Morphological Features for Parsing Morphologically-rich Languages: A Case of Arabic

Jon Dehdari, Lamia Tounsi, Josef van GenabithSPMRL@IWPT 2011 · OpenReview

Abstract

Parsing morphologically-rich languages (MRLs) like Arabic presents a significant challenge due to the high dimensionality of morphological information and resulting data sparsity. This paper investigates the optimal selection of morphological features for syntactic parsing using the Berkeley latent-variable PCFG parser (PCFG-LA). By employing heuristic search strategies—including greedy-iterative, backtracking, and simulated annealing—the researchers identify a "Goldilocks" zone of features that maximize parsing accuracy. Results demonstrate state-of-the-art improvements of up to 2.86% in F-score, particularly through the use of product models and optimized feature sets that better represent noun phrases.

1. Introduction

In morphologically-rich languages, a single word-form often encodes multiple grammatical properties that are typically expressed as separate words in languages like English. For instance, the Penn Arabic Treebank (ATB) contains 498 fine-grained Part-of-Speech (POS) tags. While this detail is linguistically accurate, it leads to severe data sparsity for statistical parsers. Conversely, reducing the tagset to a coarse mapping (e.g., the 37-tag Bikel-Bies mapping) discards vital morphological clues. This research addresses the challenge of identifying the specific subset of morphological information that maximizes parsing performance without introducing excessive noise.

2. Methodology

The study utilizes a re-implementation of the Berkeley latent-variable PCFG parser (PCFG-LA), which is an unlexicalized parser that automatically refines grammar categories. The core challenge is navigating the search space of $2^{15}$ possible combinations of 15 identified morphological features, such as Person, Number, Voice, and Definiteness.

To find the optimal feature set, several heuristic search strategies were employed:

  • Non-iterative best-first: Features are ranked individually and added sequentially.
  • Greedy-iterative: An iterative approach that retains a feature only if it yields a performance gain.
  • Backtracking: A search that can revert previous selections to escape local optima.
  • Simulated Annealing: A probabilistic technique for global optimization.
Iterative feature selection process showing F-score improvements as specific morphological features are added.

Additionally, the system incorporates a heuristic-based morphological clue extractor to handle Out-of-Vocabulary (OOV) words by mapping them to 26 pre-defined word classes.

3. Results and Analysis

The experiments were conducted on the Penn Arabic Treebank (ATB Part 3 v3.2) for both vocalized and unvocalized text. The optimized feature selection yielded significant gains over the baseline:

  • Vocalized Text: Achieved an F-score of 85.03% (a 2.17% improvement).
  • Unvocalized Text: Achieved an F-score of 83.34% (a 1.88% improvement).

3.1 The Importance of Noun Phrases

A key insight was the "NP Factor." The most effective feature sets heavily favored genitive clitics, proper nouns, and determiners—features central to Noun Phrases (NPs). Given that NPs constitute 58% of non-terminals in the treebank, improving their identification has a disproportionately large impact on overall parsing accuracy. Conversely, features like "Case" were found to hurt performance in certain contexts due to surface-level ambiguity in Arabic spelling.

3.2 Product Models and Randomness

The researchers observed that the random seed used for the parser's EM algorithm initialization could cause F-score fluctuations of over 1%. To mitigate this, "product models" were used to aggregate votes from multiple grammars trained with different seeds, leading to more stable and higher-performing results.

Comparison of F-scores across different seed counts for product models, highlighting the performance gap between vocalized and unvocalized text.

4. Conclusion

This work demonstrates that "smarter" data selection is more effective than simply increasing the volume of morphological information. By identifying optimal feature combinations, the study establishes a new state-of-the-art for Arabic parsing. While the search methods are language-independent, the results suggest that the optimal feature set is likely task-specific. Future research will explore applying these heuristics to dependency parsing and even more granular feature sets.

Summary updated

BibTeX
@inproceedings{dehdari2011morphological,
  author = {Jon Dehdari and Lamia Tounsi and Josef van Genabith},
  title = {{Morphological Features for Parsing Morphologically-rich Languages: A Case of Arabic}},
  booktitle = {SPMRL@IWPT},
  year = {2011},
  url = {https://openreview.net/forum?id=vDvfS-B-Fvc},
}
Figure