An Underspecified Segmented Discourse Representation Theory (USDRT)

Frank SchilderCOLING 1998

Sure, here's a blog-style summary of the paper "An Underspecified Segmented Discourse Representation Theory (USDRT)":

TL;DR

This paper introduces Underspecified Segmented Discourse Representation Theory (USDRT), a novel approach to modeling discourse structure that can represent ambiguities. By combining Segmented Discourse Representation Theory (SDRT) with Tree Description Grammars (TDGs), USDRT offers a more flexible and comprehensive way to capture the hierarchical relationships and potential interpretations within a text.

What Problem Does This Paper Address?

Interpreting discourse, or how sentences connect to form a coherent whole, is a complex task. Traditional theories like Discourse Representation Theory (DRT) struggle to capture the rhetorical relationships between sentences (e.g., narration, elaboration) and how these relationships influence meaning, especially when multiple interpretations are possible. Segmented Discourse Representation Theory (SDRT) was developed to address this by explicitly modeling these rhetorical relations and the hierarchical structure of discourse. However, SDRT typically generates all possible interpretations, which can be computationally expensive and doesn't directly represent the underspecified nature of ambiguity. This paper tackles the challenge of representing discourse structure in a way that acknowledges and handles ambiguity without necessarily enumerating every single possibility upfront.

Methods

The paper proposes USDRT, which builds upon SDRT by integrating it with Tree Description Grammars (TDGs).

  • SDRT (Segmented Discourse Representation Theory): This is the foundation, providing a framework for representing discourse structure using labeled Discourse Representation Structures (DRSs) connected by discourse relations (like narration, elaboration). SDRT uses a graph-like representation to show these connections and identify "open attachment sites" for new sentences.
  • TDGs (Tree Description Grammars): Inspired by quasi-trees, TDGs are a constraint-based grammar formalism. They use a logic with variables for nodes and binary relations to describe dominance and precedence relationships between subtrees. This formalism is particularly adept at representing hierarchical structures.
  • USDRT: The core innovation is the combination of SDRT's discourse relations with TDG's ability to represent underspecified structures. Instead of generating all possible interpretations, USDRT uses TDG's logic to encode the structural information, including dominance and precedence relations between discourse segments. This allows for a representation that explicitly marks ambiguity, such as multiple possible attachment sites or interpretations of discourse relations, without committing to a single reading. The paper highlights that this approach avoids the overwriting of structures seen in some SDRT updates and provides a more uniform way to handle hierarchical ordering.

Datasets

Not discussed in the paper. The paper uses example sentences and discourse fragments for illustration rather than formal datasets.

Key Insights and Ideas

  • Representing Ambiguity Directly: The most significant insight is the ability to represent underspecified discourse structures. Instead of generating all possible readings of an ambiguous sentence or discourse segment, USDRT can represent the ambiguity itself within the structure.
  • Uniform Hierarchical Description: By leveraging TDGs, USDRT offers a more uniform and consistent way to describe the hierarchical structure of discourse compared to previous SDRT representations.
  • Avoiding SDRT's Monotonicity Issues: The paper points out that standard SDRT can sometimes involve overwriting previously derived structures. USDRT, by using a more flexible tree-based representation, aims to avoid this, allowing for more robust updates.
  • Open Attachment Sites: The TDG formalism naturally encodes information about open attachment sites (where new discourse segments can connect), which is crucial for discourse processing and anaphora resolution.

Key Results

The paper demonstrates the USDRT formalism through examples, showing how it can represent ambiguous discourse structures. For instance, it illustrates how sentence (k5) in example (2) could be related to either the entire discourse segment (k1) or a more specific part of it, and USDRT can capture this ambiguity. The formalism is shown to provide a unified description of hierarchical discourse structure and the ability to express ambiguities within it.

Limitations and Future Work

  • Underspecified Discourse Representation: The paper acknowledges that SDRT, in its standard form, struggles to represent underspecified discourse; all readings must be generated. USDRT aims to solve this.
  • Monotonicity: SDRT can be non-monotonic, meaning updates might overwrite previous structures. USDRT seeks to mitigate this.
  • Interaction of Discourse Conditions: It's noted that the interaction between different SDRS conditions (discourse relations, topic relations, etc.) and their effect on discourse structure isn't always clear.
  • Formal Definition of PROMI: The paper mentions that Schilder (1997) uses a feature called PROMI for topic information but doesn't provide a formal definition. The authors stick to Asher's definition of topic.
  • Future Work: The paper sets the stage for further development by proposing this underspecified framework, suggesting that it can be used to provide a more flexible and comprehensive representation of discourse structure. The integration of TDGs is presented as a promising direction for handling ambiguity in discourse interpretation.

Summary updated

BibTeX
@inproceedings{schilder1998underspecified,
  author = {Frank Schilder},
  title = {{An Underspecified Segmented Discourse Representation Theory (USDRT)}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1998},
}

Loading related content...

Figure