This classic 1980 paper from the AAAI proceedings explores the "missing link" between how a computer understands the structure of a sentence (syntax) and its actual meaning (semantics). Authors Robert J. Bobrow and Bonnie L. Webber introduce the RUS framework and its implementation, PSI-KLONE, which treats language processing as a continuous, interactive dialogue between the parser and the knowledge base.
TL;DR
The paper proposes "Incremental Description Refinement" (IDR), a method where syntactic parsing and semantic interpretation happen in parallel rather than in sequence. By using a structured inheritance network (KL-ONE), the system uses partial semantic meaning to prune impossible grammatical paths, achieving the speed of specialized "semantic grammars" without sacrificing the flexibility of general-purpose linguistic rules.
What Problem Does This Paper Address?
In the early days of Natural Language Processing (NLP), researchers faced a dilemma. You could build a semantic grammar, which was fast but brittle because it was hard-coded for specific domains. Alternatively, you could build a modular system that separated syntax and semantics, which was flexible but often slow because the parser would waste time exploring grammatically correct but nonsensical interpretations.
Bobrow and Webber tackle the challenge of how to integrate these two worlds. They ask: How can we use our knowledge of the world to guide the parser in real-time, word-by-word, so it doesn't get lost in a sea of ambiguity?
Methods: The RUS Architecture
The authors introduce a "cascaded" architecture called RUS.
- The Cascade: As the parser (an Augmented Transition Network or ATN) scans a sentence from left to right, it doesn't wait until the end to think about meaning. Instead, every time it identifies a potential part of a sentence (like a subject or an object), it sends a "transmission" to a semantic interpreter.
- PSI-KLONE: This implementation uses the KL-ONE formalism, a "frame-like" knowledge representation where concepts are arranged in a taxonomic lattice (a hierarchy with multiple inheritance).
- Incremental Description Refinement (IDR): This is the core algorithm. The system views the knowledge base as a "discrimination net." Each new piece of information from the parser allows the interpreter to move down the lattice, narrowing down the set of possible meanings.
Key Insights: The Power of the Lattice
The most important "aha moment" is that semantics should provide feedback to syntax. If the interpreter says "a brick cannot be the Agent of a hitting action," the parser can stop pursuing that specific grammatical branch immediately.
By using a taxonomic lattice instead of a simple tree, the system can handle "order independence." It doesn't matter if it learns about the subject or the object first; both pieces of information constrain the possible "description" of the event. For example, the system distinguishes between different senses of the verb "run" based on whether the object is a "machine" or a "race."
Key Results
- Efficiency: The system achieved parsing times of approximately 0.25 CPU seconds for 20-word sentences on a DEC KL10, which was highly performant for the era.
- Reduced Backtracking: By using semantic "rejections," the authors significantly reduced the amount of backtracking the ATN parser had to perform, allowing it to approach the speed of "deterministic" parsers.
- Flexibility: Unlike previous semantic grammars, the RUS framework proved robust across different domains because the linguistic rules remained separate from the domain-specific knowledge.
Limitations and Future Work
The authors admit the system struggles with metonymy (e.g., "The hamburger is getting impatient"), as a food item cannot feel impatience. Furthermore, the implementation of KL-ONE at the time did not always distinguish between exhaustive and non-exhaustive lists of possibilities, which is crucial for logical inference. Future work was suggested to integrate "discourse inference" to understand how sentences relate to each other in a broader conversation.
Summary updated