Automatically Inducing a Part-of-Speech Tagger by Projecting from Multiple Source Languages Across Aligned Corpora

Automatically Inducing a Part-of-Speech Tagger by Projecting from Multiple Source Languages Across Aligned Corpora Victoria Fossum1 and Steven Abney2 1 Dept. of EECS, University of Michigan, Ann Arbor MI 48105 vfossum@umich.edu 2 Dept. of Linguistics, University of Michigan, Ann Arbor MI 48105 abney@umich.edu Abstract. We implement a variant of the algorithm described by Yarowsky and Ngai in [21] to induce an HMM POS tagger for an ar- bitrary target language using only an existing POS tagger for a source language and an unannotated parallel corpus between the source and tar- get languages. We extend this work by projecting from multiple source languages onto a single target language. We hypothesize that systematic transfer errors from differing source languages will cancel out, improving the quality of bootstrapped resources in the target language. Our exper- iments confirm the hypothesis. Each experiment compares three cases: (a) source data comes from a single language A, (b) source data comes from a single language B, and (c) source data comes from both A and B, but half as much from each. Apart from the source language, other condi- tions are held constant in all three cases – including the total amount of source data used. The null hypothesis is that performance in the mixed case would be an average of performance in the single-language cases, but in fact, mixed-case performance always exceeds the maximum of the single-language cases. We observed this effect in all six experiments we ran, involving three different source-language pairs and two different target languages. 1 Introduction 1.1 Background Statistical NLP techniques typically require large amounts of annotated data. Labelling data by hand is time-consuming; a natural goal is therefore to generate text analysis tools automatically, using minimal resources. Yarowsky et al. [22] present methods for automatically inducing various monolingual text analysis tools for an arbitrary target language, using only the corresponding text analysis tool for a source language and a parallel corpus between the source and target languages. Hwa et al. [15] induce a parser for Chinese text via projection from English using a similar method to that of [22]. Cucerzan and Yarowsky [8] present a method for bootstrapping a POS tagger for an arbitrary target language using R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 862–873, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Automatically Inducing a Part-of-Speech Tagger 863 only a bilingual dictionary between the source and target languages, a “basic library reference grammar” for the target language, and an existing corpus in the target language. While automatically induced text analysis tools use fewer resources, their accuracy lags behind that of more resource-intensive tools. One solution to the problem of error reduction on NLP tasks is to train multiple classifiers, then compute a consensus classifier. Combining multiple classifiers is an effective way to reduce error if the errors made by each classifier are independently distributed. Such approaches have been successfully applied to a range of NLP tasks. Brill and Wu [4], van Halteren et al. [19], and Zavrel and …

Content truncated to 500 words