Automatic diacritization of Tunisian dialect text using Recurrent Neural Network

Proceedings of Recent Advances in Natural Language Processing, pages 730–739, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_085 730 Automatic diacritization of Tunisian dialect text using Recurrent Neural Network Abir Masmoudi MIRACL Laboratory University of Sfax masmoudiabir@gmail.com Mariem Ellouze Khmekhem MIRACL Laboratory University of Sfax Mariem.Ellouze@planet.tn Lamia Hadrich Belguith MIRACL Laboratory University of Sfax l.belguith@fsegs.rnu.tn Abstract The absence of diacritical marks in the Arabic texts generally leads to morpholog- ical, syntactic and semantic ambiguities. This can be more blatant when one deals with under-resourced languages, such as the Tunisian dialect, which suffers from unavailability of basic tools and linguis- tic resources, like sufficient amount of cor- pora, multilingual dictionaries, morpho- logical and syntactic analyzers. Thus, this language processing faces greater chal- lenges due to the lack of these resources. The automatic diacritization of MSA text is one of the various complex problems that can be solved by deep neural net- works today. Since the Tunisian dialect is an under-resourced language of MSA and as there are a lot of resemblance between both languages, we suggest to investigate a recurrent neural network (RNN) for this dialect diacritization problem. This model will be compared to our previous models models CRF and SMT (24) based on the same dialect corpus. We can experimen- tally show that our model can achieve bet- ter outcomes (DER of 10.72%), as com- pared to the two models CRF (DER of 20.25%) and SMT (DER of 33.15%). 1 Introduction Modern Standard Arabic (MSA) as well as Arabic dialects are usually written without diacritics(24). It is easy for native readers to infer correct pro- nunciation from undiacritized words not only from the context but also from their grammatical and lexical knowledge. However, this is not the case for children, new learners and non-native speak- ers as they dont have a good mastery of rich lan- guage derivations. Moreover, the absence of dia- critical marks leads to ambiguity that affects the performance of NLP tools and tasks. This may generally bring a considerable awkward ambiguity at the data-processing level for NLP applications. Hence, we can notice that automatic diacritization has been shown to be useful for a variety of NLP applications, such as Automatic Speech Recogni- tion (ASR), Text-to-Speech (TTS), and Statistical Machine Translation (SMT)(24). In this paper, we present our method to auto- matic dialectical Arabic diacritization. In fact, both previous experiences and works have shown that the use of Recurrent Neural Network (RNN) could give better results for such an MSA dia- critization system as compared to the other ap- proaches, like the Lexical Language Model (16), a hybrid approach combining statistical and rule- based techniques (28). For instance, the authors (1) demonstrated in their study that the RNN gave the least DER, compared to the other MSA dia- critization works. Based on the huge similarity between MSA and Tunisian dialect, we decided to benefit from its advantages by testing the RNN performance in the automatic diacritization of Tunisian Arabic di- alects. To the best of our knowledge, it is …

Content truncated to 500 words