Machine Translation System on the Pair of Arabic / English

Transcription

1 Machine Translation System on the Pair of Arabic / English Khaireddine Bacha and Mounir Zrigui Laboratoire UTIC, Équipe de Monastir, Faculté des sciences de Monastir, Monastir, Tunisia Khairi.bacha@gmail.com, Mounir.Zrigui@fsm.rnu.tn Keywords: Abstract: Learning, Arabic-English Machine Translation, Statistical Models. Our work fits into the project entitled "TELA": an environment for learning the Arabic language computerassisted, which covers many issues related to the use of words in Arabic. This environment contains several sub-systems whose purpose is to provide an important educational function by allowing the learner to discover information beyond the scope of the phrase of the year. In these subsystems there are semantic analyzers which have several features and multifunctions (Arabic-English machine translation, Arabic- English machine translation, derivation, and conjugation, etc.). Therefore, in this article we focused upon the design of machine translation systems on the pair of Arabic / English based on statistical models. 1 INTRODUCTION In the case of Arabic writing, the search began around 1970, even before problems editing Arabic texts are completely controlled. Early work involved including lexicons. For ten years, the internationalization of the Web and the proliferation of media in Arabic, revealed a large number of Arabic NLP applications. Researchs have begun to address more diverse issues such as syntax, machine translation, automatic indexing of documents, research information, etc... The field of natural language processing has been a major revolution in recent years in machine translation and the other hand needs reliable automatic translators which are constantly increasing. Therefore, we focused on this area to develop an automatic translator based on a statistical model. today. One of the first articles about this approach is that of (Brown, 1993). The translation can take place in two modules: An analysis module, which produces a representation of the input text in source language into a language independent of any postulated pivot language, and a generation module, which builds from this even a text representation of output in the target language. The triangle shown in Fig. 1 is assigned to Vauquois. It summarizes an analysis of the translation process yet fully relevant and used today (Vauquois, 1968). To find the best translation of a source sentence (s), the translator Moses seeks the target sentence (t) that maximizes a log-linear combination of characteristic functions. The characteristic functions used in this system are: 2 TRANSLATION OF REFERENCE SYSTEMS The statistical translation comes down to find the target document with the highest probability of being the translation of a source document. The development and maintenance of such a system usually requires a significant human labor by specialists (bilingual). This section summarizes several approaches that have marked research on machine translation, the middle of last century to Figure 1: Vauquois triangle; Representation of different linguistic architecture. The score (m) in the translation table assigns to every pair of phrases (t, s). The score of a language model: our experiments use a trigram model to fold. The score of the distortion model. 347

2 KEOD2012-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment The exponential in the number of target words generated. This "characteristic function", called word penalty, simply allows the counterbalance system's tendency to prefer short phrases. Finally, the expression that must maximize Moses is: Translation system based on statistical segments which are called in English "phrase based machine translation system statistique" for both language pairs, using free software Moses (KOEHN, 2007). Among the measures most commonly used in the community of machine translation to see the quality of the translation of reference: The score BLEU (Bilingual Evaluation Understudy) is proposed by (Papineni, 2002). The main idea is to compare the output with a translator / translation reference... The score NIST (National Institute of Standards and Technology) that was proposed in International assessments of machine translation systems. The conditions and detailed results are available on the NIST website. In 2006 and 2008, the BLEU scores of these reference systems are given in Table 1. Table 1: BLEU score of the reference system Arabic / English (Schwenk, 2010). Bitextes News + ISI bitextes + données ONU Taille des bitextes 56M 189M Dev Nist06 42,69 43,51 Test Nist08 42,06 42,19 3 OVERVIEW OF THE WORK OF TRANSLATION SYSTEMS STATISTICS FOR COUPLE OF ARABIC-ENGLISH The translation from Arabic into English, two languages considered remote structure, requires more effort than the translation between two languages of similar structure (such as French and English). The need of Arabization was mainly due to the need to simplify the use of Arabic in computer technology today. This implies the establishment up of engineering of the Arabic language in the context of multilingualism required the opening of the Arab world to other cultures that carry the current scientific and technological knowledge. The Arabic language is characterized by a rather complex morphology whose complexity presents challenges for machine translation (Diab, 2007). Our investigations are based on the main existing work in this area (Lee, 2004), using a segmentation approach in roots and affixes, then (Habash, 2006), with a linguistically motivated approach to tokenization, showed that morphological preprocessing can be useful for statistical machine translation. On the other hand, first Diab and colleagues (Diab, 2004), developed in Perl by the team of Mona Diab 1 at Leland Stanford Junior University grammatical labeling is used for Arabic ASVM 2. This is an adaptation to Arabic English Yamcha 3 system based on support vector machines margin (SVM). The system is trained on an annotated corpus named Arabic Treebank. Second, researchers, (Diab, 2004) showed that the use of voyellation only leads to no improvement (partial voyellation), or even worse results (voyellation complete). Besacier, is an improvement from the first system in the enrichment of single words with syntactic information, as labels (part of speech tags).this system has been subject to IWSLT08 (Besacier, 2004). And lately, Wigdan, provides an assessment of the morphological analyzer as a tool for preprocessing for Arabic- English translation (Wigdan, 2011). This work is done according to the recommendations of the first comprehensive system of automatic translation is shown under the "Georgetown Experiment" in January 7, The system, which has a vocabulary of 250 words, manages to translate some sixty sentences carefully chosen from Russian to English. The heart of this system is based on six rules of grammar and especially a bilingual dictionary. 4 SYSTEM ARCHITECTURE FORMACHINE TRANSLATION 4.1 Location of Module Automatic Translation Arabic / English in the TELA Project Under the TELA project (Towards Environmental Learning Arabic) (Khaireddine, 2011), whose goal is the design and implementation of an environment of

