Parsing Arabic Using Treebank-based Lfg Resources

In this paper we present initial results on parsing Arabic using treebank-based parsers and automatic LFG f-structure annotation methodologies. The Arabic Annotation Algorithm (A) (Tounsi et al., 2009) exploits the rich functional annotations in the Penn Arabic Treebank (ATB) (Bies and Maamouri, 2003; Maamouri and Bies, 2004) to assign LFG f-structure equations to trees. For parsing, we modify Bikel’s (2004) parser to learn ATB functional tags and merge phrasal categories with functional tags in the training data. Functional tags in parser output trees are then “unmasked” and available to A to assign f-structure equations. We evaluate the resulting f-structures against the DCU250 Arabic gold standard dependency bank (Al-Raheb et al., 2006). Currently we achieve a dependency f-score of 77%. 1 Related Work Arabic parsing systems have been reported in (Ditters, 2001; Zabokrtsky and Smrz, 2003; Othman et al., 2003; Ramsay et Mansour, 2007). (Attia, 2008) gives an overview of an LFG rule-based analysis of Arabic using XLE (Xerox Linguistics Environment). He concentrated on short sentences and used robustness techniques to increase the coverage. All of these use hand-crafted grammars, which are time-consuming to produce and difficult to scale to unrestricted data. More recently, the Penn Arabic Treebank (ATB) has been employed to acquire wide-coverage parsing resources. The best-known Arabic statistical parser was developed by Bikel (Bikel, 2004). Bikel reports parse quality “far below” English and Chinese (Kulick et al., 2006). The main reasons cited were a significant number of POS-tag inconsistencies (in the version of the ATB available at the time) and the considerable differences between Arabic and English sentence structure. (Dieb et al., 2004) and (Habash and Rambow, 2005) present knowledgeand machine-learning-based methods for tokenisation, basic POS tagging with a reduced tagset and base phrase chunking. Bikel’s parser produces phrase-structure trees (c-structures). The main objective of our research is to automatically enrich the output of Bikel’s parser with more abstract and “deep” dependency information (in the form of LFG f-structures), using the Arabic A annotation algorithm (Tounsi et al., 2009), extending the approach of (Cahill et al., 2004), originally developed for English. 2 The Penn Arabic Treebank (ATB) Arabic is a subject pro-drop language. It has relatively free word order: mainly S(ubject) V(erb) and O(bject), with VSO and VOS also possible. Arabic is a highly inflectional and cliticizing language. The ATB consists of 23,611 parse-annotated sentences (Bies and Maamouri, 2003; Maamouri and Bies, 2004) from Arabic newswire text in Modern Standard Arabic (MSA). The ATB annotation scheme involves 497 different POS-tags with morphological information (reduced to 24 basic POS-tags by Bikel e.g. NN, NNS, JJ), 22 phrasal tags e.g. NP, VP, PP and 20 functional tags e.g. SBJ, OBJ, TPC (52 combined functional tags, as functional tags can stack). 3 The Arabic Annotation Algorithm (A) The Arabic Annotation Algorithm (Tounsi et al., 2009) is constructed adapting and revising the methodology of (Cahill et al., 2004) for English as follows: 1. Automatic extraction of the most frequent rule types from the treebank. 2. Head lexicalisation of ATB trees to identify local heads. 3. Default f-structure equations are assigned to ATB functional tags. In addition, lexical macros exploits the rich morphological information provided by the ATB. With 85% token coverage.