Integrating a Large-Scale, Reusable Lexicon with a Natural Language Generator

Hongyan Jing, Yael Dahan Netzer, Michael Elhadad, Kathy McKeownINLG 2000 · DOI

TL;DR

This paper tackles the challenge of building reusable components for natural language generation (NLG) systems by integrating a comprehensive, multi-source lexicon with the FUF/SURGE generator. This integration automates much of the lexical realization process, enabling the generation of diverse paraphrases and preventing ungrammatical output.

What Problem Does This Paper Address?

Developing natural language generation systems often requires hand-coding lexical, syntactic, and semantic knowledge for each new application. This is inefficient and leads to duplicated effort. While reusable components are common in NLG, a "reusable lexical chooser" – the part of the system that maps semantic concepts to words and phrases – has been elusive. This is primarily because lexical choosers are tightly coupled to the specific semantic representations and syntactic realizer of an application, and the conceptual elements they link to words are inherently domain-dependent. This paper aims to bridge this gap by creating a large-scale, reusable lexicon that can be integrated with existing NLG frameworks, specifically FUF/SURGE, to automate and standardize the lexical realization process.

Methods

The core of the paper is the integration of a large-scale, reusable lexicon with the FUF/SURGE syntactic realizer. This integration is achieved through a multi-step unification process.

  1. Lexicon Construction: The lexicon is built by combining multiple existing resources:
  2. WordNet: Provides a vast collection of words, senses, and lexical relations (synonymy, antonymy, etc.), indexed by synsets (semantic concepts).
  3. English Verb Classes and Alternations (EVCA): Categorizes verbs based on syntactic properties and studies verb alternations (e.g., "there-insertion").
  4. COMLEX Syntax Dictionary: Offers syntactic information for English words.
  5. Brown Corpus: Used for frequency measurements of lexical items and syntactic patterns. The lexicon focuses on verbs and is indexed by WordNet synsets, making it suitable for generation. It includes subcategorizations, alternations, frequency information, and lexical relations. The construction process is semi-automatic, involving merging resources and then aligning them with WordNet synsets.

  6. Integration with FUF/SURGE: The integration is performed via four incremental unification steps:

  7. Semantic Input: A standard semantic input format is defined, using WordNet synsets to represent concepts and numbered arguments for predicate-argument relations. This format is designed to be adaptable from existing application-specific inputs.
  8. Lexical Unification: Semantic concepts (synsets) from the input are mapped to concrete words. This step allows for the generation of lexical paraphrases by selecting different words from the same synset.
  9. Structural Unification: The chosen verb sense is used to unify with its corresponding subcategorization and alternation information from the lexicon. This adds syntactic information and enables syntactic paraphrases.
  10. Constraints on Arguments: Subcategorization constraints are used to restrict unification, ensuring that the generated sentence structure matches the verb's requirements and preventing ungrammatical output. This step also helps select the appropriate transitivity for the verb.
  11. Mapping to SURGE Input: Finally, subcategorizations and alternations are mapped to the FUF/SURGE input format, which involves matching numbered arguments to lexical roles and syntactic categories. This mapping is largely manually encoded for the 147 patterns but is reusable across applications.

Datasets

  • WordNet Lexical Database: Used as the primary source for semantic concepts and lexical relations.
  • English Verb Classes and Alternations (EVCA): Provided information on verb alternations and syntactic properties.
  • COMLEX Syntax Dictionary: Contributed syntactic information for words.
  • Brown Corpus: Used for frequency measurements.

The paper does not introduce new datasets but rather combines existing ones to create a unified lexicon.

Key Insights and Ideas

  • Reusability through Abstraction: By indexing the lexicon with WordNet synsets, the system decouples lexical knowledge from specific application domains, enabling reuse.
  • Automating Lexical Realization: The multi-step unification process automates the mapping from semantic input to SURGE input, significantly reducing the manual effort required for each new generation application.
  • Paraphrase Generation: The lexicon inherently supports generating both lexical paraphrases (by choosing different words from a synset) and syntactic paraphrases (by leveraging verb alternations).
  • Preventing Ungrammatical Output: The integration of subcategorization and argument constraints directly addresses the problem of generating grammatically incorrect sentences.
  • Standardized Semantic Input: The proposed standard semantic input format, using synsets and numbered arguments, is a crucial step towards making lexical choosers more pluggable and reusable.

Key Results

The paper demonstrates the successful integration of a large-scale lexicon with FUF/SURGE. While specific quantitative results (e.g., accuracy improvements, reduction in development time) are not explicitly detailed, the paper highlights the qualitative benefits:

  • Lexical Paraphrases: The system can generate multiple lexical variations for the same semantic concept (e.g., "He seems happy," "He looks happy," "He appears happy").
  • Syntactic Paraphrases: The system can generate syntactic variations (e.g., "Brenda agreed with Molly," "Brenda and Molly agreed," "Brenda and Molly agreed with each other").
  • Prevention of Ungrammatical Output: The system avoids generating sentences with incorrect verb argument structures or missing prepositions.
  • Wide Coverage: The final lexicon includes 5,676 verbs, over 14,100 senses, and 11,000 synsets, providing broad coverage.

Limitations and Future Work

  • Manual Encoding of Mappings: While the unification steps are automatic, the mapping of subcategorizations to SURGE input for the 147 patterns was manually encoded.
  • Human Evaluation of Paraphrases: The authors plan to validate the generated paraphrases by asking human subjects to judge their acceptability and investigate methods to filter out unacceptable paraphrases.
  • Multilingual Generation: The authors are interested in exploring the system's usage in multilingual generation.
  • Adaptability of Synsets: Users may need to adapt synsets for specific domains (e.g., removing irrelevant ones, merging, relabeling) for optimal performance.

Summary updated

BibTeX
@inproceedings{jing2000integrating,
  author = {Hongyan Jing and Yael Dahan Netzer and Michael Elhadad and Kathy McKeown},
  title = {{Integrating a Large-Scale, Reusable Lexicon with a Natural Language Generator}},
  booktitle = {INLG},
  year = {2000},
  doi = {10.3115/1118253.1118282},
}
Figure