Arabic-English Text Translation Leveraging Hybrid NER

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 124–131 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Emna Hkiri, Souheyl Mallat and Mounir Zrigui Arabic-English Text Translation Leveraging Hybrid NER Emna Hkiri Souheyl Mallat Mounir Zrigui Latice Laboratory Latice Laboratory Latice Laboratory Faculty of sciences Monastir ISIM of Monastir Faculty of sciences Monastir Emna.hkiri@gmail.com Souheyl.mallat@gmail.com Mounir.zrigui@fsm.rnu.tn Abstract Named Entities (NEs) are a very important part of a sentence and their treatment is a potentially useful preprocessing step for Statistical Machine Translation (SMT). Improper translation of NE lapse the quality of the SMT output and it can hurt sentence’s human readability considerably. Dropping NE often causes translation failures beyond the context, affecting both the morphosyntactic formedness of sentences and the word sense disambiguation in the source text. Due to peculiarities of the written Arabic language, the translation task is however rather challenging. In this work, we address the challenging issues of NEs treatment in the context of SMT of Arabic into English texts. We have experimented on three types of named entities which are: Proper names, Organization names and Location names. In this paper, we present integration between machine learning and rule based techniques to tackle Arabic NER problem in attempt to improve the final quality of the SMT system output. We show empirically that each aspect of our approach is important, and that their combination leads to the best results already after integration of NER into SMT. We show improvements in terms of BLEU scores (+4 points) and reduction of out of vocabulary words over a baseline for the News Commentary corpus. 1 Introduction Named entities recognition is essential for many tasks of natural language processing, whether monolingual or multilingual, as information retrieval or machine translation. In this work, we are interested in the processing of NE in the context of statistical machine translation from Arabic into English, in which the processing of NE poses particular problems. A statistical machine translation (SMT) system learns to translate based on examples of translations already made, extracted from parallel corpus. In so far as these training corpus are relatively small in size, this raises the question of translation of words that are not seen during training(Hkiri et al., 2015a). This is particularly critical in case of Arabic texts. Arabic is indeed a morphologically complex language (Habash, 2011), many possible forms are rarely observed in the corpus (Heintz, 2008). This requires, at least, implementing morphological analysis to define the source inventory units of the translation system. A study carried out by Habash (2008) on unknown words in a journalistic corpus for Arabic-English language reports that about 40% of unknown words correspond to proper names. The SMT systems use a default strategy to treat these unknown words; it copies their forms in the output target language. This strategy is sometimes operative especially for person names when the source and target languages use the same alphabet. Unfortunately, this strategy is unsuitable in the case of Arabic into English translation. To overcome this …

Content truncated to 500 words