Disambiguating vectors for bilingual lexicon extraction from comparable corpora

This paper presents an approach to enhance the extraction of translation equivalents from comparable corpora by plugging in bilingual lexico-semantic knowledge harvested from a parallel corpus. First, the bilingual lexicon obtained from word-aligning the parallel corpus replaces an external seed dictionary, making the approach knowledge-light and portable. Next, instead of using simple 1:1 mappings between the source and the target language, translation equivalents are clustered into sets of synonyms based on contextual similarities, enabling us to expand the translation of vector features with several translation variants. And last but not least, the vector features are disambiguated and translated only with the translation variants from the most appropriate cluster, thus producing less noisy vectors that allow for a more successful cross-lingual comparison of the vectors compared to simpler methods. Razdvoumljanje vektorjev za izboljšanje luščenja dvojezičnih leksikonov iz primerljivih korpusov V prispevku predstavljamo pristop za izboljšanje luščenja prevodnih ustreznic iz primerljivih korpusov z dodatnim virom leksikosemantičnega znanja, izluščenega iz vzporednega korpusa. Za razliko od večine sorodnih pristopov dvojezični leksikon, potreben za prevajanje kontekstnih vektorjev, izdelamo avtomatsko iz vzporednega korpusa. Tako pristop ni več odvisen od slovarja, potrebnega za prevajanje kontekstnih vektorjev in je tako prenosljiv na številne jezikovne pare in strokovna področja. V naslednjem koraku prevodne ustreznice v dvojezičnem leksikonu razvrstimo v gruče, kar nam omogoča, da lastnosti v kontekstnih vektorjev, izdelanih iz primerljivih korpusov, prevajamo z več kot eno prevodno različico. To nam olajša primerjavo kontekstnih vektorjev v izvornem in ciljnem jeziku. Tretja izboljšava, ki jo v prispevku predstavljamo, pa je razdvoumljanje večpomenski lastnosti kontekstnih vektorjev iz primerljivega korpusa z gručami, generiranimi iz dvojezičnega leksikona, ki omogoča natančnejše prevajanje vektorjev in izboljša njihovo primerjavo z vektorji v ciljnem jeziku. 1. Motivation and related work Due to the lack of general language parallel corpora, finding translations in comparable corpora has become a very active area of research. The main idea behind this approach is the assumption that a source word and its translation appear in similar contexts, so that in order to identify them their contexts are compared via a seed dictionary (Fung, 1998; Rapp, 1999). The biggest advantage of the approach is that it offers low-resourced language pairs and domains a fast and affordable way to construct bilingual lexica. However, it also presupposes the availability of a bilingual dictionary to translate vector features, which is not the case for many language pairs or domains. In addition, the original approach and most of its extensions (Shao and Ng, 2004; Otero, 2007; Yu and Tsujii, 2009; Marsi and Krahmer, 2010) neglect polysemy and consider a translation candidate as correct if it is an appropriate translation for at least one possible sense of the source word, which will often be the most frequent sense of the word due to the way context vectors are built. The goal of this paper is twofold: (1) we eliminate the need for an external knowledge source by automatically extracting a bilingual lexicon from a parallel corpus, and (2) we propose a way of disambiguating polysemous features in the context vectors, as these features may be translated differently according to the sense in which they are used in a given context. The need to bypass pre-existing dictionaries has been addressed by Koehn and Knight (2002) who built the initial seed dictionary automatically, based on identical spelling features. Cognate detection has also been used by Saralegi et al. (2008). Both approaches have been successfully combined by Fišer and Ljubešić (2011) who showed that the results with an automatically created seed lexicon, based on the similarity between the languages, can be as good as with a pre-existing dictionary. But all these approaches cannot be used as successfully for language pairs with little lexical overlap, such as English (EN) and Slovene (SL), which is the case in this experiment. We believe we can produce less noisy vectors and improve their comparison across languages by using contextual information to disambiguate their features. A similar idea has been implemented by Kaji (2003) who clustered synonymous Japanese translations of English words in comparable corpora using pre-defined bilingual dictionaries. In addition, instead of providing one translation for each disambiguated feature, we translate it with all translation equivalents that belong to the assigned cluster similar to Déjean et al. (2005) who use a bilingual thesaurus instead of a lexicon. The contribution of this paper is a language independent and fully automated corpus-based approach to bilingual lexicon extraction from comparable corpora that does not rely on any external knowledge sources to determine word senses or translation equivalents. The rest of the paper is organized as follows: In the next section we present the resources that were used in our experiments. In Section 3, we describe the approach and the experimental setup in detail. Evaluation and discussion of the obtained results are given in Section 4, after which the paper is wrapped up with some concluding remarks and ideas for future work.