The Effects of Syntactic Features in Automatic Prediction of Morphology

Wolfgang Seeker, Jonas KuhnEMNLP 2013 · DOI

Introduction

In computational linguistics, morphological tagging is the task of assigning grammatical categories—such as gender, number, and case—to words in a sentence. While this is straightforward for some languages, others exhibit high levels of syncretism, where a single word form can represent multiple distinct grammatical categories. For instance, in German, the word Regionen (regions) can be nominative, genitive, dative, or accusative depending on its context.

Standard taggers typically rely on a "linear window" approach, examining immediate neighbors to resolve ambiguity. However, the information required to disambiguate a word often resides in its structural relationship with distant words, such as a verb agreeing with its subject. This paper explores how incorporating syntactic dependency information can significantly enhance morphological prediction in syncretic languages like German and Czech.

Figure 1: A dependency tree illustrating the complex relationships in a German noun phrase. The word 'Regionen' is disambiguated through its structural links.

Methodology: A Bootstrapping Approach

The authors propose a multi-stage "bootstrapping" architecture to integrate syntax into morphology prediction:

  1. Baseline Tagger: A standard tagger is built using local features like prefixes, suffixes, and neighboring words.
  2. Dependency Parser: A graph-based dependency parser is trained using the output of the baseline tagger as input.
  3. Syntactic Feature Integration: The morphological tagger is retrained, now including features derived from the parser’s output. These features include the lemma and tag of a word's syntactic "head" (the word it depends on), the direction of that head, and the tag of the word's leftmost dependent.
  4. Jackknifing: To ensure the model learns to handle the imperfect syntax it will encounter in real-world applications, a 10-fold jackknifing procedure is used during training.

Experimental Setup

The study evaluates four typologically diverse languages to test the limits of syntactic features:

  • German (TiGer treebank): High syncretism, relatively free word order.
  • Czech (CoNLL 2009): Highly inflected with significant ambiguity.
  • Hungarian (Szeged Treebank): An agglutinating language where grammatical categories are usually marked by unique morphemes.
  • Spanish (CoNLL 2009): A Romance language with more restrictive word order and lower baseline ambiguity.

Results and Discussion

The results confirm that syntactic features are most effective for languages with high syncretism that can be resolved through agreement.

Accuracy Gains

For German, accuracy improved from a baseline of 89.11% to 90.10%. Czech saw a similar boost, rising from 92.29% to 92.82%. In contrast, Hungarian showed almost no improvement because its word forms are rarely ambiguous. Spanish, already having high baseline accuracy, left little room for syntactic features to provide additional value.

The Data Plateau

An important finding is the relationship between the amount of syntactically annotated data and the resulting morphological accuracy. The authors observed that the benefits of syntactic features plateau relatively quickly.

Figure 3: Accuracy of morphology prediction for German and Czech as a function of the number of sentences in the training data for the syntactic parser.

As shown in the graphs, the performance gain is realized with as few as 1,000 to 5,000 training sentences. This suggests that even small treebanks can be leveraged to improve morphological tagging significantly.

Agreement and Unknown Words

The use of syntax was particularly potent for capturing long-distance relationships. When using "gold" (perfect) syntax, subject-verb agreement accuracy in German jumped to 97.15%. Furthermore, syntactic context proved invaluable for "out-of-vocabulary" words, where the model cannot rely on memorized lexical patterns and must instead depend on the word's role in the sentence structure.

Conclusion

This research demonstrates that syntax is a powerful tool for resolving morphological ambiguity, particularly in languages like German and Czech. While the pipeline approach introduces potential for error propagation, the "orthogonal" information provided by dependency structures offers a clear advantage over purely linear models. Future work may focus on joint models that predict syntax and morphology simultaneously to further refine these results.

Summary updated

BibTeX
@inproceedings{seeker2013effects,
  author = {Wolfgang Seeker and Jonas Kuhn},
  title = {{The Effects of Syntactic Features in Automatic Prediction of Morphology}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2013},
  doi = {10.18653/v1/d13-1033},
}
Figure