High Performance Word Sense Alignment by Joint Modeling of Sense Distance and Gloss Similarity

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 245–256, Dublin, Ireland, August 23-29 2014. High Performance Word Sense Alignment by Joint Modeling of Sense Distance and Gloss Similarity Michael Matuschek ‡ and Iryna Gurevych †‡ † Ubiquitous Knowledge Processing Lab (UKP-DIPF), German Institute for Educational Research and Educational Information Schloßstr. 29, 60486 Frankfurt, Germany ‡Ubiquitous Knowledge Processing Lab (UKP-TUDA), Department of Computer Science, Technische Universit¨at Darmstadt Hochschulstr. 10, 64289 Darmstadt, Germany http://www.ukp.tu-darmstadt.de Abstract In this paper, we present a machine learning approach for word sense alignment (WSA) which combines distances between senses in the graph representations of lexical-semantic resources with gloss similarities. In this way, we significantly outperform the state of the art on each of the four datasets we consider. Moreover, we present two novel datasets for WSA between Wiktionary and Wikipedia in English and German. The latter dataset in not only of unprecedented size, but also created by the large community of Wiktionary editors instead of expert annotators, making it an interesting subject of study in its own right as the first crowdsourced WSA dataset. We will make both datasets freely available along with our computed alignments. 1 Introduction Lexical-semantic resources (LSRs) are an important foundation for numerous natural language process- ing (NLP) tasks such as word sense disambiguation (WSD) or information extraction (IE). However, large-scale LSRs are only available for a few languages. The Princeton WordNet (Fellbaum, 1998) is commonly used for English, but for most languages such resources are small or missing altogether. Another problem is that, even for English, there is no single LSR which is suitable for all different application scenarios, because the resources contain different words, senses or even information types. Recently, it has been argued that collaboratively constructed resources (e.g. Wiktionary (Meyer and Gurevych, 2012))) are a viable alternative, especially for smaller languages (Matuschek et al., 2013), but there are still considerable drawbacks in coverage which make their usage challenging. These observations have led to the insight that word sense alignment (WSA), i.e. linking at the level of word senses, is key for the efficient exploitation of LSRs, and it was shown that the usage of linked resources can indeed yield performance improvements. Examples include WSD using aligned Word- Net and Wikipedia (Navigli and Ponzetto, 2012a), semantic role labeling using PropBank, VerbNet and FrameNet (Palmer, 2009), the construction of a semantic parser using FrameNet, WordNet, and VerbNet (Shi and Mihalcea, 2005) and IE using WordNet and Wikipedia (Moro et al., 2013). Cholakov et al. (2014) address the special task of verb sense disambiguation. They use the large-scale resource UBY (Gurevych et al., 2012) which contains nine resources in two languages, mapped to a uniform represen- tation using the LMF standard for interoperability (Eckle-Kohler et al., 2012), and also (among others) sense alignments between WordNet, FrameNet, VerbNet and Wiktionary which are exploited in their approach. However, WSA is challenging because of word ambiguities, different sense granularities and informa- tion types (Navigli, 2006), so that past efforts mostly …

Content truncated to 500 words