Cognate Identification for a French - Romanian Lexical Alignment System: Empirical Study

Cognate Identification for a French - Romanian Lexical Alignment System: Empirical Study Mirabela Navlea Linguistique, Langues, Parole (LiLPa) Université de Strasbourg 22, rue René Descartes BP, 80010, 67084 Strasbourg, cedex navlea@unistra.fr Amalia Todira��� Linguistique, Langues, Parole (LiLPa) Université de Strasbourg 22, rue René Descartes BP, 80010, 67084 Strasbourg, cedex todiras@unistra.fr �Abstract This paper describes a cognate identifica- tion method, used by a lexical alignment system for French and Romanian. We combine statistical techniques and lin- guistic information to extract cognates from lemmatized, tagged and sentence- aligned parallel corpora. We evaluate the cognate identification model and we compare it to other methods using pure statistical techniques. We show that the use of linguistic information in the cog- nate identification system improves sig- nificantly the results. 1 Introduction We present a new cognate identification module required for a French - Romanian lexical align- ment system. This system is used for French - Romanian law corpora. Cognates are translation equivalents presenting orthographic or phonetic similarities (common etymology, borrowings, and calques). They represent very important ele- ments in a lexical alignment system for legal - texts for two reasons: French and Romanian are two close - Romance languages with a rich morphology; Romanian language borrowed and cal- qued legal terminology from French. So, cog- nates are very useful to identify bilingual legal terminology from parallel corpora, while we do not use any external terminological resources for these languages. Cognate identification is one of the main steps applied for lexical alignment for MT systems. If we have several efficient tools for several Euro- © 2011 European Association for Machine Translation. pean languages, few lexically aligned corpora or lexical alignment tools (Tufi�� ��� ����� ������ ���� available for Romanian - English or Romanian - ����������������������������������������������� few linguistic resources and tools for Romanian (dictionaries, parallel corpora, terminological data bases, MT systems) are currently available. Some MT systems use resources for the English - Romanian language pair (Marcu and Munteanu, ������ �������� ������ �������� ������� ������ ��� systems develop resources for German - Roma- ����� ���������� ������ ������� ���� ��������� ������ or for French - Romanian (N��������������������� 2010). Most of the cognate identification mod- ules used by these systems were purely statistic- al. No cognate identification method is available for the studied languages. Cognate identification is a difficult problem, es- pecially to detect false friends. Inkpen et al. (2005) classify bilingual words pairs in several categories such as: - cognates (reconnaissance (FR) - reco- gnition (EN)); - false friends (blesser �������������������- bless (EN)); - partial cognates (facteur (FR) - factor or mailman (EN)); - genetic cognates (chef (FR) - head (EN)); - unrelated pairs of words (glace (FR) - ice (EN) and glace (FR) - chair (EN)). In our method, we rather identify cognates and partial cognates to improve lexical alignment. Thus, we aim to obtain a high precision of our method and to eliminate some false friends using statistical techniques and linguistic information. To identify cognates from parallel corpora, sev- eral approaches exploit the orthographic similari- ty …

Content truncated to 500 words