Automatic Lexicon Extraction on RandomIndexing Word Spaces using Small Seed Lexica

Abstract: Automatic bilingual lexicon extraction has many applications in Natural Language Processing, but often times requires highly structured,parallel, data or extensive bilingual seed lexicas to get reasonably good performance. Random Indexing models with a small bilingual seed lexicon could be used to perform (semi-)automatic lexicon extraction using only separate monolingual data. This thesis explores, explains and evaluatessuch a method of (semi-)automatic lexicon extraction on Random Indexing models using a small lexicon. The main idea is to construct alinear transformation that aligns the vector representation of the words in the lexicon. Necessitated by the kind of transformation used, a slight modification of the cosine similarity measure is presented. By evaluating the method against a bilingual sentiment lexicon it was found that while the method worked well in a same language setting between comparable corpora, performance was greatly reduced in an interlanguage setting.In conclusion the method proposed is altogether inadequate as a method of (semi-)automatic lexicon extraction, but might be improved upon by further dimension reduction techniques or larger seed lexica.