Using Deep Morphology to Improve Automatic Error Detection in Arabic Handwriting Recognition

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 875–884, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Using Deep Morphology to Improve Automatic Error Detection in Arabic Handwriting Recognition Nizar Habash and Ryan M. Roth Center for Computational Learning Systems Columbia University {habash,ryanr}@ccls.columbia.edu Abstract Arabic handwriting recognition (HR) is a challenging problem due to Arabic’s con- nected letter forms, consonantal diacritics and rich morphology. In this paper we isolate the task of identification of erroneous words in HR from the task of producing corrections for these words. We consider a variety of linguistic (morphological and syntactic) and non-linguistic features to automatically iden- tify these errors. Our best approach achieves a roughly ∼15% absolute increase in F-score over a simple but reasonable baseline. A de- tailed error analysis shows that linguistic fea- tures, such as lemma (i.e., citation form) mod- els, help improve HR-error detection precisely where we expect them to: semantically inco- herent error words. 1 Introduction After years of development, optical character recog- nition (OCR) for Latin-character languages, such as English, has been refined greatly. Arabic, however, possesses a complex orthography and morphology that makes OCR more difficult (Märgner and Abed, 2009; Halima and Alimi, 2009; Magdy and Dar- wish, 2006). Because of this, only a few systems for Arabic OCR of printed text have been devel- oped, and these have not been thoroughly evalu- ated (Märgner and Abed, 2009). OCR of Arabic handwritten text (handwriting recognition, or HR), whether online or offline, is even more challenging compared to printed Arabic OCR, where the unifor- mity of letter shapes and other factors allow for eas- ier recognition (Biadsy et al., 2006; Natarajan et al., 2008; Saleem et al., 2009). OCR and HR systems are often improved by per- forming post-processing; these are attempts to eval- uate whether each word, phrase or sentence in the OCR/HR output is legal and/or probable. When an illegal word or phrase is discovered (error detec- tion), these systems usually attempt to generate a le- gal alternative (error correction). In this paper, we present a HR error detection system that uses deep lexical and morphological feature models to locate possible "problem zones" – words or phrases that are likely incorrect – in Arabic HR output. We use an off-the-shelf HR system (Natarajan et al., 2008; Saleem et al., 2009) to generate an N-best list of hy- potheses for each of several scanned segments of Arabic handwriting. Our problem zone detection (PZD) system then tags the potentially erroneous (problem) words. A subsequent HR post-processing system can then focus its effort on these words when generating additional alternative hypotheses. We only discuss the PZD system and not the task of new hypothesis generation; the evaluation is on er- ror/problem identification. PZD can also be useful in highlighting erroneous text for human post-editors. This paper is structured as follows: Section 2 pro- vides background on the difficulties of the Arabic HR task. Section 3 presents an analysis of HR er- rors …

Content truncated to 500 words