FrameNet and Typology

Michael Ellsworth, Collin Baker, Miriam R. L. PetruckNAACL 2021

TL;DR

This paper proposes using FrameNet's semantic frame annotations on parallel texts as a powerful new tool for computational semantic typology. By analyzing how different languages express similar events and situations, researchers can uncover universal semantic patterns and language-specific exceptions, advancing our understanding of linguistic diversity.

What Problem Does This Paper Address?

Linguistic typology has long sought to identify recurring patterns and differences across languages, particularly in how semantic information is encoded. While syntactic typology is well-established, semantic typology has lagged behind. Researchers have observed cross-linguistic patterns in areas like motion event descriptions (e.g., Talmy's verb-framing vs. satellite-framing) and argument structure, but a systematic, large-scale approach to uncovering these semantic universals and exceptions has been missing. Existing resources like PropBank and Unified Meaning Representation (UMR) don't fully capture the fine-grained, predicate-specific semantic roles that FrameNet provides. The paper argues that FrameNet, with its rich semantic frames and extensive multilingual extensions, is an underutilized resource that can bridge this gap and significantly contribute to computational semantic typology.

Methods

The core method proposed is the application of FrameNet's semantic frame annotation to parallel texts across multiple languages. This involves:

  • Frame Semantics: Utilizing the theory of Frame Semantics, which posits that words evoke "frames" – structured representations of concepts, events, or situations – and these frames are populated by "frame elements" (FEs) representing participants or roles.
  • FrameNet and Multilingual FrameNet (MLFN): Leveraging the existing FrameNet resource (for English) and its extensions to various languages (MLFN). These MLFN projects build frames based on their own linguistic data, allowing for cross-linguistic comparison.
  • Parallel Text Annotation: Annotating parallel texts (texts that are translations of each other) with semantic frames. This allows for direct comparison of how the same meaning is expressed in different languages.
  • Frame Alignment and Visualization: Developing methods and tools (like ViToXF) to align frames across languages and visualize similarities and differences. This involves comparing frame structures, lexical units (LUs), and frame elements.
  • Frame Shift Analysis: Identifying instances where a translated phrase evokes a different semantic frame than the original, a phenomenon termed "frame shift." This is seen as a key indicator of typological differences.

Datasets

The paper discusses the use of several datasets and resources:

  • FrameNet (BFN): The original English FrameNet resource, serving as a foundational element.
  • Multilingual FrameNet (MLFN) Projects: Resources for languages including Chinese, Portuguese (Brazil), French, German, Japanese, and Spanish. Table 2 provides a snapshot of the number of frames and lexical units in these resources.
  • Parallel Texts:
    • The Hound of the Baskervilles: A chapter from this novel was annotated in English, Japanese, Spanish, and German to study motion and location vocabulary.
  • TED Talk ("Do Schools Kill Creativity?"): Portions of this TED talk were annotated in English, Portuguese, Japanese, and French by Global FN teams. This was done using only BFN frames to facilitate comparison, with discrepancies noted.

Key Insights and Ideas

  • Frame Semantics as a Typological Tool: Frame annotation provides a fine-grained, semantically-driven approach to typology, complementing syntactic methods and revealing patterns that might otherwise be missed.
  • Universality vs. Specificity: While many basic frames (e.g., for eating, drinking) appear to be universal, there are also language-specific exceptions and variations in how these frames are realized.
  • Frame Shifts Reveal Typological Differences: The phenomenon of "frame shift" (where a translation evokes a different frame) is a powerful indicator of deeper conceptual and typological differences between languages, particularly in how they structure events and situations.
  • MLFN is Crucial for Typology: The development of FrameNet resources for diverse languages is essential for enabling cross-linguistic semantic analysis and computational typology.
  • ViToXF for Visualization: Tools like ViToXF are vital for making sense of large-scale cross-linguistic frame comparisons and identifying patterns.

Key Results

  • Motion and Location Differences: An analysis of The Hound of the Baskervilles revealed typological differences in motion and location vocabulary, with English exhibiting satellite-framing tendencies while Spanish and Japanese showed verb-framing tendencies.
  • Visibility Saliency: The study highlighted how Japanese explicitly encodes visibility in ways that are often implicit in English, suggesting a potential cline of "saliency of visibility" comparable to "saliency of manner."
  • Frame Shifts in TED Talk: The TED talk annotation study reinforced the finding of frame shifts, demonstrating how even seemingly equivalent phrases can evoke different frames across languages, impacting subtitle generation and cross-linguistic understanding. For example, "children starting school" was mapped to Activity_start in English but Becoming_a_member in Japanese.
  • Frame Similarity Metrics: The MLFN team developed quantitative measures (e.g., using LU vector embeddings) to assess frame similarity across languages, revealing interesting conceptual differences.

Limitations and Future Work

  • Annotation Consistency: The TED talk annotation highlighted challenges in achieving consistent "complete annotation" across different language teams and the impact of subtitle timing constraints on translation brevity.
  • Frame Equivalence: While many frames are applicable across languages, cultural differences can preclude direct equivalence for certain concepts (e.g., religious or legal). The paper notes that frame relations like Perspective_on can help capture different viewpoints.
  • Ambiguity: The ambiguity of words taken out of context poses a challenge for frame similarity metrics.
  • Future Work: The authors suggest that frame annotation can help typologists confirm, refute, or refine hypotheses about semantic patterns. Work is underway to develop systems that can predict frame shifts based on annotation data. The paper also points to the need for a robust framework that structures relationships between meanings to reveal commonalities and differences.

Summary updated

BibTeX
@inproceedings{ellsworth2021framenet,
  author = {Michael Ellsworth and Collin Baker and Miriam R. L. Petruck},
  title = {{FrameNet and Typology}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure