Natural Language Generation (NLG) is traditionally divided into two distinct stages: the strategic component, which determines what to say, and the tactical component, which determines how to say it. While strategic planning is complex, the tactical phase often presents a significant computational bottleneck. Mapping abstract semantic representations to grammatical surface strings requires navigating massive search spaces within broad-coverage grammars.
In this research, Günter Neumann proposes a method to bypass this complexity by applying Explanation-based Learning (EBL) to automatically extract and utilize efficient "subgrammars" tailored to specific domains, resulting in generation speeds 10–20 times faster than traditional methods.
The Subgrammar Processing Architecture
The core of the proposed system is the Subgrammar Processing Component (SGP), which operates alongside a standard linguistic generator (the Linguistic Core Component, or LCB). The SGP acts as an intelligent cache that stores and generalizes successful grammatical derivations.
As shown in the architecture diagram, the system functions in two primary phases:
- Training Phase: The system processes a semantic input (Minimal Recursion Semantics, or MRS) using the full source grammar. It extracts the resulting derivation tree—the "explanation"—and generalizes it. This generalized template is then indexed in a subgrammar for future use.
- Application Phase: When a new semantic request arrives, the system first queries the subgrammar. If a matching template is found, it is instantiated with new lexical items, producing a result immediately and bypassing the expensive search process of the grammatical processor.
Generalization and Template Retrieval
A key innovation of this approach is the use of Type Subsumption and generalization. Rather than storing "canned text," the system generalizes specific lexical entries into their semantic types. For example, a specific name like "Sandy" is generalized to a generic "Named Entity" type.
By storing templates based on semantic types (as seen in the generalized tree above), a single training instance can generate thousands of variations. For instance, a template derived from "Sandy gives a chair to Kim" can be instantly reused to generate "Noam donates a book to Peter," provided the semantic structures match the generalized template.
Furthermore, the system supports Partial Matching. If a complete sentence template is not available, the SGP can break the semantic input into chunks, retrieve "phrasal templates" for sub-parts of the sentence, and combine them. This ensures the system remains robust even when encountering novel sentence combinations.
Key Results and Performance
The method was evaluated using a broad-coverage HPSG-based English grammar containing over 2,000 lexical entries. Testing on a set of 179 sentences demonstrated a 10 to 20-fold speedup compared to generating from scratch. The average time to index and instantiate a sentence-level template was reduced to approximately one second.
Analysis of the remaining computation time revealed that 70% was spent on "terminal matching"—the final stage of lexical lookup and unification. This suggests that while EBL effectively solves the structural search problem, the "last mile" of selecting exact word forms remains a primary area for future optimization.
Conclusion
By treating tactical generation as the inverse of parsing, Neumann demonstrates that EBL can significantly prune the search space for NLG. This approach provides the speed of a template-based system with the flexibility and correctness of a full-scale grammar. Future work points toward unified models where systems learn to generate language by "listening" to and generalizing from parsed examples, further bridging the gap between natural language understanding and generation.
Summary updated