Lexicon Acquisition for Dialectal Arabic Using Transductive Learning

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 399–407, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Lexicon Acquisition for Dialectal Arabic Using Transductive Learning Kevin Duh Dept. of Electrical Engineering University of Washington Seattle, WA, USA duh@ee.washington.edu Katrin Kirchhoff Dept. of Electrical Engineering University of Washington Seattle, WA, USA katrin@ee.washington.edu Abstract We investigate the problem of learn- ing a part-of-speech (POS) lexicon for a resource-poor language, dialectal Arabic. Developing a high-quality lexicon is often the first step towards building a POS tag- ger, which is in turn the front-end to many NLP systems. We frame the lexicon ac- quisition problem as a transductive learn- ing problem, and perform comparisons on three transductive algorithms: Trans- ductive SVMs, Spectral Graph Transduc- ers, and a novel Transductive Clustering method. We demonstrate that lexicon learning is an important task in resource- poor domains and leads to significant im- provements in tagging accuracy for dialec- tal Arabic. 1 Introduction Due to the rising importance of globalization and multilingualism, there is a need to build natu- ral language processing (NLP) systems for an in- creasingly wider range of languages, including those languages that have traditionally not been the focus of NLP research. The development of NLP technologies for a new language is a chal- lenging task since one needs to deal not only with language-specific phenomena but also with a po- tential lack of available resources (e.g. lexicons, text, annotations). In this study we investigate the problem of learning a part-of-speech (POS) lexi- con for a resource-poor language, dialectal Arabic. Developing a high-quality POS lexicon is the first step towards training a POS tagger, which in turn is typically the front end for other NLP appli- cations such as parsing and language modeling. In the case of resource-poor languages (and dialec- tal Arabic in particular), this step is much more critical than is typically assumed: a lexicon with too few constraints on the possible POS tags for a given word can have disastrous effects on tag- ging accuracy. Whereas such constraints can be obtained from large hand-labeled corpora or high- quality annotation tools in the case of resource- rich languages, no such resources are available for dialectal Arabic. Instead, constraints on possible POS tags must be inferred from a small amount of tagged words, or imperfect analysis tools. This can be seen as the problem of learning complex, structured outputs (multi-class labels, with a dif- ferent number of classes for different words and dependencies among the individual labels) from partially labeled data. Our focus is on investigating several machine learning techniques for this problem. In partic- ular, we argue that lexicon learning in resource- poor languages can be best viewed as transduc- tive learning. The main contribution of this work are: (1) a comprehensive evaluation of three trans- ductive algorithms (Transductive SVM, Spectral Graph Transducer, and a new technique called Transductive Clustering) as well as an inductive SVM on this task; and (2) a demonstration that lexicon learning …

Content truncated to 500 words