3 MachineTranslationSystemonthePairofArabic/English computer-assisted learning, covering many issues related to the use of words Arabic. A platform that allows teachers to create language learning activities based on technologies from the TAL, which notably includes automatic evaluations. We have designed and built a prototype dictionary (Khaireddine, 2012). The latter two roles, firstly it is the source of data on which are based exercises. that maximizes p (s / c), the probability that a sentence is the translation of c s (we always translate Arabic into English s in this c following): Source Text S Decoder Argmax p(e)*p(f/e) P(S/c) P(c) Models Translation model Language model Phase of training Arabic Corpus Anglish Corpus c Target Text S : Arabic sentence Figure 3: Machine statistical machine translation Arabic / English. Figure 2: General architecture of semantic analyzer multifunction. This return is also generated automatically by synthesizing the information in the dictionary. It has an important didactic function by allowing the learner to discover information beyond the scope of the phrase of the year. Second, it is an electronic dictionary trilingual (Arabic, French and English) one-way. Its macrostructure consists on a single volume. It can be seen as a structure composed of linguistic objects. Among these objects, we can find: the headword, pronunciation, grammatical,اسم,فعل) categories that can have this headword examples, etc..), Definitions, translations, صفة,مصدر collocations, an etymology, meanings, glosses, plans lexical, lexical functions, etc.. We first started completed the module Arabic- English machine translation. Construct a statistical translation system for a language of rich morphology such as Arabic requires several pretreatments. In this section, we will implement and test a machine translation system that manipulates data enriched with morphosyntactic information for the Arabic / English and propose some solutions. 4.2 Methdoologies: Approach used to Build Our Own Arabic Tagger Recall that the approach to statistical machine translation is as followed. By giving an Arabic sentence s, we may seek the English translation of c 3 shows the main components of probabilistic machine translation system. The decoder takes as input the source text, the translation model and the language model for outputting the translated text. Note that the language into which we want to translate is called "target language" 4.3 The Implementation Stages The initial input of our system, as described in Figure 3, has two bodies, one body and another Arabic English. These two bodies are structured so that each row i in the Arabic corpus is aligned with the line i in the English corpus, so that row i in the corpus is the English translation of the line i in the Arabic corpus. To obtain more flexibility in our translation table, we removed the short vowels in Arabic corpus. We find the same word being vowelized differently in our training corpus, that is to say, sometimes it is not vowelized and sometimes the first or last letter is vowelized. The effect of enriching words by morphosyntactic categories appears to improve translation quality of these examples (remove false alignments that existed in the classical model, filter model translation and generates more opportunities correct that no longer existed in the classical model). 4.4 Results and Evaluation of the Analyzer in Machine Translation Games development and test each consist of

