An Underspecified HPSG Representation for Information Structure

Jonas KuhnCOLING 1996

TL;DR

This paper proposes a novel, underspecified representation for information structure (IS) that can be efficiently composed within HPSG grammar. This approach aims to overcome the computational challenges of traditional, disjunctive IS representations by offering a compact graphical model that can be resolved using contextual information, making it more suitable for applications like text-to-speech and machine translation.

What Problem Does This Paper Address?

Information structure (IS), particularly how focus is marked in speech (prosody), is crucial for natural language processing applications like text-to-speech and machine translation. However, prosodic marking of focus is often ambiguous, leading to multiple possible interpretations of a sentence's IS. Existing theories, like those based on HPSG, often generate a disjunction of alternative sentence analyses to account for this ambiguity. This disjunctive approach becomes computationally intractable when combined with other linguistic ambiguities, hindering practical applications. The paper addresses this by proposing a more efficient, underspecified representation that avoids explicit disjunction and can be resolved using context.

Methods

The core contribution is a new, underspecified representation for information structure. This representation is graphical, with nodes representing syntactic constituents and directed edges indicating relationships between them and the focus (FOC) and background (BG) partitions of IS.

Key methodological aspects include:

  • Underspecification: Instead of committing to a single, fully specified IS analysis, the representation allows for ambiguity to persist until context-driven resolution.
  • Graphical Representation: A graph-based model is used, where nodes are semantic values of constituents and edges represent accessibility relations. Solid lines denote obligatory inclusion in FOC/BG, while dashed lines represent defeasible inclusion.
  • HPSG Integration: The representation is designed to be composed using HPSG principles. Specific HPSG principles are proposed for linking semantic objects to FOC/BG (Focus Linking, Background Linking) and for projecting focus from heads to their daughters (Head F-Projection, Indirect Head F-Marking).
  • F-Skeleton: A crucial component for resolution is the "F-skeleton," which captures the structure of potential focus markings. This allows for checking contextual entailment without needing a fully specified IS beforehand.
  • Resolution Routine: A context-based resolution routine is presented. It leverages the F-skeleton and a "maximality assumption" to determine givenness and resolve underspecification by adding solid line arrows to the graph, effectively collapsing the underspecified representation into a specific IS.

Datasets

Not discussed in the paper. The paper focuses on linguistic principles and theoretical representation, using example sentences to illustrate its points rather than relying on corpus data for training or evaluation.

Key Insights and Ideas

  • Computational Tractability: The primary insight is that underspecification, rather than explicit disjunction, is key to making IS computationally tractable for NLP applications. This avoids the combinatorial explosion of alternative analyses.
  • Graphical Representation as a Unifying Framework: The proposed graph structure elegantly captures the interplay between syntactic constituents, prosodic marking, and contextual information, providing a unified view of IS composition and resolution.
  • F-Skeleton for Contextual Resolution: The F-skeleton is a clever mechanism that allows for checking contextual entailment without requiring a fully specified IS, thus enabling resolution before full specification.
  • Maximality Assumption Simplifies Resolution: The maximality assumption (checking for the maximal potentially F-marked subconstituents) simplifies the resolution process by reducing the number of entailment checks needed.
  • Avoidance of Pragmatic Filters: The underspecified approach, combined with contextual resolution, makes pragmatic filters like Schwarzschild's "Avoid F" redundant, streamlining the process.

Key Results

The paper presents a theoretical framework and demonstrates its application through examples. The "results" are primarily qualitative and conceptual:

  • A method for composing an underspecified IS representation within HPSG is formally outlined.
  • A resolution routine is presented that effectively resolves underspecification using contextual information and the F-skeleton.
  • The proposed system is shown to handle examples of focus ambiguity and lexical ambiguity (like the word "Besuch") more efficiently than disjunctive approaches.
  • The paper argues that its approach is more computationally efficient and conceptually cleaner than existing theories that rely on explicit disjunction or complex pragmatic filters.

Limitations and Future Work

  • Formalization of "Integration": The paper notes that the formulation of the Indirect Head F-Marking Principle (29) might need refinement once a clearer concept of "integration" is established in linguistic theory.
  • Status of "Avoid F" Filter: The authors express some reservations about the pragmatic nature and necessity of Schwarzschild's "Avoid F" filter, suggesting their underspecified approach makes it redundant.
  • Future Work: The paper suggests that the proposed representation and resolution routine can be directly implemented in HPSG. Further work could explore the practical implementation in NLP systems and test its performance on larger datasets. The authors also hint at the possibility of using heuristics to trigger "Focus Closure" earlier in applications to optimize performance.

Summary updated

BibTeX
@inproceedings{kuhn1996underspecified,
  author = {Jonas Kuhn},
  title = {{An Underspecified HPSG Representation for Information Structure}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1996},
}
Figure