ParsPer: A Dependency Parser for Persian

Mojgan Seraji, Bernd Bohnet, Joakim NivreDepLing 2015 · OpenReview

Abstract

Mojgan Seraji, Bernd Bohnet, Joakim Nivre. Proceedings of the Third International Conference on Dependency Linguistics (Depling 2015).

2015.

Introduction

Dependency parsing—the process of mapping grammatical relationships between words in a sentence—is a foundational task in Natural Language Processing (NLP). However, the Persian language presents a unique "orthographic minefield" for standard parsers. Persian features a high degree of free word order and complex morphology, but its primary challenge lies in inconsistent writing standards. Clitics and "light verb constructions" (LVCs) are sometimes written as single words, sometimes separated by spaces, and sometimes joined by a "zero-width non-joiner" (ZWNJ). ParsPer is the first high-performance, open-source dependency parser specifically optimized to navigate these complexities.

Methods and Datasets

The researchers developed ParsPer using a two-stage approach:

  1. Feature Optimization: They first used MaltParser (a transition-based parser) and MaltOptimizer to find the best data representation for Persian. This involved experiments with "coarsening" data, such as reducing 31 fine-grained part-of-speech (POS) tags down to 15 broader categories to reduce data sparsity.
  2. Model Selection: After optimizing the data representation, they evaluated several parsing algorithms, including MSTParser, TurboParser, and the Mate Tools suite. The Mate graph-based parser emerged as the winner, proving most resilient to Persian’s flexible word order.

The models were trained and validated using the Uppsala Persian Dependency Treebank (UPDT), which consists of 6,000 annotated sentences. They also utilized the Uppsala Persian Corpus (UPC) for initial tagging and the Hamshahri Corpus for out-of-domain evaluation.

Key Insights: Optimization and Trade-offs

A central finding of the research was that simplifying certain linguistic features actually improved parsing performance. By merging sub-categories into coarse-grained POS tags (CPOS), the parser could focus more effectively on the structural backbone of the sentence. Similarly, the authors found that merging various complex LVC labels into a single lvc tag improved overall accuracy, even if it sacrificed some granular detail about the specific type of verb construction.

Performance improvements observed during the optimization phase by applying coarse POS tags (CPOS) and merging LVC labels (1 LVC).

The study also highlighted a critical "tokenization bottleneck." While POS tagging errors only degraded parsing results by about 1%, tokenization errors (incorrectly splitting or joining words) caused a massive 5% drop in accuracy. This indicates that for Persian, the pre-processing stage is as critical as the parsing algorithm itself.

Key Results

ParsPer achieved impressive benchmarks on the UPDT dataset, reaching a Labeled Attachment Score (LAS) of 82.58% and an Unlabeled Attachment Score (UAS) of 86.69%. When tested on out-of-domain text from the Hamshahri corpus, the LAS remained robust at approximately 78% (using manual segmentation).

Detailed analysis of dependency relations shows a hierarchy of difficulty. While relations like pobj (prepositional object) are identified with very high precision and recall, others like dobj (direct object) remain more challenging for the system.

Precision (P) and recall (R) metrics for various dependency relations in the optimized model, including the merged 'lvc' category.

Limitations and Future Work

The primary limitation remains the orthographic inconsistency of Persian. Because the parser relies on pre-tokenized text, errors made by the initial segmenter create "broken" inputs that the parser cannot easily correct. The authors suggest that the next frontier for Persian NLP is joint segmentation and tagging. By handling these tasks simultaneously, a system could use syntactic clues to resolve ambiguous word boundaries, potentially pushing accuracy even higher.

Summary updated

BibTeX
@inproceedings{seraji2015parsper,
  author = {Mojgan Seraji and Bernd Bohnet and Joakim Nivre},
  title = {{ParsPer: A Dependency Parser for Persian}},
  booktitle = {DepLing},
  year = {2015},
  url = {https://openreview.net/forum?id=L9upxoJGNK7},
}
Figure