using structured sources of lexical knowledge (Lesk, 1986; Moro et al., 2014; Agirre et al., 2014; Weissenborn et al., 2015). On one hand, knowledge-based systems are easier to setup, not requiring annotated corpora; on the other hand, super- vised models currently yield to state-of-the-art results in the most popular benchmarks (Raganato et al., 2017). Recent supervised WSD models are frequently based on the “It Makes Sense” (IMS) framework (Zhong and Ng, 2010), whose original implementation is also studied when aug- mented with Word Embeddings (WEs) to generate a dis- tributed representations of the local context around the tar- get word (Taghipour and Ng, 2015c; Rothe and Sch¨utze, 2015). Other WSD approaches are based on Recurrent Neural Networks (RNNs) (Yuan et al., 2016; Melamud et al., 2016), such as the Context2Vec (C2V) model (Mela- …

Content truncated to 500 words