Abstract
This paper presents a deterministic natural language analyzer that translates Italian commands directly into relational database operations. By extending Mitchell Marcus’s "determinism" hypothesis, the system avoids backtracking and intermediate representations, instead using a cooperative frame-based architecture to bridge the gap between syntax and semantics.
Introduction
In the early 1980s, natural language processing (NLP) typically relied on multi-stage pipelines where a syntactic parser built a tree before a semantic component interpreted it. This approach often led to "backtracking"—where the system hits a dead end and must undo previous work—or the creation of complex intermediate structures that did not map easily to computer-readable data. The authors sought to build a system capable of taking an Italian query (e.g., "Give me the value of total bilirubin for patients with jaundice") and translating it immediately into a sequence of algebraic operations, such as SELECT and JOIN, on a relational database. The core challenge was achieving this "deterministically," ensuring the system makes the correct decision at every step without guessing.
Methodology
The system relies on three core components working in tandem: a lookahead buffer, syntactic rule packets (SRP), and a frame-based representation.
Inspired by Mitchell Marcus’s work, the analyzer uses a small buffer of upcoming words to "see into the future" just enough to decide which grammatical rule to apply. Instead of a monolithic list of rules, the grammar is organized into packets containing "pattern-action" rules. A discriminant net identifies which rule matches the current buffer state.
The system instantiates "frames" (such as ACTIONFR or CONCEPTFR) with slots filled during the parse. These frames utilize cooperative processing: check procedures attached to slots can "suspends" themselves if they require information from a slot not yet reached, resuming automatically once the data appears.
Semantic Integration and Insights
A key innovation is the shift from traditional parts of speech (noun, verb) toward semantically biased categories like "concept," "property," and "identifier." This bias smooths the transition from grammar to meaning. The authors define "determinism" as a process where all structures built during the parse "concur" to build the final representation simultaneously.
Furthermore, the system is designed to handle implicit information. For instance, in the phrase "patients with jaundice," the system automatically fills the "value" slot of the jaundice symptom with "present," even if the word is absent from the input. Control flow moves automatically between the lexicon, rules, and semantic checks without a central master algorithm.
Results and Relational Mapping
The system successfully demonstrated surface structure invariance: distinct Italian sentence structures—such as direct commands or questions—can result in the same internal frame representation.
The authors established a clear path from natural language to database logic. Pointers between frames correspond to JOIN operations, while frame restrictions correspond to SELECT operations. Implemented in LISP on a DEC-10 computer, the lexical and syntactic components were fully tested, while the semantic check procedures were in the final stages of debugging at the time of publication.
Limitations and Future Work
While effective for medical consultation, the system's reliance on a specialized lexicon and prototype frames suggests significant manual engineering would be required for other domains. Additionally, while the lookahead buffer proved sufficient for the tested grammar, its performance with more complex or ambiguous linguistic constructions remains to be seen. The complexity of the cooperative knowledge source interaction also presented challenges in stabilizing the procedural semantic checks.
Summary updated