Abstract
This paper presents the control architecture for PHLIQA (PHiLIps Question Answering), a natural language processing system designed to translate English queries into database-executable expressions. The system addresses the inefficiencies of traditional depth-first search and backtracking by introducing a "flexible decision tree strategy." By deferring ambiguous interpretations until sufficient context is available and utilizing symmetric programming in standard algorithmic languages, PHLIQA achieves a more structured and efficient approach to natural language understanding.
Introduction
In the early 1970s, the primary challenge in natural language processing was managing local ambiguities. A single word or phrase could often be interpreted in multiple ways, leading to a combinatorial explosion of possibilities. Traditional AI systems of the era typically relied on depth-first search: picking one interpretation and backtracking upon failure. This often resulted in inefficient "spaghetti code" and high computational overhead. P. Medema of Philips Research Laboratories proposed a more disciplined control structure to handle these complexities within the PHLIQA system.
Methodology
The PHLIQA architecture is built upon three primary innovations designed to streamline the interpretation process.
1. Multilevel Interpretation via Converters
Rather than attempting a direct translation from English to a database query, PHLIQA employs a series of intermediate languages. "Converters" are used to transform tree structures from one level of abstraction to the next, allowing for a modular and manageable translation pipeline.
2. The Flexible Decision Tree Strategy
The core contribution of the paper is a strategy that allows decision points to remain "dangling" rather than forcing an immediate choice. The system utilizes three-valued logic (True, False, Unknown) to evaluate transformation rules. If a rule's applicability is "Unknown" due to missing context, the system proceeds to other parts of the tree. Once neighboring nodes are transformed, they provide the necessary context to re-evaluate and resolve the dangling decisions, significantly reducing unnecessary branching.
3. Symmetric Programming
Medema advocates for the use of standard algorithmic languages (such as ALGOL68 or PL/I) over specialized AI list-processing languages. By employing a top-down design with recursive procedures, the system leverages the hardware's built-in stack mechanism to handle storage and backtracking implicitly. This approach ensures that the code remains readable and efficient for compilers.
Evaluation and Results
The paper demonstrates the efficiency of the flexible strategy through qualitative analysis of specific parse examples.
In a sample parse of the question "Does Philips use computers?" (see Figure 1), the flexible strategy performed the reduction of "COMPUTERS" to "NOUN" only once. In contrast, a conventional left-to-right depth-first parser would have been forced to perform and potentially backtrack through that same transformation six times. This context-driven re-evaluation mimics human problem-solving, where solving one part of a puzzle provides the clues needed for the next.
Furthermore, the modular design using independent converters allowed the system to remain largely independent of the specific database content, enhancing its potential for adaptability.
Limitations and Future Work
Despite its improvements, Medema identifies several areas for future development:
- Linearity: The model still assumes a largely linear ordering of decisions. Future work could explore processing mutually independent chains of decisions in parallel.
- Intelligent Backtracking: The current system retraces every minor step during failure. An "intelligent" system should ideally identify the specific high-level decision that caused the contradiction and jump back to that point directly.
- Scope: The system was developed for a restricted "universe of discourse" (computer configurations) and was not intended as a general-purpose natural language processor.
Conclusion
The PHLIQA control structure represents a significant step toward more efficient natural language interpretation. By prioritizing context and leveraging standard programming paradigms, the system demonstrated that complex AI tasks could be handled with structured, modular, and computationally efficient architectures.
Summary updated