Graph Based Semi-Supervised Learning Approach for Tamil POS tagging

Graph Based Semi-Supervised Learning for Tamil POS Tagging Mokanarangan Thayaparan, Surangika Ranathunga, Uthayasanker Thayasivam Dept. of Computer Science and Engineering University of Moratuwa, Katubedda 10400, Sri Lanka {mokanarangan.11, surangika, rtuthaya }@cse.mrt.ac.lk Abstract Parts of Speech (POS) tagging is an important pre-requisite for various Natural Language Processing tasks. POS tagging is rather challenging for morphologically rich languages such as Tamil. Being low-resourced, Tamil does not have a large POS annotated corpus to build good quality POS taggers using supervised machine learning techniques. In order to gain the maximum out of the existing Tamil POS tagged corpora, we have developed a graph-based semi-supervised learning approach to classify unlabelled data by exploiting a small sized POS labelled data set. In this approach, both labelled and unlabelled data are converted to vectors using word embeddings and a weighted graph is constructed using Mahalanobis distance. Then semi-supervised learning (SSL) algorithms are used to classify the unlabelled data. We were able to gain an accuracy of 0.8743 over an accuracy of 0.7333 produced by a CRF tagger for the same limited size corpus. Keywords: Semi-Supervised Learning, Low-resourced languages, Graph-based SSL, Word Embedding, POS tagging 1. Introduction In the recent past, supervised learning methods have pro- duced high accuracies for Parts-of-Speech (POS) tag- ging (Gimenez and Marquez, 2004). In particular, sequence models such as hidden Markov models (HMM) and condi- tional random fields (CRF) have given good results (Huang et al., 2015). However, these techniques rely on the avail- ability of relatively large amounts of annotated data. Hence, building an accurate domain insensitive POS tagger is chal- lenging for low resourced languages. Tamil is one such low resourced language, which is widely used in South India and Sri Lanka. There have been sev- eral POS taggers developed for Tamil language using super- vised learning techniques (Dhanalakshmi et al., 2009)(Pan- dian and Geetha, 2009). Since the annotated corpora used in this research have been of small size and from a single domain, these supervised techniques greatly suffer from ac- curacy and domain adaptability (Rani et al., 2016). For ex- ample, FIRE corpus (Forum for Information Retrieval Eval- uation, 2014), a widely used freely available Tamil POS annotated corpus contains only 80k words. In contrast, the Wall Street corpus, which is an English POS-annotated cor- pus has a word count of 1,173K words (Gimenez and Mar- quez, 2004), meaning that the size of the FIRE corpus is approximately 15 times smaller than the Wall Street cor- pus. Thus, when using a small corpus such as FIRE, we cannot expect similar accuracy to that of English when su- pervised techniques are used. Moreover, these approaches depend on language dependent features such as morpho- logical tags (Dhanalakshmi et al., 2009) thus limiting the scalability for adapting to other low resourced languages. In contrast to supervised approaches, semi-supervised ap- proaches such as graph based semi-supervised learning and manifold regularization (Niyogi, 2013) use both labeled and unlabelled data for their classification, and have proven to work with a small data sets (Zhu et …

Content truncated to 500 words