Automatic Morphological Enrichment of a Morphologically Underspecified Treebank

Proceedings of NAACL-HLT 2013, pages 460–470, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Automatic Morphological Enrichment of a Morphologically Underspecified Treebank Sarah Alkuhlani, Nizar Habash and Ryan Roth Center for Computational Learning Systems Columbia University {salkuhlani,habash,ryanr}@ccls.columbia.edu Abstract In this paper, we study the problem of auto- matic enrichment of a morphologically under- specified treebank for Arabic, a morpholog- ically rich language. We show that we can map from a tagset of size six to one with 485 tags at an accuracy rate of 94%-95%. We can also identify the unspecified lemmas in the treebank with an accuracy over 97%. Fur- thermore, we demonstrate that using our au- tomatic annotations improves the performance of a state-of-the-art Arabic morphological tag- ger. Our approach combines a variety of tech- niques from corpus-based statistical models to linguistic rules that target specific phenomena. These results suggest that the cost of treebank- ing can be reduced by designing underspec- ified treebanks that can be subsequently en- riched automatically. 1 Introduction Collections of manually-annotated morphological and syntactic analyses of sentences, or treebanks, are an important resource for building statistical parsing models or for syntax-aware approaches to applications such as machine translation. Rich tree- bank annotations have also been used for a variety of natural language processing (NLP) applications such as tokenization, diacritization, part-of-speech (POS) tagging, morphological disambiguation, base phrase chunking, and semantic role labeling. The development of a treebank with rich annota- tions is demanding in time and money, especially for morphologically complex languages. Consequently, the richer the annotation, the slower the annotation process and the smaller the size of the treebank. As such, a tradeoff is usually made between the size of the treebank and the richness of its annotations. In this paper, we investigate the possibility of automatically enriching the morphologically un- derspecified Columbia Arabic Treebank (CATiB) (Habash and Roth, 2009; Habash et al., 2009) with the more complex POS tags and lemmas used in the Penn Arabic Treebank (PATB) (Maamouri et al., 2004). We employ a variety of techniques that range from corpus-based statistical models to handwrit- ten rules based on linguistic observations. Our best method reaches accuracy rates of 94%-95% on full POS tag identification. We can also identify the un- specified lemmas in CATiB with an accuracy over 97%. 37% of our POS tag errors are due to gold tree or gold POS errors. A learning curve experiment to evaluate the dependence of our method on annotated data shows that while the quality of some compo- nents may reduce sharply with less data (12% abso- lute reduction in accuracy when using 1 32 of the data or some 10K annotated words), the overall effect is a lot smaller (2% absolute drop). These results sug- gest that the cost of treebanking can be reduced by designing underspecified treebanks that can be sub- sequently enriched automatically. The rest of this paper is structured as follows: Section 2 presents related work; Section 3 details various language background facts about Arabic and its …

Content truncated to 500 words