Incorporating Inner-word and Out-word Features for Mongolian Morphological Segmentation

Proceedings of the 28th International Conference on Computational Linguistics, pages 4638–4648 Barcelona, Spain (Online), December 8-13, 2020 4638 Incorporating Inner-word and Out-word Features for Mongolian Morphological Segmentation Na Liu1,2,3, Xiangdong Su1,2,∗, Haoran Zhang1,2, Guanglai Gao1,2, Feilong Bao1,2 1Inner Mongolia Key Laboratory of Mongolian Information Processing Technology, China 2College of Computer Science Inner Mongolia University, China 3Department of Mathematics and Computer Science Hetao University, China cssxd@imu.edu.cn Abstract Mongolian morphological segmentation is regarded as a crucial preprocessing step in many Mon- golian related NLP applications and has received extensive attention. Recently, end-to-end seg- mentation approaches with long short-term memory networks (LSTM) have achieved excellent results. However, the inner-word features among characters in the word and the out-word features from context are not well utilized in the segmentation process. In this paper, we propose a neural network incorporating inner-word and out-word features for Mongolian morphological segmen- tation. The network consists of two encoders and one decoder. The inner-word encoder uses the self-attention mechanisms to capture the inner-word features of the target word. The out-word encoder employs a two layers BiLSTM network to extract out-word features in the sentence. Then, the decoder adopts a multi-head double attention layer to fuse the inner-word features and out-word features and produces the segmentation result. The evaluation experiment compares the proposed network with the baselines and explores the effectiveness of the sub-modules. 1 Introduction Mongolian is a morphologically rich language and its words are formed by attaching suffixes to roots (Kullmann and Tserenpil, 2008). Each word has a root and zero or more suffixes, which are called Mongolian morphemes. The morphemes in a word indicate the basic word features and provide gram- matical and semantic relations among words in the sentence. Mongolian morphological segmentation aims to split Mongolian words into their morphemes, which facilitates the Mongolian NLP tasks, such as name entity recognition (Wang et al., 2016; Wang et al., 2019), information retrieval (Liu et al., 2012), machine translation (Fan et al., 2017; Yang et al., 2016), and speech synthesis (Liu et al., 2017). There are about 60 thousand of morphemes in Mongolian, and the number of their formed words is more than 7 million. It becomes a tendency to process Mongolian text on morphemes rather than on words to make full use of the morpheme information in Mongolian NLP tasks. Besides, Segmenting Mongolian words into morpheme can alleviate the data-sparse problem and out-of-vocabulary (OOV) problem. Therefore, Mongolian morphological segmentation is an essential preprocessing step and effects the downstream Mongolian NLP tasks. Mongolian morphological segmentation is closely related to the words themselves and their context. Table 1 shows several morphological segmentation examples. In usual, a Mongolian word corresponds to only one segmentation, such as the target words “ (bariba)” and “ (barigulba)” in sentences I and II. But in some cases, parts of Mongolian words correspond to different segmentation results according to the context where they appear. For example, the target word “ (higed)” in the sentences V and VI is segmented into different morphemes due to its …

Content truncated to 500 words