Context Independent Term Mapper for European Languages

Proceedings of Recent Advances in Natural Language Processing, pages 562–570, Hissar, Bulgaria, 7-13 September 2013. Context Independent Term Mapper for European Languages Mārcis Pinnis University of Latvia 19 Raina Blvd., Riga, Latvia marcis.pinnis@gmail.com Tilde Vienības gatve 75a, Riga, Latvia marcis.pinnis@tilde.lv Abstract In this paper the author presents a new context independent method for bilingual term map- ping using maximised character alignment maps. The method tries to particularly address mapping of multi-word terms and compound terms that are extracted from comparable cor- pora. The method allows integrating linguistic resources (e.g., probabilistic dictionaries and character based transliteration systems) that significantly increase the mapping recall while maintaining a stable precision. The term map- ping method has been automatically evaluated using the EuroVoc thesaurus with varying availability of linguistic resources and on terms extracted from Latvian-English medical domain comparable corpus collected from the Web. The paper shows that the results signifi- cantly outperform previously reported results on the same evaluation corpus. 1 Introduction Multi-lingual terminology is a valuable re- source not only in human and machine transla- tion (MT), but also in many other application domains, for instance, information retrieval, se- mantic analysis, question answering and others. Multi-lingual term glossaries can be automatical- ly acquired from existing resources (monolingual lists of terms, parallel or comparable corpora, etc.) with the help of term mapping. Term map- ping methods according to previous research in the field can be divided in two categories – con- text dependent methods and context independent methods. The context dependent methods are applicable in situations when there is enough context from which to draw statistics. The necessary amount of context can differ depending on the methods. For instance, for term mapping in parallel data it can be enough to simply have one parallel docu- ment pair or a sentence-aligned parallel corpus (Federmann et al., 2012; Wolf et al., 2011; Lefe- ver et al., 2009; Gaussier et al., 2000). For under-resourced languages and numerous domains, however, parallel resources are scarce and not always available. Therefore, a more promising resource is comparable corpora, which has recently received much attention in the scien- tific community for its applicability in MT (Skadiņa et al., 2012). Most of the context- dependent methods designed for term mapping in comparable corpora, however, require relatively large corpora (e.g., hundreds or even thousands of documents) in order to calculate reliable cross-lingual association measures (Fung and Yee, 1998; Rapp, 1999; Shao and Ng, 2004; Morin and Daille, 2010). The proposed methods have also been focussed on language pairs with relatively simple morphology (e.g., German- English, French-English), but have not been thoroughly investigated for more complex lan- guages (e.g., Finnish, Latvian, etc.). A recent study in the European Commission financed pro- ject TTC (2013) revealed that while the context- dependent methods by Morin et al. (2010) per- form well for English-French, their applicability for English-Latvian is questionable because of a term mapping precision of below 5%. Laroche and Langlais (2010) also reported a relatively low precision (far below 50%) using context- dependent methods. …

Content truncated to 500 words