Deduction Automatique Et Systemes Transformationnels

J. ChaucheCOLING 1986

TL;DR

This paper explores the application of "transformational systems" for automated reasoning and knowledge representation, contrasting them with traditional deductive AI approaches. It argues that transformational systems offer a more object-oriented and potentially more powerful way to handle complex data structures and knowledge, particularly in natural language processing.

What Problem Does This Paper Address?

The paper addresses the challenge of building intelligent systems that can perform reasoning and manage knowledge bases, particularly in the context of natural language processing. Traditional AI approaches often rely on deductive systems that operate on formal logic and predicate calculus, which can be rigid and computationally intensive when dealing with the nuances of human language. The authors propose "transformational systems" as an alternative, suggesting they offer a more flexible and object-oriented paradigm for representing and manipulating complex information. The core problem is to demonstrate how these transformational systems can effectively perform tasks like deduction and knowledge representation, drawing parallels and contrasts with established logical systems like Prolog.

Methods

The paper introduces and contrasts two main approaches to AI: deductive systems and transformational systems.

  • Deductive Systems (e.g., Prolog): These systems operate on a formal, predicative representation of language. They rely on logical inference and are procedural, focusing on the structure of predicates. The paper highlights Prolog's use of predicates like append for list manipulation and its bottom-up analysis approach.
  • Transformational Systems (e.g., SYGMART): These systems are object-oriented, focusing on manipulating objects within a defined universe. They are characterized by local transformations and a more global view of the data. The paper emphasizes their ability to handle complex structures and associated properties within "labels" (étiquettes). The core mechanism involves applying transformation rules to modify objects, simulating deduction.

The paper illustrates these concepts through examples of natural language parsing and knowledge base querying, translating Prolog examples into their transformational system equivalents. It also discusses how transformational systems can handle algorithms like sorting and the Towers of Hanoi by defining objects and transformation rules.

Datasets

Not discussed in the paper. The paper uses illustrative examples and hypothetical knowledge bases rather than specific, named datasets for training or evaluation.

Key Insights and Ideas

  • Object-Oriented vs. Predicative Representation: Transformational systems offer an object-oriented perspective, where complex objects with associated properties are manipulated directly. This contrasts with predicative systems that break down information into simpler atoms and predicates.
  • Local vs. Global Transformations: Transformational systems employ local transformations, which can be more manageable and flexible than the global analysis often required in deductive systems.
  • Richness of Labels: In transformational systems like SYGMART, "labels" can encapsulate a significant amount of information about an object, providing a richer representation than individual variables in a predicative system.
  • Determinism in Transformational Systems: Unlike some deductive systems that might explore multiple solution paths (e.g., Prolog's "cut"), transformational systems are inherently deterministic. Any choice or branching logic must be explicitly defined within the transformation rules.
  • Simulating Deduction: The paper argues that transformational rules (A → B) can directly simulate deductive steps, where state B is deduced from state A.
  • Flexibility in Object Manipulation: Transformational systems can operate on diverse and complex object classes, not limited by predefined structures as in some deductive approaches.

Key Results

The paper doesn't present quantitative results in terms of accuracy or performance metrics. Instead, it demonstrates the equivalence and potential advantages of transformational systems through qualitative examples:

  • Natural Language Parsing: The paper shows how a sentence like "the man eats the apple" can be represented and transformed in both Prolog and SYGMART, highlighting the richer information captured by SYGMART's labels.
  • Knowledge Base Querying: An example of a "sister" query in Prolog is translated into a transformational system, demonstrating how knowledge base lookups and deductions can be achieved through label manipulation and rule application.
  • Algorithmic Simulation: The paper outlines how algorithms like bubble sort and the Towers of Hanoi can be expressed using transformational rules, illustrating the system's ability to handle procedural logic.

Limitations and Future Work

  • Complexity of Representation: While transformational systems can handle complex objects, defining these objects and their associated transformations can be intricate. The paper acknowledges that the complexity of a "key" in a knowledge base influences the number of transformations needed.
  • Explicit Control of Choices: The deterministic nature of transformational systems means that any non-deterministic choices or exploration of multiple solutions must be explicitly programmed into the grammar or rules, which can add complexity.
  • External Projection/Inverse Functions: For practical applications, external functions are needed to project data into the system's object representation and to translate the system's output back into a human-readable format. The paper mentions these are often tied to morphological analysis/synthesis in SYGMART for natural language processing.
  • Future Work: The paper implies that further exploration into defining complex object structures and efficient transformation rule sets for various AI tasks, especially in natural language processing, is a promising direction. The SYGMART system itself is presented as a tool for such exploration.

Summary updated

BibTeX
@inproceedings{chauche1986deduction,
  author = {J. Chauche},
  title = {{Deduction Automatique Et Systemes Transformationnels}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1986},
}

Loading related content...

Figure