Universal Feature-based Morphological Trees

Federica Gamba, Abishek Stephen, Zdeněk ŽabokrtskýCOLING 2024

TL;DR

This paper introduces a novel way to represent word morphology within Universal Dependencies (UD) syntactic trees, creating "morphological trees." This approach aims to improve cross-lingual comparability by explicitly encoding internal word structure and aligning it with syntactic dependencies, offering a richer representation for linguistic analysis.

What Problem Does This Paper Address?

Universal Dependencies (UD) provides a standardized framework for syntactic annotation across many languages, enabling cross-lingual comparisons. However, the standard UD tokenization, often based on whitespace, can obscure crucial internal word structure, especially in morphologically rich languages. This can lead to dissimilar syntactic tree structures for semantically equivalent sentences across languages, hindering true comparability. For instance, a single Czech word might correspond to multiple English words, or vice versa, making direct alignment difficult. This paper tackles this by proposing a method to integrate morphological segmentation directly into UD syntactic trees.

Methods

The core of this work is a novel data representation that merges UD syntactic trees with the internal morphological structure of words. The authors developed a workflow to achieve this:

  • Morphological Segmentation: They leverage existing resources like UniSegments and UniMorph to segment words into their constituent morphs (roots, prefixes, suffixes). This process involves a series of steps to handle inflected and uninflected forms, and cases where lemmas are or aren't segmented in the available resources.
  • Feature Alignment: They investigate the correlation between morphs and UD's universal morphological features. Using AP (Association Probability) scores, they quantify the strength of the relationship between specific morphs and features, allowing them to assign relevant features to individual morphs.
  • Integration into UD: The segmented morphs are integrated into UD trees as Multi-Word Tokens (MWTs), with specific deprel subtypes (e.g., case:morph, conj:morph) and POS tags (X) to denote their morphological role.

Datasets

The study utilizes several existing resources:

  • UniSegments: A collection of harmonized segmentation resources for 32 languages.
  • UniMorph: Provides normalized morphological paradigms for numerous languages.
  • SIGMORPHON: Specifically, Czech gold-annotated data for morpheme segmentation.
  • Universal Dependencies (UD) Treebanks: Version 2.12 of UD treebanks for 10 languages, including Parallel UD (PUD) treebanks where available.

No new benchmark datasets were introduced; instead, the focus was on integrating existing resources.

Key Insights and Ideas

  • Morphs Carry Grammatical Meaning: The statistical analysis (AP scores) reveals a strong correlation between specific morphs and grammatical features, suggesting that morphs are not just arbitrary segments but often encode specific linguistic functions.
  • Cross-Lingual Comparability Enhancement: By explicitly representing internal word structure and aligning it with syntax, the proposed "morphological trees" offer a more granular and potentially more comparable representation of sentence structures across languages, especially those with differing morphological complexity.
  • Leveraging Existing Resources: The methodology effectively combines multiple linguistic resources (syntactic annotations, segmentation lexicons, morphological paradigms) to create a richer representation, demonstrating the power of resource integration.
  • UD as a Flexible Framework: The work shows how UD, with its extensibility (e.g., deprel subtypes), can accommodate more complex linguistic phenomena like internal word structure.

Key Results

  • Improved Segmentation Quality: A small-scale manual evaluation on French words indicated that combining UniSegments and UniMorph using their algorithm led to a significant improvement in segmentation quality (48% correctly identified morphs) compared to using either resource alone (14% and 32% respectively).
  • Empirical Evidence for Morph-Feature Correlation: The AP scores provided quantitative evidence for the association between morphs and features across languages like English, German, Hungarian, and Latin. For instance, the morph "ing" in English showed a very strong correlation with the feature VerbForm=Ger (AP score of 0.966).
  • Demonstration of Morphological Trees: The paper presents visual examples of these "morphological trees" for English, French, and Finnish sentences, illustrating how segmented words are integrated into the UD syntactic structure.

Limitations and Future Work

  • Resource Dependency: The quality of the proposed representation is inherently tied to the quality and limitations of the exploited resources. Discrepancies between resources can indirectly affect the output.
  • Ambiguity in Segmentation: The current approach focuses on a single best segmentation and does not fully handle potential ambiguities, especially in highly ambiguous languages.
  • Limited Evaluation: The evaluation of segmentation quality was based on a very small sample.
  • Future Work: The authors plan to:
  • Improve the workflow and expand the collection of morphological treebanks to more languages.
  • Explore the integration of these morphological trees with other linguistic tools like DeriNet.
  • Investigate the use of morphological features for multilingual morphological analyzers.
  • Incorporate phenomena like non-concatenative morphology.
  • Develop methods to estimate the quality of the resulting trees.

Summary updated

BibTeX
@inproceedings{gamba2024universal,
  author = {Federica Gamba and Abishek Stephen and Zdeněk Žabokrtský},
  title = {{Universal Feature-based Morphological Trees}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure