A Framework for Lexical Selection in Natural Language Generation

Sergei Nirenburg, Irene NirenburgCOLING 1988

TL;DR

This paper introduces a novel framework for lexical selection in natural language generation (NLG) systems, focusing on how to choose the "right" word from a vast vocabulary. It addresses the challenge of generating fluent and contextually appropriate text by incorporating meaning, collocations, and non-lexical realization strategies like anaphora and ellipsis.

What Problem Does This Paper Address?

Generating natural language from structured meaning representations is a complex task. While many NLG systems excel at syntactic structure, the crucial step of selecting the most appropriate open-class lexical items (nouns, verbs, adjectives, adverbs) often remains a significant challenge. This is particularly true for systems that need to produce a wide range of outputs, such as those used in machine translation or text summarization, where the lexicon is large and context is paramount. The authors argue that current systems often "don't use words at all" in a sophisticated sense, leading to less fluent output. They aim to bridge this gap by developing a robust lexical selection module that can handle varying contextual circumstances and leverage rich linguistic knowledge.

Methods

The paper proposes a multi-faceted approach to lexical selection within the DIOGENES generation system. Key components include:

  • Meaning-Based Selection: An initial step involves matching input meaning units to candidate lexical items based on semantic correspondences. A "context-independent meaning matching metric" is used for this.
  • Collocational Constraints: Recognizing that word choice is heavily influenced by co-occurrence patterns, the system incorporates a "generation lexicon" that stores collocational information. This allows the system to filter candidate words based on what typically appears together (e.g., "high voltage" vs. "large voltage"). This includes both syntagmatic (word-to-word) and paradigmatic (set-complement) collocations.
  • Non-Lexical Realization: The framework explicitly handles cases where a word is not the best choice. This includes:
  • Pronominalization: Using pronouns (e.g., "he," "it") when a concept has already been introduced.
  • Definite Descriptions: Using phrases like "the book" to refer to a concept.
  • Ellipsis: Omitting words or phrases when their meaning is clear from context.
  • Confidence-Based Decision Making: The algorithm uses confidence scores to decide whether to proceed with a chosen lexical item or explore alternative realizations, including non-lexical ones.

Datasets

Not discussed in the paper. The paper focuses on the framework and its components, rather than specific empirical evaluations on large datasets. The generation lexicon is described as a knowledge source that needs to be acquired.

Key Insights and Ideas

  • Collocations are Crucial for Fluency: The paper strongly emphasizes that simply matching meaning is insufficient. Collocational knowledge, which captures idiosyncratic word co-occurrence patterns, is essential for generating natural-sounding language.
  • Non-Lexical Realization is Integral: Lexical selection isn't just about picking words; it's also about knowing when not to pick a word and to use pronouns, definite descriptions, or ellipsis instead. This is a key aspect of human-like language generation.
  • Context is King (and Queen): The framework is designed to be highly context-aware, using both semantic and collocational information to make informed decisions.
  • Lexicon Acquisition is a Bottleneck: The authors acknowledge that building a comprehensive generation lexicon, especially with detailed collocational information, is a labor-intensive task.

Key Results

The paper describes the implementation of the DIOGENES system's lexical selection module, including the blackboard architecture, meaning matching, and collocation treatment. The anaphora treatment is implemented for pronominalization. While quantitative results are not presented, the framework is designed to improve the fluency and appropriateness of generated text by addressing the limitations of purely meaning-based lexical choice.

Limitations and Future Work

  • Lexicon Acquisition: The primary limitation is the labor-intensive nature of acquiring the generation lexicon, particularly its collocational information. The authors suggest that intelligent interactive aids can help accelerate this process.
  • Anaphora Treatment: The current anaphora treatment is limited to pronominalization, and the authors state that the number of rules needs to be increased.
  • Future Work: The authors plan to extend their knowledge base maintenance system to facilitate the acquisition of generation lexicons and to further develop the anaphora treatment. They also aim to conduct extensive testing of the collocation treatment module.

Summary updated

BibTeX
@inproceedings{nirenburg1988framework,
  author = {Sergei Nirenburg and Irene Nirenburg},
  title = {{A Framework for Lexical Selection in Natural Language Generation}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure