Abstract
Traditional natural language processing (NLP) systems often rely on a "syntax-first" approach, where grammatical structures are built before semantic constraints are applied. This paper challenges that convention, proposing a "semantics-first" method that uses domain-specific world knowledge to drive the parsing process. Empirical evaluations across multiple corpora demonstrate that prioritizing semantic attachments leads to significant gains in processing speed and more favorable scaling as sentence length increases.
Introduction: The Interaction Problem
In early NLP research, the dominant paradigm involved a sequential process: a parser would first construct all possible syntactic interpretations of a sentence and then use semantic information to prune nonsensical results. However, syntactic ambiguity—where a single sentence can have numerous valid but meaningless grammatical structures—makes this approach computationally expensive. Steven L. Lytinen addresses this "interaction problem" by investigating how syntax and semantics can coordinate more efficiently.
Methodology: Flipping the Script
The study compares two versions of the LINK unification-based NLP system:
- Interleaved Syntax-First: A baseline bottom-up chart parser where grammar rules (e.g., NP followed by VP) trigger the search. Semantics are checked only after a rule is proposed.
- Semantics-First: A novel approach that first identifies "desirable semantic attachments" based on domain knowledge. For example, the system asks if a 'CITY' can be the destination of 'DRIVE' before looking for the grammar rules that could realize that connection.
Both methods utilize the same underlying unification grammar, ensuring that the performance differences stem solely from the order of application rather than the depth of linguistic knowledge.
Experimental Setup
The approaches were tested on 50 randomly selected sentences from three corpora:
- Automobile Repair: Terse, real-world symptom descriptions (e.g., "ENGINE QUITS AT CRUISING SPEEDS").
- Assembly Line: Short task descriptions (e.g., "WALK TO FRONT OF JOB").
- Grammatical Auto Repair: A version of the first corpus rewritten into standard English to control for the effects of "telegraphic" input.
Key Results and Scaling
The semantics-first approach consistently outperformed the baseline across all metrics:
- Efficiency: It was approximately 26% faster on average.
- Reduced Search Space: In the auto repair domain, the semantics-first approach required 27% fewer rule applications (46 vs. 63).
- Linear Scaling: While the syntax-first approach's workload grew at a higher-order polynomial rate relative to sentence length, the semantics-first approach exhibited a nearly linear relationship, making it much more scalable for complex input.
Discussion and Limitations
The primary insight is that "efficiency is an algorithmic property." By avoiding the construction of semantically anomalous constituents, the parser bypasses the inherent ambiguity of general grammar. However, the author notes that this approach is most effective in narrow domains where semantic constraints are strong. In broader domains, the complexity of semantic search might increase, potentially narrowing the performance gap between the two strategies. Future work would require more advanced general-purpose semantic theories to handle unrestricted text.
Summary updated