4 KEOD2012-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment pairs of sentences. Measures used for evaluation are BLUE and NIST. The results which are respectively given in Table 2. Table 2: Blue scores and results obtained NILT system. ASVM Dev Test 06 Test08 Score BLEU Score NIST We can notice that the optimization is very important for the system, since it has increased significantly Bleu and NIST scores after adjustment system Arabic / English with texts in Arabic. The adapted system clearly produced good translations for these examples. There is of course a few mistakes in these sentences, but the quality of translations can largely understand the meaning of sentences. To be able to test and validate the overall approach, a first version of statistical machine translation system was performed. It allows the creation of virtually any type of words and phrases, without collection and use of the trace of their implementation by the learners. The screen shot below shows the current state of the "visible" system developed so far. The ergonomics and design interfaces are provisional and should evolve in the future... We want to have a table near the window. نريد مائدة بجانب النافذة. Figure 4: Translation Arabic / English. 5 CONCLUSIONS AND OUTLOOK The statistical approach to machine translation is now used to quickly build translation systems for many language pairs. This system has been integrated into a semantic analyzer multifunctional; its realization is part of a learning environment of the Arabic language computer-assisted "TELA". As perspective, in addition to the experimental validation on a larger scale of our results, it would be interesting to determine and integrate our machine translation system on complete pair of Arabic / English in the semantic analyzer and add a feature that calculates the similarity between the source word and its translation to avoid losing the meaning of the text, because even for commercial systems such as SYSTRAN translation, there are translations that have nothing to do with the source. REFERENCES Besacier, L., Benyoussef Atef and Blanchon, (October 2008). H. The LIG Arabic / English Speech translation System at IWSLT08, pp , Hawaii Brown, P., Pietra, S. D., Pietra, V. D. and Mercer, R. (1993), The mathematics of machine translation: Parameter estimation. in Computational Linguistics, pp Diab M., Ghoneim M. and Habash N. (2007). Arabic diacritization in the context of statistical machine translation. In Proceedings of MT-Summit. Habash, Nizar, Bonnie Dorr and Christof Monz. Challenges in Building an Arabic Generation-heavy Machine Translation System and Extending it with Statistical Components (2006). In Proceedings of the Association for Machine Translation in the Americas (AMTA-2006), Boston Diab, M., Hacioglu, K. and Jurafsky, D. (2004), Automatic tagging of arabic text: From raw text to base phrase chunks, in 5th Meeting of the North American Chapter of the Association for Computational Linguistics/Human Language Technologies Conference (HLT-NAACL04), Boston, MA. USA Diab M., Ghoneim M. and Habash N. (2007). Arabic diacritization in the context of statistical machine translation. In Proceedings of MT-Summit Lee Y. (2004). Morphological analysis for statistical machine translation. In Proceedings of HLT-NAACL 2004, p : Association for Computational Linguistics. Khaireddine. B., Mounir. Z., Mohamed Amine. N., Anis. Z., 2011, TELA: Towards Environmental Learning Arabic, The International Conference on Artificial Intelligence (ICAI'11), 2011, WORLDCOMP'11, Las Vegas, Nevada, USA, 6 pages. download/worldcomp%2711/2011%20cd%20pa pers/eee4685.pdf Khaireddine. B and Mounir. Z 2012, Design of a Synthesizer and a Semantic Analyzer's Multi Arabic, for use in Computer Assisted Teaching, International 350

5 MachineTranslationSystemonthePairofArabic/English Journal of Information Sciences and Application (IJISA) pp , /volume/ijisav4n1.htm Koehn P. and Schroeder J. (2007). Experiments in domain adaptation for statistical machine translation. In Second Workshop on SMT, p Schwenk H. (2010) Adaptation d un Système de Traduction Automatique Statistique avec des Ressources monolingues- Montréal. Papineni, K., Roukos, S., Ward, T., & Zhu, W. (2002). BLEU: a Method for Automatic Evaluation of Machine Translation. Proceedings of COLING-ACL 02, Philadelphia, USA, Vauquois, B. (1968). A Survey of Formal Grammars and Algorithms for Reconition and Translation. FIP Congress-68, Edinburg, Wigdan. M, J. Gosme, F. Debili, Y. Lepage, N. Lucas (2011) Évaluation de G-LexAr pour la traduction automatique statistique - Montpellier 351