Automatic Treebank-Based Acquisition of Arabic LFG Dependency Structures

Lamia Tounsi, Mohammed Attia, Josef van GenabithSEMITIC@EACL 2009 · OpenReview

Introduction

In the field of computational linguistics, the gap between "shallow" parsing (identifying parts of speech and basic phrase structures) and "deep" syntactic analysis remains a significant hurdle for meaning-sensitive applications. While shallow parsers are robust, they often fail to capture the underlying "who-did-what-to-whom" relationships—the dependencies—that are crucial for tasks like high-quality machine translation or complex question answering.

Lexical Functional Grammar (LFG) provides a framework for these deep structures, known as f-structures. However, hand-coding LFG grammars is a labor-intensive process. This paper, originally presented at the EACL 2009 Workshop on Computational Approaches to Semitic Languages, details a robust method for automatically inducing these structures from the Penn Arabic Treebank (ATB). By adapting methodologies previously successful for English, the authors address the unique linguistic challenges of Arabic to create a high-coverage, high-accuracy resource.

Methodology: The Annotation Pipeline

The core of the research is an eight-stage annotation algorithm designed to transform phrase-structure trees into rich LFG dependency structures. This pipeline is visualized in Figure 1.

The eight-stage annotation pipeline for extracting LFG structures from the Arabic Treebank.

The process begins with Conflation, where the massive set of over 460 morphological tags in the ATB is simplified into generic categories. This reduction is critical for making the subsequent rule-writing manageable; by temporarily simplifying these tags, the researchers were able to reduce nearly 46,000 phrase-structure rules down to approximately 9,700. Following Head Lexicalization, which identifies the semantic core of each phrase, the system employs Lexical Macros to map the original complex morphological data back into specific LFG features (such as gender, number, and case).

The algorithm then leverages existing Functional Tags in the ATB (like -SBJ for subject) and uses Left-Right Context Rules to determine the roles of untagged constituents based on their position relative to the head. Specialized modules handle Coordination and Traces (resolving long-distance dependencies), while a Catch-All & Clean-Up phase ensures consistency and fixes common errors in multi-word expressions.

Arabic presents several hurdles that make simple adaptation of English-centric models difficult. Unlike English, which relies heavily on word order (configuration) to signal grammatical roles, Arabic is morphologically rich and features flexible word order (VSO, SVO, VOS).

Morphology over Configuration

The researchers found that in Arabic, morphological tags and the functional labels already present in the treebank are far more reliable indicators of syntactic roles than constituent position. By prioritizing these features, the system maintains high accuracy despite the language's structural fluidity.

The "Null Be" Phenomenon

A distinctive feature of Arabic is the frequent use of verbless (equational) clauses, such as "The student [is] creative." To maintain a consistent logical structure across different languages and sentence types, the algorithm automatically inserts a null_be predicate into the resulting f-structure.

An example f-structure for an Arabic verbless clause, featuring the automatically inserted 'null_be' predicate.

As shown in Figure 4, this null_be predicate takes the subject and the predicate (XCOMP) as its arguments, ensuring the dependency structure remains semantically complete even when the verb is phonologically absent.

Results and Evaluation

The system was evaluated using the Penn Arabic Treebank (approximately 23,611 sentences) and a manually verified "Gold Standard" of 250 f-structures.

  • Coverage: The pipeline achieved a remarkable 98% coverage, meaning it successfully produced a complete, connected dependency structure for nearly every tree in the ATB. The 2% failure rate was primarily due to feature-value clashes in the underlying logic.
  • Accuracy: When compared against the Gold Standard, the system achieved an overall F-score of 94.96%.
  • Strengths and Weaknesses: The system excelled at identifying oblique arguments (100% precision) and possessives (96% precision). However, it faced more challenges with subjects (89% precision, 68% recall), largely due to "pro-drop" instances where the subject is encoded within the verb rather than appearing as a separate noun phrase.

Conclusion

This research demonstrates that deep syntactic structures can be successfully and automatically acquired for morphologically rich languages like Arabic. By combining morphological analysis with existing treebank annotations, the authors achieved high coverage and accuracy. This automated pipeline provides a scalable way to create the training data necessary for the next generation of wide-coverage Arabic parsers and translation systems. Future work will focus on improving subject-verb agreement handling to further refine the identification of omitted subjects.

Summary updated

BibTeX
@inproceedings{tounsi2009automatic,
  author = {Lamia Tounsi and Mohammed Attia and Josef van Genabith},
  title = {{Automatic Treebank-Based Acquisition of Arabic LFG Dependency Structures}},
  booktitle = {SEMITIC@EACL},
  year = {2009},
  url = {https://openreview.net/forum?id=RC_5YkgU8N},
}
Figure