Clausal parsing helps data-driven dependency parsing: Experiments with Hindi

Samar Husain, Phani Gadde, Joakim Nivre, Rajeev SangalIJCNLP 2011 · OpenReview

Abstract: Samar Husain, Phani Gadde, Joakim Nivre, Rajeev Sangal. Proceedings of 5th International Joint Conference on Natural Language Processing.

2011.

Introduction

Dependency parsing—the task of mapping out the grammatical relationships between words in a sentence—is notoriously difficult for languages like Hindi. Unlike English, where word order is relatively fixed, Hindi is morphologically rich and allows for flexible word positioning.

Traditional "single-stage" data-driven parsers often struggle with these languages due to error propagation: a mistake made early in the parsing process ripples outward, leading to incorrect global structures. Furthermore, syntactic cues in Hindi are often distributed across complex case markers that can be highly ambiguous. The authors set out to see if the linguistic concept of a "clause" could serve as a natural boundary to simplify the parsing search space and reduce these errors.

Methodology

The authors propose a two-stage parsing framework based on the observation that different types of grammatical relations occur at different "depths" and "arc lengths."

  1. Stage 1 (Intra-clausal): The parser identifies dependencies within individual clauses (e.g., a subject and its verb).
  2. Stage 2 (Inter-clausal): The parser identifies dependencies between clauses (e.g., a main clause and a subordinate clause).
Figure 1: Distribution of dependency labels across intra-clausal and inter-clausal relations.

They tested two specific implementation strategies using MaltParser:

  • 2-Hard (Hard Constraints): The output of the first stage is fixed. The second-stage parser is initialized with these partial trees and is forbidden from changing them; it can only add new relations between the existing sub-trees.
  • 2-Soft (Soft Constraints/Parser Stacking): The output of the first stage is used as a set of "guide features" for the second stage. The second-stage parser performs a complete parse but "listens" to the suggestions from the first stage, allowing it the flexibility to override initial mistakes.

Datasets

The study utilized the Hyderabad Dependency Treebank (HDT) for Hindi, specifically the version released for the ICON10 tools contest.

  • Training set: 2,972 sentences.
  • Development/Test sets: 543 and 321 sentences, respectively.
  • The experiments were validated using 5-fold cross-validation to ensure the robustness of the results.

Key Insights

The authors proved that intra-clausal relations (like 'agent' or 'genitive') are overwhelmingly short-distance, while inter-clausal relations (like 'conjunction' or 'relative clause modification') are long-distance. Treating the clause as a unit naturally aligns the parser with the sentence's actual linguistic structure.

Figure 2: Relation between arc length and the type of dependency (intra vs inter-clausal).

By focusing on the clause first, the model could better identify the specific verbal arguments of ambiguous postpositions (like 'se', which can mean 'by', 'from', or 'with'). Furthermore, by identifying sub-trees in Stage 1, the Stage 2 parser effectively works with "shallower" trees. This significantly reduces the chance of error propagation, which is a known weakness of transition-based parsers like MaltParser.

Key Results

The 2-Hard approach was the clear winner, outperforming the single-stage baseline across all major metrics:

  • Overall Accuracy: 2-Hard achieved a Labeled Attachment Score (LAS) of 75.65% and an Unlabeled Attachment Score (UAS) of 89.1%, both statistically significant improvements over the baseline (75.02% LAS / 88.82% UAS).
  • Complex Sentences: On sentences with more than one finite clause, the improvement was even more pronounced, with LAS jumping from 74.87% to 75.83%.
  • Inter-clausal Relations: The 2-Hard model was particularly dominant here, showing a massive improvement in LAS for relations between clauses (87.71% vs. 85.43% for the baseline).
  • Arc Length: As the distance between words (arc length) increased, the advantage of the 2-Hard approach became more significant, proving its efficacy for long-distance dependencies.

Conclusion

While the clausal approach did not significantly help with non-projective relations, it demonstrated that breaking down dependency parsing into a two-stage process significantly improves accuracy for morphologically rich, free-word-order languages. Future work may explore separating the tasks of identifying attachments and labels into different feature models to further refine accuracy.

Summary updated

BibTeX
@inproceedings{husain2011clausal,
  author = {Samar Husain and Phani Gadde and Joakim Nivre and Rajeev Sangal},
  title = {{Clausal parsing helps data-driven dependency parsing: Experiments with Hindi}},
  booktitle = {IJCNLP},
  year = {2011},
  url = {https://openreview.net/forum?id=xIVLEolWVx5},
}
Figure