Generating Referring Quantified Expressions

James Shaw, Kathleen R. McKeownINLG 2000 · DOI

TL;DR

This paper presents an algorithm for generating natural language quantified expressions in text generation systems. By leveraging discourse and ontological information, the system aims to produce more concise and fluent text by replacing explicit entity references with quantified phrases, while also resolving ambiguities inherent in natural language quantification.

What Problem Does This Paper Address?

Text generation systems often aim for conciseness and fluency. While techniques like ellipsis are common, this paper focuses on a specific challenge: how to automatically generate quantified expressions (e.g., "each patient," "both arms") to refer to multiple entities in a way that is clear and unambiguous. Natural language quantification is notoriously ambiguous, with issues like distributive vs. collective readings (e.g., "all nurses inspected the patient" - did they do it individually or as a group?) and scope ambiguity (e.g., "a nurse inspected each patient" - does each patient have a unique nurse, or is there one nurse for all patients?). Existing systems often struggle to generate these expressions correctly, especially when dealing with multiple quantifiers or when the intended meaning needs to be precisely conveyed. This paper tackles the generation side of this problem: given an unambiguous representation of facts, how can a system produce a quantified sentence that clearly conveys the intended meaning?

Methods

The authors propose an algorithm that operates within the sentence planner module of a text generation system. The core steps involve:

  1. Identifying Quantifiable Entities: The system first identifies distinct entities within input propositions that can be candidates for quantification.
  2. Generalization and Ontology Lookup: For these entities, the system consults an ontology (specifically, a CLASSIC-based system integrated with WordNet and a medical dictionary) to find a generalized concept that encompasses them. This generalization is crucial for forming a quantified phrase.
  3. Quantification Appropriateness and Type Selection: Based on discourse history and ontological information, the system determines if quantification is appropriate. It then selects the most specific class that covers the entities. The system also considers contexts like discourse (previous mentions) and domain ontology to compute the set of entities (set D) that the quantifier will refer to.
  4. Ambiguity Resolution:
    • Distributive vs. Collective: For universal quantifiers, the system prioritizes avoiding ambiguity. It notes that its input propositions, often derived from combining distinct database entries, naturally lead to distributive readings. The quantifier "each" is favored as it inherently implies a distributive reading. "Every" and "all" are used when a collective reading is unlikely (e.g., for state propositions) or when the lexicon marks the concept as distributive.
  5. Scope Orderings: The algorithm generates different expressions for different scope orderings of multiple quantifiers. It prioritizes universal quantifiers having wider scope than existential ones, as this is often more natural and less ambiguous.
  6. Quantifier Selection: The system has a set of predefined realizations for universal quantifiers (e.g., "each," "every," "all," "both," "the," "any") and existential quantifiers (e.g., "a/an," cardinal numbers). The choice depends on factors like the number of entities, discourse context, and whether negation is involved.

Datasets

Not discussed in the paper. The system relies on an ontology (CLASSIC, WordNet, medical dictionary) and instance-based representations from database tuples.

Key Insights and Ideas

  • Generation vs. Understanding: The paper highlights that while understanding quantifier scope is about disambiguating multiple possible readings, generation is about selecting the right expression to avoid ambiguity from an unambiguous input.
  • Ontology as a Key Resource: The integration with an ontology is critical for finding generalized concepts that allow for quantification and for determining the scope of quantifiers based on domain knowledge.
  • Distributive Reading as Default: The system's architecture, which combines information from multiple sources, naturally leads to distributive readings. This simplifies the choice of universal quantifiers, favoring "each."
  • Context is King: Discourse history and domain-specific knowledge are essential for selecting the most appropriate quantifier and resolving potential ambiguities. For instance, "both" is used when exactly two entities are involved and have been previously mentioned.
  • Prioritizing Clarity over Imprecision: The authors deliberately avoid imprecise quantifiers (like "some," "many") because they don't guarantee that the reader can recover the exact entities, which is crucial in their medical domain.

Key Results

The paper demonstrates an algorithm that can generate quantified expressions, such as "After each critical point, Mrs. Doe had tachycardia" or "Alice has an IV in each arm." It shows how different quantifiers are selected based on context (e.g., "each" for distributive, "both" for two entities, "any" under negation). The system can also generate sentences with multiple quantifiers, typically with the universal quantifier taking wider scope (e.g., "Each patient is given a high severity rating").

Limitations and Future Work

  • Imprecise Quantifiers: The system currently does not handle imprecise quantifiers (some, most, at least, few) because they lack the property of allowing readers to recover exact entities, which is important in their medical application. Future work could explore how to incorporate these if the domain allows for less precision.
  • More Sophisticated Context: While the system uses discourse and ontology, it acknowledges that a more developed pragmatic module could handle more complex interactions, such as those involving referring expression generation modules for more nuanced descriptions.
  • Scope Ambiguity with Existential First: The paper primarily focuses on cases where universal quantifiers have wider scope. While it briefly touches on the reverse, a more comprehensive handling of existential quantifiers taking wider scope could be an area for future development.
  • Domain Independence: While the goal is domain independence, the specific ontology and the decision to avoid imprecise quantifiers are tied to their medical application. Extending this to other domains might require adjustments.

Summary updated

BibTeX
@inproceedings{shaw2000generating,
  author = {James Shaw and Kathleen R. McKeown},
  title = {{Generating Referring Quantified Expressions}},
  booktitle = {INLG},
  year = {2000},
  doi = {10.3115/1118253.1118268},
}
Figure