Towards a Unified Taxonomy of Deep Syntactic Relations

Kira Droganova, Daniel ZemanCOLING 2024

TL;DR

This paper tackles the challenge of inconsistent deep syntactic role labeling across different linguistic frameworks by proposing a unified taxonomy. By analyzing Meaning-Text Theory, Functional Generative Description, and PropBank, the authors aim to create a standardized set of semantic role labels that can be mapped from existing datasets, fostering better cross-lingual and cross-framework interoperability in natural language processing.

What Problem Does This Paper Address?

Natural language processing (NLP) systems strive to understand the meaning behind language, not just its surface structure. While Universal Dependencies (UD) has standardized morphological and surface syntactic annotation across many languages, a deeper layer of semantic understanding remains fragmented. Different linguistic theories and annotation frameworks (like Meaning-Text Theory, Functional Generative Description, and PropBank) use their own sets of "deep syntactic" or "semantic" role labels to capture predicate-argument relationships. This lack of uniformity makes it difficult to compare findings, build cross-lingual resources, and develop robust NLP models that can generalize across different annotation schemes. The paper addresses this by proposing a unified taxonomy to bridge these diverse representations.

Methods

The core of this paper's methodology involves a comparative analysis of three prominent deep syntactic frameworks: Meaning-Text Theory (MTT), Functional Generative Description (FGD), and PropBank. The authors systematically surveyed the semantic role labels used within each framework, focusing on their definitions, typical syntactic realizations (often mapped to Universal Dependencies labels), and the phenomena they capture. Based on this comprehensive comparison, they propose a hierarchical, unified taxonomy of semantic role labels. This taxonomy is designed to allow for mapping from the source frameworks to the unified set with minimal information loss, and ideally, to enable mapping back from the unified set to the source frameworks. The proposed taxonomy is structured with main categories and optional subcategories to preserve finer distinctions where available.

Datasets

The research utilizes existing annotated datasets from the following languages and frameworks:

  • Spanish and Catalan: AnCora corpus (Taulé et al., 2008)
  • Czech: Prague Dependency Treebank (PDT) (Hajič et al., 2006, 2013, 2017)
  • English: Penn Treebank (Marcus et al., 1993) annotated with PropBank roles (Palmer et al., 2005)
  • Finnish: Finnish Proposition Bank (Haverinen et al., 2015)

The paper also mentions that updated datasets with the new universal semantic role labels are publicly available as a result of their work. Their ultimate goal is to apply a subset of these universal labels to the full Universal Dependencies (UD) data.

Key Insights and Ideas

  • The "Deep Syntax" Landscape is Fragmented: A major takeaway is the significant divergence in how different linguistic theories and annotation efforts define and label semantic roles, even for seemingly similar concepts. This highlights the fundamental challenge in achieving cross-linguistic and cross-framework consistency.
  • UD as a Bridge, But Not a Solution: While UD provides a valuable surface-level standardization, it doesn't capture the deeper semantic relationships that are crucial for true language understanding. The paper positions its work as a necessary step beyond UD.
  • Hierarchical Taxonomy is Key: The proposed unified taxonomy's hierarchical structure is a crucial design choice. It allows for both coarse-grained, universally applicable labels and finer-grained distinctions that can be preserved when mapping from more detailed source frameworks. This flexibility is essential for practical application.
  • Mapping is Not Always One-to-One: The authors acknowledge that a perfect, lossless mapping between all labels across frameworks is not always guaranteed. This implies that some information might be generalized or lost when moving between systems, a realistic limitation of unification efforts.
  • The Importance of "Inner Participants": The paper emphasizes the core arguments (inner participants) of verbs as a primary focus for unification, recognizing their central role in sentence meaning.

Key Results

The paper's primary result is the proposal of a unified taxonomy of deep syntactic relations. This taxonomy aims to provide a standardized set of labels that can encompass the distinctions found in Meaning-Text Theory, Functional Generative Description, and PropBank. While the paper doesn't present quantitative results in terms of accuracy or performance metrics (as it's a proposal for a taxonomy), it demonstrates the feasibility of mapping labels from the analyzed frameworks to this unified set. The authors have made updated datasets with these new labels publicly available, facilitating future research and application.

Limitations and Future Work

  • Language Typology: The authors acknowledge that their current language pool is not very typologically varied (four Indo-European and one Uralic language). They suggest that future work could explore how their proposal would fare with ergative languages, for example.
  • Ideal Mapping: The paper states that while they aim for a mapping without loss of information, there's no guarantee this ideal goal is achievable.
  • Application to UD: A key future direction is the application of this unified label set to the full Universal Dependencies (UD) data. This involves combining cross-lingual projection with heuristics and potentially valency frame lexicons.
  • Enhanced Graphs: The authors intend to apply their semantic labels to enhanced graphs in future work, beyond the basic trees currently used.
  • Noise in Automatic Annotation: They anticipate that automatically derived annotations using this unified set will likely contain noise, but the unified labels will at least allow for comparable annotation across datasets.

Summary updated

BibTeX
@inproceedings{droganova2024towards,
  author = {Kira Droganova and Daniel Zeman},
  title = {{Towards a Unified Taxonomy of Deep Syntactic Relations}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}

Loading related content...

Figure