Goal Oriented Parsing: Improving the Efficiency of Natural Language Access to Relational Data Bases

Giovanni GuidaCOLING 1980

TL;DR

This paper introduces "goal-oriented parsing," a novel approach to natural language understanding (NLU) for database querying. By focusing on specific user goals rather than full linguistic comprehension, the system significantly improves efficiency and allows for more practical, specialized NLU interfaces.

What Problem Does This Paper Address?

Interfacing with relational databases using natural language has been a long-standing challenge in AI. Traditional NLU systems often aim for a comprehensive understanding of language, which is computationally expensive and often overkill for specific tasks like database querying. This paper addresses the inefficiency and complexity of general-purpose NLU by proposing a more focused approach. The authors argue that for many applications, such as querying a database, a "bounded scope, restricted, goal-oriented understanding" is sufficient and can dramatically improve system performance. This means the system only needs to extract the information relevant to achieving a specific goal (e.g., retrieving specific data from a database) and can ignore nuances not pertinent to that goal.

Methods

The core of the paper is the "goal-oriented parsing" approach, implemented in their system called NLI (Natural Language Interface). NLI-2, the focus of the paper, is built on a modular architecture with distinct components for data (vocabulary, database model, formal query language structure) and programs (analyzer, generator, vocabulary management).

Key methodological aspects include:

  • Goal-Oriented Understanding: The system's comprehension is driven by a predefined goal, in this case, translating Italian natural language queries into the QBE (Query-By-Example) formal query language.
  • Semantics-Directed Parsing: The parsing process prioritizes semantic meaning over syntactic structure, aiming to extract the core intent of the query.
  • Modular Architecture: The system is divided into modules for analysis, vocabulary management, and generation, allowing for flexibility and easier development.
  • Lexicon and Heuristics: Linguistic information is stored in a lexicon (words and elementary constructs) and a base of heuristics (for handling ambiguities and complex cases).
  • Generalized Parsing Algorithm: NLI-2 employs a parsing algorithm designed to be independent of the specific lexicon, database structure, or natural language used, promoting reusability.
  • Interactive Vocabulary Management: The system allows for interactive definition and dynamic management of its vocabulary and heuristics, facilitating adaptation to new applications.

Datasets

The paper uses a "toy relational data base" representing a department store's catalog for its sample application. This database is described in an Appendix and is the same one used for illustrating QBE. No specific benchmark datasets are mentioned; the focus is on demonstrating the system's capabilities with a custom, albeit representative, dataset.

Key Insights and Ideas

  • Efficiency through Focus: The most significant insight is that aiming for complete natural language understanding is often unnecessary and inefficient for specific tasks. By defining a clear goal, systems can be made much more performant.
  • Many-to-One Mapping: The paper highlights that different natural language expressions can map to the same formal representation if they serve the same goal. This simplifies the parsing process.
  • Goal Dictates Meaning: The same sentence can have different meanings depending on the user's underlying goal. This is a fundamental principle of goal-oriented parsing.
  • Vocabulary Design is Crucial: The authors note that simply increasing the size of the vocabulary doesn't always improve understanding capabilities; the content and organization of the vocabulary are key.
  • Heuristics for Ambiguity: Resolving ambiguities in natural language queries is critical and is handled through a base of heuristic rules, often invoked in a pattern-directed manner.

Key Results

The paper focuses on the architecture and methodology of NLI-2, illustrating its functionality through several parsing examples. While specific quantitative results (like accuracy percentages or speedup figures) are not explicitly detailed, the paper implies significant efficiency gains due to the goal-oriented approach. The examples demonstrate the system's ability to correctly translate Italian queries into QBE, handling various complexities like identifying relations, domains, and logical connections. The paper also shows how the system can generate diagnostic messages when parsing fails, aiding in debugging and system tuning.

Limitations and Future Work

The authors acknowledge that the current system's vocabulary is strictly tied to the particular application. This rigid bound to the application's purpose is a direct consequence of the goal-oriented understanding concept.

Promising directions for future research include:

  • Developing system architectures that allow for more flexible handling of the comprehension purpose and domain.
  • Defining performance evaluation criteria for measuring the understanding capabilities of natural language systems.
  • Evaluating the complexity of parsing algorithms in relation to input sentence complexity and comprehension scope.
  • Investigating the system's flexibility with small variations in the data base structure, content, and the natural language used.

Summary updated

BibTeX
@inproceedings{guida1980goal,
  author = {Giovanni Guida},
  title = {{Goal Oriented Parsing: Improving the Efficiency of Natural Language Access to Relational Data Bases}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1980},
}
Figure