Tiny Corpus Applications with Transformation-Based Error-Driven Learning : Evaluations of Automatic Grammar Induction and Partial Parsing of SaiSiyat

Tiny Corpus Applications with Transformation-Based Error-Driven Learning: Evaluations of Automatic Grammar Induction and Partial Parsing of SaiSiyat Zhemin LIN Graduate Institute of Linguistics, National Taiwan University 1, Roosevelt road sec. 4 Taipei, R.O.C. lin.zhemin@gmail.com Li-May SUNG Graduate Institute of Linguistics, National Taiwan University 1, Roosevelt road sec. 4 Taipei, R.O.C. limay@ntu.edu.tw Abstract This paper reports a preliminary result on automatic grammar induction based on the frame- work of Brill and Markus (1992) and binary-branching syntactic parsing of Esperanto and SaiSiyat (a Formosan language). Automatic grammar induction requires large corpus and is found implausible to process endangered minor languages. Syntactic parsing, on the contrary, needs merely tiny corpus and works along with corpora segmented by intonation-unit which results in high accuracy. 1 Introduction SaiSiyat is a Formosan Austronesian language with less than 4,677 speakers (1995 census data). It is an SOV language with four verbal voices, six case markers, but without declensions (Yeh (2000)). As other Austronesian languages in Taiwan, SaiSiyat writing system is just officially standardised. 1 Few written materials are published in this language and the main source of its corpora is linguistic fieldwork in form of transcription of oral narration and conversation. The tiny scale of corpora makes it hard to do probabilistic natural language processing. Other affordable methods to build a syntactically tagged treebank are thus subjects to our work. SaiSiyat parallels to ancient Egyptian in terms of the description of Rosmorduc (published on Internet). Part of its grammar is still unsure. Grammatical errors are found in texts. The absence of punctuation makes the corpus impossible to be proceeded at sentence level. In order to partially parse this language, the applications of Kullback-Leibler divergence and transformation-based error-driven learning (TBL) are evaluated in the paper. NTU SaiSiyat corpus (?))contains 27 texts, 3702 intonation units (IUs), 12065 words. Its notation follows the convention of Du Bois (1993). Sixteen narrations are composed in the corpus, including 4 Pear Stories (a colour mute film), 8 Frog Stories (a sketchbook by Mayer (1980)) and 4 indigenous legends. The corpus is tagged with a TBL tagger in reduced Penn Treebank Tagset. The overall accuracy is 88.11%. (Lin (2004)) Additional collected texts are added in our experiment to enlarge the corpus. An example of original and tagged data segment follows: 1. kor-koring min-a’rem korkoring/NN mina’rem/VB Red-discipline MIN-rest "A child was asleep." Esperanto is planned as an international help language in 1887 by L. Zamenhof. 2 Large corpora of authentic journals, translated works and archives of Yahoo!Groups are available online for free. Its declension and conjugation are regular, permitting us to tag the texts easily, quickly and correctly. We 1A standardised spelling system of Formosan Austronesian languages is published by the Council of Indigenous Peoples, Executive Yuan. Diversity still exists among their users. 2Cf. http://eo.wikipedia.org/wiki/Esperanto PACLIC 18, December 8th-10th, 2004, Waseda University, Tokyo - 197 - choose “Monato” archive, a periodic written in the language, as a sentence-based contrast. See table 1 for corpora statistics. Table 1: Corpora data Language Size Words Vocab. Tags Sentence …

Content truncated to 500 words