Modelling Tibetan Verbal Morphology

Modelling Tibetan Verbal Morphology Qianji Di Ekaterina Vylomova Timothy Baldwin The University of Melbourne, Melbourne, Australia Parkville, Victoria 3010, Australia qdi@student.unimelb.edu.au {ekaterina.vylomova,tbaldwin}@unimelb.edu.au Abstract The Tibetan language, despite being spoken by 8 million people, is a low- resource language in NLP terms, and research to develop NLP tools and re- sources for the language has only just be- gun. In this paper, we focus on Tibetan verbal morphology — which is known to be quite irregular — and introduce a novel dataset for Tibetan verbal paradigms, comprising 1,433 lemmas with corres- ponding inflected forms. This enables the largest-scale NLP investigation to date on Tibetan morphological reinflection, wherein we compare the performance of several state-of-the-art models for mor- phological reinflection, and conduct an extensive error analysis. We show that 84% of errors are due to the irregularity of the Tibetan language. 1 Introduction Tibetan is the language used by Tibetan people, in Tibetan areas in China — Tibet Autonomous Region, Qinghai Province, Sichuan Province, Gansu Province, and Yun- nan Province — as well as parts of Nepal, Bhutan, India, and Pakistan. Among the lan- guages of the Sino-Tibetan language fam- ily and Chinese national writing systems, the Tibetan language has one of the longest histories and most extensive bodies of lit- erature. Although Tibetan is roughly di- vided into three dialects (Weizang, Khampa, and Amdo), the written language is uniform across all regions (Jitaiga, 2018). The earliest work on Tibetan information processing began in the 1980s, focusing on fonts, character encoding support, and text input methods. This early work resulted in the ability to develop and share digital text re- sources in Tibetan, which benefited Tibetan scholars and the Tibetan diaspora (Zhijie, 2009). Tibetan information processing tech- nology is rapidly developing, but some key problems remain, including the analysis of Tibetan verbs. Recent advances in Tibetan NLP have included the ability to automatic- ally identify verbs, analyze the rules govern- ing word form changing processes, and re- veal various linguistic phenomena of Tibetan verbs (Zhijie and Rangzhuoma, 2010; Zhijie, 2005). In this paper, we specifically focus on the Tibetan verbal inflection system, which is known for its irregularity (Suonanjiancuo, 2013). Specifically, we make use of the mor- phological reinflection task recently intro- duced under the umbrella of SIGMORPHON (Cotterell et al., 2018). We train several state-of-the-art machine learning models for reinflection and provide an extensive error analysis. We find that the original experi- mental data contained some errors, and the models do not cater to idiosyncrasies of the Tibetan language. After correcting errors in the data, experimental results improve. We also develop a new dataset for Tibetan verbs comprising 1,433 verbal lemmas with their present, past, future, and imperative forms.1 2 Tibetan Language The Tibetan language belongs to the Tibetan branch of the Tibeto-Burman language group of the Sino-Tibetan language family. The Tibetan script is an abugida or alphasyl- labary, whereby consonant–vowel sequences are written as a single unit. There are two schools of thought regarding the ori- gin of Tibetan literature: …

Content truncated to 500 words