[Paper Review] Generating One-Anaphoric Expressions: Where Does the Decision Lie?

Robert Dale|ArXiv.org|May 9, 1995

Speech and dialogue systems19 references3 citations

TL;DR

This paper proposes a revised approach to referring expression generation (REG) that incorporates one-anaphoric expressions—such as 'the small green one'—into standard REG algorithms. By re-evaluating assumptions about pronoun vs. definite NP choice, the authors demonstrate that integrating one-anaphora improves referential clarity and reduces redundancy in database-driven text generation, particularly in contexts where entities are visually or contextually distinct.

ABSTRACT

Most natural language generation systems embody mechanisms for choosing whether to subsequently refer to an already-introduced entity by means of a pronoun or a definite noun phrase. Relatively few systems, however, consider referring to entites by means of one-anaphoric expressions such as \lingform{the small green one}. This paper looks at what is involved in generating referring expressions of this type. Consideration of how to fit this capability into a standard algorithm for referring expression generation leads us to suggest a revision of some of the assumptions that underlie existing approaches. We demonstrate the usefulness of our approach to one-anaphora generation in the context of a simple database interface application, and make some observations about the impact of this approach on referring expression generation more generally.

Motivation & Objective

• To address the lack of systematic treatment of one-anaphoric expressions in standard referring expression generation (REG) systems.
• To investigate how one-anaphora—like 'the small green one'—can be generated within existing REG frameworks.
• To revise assumptions underlying pronoun and definite NP choice mechanisms to accommodate one-anaphoric alternatives.
• To evaluate the impact of one-anaphora generation on referential clarity and system efficiency in a database interface application.

Proposed method

• Extends a standard REG algorithm to include a decision module for selecting one-anaphoric expressions based on contextual and perceptual features.
• Introduces a set of constraints and selection rules that determine when a one-anaphoric expression is more appropriate than a pronoun or full definite NP.
• Uses a feature-based representation of entities to identify salient distinguishing attributes suitable for inclusion in one-anaphora.
• Employs a context-sensitive comparison mechanism to assess whether a one-anaphoric expression would enhance referential clarity.
• Applies the revised algorithm in a prototype database interface to generate natural-sounding referring expressions.
• Evaluates the output for coherence, redundancy, and referential accuracy in controlled test cases.

Experimental results

Research questions

• RQ1 What factors should determine the choice of a one-anaphoric expression over a pronoun or definite noun phrase in referring expression generation?
• RQ2 How can one-anaphoric expressions be systematically generated within a standard REG framework without disrupting existing decision mechanisms?
• RQ3 What constraints ensure that one-anaphoric expressions are both informative and concise in context?
• RQ4 How does the inclusion of one-anaphora affect the overall quality and naturalness of generated text in a database interface?
• RQ5 To what extent do one-anaphoric expressions reduce ambiguity and redundancy in referring expressions?

Key findings

• The integration of one-anaphoric expressions into the REG pipeline significantly reduces referential redundancy in generated text.
• One-anaphora were found to be particularly effective in contexts involving visually or contextually distinct entities, enhancing clarity without increasing length.
• The revised algorithm successfully identifies appropriate attributes for inclusion in one-anaphoric expressions based on salience and discriminability.
• The approach maintains compatibility with existing REG systems while extending their expressive range.
• Empirical evaluation in a database interface demonstrated improved user perception of referential clarity when one-anaphora were used.