Exploring self training for Hindi dependency parsing

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1452–1456, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Exploring self training for Hindi dependency parsing Rahul Goutam Bharat Ram Ambati Language Technologies Research Centre Language Technologies Research Centre IIIT-Hyderabad, India IIIT-Hyderabad, India rahul.goutam@research.iiit.ac.in ambati@research.iiit.ac.in Abstract In this paper we explore the effect of self- training on Hindi dependency parsing. We consider a state-of-the-art Hindi dependency parser and apply self-training by using a large raw corpus. We consider two types of raw corpus, one from same domain as of training and testing data and the other from different domain. We also do an experiment, where we add small gold-standard data to the training set. Comparing these experiments, we show the impact of adding small, but gold-standard data to training data versus large, but automat- ically parsed data on Hindi parser. 1 Introduction Parsing morphologically rich free-word-order languages like Czech, Hindi, Turkish, etc., is a challenging task. Unlike English, most of the parsers for such languages have adopted the de- pendency grammatical framework. It is well known that for these languages, dependency framework is better suited (Shieber, 1985; Mel’čuk, 1988, Bharati et al., 1995). Due to the availability of annotated corpora in recent years, data driven dependency parsing has achieved considerable success. In spite of availability of annotated treebanks, state-of-the-art parsers for these languages have not reached the perfor- mance obtained for English (Nivre et al., 2007a). Frequently stated reasons for low performance are small treebank size, complex linguistic phe- nomenon, long distance dependencies, and non- projective structures (Nivre et al., 2007a; Nivre et al., 2007b; Bharati et al., 2008). In this paper, we try to address the problem of small treebank size. We have lots of un- annotated data. One way to increase treebanks’ size is to manually annotate this data. But it is very time consuming task. Other way is to auto- matically parse this data and consider highly re- liable parses. But, what criteria should be used for extracting reliable parses is a really challeng- ing task. In this paper, we explore a bootstrap- ping technique called self training and see its impact on dependency parsing accuracy. We consider a state-of-the-art Hindi dependency parser and analyze its performance using self- training. We consider two types of raw corpus, one from the same domain as of training and testing data and the other from a different do- main. We also show the impact of adding small, but gold-standard data to training data versus large, but automatically parsed data on Hindi dependency parsing. The paper is arranged as follows. In section 2, we describe the related work in this field. In sec- tion 3, we present the current state-of-the-art of Hindi dependency parser. Section 3, talks about different experiments conducted and presents the results. We conclude with possible future work in section 4. 2 Related Work In this section, we briefly describe major works on bootstrapping using statistical parsers. Steedman et al. (2003) did experiments to …

Content truncated to 500 words