Bilingual Word Embeddings for Bilingual Terminology Extraction from Specialized Comparable Corpora

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 685–693, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Bilingual Word Embeddings for Bilingual Terminology Extraction from Specialized Comparable Corpora Amir Hazem1 Emmanuel Morin1 1 LS2N - UMR CNRS 6004, Universit´e de Nantes, France {amir.hazem,emmanuel.morin}@univ-nantes.fr Abstract Bilingual lexicon extraction from compa- rable corpora is constrained by the small amount of available data when dealing with specialized domains. This aspect pe- nalizes the performance of distributional- based approaches, which is closely re- lated to the reliability of word’s cooccur- rence counts extracted from comparable corpora. A solution to avoid this limitation is to associate external resources with the comparable corpus. Since bilingual word embeddings have recently shown efficient models for learning bilingual distributed representation of words, we explore dif- ferent word embedding models and show how a general-domain comparable corpus can enrich a specialized comparable cor- pus via neural networks. 1 Introduction Bilingual lexicon extraction from comparable cor- pora has shown substantial growth since the sem- inal work of Fung (1995) and Rapp (1995). Com- parable corpora, which are comprised of texts sharing common features such as domain, genre, sampling period, etc. and without having a source text/target text relationship (McEnery and Xiao, 2007), are more abundant and reliable resources than parallel corpora. On the one hand, parallel corpora are difficult to obtain for language pairs not involving English. On the other hand, as par- allel corpora are comprised of a pair of trans- lated texts, the vocabulary appearing in the trans- lated texts is highly influenced by the source texts. These problems are aggravated in specialized and technical domains. Although it is easier to build large general- domain comparable corpora, specialized compa- rable corpora are often of modest size (around 1 million words) due to the difficulty to obtain many specialized documents in a language other than English. Consequently, word co-occurrence counts of the historical context-based projection approach, known as the standard approach (Fung, 1995; Rapp, 1995), dedicated to bilingual lexi- con extraction from comparable corpora are un- reliable in specialized domain. This problem per- sists with other paradigms such as Canonical Cor- relation Analysis (CCA) (Gaussier et al., 2004), Independent Component Analysis (ICA) (Hazem and Morin, 2012) and Bilingual Latent Dirichlet Allocation (BiLDA) (Vuli´c et al., 2011). A solution to avoid this limitation and to in- crease the representativity of distributional repre- sentations is to associate external resources with the specialized comparable corpus. These re- sources can be lexical databases such as Word- Net which allows the disambiguation of trans- lations of polysemous words (Bouamor et al., 2013) or general-domain data to improve word co- occurrence counts of specialized comparable cor- pora (Hazem and Morin, 2016). Our work is in this line and attempts to find out how a general-domain data can enrich a special- ized comparable corpora to improve bilingual ter- minology extraction from specialized comparable corpora. Since bilingual word embeddings have recently provided efficient models for learning bilingual distributed representation of words …

Content truncated to 500 words