Error Analysis of Uyghur Name Tagging: Language-specific Techniques and Remaining Challenges

Error Analysis of Uyghur Name Tagging: Language-specific Techniques and Remaining Challenges Abudukelimu Halidanmu1, Abulizi Abudoukelimu1, Boliang Zhang2, Xiaoman Pan2, Di Lu2, Heng Ji2, Yang Liu1 1 Tsinghua University, 2 Rensselaer Polytechnic Institute 1 {abdklmhldm}@gmail.com, 1 {keram1106}@163.com, 1 {liuyang2011}@tsinghua.edu.cn, 2 {zhangb8,panx2,lud2,jih}@rpi.edu Abstract Regardless of numerous efforts at name tagging for Uyghur, there is limited understanding on the performance ceiling. In this paper, we take a close look at the successful cases and perform careful analysis on the remaining errors of a state-of-the-art Uyghur name tagger, systematically categorize challenges, and propose possible solutions. We conclude that simply adopting a machine learning model which is proven successful for high-resource languages along with language-independent superficial features is unlikely to be effective for Uyghur, or low-resource languages in general. Further advancement requires exploiting rich language-specific knowledge and non-traditional linguistic resources, and novel methods to encode them into machine learning frameworks. Keywords: Low-resource Languages, Name Tagging, Error Analysis 1. Introduction Uyghur is a language spoken by 8.2 million people, pri- marily by the Uyghur people in the Xinjiang Uyghur Au- tonomous Region of Western China. In terms of the num- ber of native speakers, it’s ranked at the 94th among all the languages in the world 1, but it has extremely low linguis- tic resources. There are very few Natural Language Pro- cessing (NLP) tools, standard annotated corpora, or lan- guage universal resources (e.g., World Atlas of Linguis- tic Structure (WALS) database (Haspelmath et al., 2005; Dryer and Haspelmath, 2013)) available. Even for natu- rally existing noisy annotations such as Wikipedia markups, Uyghur is ranked very low (the 195th 2). There are only 2,566 Uyghur pages in Wikipedia, much fewer than its re- lated languages such as Turkish (277,547 pages) and Uzbek (128,664 pages). Most Uyghur Wikipedia pages contain much less content than their counterparts in Turkish, Uzbek and English. The cross-lingual links are not carefully vali- dated and thus contain many errors. It’s certainly important to develop automatic NLP tools for Uyghur, so as to distill information from textual documents written in Uyghur, as well as preserve their unique culture, music, art and the long history of which the Uyghurs are deeply prideful of. Unfortunately, the striking fact is that very little Uyghur NLP work has been published to catch the attention of the wider international NLP research com- munity. Using Uyghur name tagging as a case study, some previous studies (Li et al., 2011; Arkin et al., 2013b; Rozi et al., 2013; Turhun et al., 2012; Li, 2014; Arkin et al., 2013c; Arkin et al., 2013a; Maihefureti et al., 2014; Zhang, 2014; Zhang et al., 2015; Yu et al., 2015; Nizamidin et al., 2016) have adopted popular machine learning methods which were ef- fective for other high-resource languages. The features sug- 1https://en.m.wikipedia.org/wiki/List_of_languages_by_num- ber_of_native_speakers 2https://meta.wikimedia.org/wiki/List_of_Wikipedias gested by these previous papers include: numbers, shape, stem, suffix, the number of suffixes, first syllable, last sylla- ble, the number of syllables, Part-of-speech tags, the closest verb, word length, position in the sentence and special rules to identify Chinese …

Content truncated to 500 words