A Neural Network Based Model for Loanword Identification in Uyghur

A Neural Network Based Model for Loanword Identification in Uyghur Chenggang Mi†‡, Yating Yang†‡, Lei Wang†‡, Xi Zhou†‡, Tonghai Jiang†‡ †The Xinjiang Technical Institute of Physics & Chemistry of Chinese Academy of Sciences, Urumqi, China ‡Key laboratory of speech language information processing of Xinjiang, Urumqi, China {micg,yangyt,wanglei,zhouxi,jth}@ms.xjb.ac.cn Abstract Lexical borrowing happens in almost all languages. To obtain more bilingual knowledge from monolingual corpora, we propose a neural network based loanword identification model for Uyghur. We build our model on a bidirectional LSTM - CNN framework, which can capture past and future information effectively and learn both word level and character level features from training data automatically. To overcome data sparsity that exists in model training, we also suggest three additional features , such as hybrid language model feature, pronunciation similarity feature and part-of-speech tagging feature to further improve the performance of our proposed approach. We conduct experiments on Chinese, Arabic and Russian loanword detection in Uyghur. Experimental results show that our proposed method outperforms several baseline models. Keywords: Loanword Identification, Uyghur Language, BiLSTM-CNN, Language Model, Pronunciation Similarity 1. Introduction Lexical borrowing is very common between languages (Taylor and Grant, 2015). It is a phenomenon of cross- linguistic influence. If loanwords in resource-poor lan- guages (e.g. Uyghur) can be identified effectively, we can use the donor-receipt word pairs to extend bilingual dictio- nary. And the bilingual dictionary plays a very important role in many cross-lingual areas in natural language pro- cessing (NLP), such as machine translation (Tsvetkov and Dyer, 2015). In this paper, we describe a novel method to identify loan- words in Uyghur texts to alleviate the data sparsity that ex- ists in Uyghur related NLP tasks. Our loanword identifica- tion model is based on a bidirectional long-short term mem- ory (BiLSTM) - convolutional neural networks (CNNs) framework (Chiu and Nichols, 2016). The BiLSTM have achieved state-of-the-art performance in varies sequence- to-sequence learning tasks, a very important reason is that it can capture past (from previous tagged words) and fu- ture (from next untagged words) information effectively. We use it to model word level features. CNNs have been used in several character level natural language process- ing (NLP) tasks; we use it to model character level fea- tures. Therefore, our model can learn both word level and character level feature from training data, automatically. Additionally, we also propose three features (hybrid lan- guage model, pronunciation similarity and part-of-speech tagging) to argument the BiLSTM-CNN model by exploit knowledge learned from monolingual corpus. We conduct experiments on Chinese, Arabic and Russian loanwords de- tection in Uyghur, respectively. Experimental results show that our model outperforms several baseline models. 2. Loanwords in Uyghur Language 2.1. An Introduction of Uyghur Language Uyghur is an official language of the Xinjiang Uyghur Au- tonomous Region in China, and is widely used in both so- cial and official spheres, as well as in print, radio and tele- vision, and is mostly used as a lingua franca by other eth- nic minorities in Xinjiang. Uyghur belongs to the Turkic …

Content truncated to 500 words