Guessing the Grammatical Function of a Non-Root F-Structure in LFG

Anton Bryl, Josef van Genabith, Yvette GrahamIWPT 2009 · OpenReview

Abstract

This paper investigates the use of a Naive Bayes classifier to predict missing grammatical function labels in Lexical-Functional Grammar (LFG) f-structures. By leveraging internal word attributes such as number, gender, and case, the proposed model achieves high accuracy (82–91%) in restoring functions like Subject or Object. The approach is shown to significantly improve the performance of transfer-based machine translation systems, providing a robust solution for handling fragmented or incomplete linguistic data.

Introduction

In computational linguistics, Lexical-Functional Grammar (LFG) represents sentences through "f-structures"—hierarchical maps detailing the relationships between words (e.g., identifying the subject of a verb). However, these structures often suffer from missing labels, particularly in transfer-based machine translation or during parsing of complex sentences. When a system identifies a dependency but fails to determine the specific grammatical function, the resulting "broken" output makes coherent sentence generation nearly impossible.

An f-structure with a missing grammatical function label (UNKNOWN).

This research addresses the challenge of labeling these unlabeled dependencies. For instance, if "Parliament" is known to depend on "adopted," the system must determine if it serves as the Subject or the Object.

Methods

The authors frame the task as a supervised classification problem using a Naive Bayes classifier. The central hypothesis is that the internal attributes of a dependent f-structure provide sufficient cues to predict its grammatical function within a parent structure.

The study explored three feature engineering strategies:

  1. Attribute Presence: Identifying if specific attributes (e.g., CASE, ATYPE) exist.
  2. Atomic Values: Using specific values (e.g., NUM=sg).
  3. Combined Features: Utilizing both attribute names and their values.

To handle linguistic nuances, the authors modified the classifier to account for "set-valued" functions like ADJUNCT, which can occur multiple times. Furthermore, they implemented a joint probability approach to resolve multiple missing labels within a single sentence, ensuring that guesses remain linguistically plausible (e.g., avoiding the assignment of two Subjects to one verb).

Experimental Setup

The experiments utilized the German-English portion of the Europarl corpus. Sentences between 5 and 15 words were parsed into f-structures using the XLE parser with established LFG grammars. For each language, the models were trained on 2,000 sentences and evaluated on a test set of 1,000 sentences.

Results and Discussion

The Naive Bayes approach proved remarkably effective, significantly outperforming the "most-frequent" (MF) baseline.

Accuracy of grammatical function guessing across different feature sets.

As shown in the results, the model using both attribute names and values (NB-N&V) achieved 91.6% accuracy for English and 82.5% for German. Key insights include:

  • Attribute Predictivity: The mere presence of attributes is highly informative; for example, the presence of ATYPE strongly indicates an ADJUNCT.
  • Beyond Case: While CASE is traditionally viewed as the primary indicator for subjects and objects, the inclusion of all available attributes yielded superior performance.
  • Language Variance: The lower performance in German likely reflects its higher grammatical complexity or differences in grammar design.
  • Practical Impact: Integrating the guesser into a Statistical Machine Translation (SMT) pipeline improved the BLEU score by 1.24%, a significant gain in translation quality.

Conclusion

This study demonstrates that simple probabilistic models can effectively reconstruct missing grammatical information in dependency structures. While currently tested on LFG, the method is representation-agnostic and potentially applicable to various NLP frameworks. Future work may explore the integration of semantic data or word-level "PRED" values to further enhance accuracy.

Summary updated

BibTeX
@inproceedings{bryl2009guessing,
  author = {Anton Bryl and Josef van Genabith and Yvette Graham},
  title = {{Guessing the Grammatical Function of a Non-Root F-Structure in LFG}},
  booktitle = {IWPT},
  year = {2009},
  url = {https://openreview.net/forum?id=A-X6oef6hLw},
}
Figure