Representation of Morphosyntactic Units and Coordination Structures in the Turkish Dependency Treebank

Umut Sulubacak, Gülşen EryiğitEMNLP 2013

TL;DR

This paper introduces novel annotation schemes for Turkish dependency parsing, focusing on representing morphosyntactic units (inflectional groups) and coordination structures. By improving how complex word structures and sentence conjunctions are marked, the authors achieve better parsing accuracy and create a more robust framework for handling the complexities of agglutinative languages like Turkish, especially for unedited web data.

What Problem Does This Paper Address?

Dependency parsing, a method for analyzing sentence structure by identifying relationships between words, has seen significant advancements. However, languages with highly agglutinative morphology, like Turkish, present unique challenges. These languages form words by adding numerous suffixes, leading to a massive number of potential word forms and complex morphological tags. Traditional parsing methods struggle with this, often resulting in lower accuracy compared to languages with simpler morphology.

This paper specifically tackles two main issues within Turkish dependency parsing:

  1. Representing Morphosyntactic Units (Inflectional Groups - IGs): Turkish words can be broken down into smaller units called inflectional groups (IGs), often separated by derivational suffixes. Accurately representing these IGs and their relationships is crucial for parsing. Existing methods in treebanks often hide the actual lexical information of these units to represent them, hindering their use in parsing decisions.
  2. Representing Coordination Structures (CSs): Coordinating structures (e.g., "A and B") are common in sentences but can be tricky to annotate consistently. The existing Turkish Treebank's approach to CSs is problematic, requiring intermediate conjunctions and potentially leading to incorrect dependency assignments and lower recall.

The authors aim to develop annotation schemes that are more accurate, easier to use, and better suited for handling the specific linguistic characteristics of Turkish, particularly for parsing less structured data like web text.

Methods

The researchers employ MaltParser, a data-driven dependency parser that uses a Support Vector Machine (SVM) learner. They utilize the Nivre's Arc-Standard parsing algorithm and optimize parameters based on previous work. The core of their contribution lies in proposing new annotation schemes:

  1. Novel IG Representation (Model II): Instead of hiding lexical information, they propose a new scheme that uses an additional "IG" column. This column indicates whether an IG is a non-final part of a multi-IG word. This allows the original "FORM" and "LEMMA" columns to retain surface lexical forms and previous IG forms, respectively, making lexical information available for parsing.
  2. Proposed Coordination Structure Representation (Swedish Style): They compare three CS annotation schemes (Turkish Treebank original, Swedish Style, and Stanford Style) and advocate for the Swedish Style. This style simplifies the representation by directly connecting conjuncts to the external head, bypassing the need for intermediate conjunctions to be explicitly linked. This approach is more robust to missing conjunctions and reduces the complexity of dependency relations.

Datasets

The primary dataset used is the Turkish Treebank. The authors start with an existing version and then apply manual and automatic transformations to create experimental datasets for their new annotation schemes. They use a 10-fold cross-validation approach for evaluation due to the lack of a separate validation set in their initial data.

Key Insights and Ideas

  • Lexical Information is Key: Retaining lexical information for morphosyntactic units (IGs) is crucial. Hiding it, as in some previous schemes, limits its utility for parsing decisions, even if it simplifies the representation.
  • Simpler CS Annotation is Better: The "Swedish Style" for coordinating structures significantly improves parsing performance, especially recall, by simplifying the annotation process and making it more robust to variations in sentence structure.
  • Agglutination Requires Specialized Handling: The success of their proposed schemes highlights that languages with rich morphology require tailored annotation strategies to achieve high parsing accuracy.
  • Web Data Compatibility: The new schemes are designed to be more compatible with unedited web data, which often lacks perfect grammatical structure and explicit conjunctions.

Key Results

  • IG Representation (LexedIG Model): The new IG representation preserves the perfect precision and recall for DERIV relations. While overall accuracy changes are not statistically significant, the ability to use lexical information is seen as a superior approach for future parsing of web data.
  • Coordination Structure Representation:
    • The Swedish Style and Stanford Style models show significant improvements over the baseline, particularly in the COORDINATION relation.
  • Recall for the COORDINATION relation increased from 53.70% (baseline) to 58.68% (Stanford Style) and 64.64% (Swedish Style).
  • The Swedish Style and Stanford Style models also showed slight improvements in accuracy for dependency types like SUBJECT, OBJECT, and DATIVE, INSTRUMENTAL, and ABLATIVE.ADJUNCTS.
  • Table 3 shows that the Swedish Style achieved a general parsing accuracy (ASL) of 73.2% ± 0.2, which is an improvement over the baseline (74.5% ± 0.2) and the corrected baseline (74.5% ± 0.1), though the ASU score is slightly lower. (Note: The paper seems to have a slight discrepancy in reporting ASU for Swedish Style vs. Baseline in Table 3, but the trend indicates improvement in ASL).

Limitations and Future Work

  • Lexical Data Sparsity: The authors note that the reconstructed lexical information in the LexedIG model caused only slight changes in overall accuracy, which they attribute to the sparsity of lexical data. They anticipate that this will become less of an issue with unsupervised parsing of web data.
  • Manual Correction: While the automatic conversion for CSs led to improvements, a significant amount of manual correction was initially needed for the corpus.
  • Future Work: The authors intend to continue their research by developing better annotation schemes for other syntactic structures, such as copulae and modifier sub-types, to serve not only Turkish but also other languages with rich morphology. They also foresee that the ability to unsupervisedly parse large amounts of web data will make lexical forms even more valuable.

Summary updated

BibTeX
@inproceedings{sulubacak2013representation,
  author = {Umut Sulubacak and Gülşen Eryiğit},
  title = {{Representation of Morphosyntactic Units and Coordination Structures in the Turkish Dependency Treebank}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2013},
}
Figure