Introduction
In the mid-1980s, natural language processing (NLP) faced a significant divide. Researchers were largely split between two methodologies: full syntactic grammars, which were precise but brittle, and "pure" semantic parsing (notably from the Schankian school), which offered robustness at the cost of immense manual labor and difficulty with complex sentence structures. The challenge was to create a parser capable of navigating "real-world" messy language—characterized by misplaced words and varying orders—without the overhead of a massive formal grammar, while still untangling deeply nested clauses.
The SABA System: A Dual Frame Approach
To address this gap, the SABA (Semantic Analyser, Backward Approach) system was developed. At its core is the "Dual Frames" method, which separates a word's semantic requirements from the roles it can play in a sentence.
Elists and Tlists
The system relies on two primary data structures:
- Elists (Complementee Lists): These function as "case frames" for every meaningful term, including nouns, adjectives, and adverbs, not just verbs. They define what a word "wants." For instance, the verb "to bark" would have an Elist requiring a "canine agent."
- Tlists (Complement Lists): These define what a word "can be" based on its context. A noun’s Tlist is often inherited from the preposition preceding it; for example, a noun following "in" might inherit a "SITUATION" role.
A dependency is established only when a Tlist role matches an Elist requirement. This intersection allows the parser to build semantic links based on fit rather than rigid syntactic position.
Fragmentation and Semantic Operations
To handle complexity, SABA employs a Fragmentation Strategy. This bottom-up approach identifies the innermost clause, establishes its semantic dependencies, and then replaces that entire clause with a simplified non-terminal symbol. This iterative "shrinking" continues until the entire structure is resolved.
Furthermore, the system utilizes mathematical frame operations such as Union (to handle ambiguity, like a pronoun that could be a subject or an object), Intersection (for coordinate structures), and Merging (for hierarchical inheritance).
To resolve ambiguities where semantic constraints alone are insufficient, SABA incorporates semantic rules to identify the "subject" based on voice and specific markers rather than just word position.
Datasets and Implementation
The SABA prototype, implemented in MACLISP and ZETALISP, was evaluated on a corpus of approximately 125 French sentences. These sentences were specifically selected to test various levels of linguistic complexity, including nested relative clauses and different verb voices.
Key Results
The system demonstrated a high degree of proficiency in parsing complex French sentences. A notable example processed by the system is: "The big black dog barks furiously when children that he doesn’t know are playing in the garden of the neighbour."
The fragmentation method proved efficient for processing these deeply nested structures. Additionally, by using an intermediate semantic network formalism called SF (Sentence Formalism), the parser remains decoupled from specific output applications, enhancing its portability.
Discussion and Key Insights
The Dual Frames approach offers several advantages over traditional syntactic-heavy models:
- Uniformity: By assigning Elists to all parts of speech, the system creates a more uniform parsing mechanism.
- Robustness: Because it prioritizes semantic "fit," SABA can successfully parse sentences with minor grammatical errors, such as misplaced negations, which would typically cause a syntactic parser to fail.
- Conciseness: The use of inheritance (e.g., a "programmer" inheriting requirements from a general "action" class) significantly reduces the manual labor required to build dictionaries for new domains.
Limitations and Future Work
Despite its successes, the SABA system has limitations. The current research does not address how the system scales to massive, general-purpose vocabularies beyond the 125-sentence test corpus. Several high-level issues remain for future exploration, including lexical and attachment ambiguities, quantifier scope, and complex pronoun reference (anaphora). Finally, while the underlying principles are general, the specific "Subject Identification Rules" were tuned for French and would require adaptation for languages with different morphological or syntactic structures.
Summary updated