On the Role of Morphosyntactic Features in Hindi Dependency Parsing

Bharat Ram Ambati, Samar Husain, Joakim Nivre, Rajeev SangalSPMRL @ NAACL 2010 · OpenReview

Abstract: Bharat Ram Ambati, Samar Husain, Joakim Nivre, Rajeev Sangal. Proceedings of the NAACL HLT 2010 First Workshop on Statistical Parsing of Morphologically-Rich Languages.

2010.

Introduction

Dependency parsing—the task of mapping grammatical relationships between words—is particularly challenging for Morphologically-Rich Languages (MRLs) like Hindi. Unlike English, which relies heavily on word order, Hindi features relatively free word order and a complex system of suffixes and postpositions known as case markers. This paper explores how specific linguistic features, particularly those related to case and verb semantics, can significantly boost parser performance.

By systematically evaluating a pool of linguistic features, the authors demonstrate that combining case markers on nouns with Tense, Aspect, and Modality (TAM) markers on verbs provides the single largest boost to Hindi dependency parsing. Their optimized transition-based model achieved a state-of-the-art labeled attachment score (LAS) of 76.5%, a 2% absolute improvement over previous systems.

Methodology

The researchers employed a transition-based dependency parsing approach using MaltParser. This method treats parsing as a series of actions (e.g., "Shift" or "Reduce") guided by a classifier that predicts the next transition based on the current state.

The core methodology involved:

  1. Feature Pool Construction: Merging feature sets from two previous top-performing Hindi parsers.
  2. Incremental Feature Selection: Grouping features into 10 categories (e.g., POS tags, lemmas, case markers) and adding them sequentially to a baseline model to measure individual impact.
  3. SVM Classification: Using Support Vector Machines with a polynomial kernel to predict transitions.
  4. Pseudo-Projective Parsing: Handling non-projective dependencies (crossing arcs) by "flattening" them during training and reconstructing them during testing.

The study utilized the Hindi Dependency Treebank (ICON09), which uses the karaka framework to focus on syntactico-semantic relations rather than pure syntax.

Key Insights: The Impact of Morphosyntax

The incremental feature selection revealed critical insights into which linguistic cues drive parsing accuracy in Hindi.

  • The Power of CTAM: The most significant finding was the impact of the CTAM feature (Case and TAM markers). As shown in Figure 1, adding this feature (Experiment 9) resulted in a massive 10% jump in LAS. In Hindi, the verb's tense and aspect often dictate the required case markers on subjects and objects; capturing this relationship is vital for correct attachment.
  • Agreement Complexity: Surprisingly, features for gender, number, and person did not improve accuracy. This is likely due to Hindi's complex agreement rules, where verbs may agree with objects rather than subjects depending on the aspect.
  • Structural Context: Utilizing POS tags of nodes already in the partially built dependency tree (Experiment 4) provided a 6% boost, highlighting the importance of local structural context in making future attachment decisions.
Performance improvement (UAS and LAS) across ten incremental feature selection experiments.

Results and Error Analysis

The optimized model achieved a Labeled Attachment Score (LAS) of 76.5% and an Unlabeled Attachment Score (UAS) of 91.1%, both setting new benchmarks for Hindi parsing.

However, granular analysis reveals persistent challenges. The parser frequently confuses k1 (agent/subject) and k2 (theme/object) labels, particularly when explicit case markers are absent. Figure 2 illustrates the precision and recall for these and other dependency labels, showing the relative difficulty in correctly identifying core arguments compared to other relations.

Precision and recall for various dependency labels, highlighting the performance on core arguments like k1 and k2.

Furthermore, error analysis confirmed that performance drops sharply as the distance between a head and its dependent increases, especially in relative clauses and coordinate structures.

Limitations and Future Directions

Despite the gains, several hurdles remain for Hindi dependency parsing:

  • Data Scarcity: The relatively small size of the treebank (~28.5k words) likely limits the parser's ability to learn complex patterns.
  • Non-Projectivity: The parser struggled to identify non-projective arcs in the test set, highlighting a major area for improvement in handling Hindi's flexible word order.
  • Semantic Ambiguity: Incorporating "minimal semantics" (e.g., identifying if a noun is human or animate) could help resolve k1-k2 ambiguities where morphological cues are missing.
  • Clausal Awareness: Future work may involve a two-stage parsing process that first identifies clause boundaries to better handle long-distance dependencies.

Summary updated

BibTeX
@inproceedings{ambati2010role,
  author = {Bharat Ram Ambati and Samar Husain and Joakim Nivre and Rajeev Sangal},
  title = {{On the Role of Morphosyntactic Features in Hindi Dependency Parsing}},
  booktitle = {SPMRL @ NAACL},
  year = {2010},
  url = {https://openreview.net/forum?id=c1NXl9o3tjY},
}
Figure