Supervised Word Sense Disambiguation on Taiwan Hakka Polysemy with Neural Network Models: A Case Study of BUN, TUNG and LAU

1 Supervised Word Sense Disambiguation on Taiwan Hakka Polysemy with Neural Network Models: A Case Study of BUN, TUNG and LAU Huei-Ling Lai Hsiao-Ling Hsu Jyi-Shane Liu National Chengchi University National Chengchi University National Chengchi University hllai@nccu.edu.tw heidimavishsu@gmail.com jsliu@cs.nccu.edu.tw Chia-Hung Lin Yanhong Chen National Chengchi University National Chengchi University. linch0520@gmail.com andy83918@gmail.com Abstract This research aims to explore an optimal model for automatic word sense disambiguation for highly polysemous markers BUN, TUNG and LAU in Taiwan Hakka, a low-resource language. The performance of word sense disambiguation tasks is carried out by examining DNN, BiLSTM and CNN models under different window spans. The results show that the CNN model can achieve the best performance with a multiple sliding window of L2R2+ L3R3 and L5R5. Keywords polysemy, word sense disambiguation, Taiwan Hakka, neural network models, CNN 1 Introduction Polysemous phenomenon leading to ambiguity is one of the crucial problems that need to be resolved for natural language processing. Extensive studies on word sense disambiguation (WSD) that engage in solving polysemous problems have provided valuable findings (Iacobacci et al., 2016; Kågebäck & Salomonsson, 2016; Raganato et al., 2017; Liu & Nguyen, 2018; Li et al., 2019). While most put emphasis on a few dominant languages like English and Chinese, low-resource languages like Taiwan Hakka still gain relatively scanty attention because of the unavailability of data, neither raw nor labeled. Lai et al. (2020), employing the DNN and the BiLSTM models, is an endeavor that investigates what information is needed to achieve the best performance of automatic polysemous word sense disambiguation in Taiwan Hakka. As a follow-up study, this research, incorporating the DNN, the BiLSTM, and the CNN in the experiment, aims to further explore what model can achieve the best performance for three semantically and syntactically intertwining polysemous markers. Since Taiwan Hakka is a low-resource language, the characteristics of raw data employed for training and testing these models is one of the core challenges. The quantity and the quality of the raw data play a central role in the performance of the experiments featured on deep learning. In this research, a workable coding framework is schematized encompassing the following procedure: integrating and modifying the findings of previous studies on Taiwan Hakka polysemous phenomena, manually annotating the data to ensure the reliability of the labeled data, and then applying the three models to the 2 massive corpus data. 2 Related Work 2.1 Polysemous BUN 分, TUNG 同 and LAU 摎 The focal point of this study is to differentiate the polysemous BUN, TUNG and LAU from different syntactic structures they occur and the various syntactic elements they are surrounded by. Drawing from findings of extant literature (Lai, 2001, 2003a, 2003b, 2004, 2015; Chiang, 2006; Huang, 2012, 2014, 2015), the coding schemes of BUN, TUNG, and LAU for human annotators are illustrated in Table 1, Table 2 and Table 3, respectively. The human annotators are well trained in linguistics and all the annotations are double checked to reach final agreement. Four labels are applied to the …

Content truncated to 500 words