Computing Parallelism in Discourse

IJCAI 1997

TL;DR

This paper introduces a novel computational theory for identifying parallel elements within discourse, bridging the gap between linguistic theories of parallelism and formal approaches like Higher-Order Unification (HOU). By integrating a sort-based abductive calculus with HOU, the framework can predict parallel structures, offering a more robust account of discourse semantics than previous methods.

What Problem Does This Paper Address?

Discourse analysis often relies on the assumption that parallel structures (like repeated phrases or similar sentence constructions) are readily identifiable. However, this assumption, prevalent in theories of ellipsis and discourse coherence, undermines the predictive power of semantic theories. If parallel elements are simply given, it becomes difficult to determine when two utterances are truly parallel and what constraints govern this determination. Existing approaches, while acknowledging parallelism, often presuppose its identification rather than computing it. This paper aims to fill this gap by providing a formal, computational theory that actively predicts which elements in two utterances are parallel to each other.

Methods

The core of the paper's contribution lies in the integration of two key components:

  • Sort-Based Abductive Calculus for Parallelism: The authors propose a novel calculus that uses a hierarchy of sorts (representing cognitive concepts) to model similarity and contrast between linguistic elements. This calculus is "abductive," meaning it generates plausible explanations for why elements might be parallel. It manipulates two types of equality: one for similarity (=s) and one for contrastive parallelism (=c). This approach allows the system to predict parallel elements, unlike previous theories that took them as given.
  • Higher-Order Coloured Unification (HOCU): The paper leverages a restricted form of HOU, known as HOCU, which is more suitable for natural language semantics than pure HOU. HOCU uses "colors" to annotate terms, allowing for constraints on how substitutions can occur. This is crucial for controlling over-generation and ensuring linguistically valid readings.

The proposed framework, called ARP (Abductive Reconstruction of Parallelism), combines these two by treating parallelism and similarity as special equations within a HOCU framework. This integration allows the system to handle the interaction between parallelism and other discourse phenomena like ellipsis and gapping.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Predicting Parallelism: The most significant insight is the shift from assuming parallelism to computing it. The abductive calculus provides a mechanism to derive parallel elements based on sort hierarchies and similarity/contrast relations.
  • Bridging Linguistic and Formal Theories: The paper successfully integrates linguistic intuitions about parallelism (similarity, contrast) with formal logical tools (HOU, sorted logic). This creates a more unified and powerful framework for discourse semantics.
  • Controlling Ambiguity with Colors: The use of HOCU's "colors" is a clever way to enforce constraints on parallelism, effectively implementing restrictions like the "Primary Occurrence Restriction" (POR) used in previous ellipsis theories. This helps rule out linguistically invalid interpretations.
  • Modeling Gapping and Ellipsis: The framework demonstrates its capability by correctly predicting the ambiguity in gapping constructions (e.g., "Jon likes golf, and Mary too") and explaining why one reading is typically preferred over another, based on the strength of the parallel elements.

Key Results

The paper presents a theoretical framework and demonstrates its application through examples, rather than reporting quantitative results on specific datasets. The key "results" are qualitative:

  • The proposed theory correctly predicts parallel elements in discourse.
  • It integrates Higher-Order Unification with discourse theories of parallelism.
  • It provides a natural framework for modeling the effect of parallelism on discourse semantics.
  • It correctly captures the interaction of VP-ellipsis and gapping with parallelism.
  • It offers a more principled way to handle ambiguity in constructions like gapping, explaining preferred readings based on the strength of parallelism.

Limitations and Future Work

  • Intractability of Naive Implementation: The authors acknowledge that a naive implementation of the ARP calculus would be intractable due to the large number of possible abducibles. They suggest that controlling the search for abducibles through a quality measure (e.g., conceptual distance of sorts) is necessary.
  • Generalizing the Framework: The current coding for parallelism in gapping constructions is noted as not being general enough for cases with more than one parallel element in the target utterance. Further work is needed to develop a general treatment.
  • Interaction with Other Discourse Phenomena: While the framework is shown to handle ellipsis and gapping, future work will investigate its interaction with anaphora and quantification in more detail.
  • Discourse Relations: The proposal needs to be extended to cover discourse relations like exemplification and generalization to fully account for the incremental effect of parallelism on semantic interpretation.
  • Sophisticated Measure Functions: The authors suggest that more sophisticated measures for abducibles, beyond simple sort distance, might be needed, potentially incorporating concepts like the specificity of the solution.

Summary updated

BibTeX
@inproceedings{1997computing,
  title = {{Computing Parallelism in Discourse}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure