Using Parallel Corpora for Word Sense Disambiguation

Proceedings of Recent Advances in Natural Language Processing, pages 336–341, Hissar, Bulgaria, 7-13 September 2013. Using Parallel Corpora for Word Sense Disambiguation Ahmad R. Shahid1,2 1Department of Computer Science COMSATS Institute of IT Islamabad, Pakistan ahmadrshahid@comsats.edu.pk Dimitar Kazakov2 2Department of Computer Science University of York York, United Kingdom dimitar.kazakov@york.ac.uk Abstract This paper presents a method of lexical se- mantic disambiguation in multilingual corpora and describes the construction of an artifi- cial word-aligned and lexically disambiguated gold-standard corpus from an existing mul- tilingual resource. The suggested approach uses sets of aligned words and phrases across languages as unique semantic tags similar to WordNet synsets that can be used as a part of unsupervised natural language processing and information retrieval tasks. The approach goes beyond one-to-one word alignment, and uses an algorithm for the aggregation of re- sults of pair-wise word alignment when the corpus contains several languages. When ap- plied to the new corpus, this methodology has proven capable of reducing the ambiguity of a polysemous word by one third on average. 1 Introduction This is a study of the specific potential that paral- lel corpora provide for word and phrase sense dis- ambiguation (WPSD). We do not discuss any of the methods that can be applied to monolingual texts, as these can be considered complementary approaches that are not mutually exclusive, but, rather, can al- ways be combined together. We focus instead on the specific contribution that the availability of multiple translations of the same text can make towards re- jecting some of the alternative senses of the words and phrases in the corpus for any of the individ- ual languages represented in it. We describe an approach in which the N translations in a paral- lel corpus are word-aligned, and the result used to group words and phrases that are translations of each other into N-tuples that can be seen as mul- tilingual synsets akin to the sets of synonyms used in WordNet (Fellbaum, 1998). These synsets can then be used as semantic tags for word and phrase sense disambiguation. The approach was applied to a large, real-world parallel corpus, namely, Eu- roparl (Koehn, 2005). In this setting the full potential of the idea can be obscured by errors introduced by one pre-processing step, such as imperfect word alignment, or the lack of another, e.g., morpholexical analysis. We therefore use an existing multilingual lexical re- source (Lefever, 2009) to develop a large, artifi- cial parallel corpus containing semantically disam- biguated polysemous words, and use it to calculate the maximum contribution that parallel corpora can make towards WPSD under ideal conditions, when all other processing steps are 100% accurate and therefore do not introduce any noise to the process. This result gauges the potential contribution of mul- tiple translations to WPSD, providing its upper limit for the data studied. The multilingual synsets produced in this frame- work represent a potentially valuable resource on its own, which could be used (as is, or after filtering out the errors) as a translation …

Content truncated to 500 words