In tactical natural language generation—the process of transforming a semantic representation into a grammatically correct sentence—traditional approaches often suffer from a lack of flexibility. Frequently, the semantic input is "hard-coded" to match a specific grammar's internal logic, creating a bottleneck when adapting the generator for diverse applications like machine translation (MT). This paper, published in 1991 by Stephan Busemann, proposes a modular approach that decouples syntactic grammar from semantic input through the use of "Pattern-Action" (PA) rules.
System Architecture
The proposed system was designed to be integrated into larger translation frameworks, such as the Berlin Machine Translation system. By separating the linguistic grammar from the application-specific semantic data, the generator becomes more adaptable to different source and target languages.
The architecture follows a clear pipeline: analysis of the source language (SL) leads to a semantic transfer, which then serves as the input for the generation of the target language (TL) sentence.
Methodology: Mapping Semantics to Syntax
The core of Busemann's approach involves bridging two distinct formalisms: Functor Argument Structures (FAS) and Generalized Phrase Structure Grammar (GPSG).
- FAS (Functor Argument Structures): This serves as the semantic input, organized as a tree where nodes contain features such as voice, tense, and thematic roles.
- GPSG (Generalized Phrase Structure Grammar): This is the syntactic output format. It uses a unification-based formalism with Immediate Dominance (ID) rules for parent-child relationships and Linear Precedence (LP) rules for word order.
As shown in Figure 2, the generation process must map the abstract semantic nodes of the FAS expression to the concrete syntactic categories of the GPSG structure.
Pattern-Action (PA) Rules
The "bridge" between these two structures is a set of Pattern-Action rules. The system traverses the semantic FAS tree from the top down. When a rule matches a specific pattern in the FAS tree, its "Action" component either triggers a syntactic ID rule or stores information in an intermediate two-dimensional array. This intermediate storage is crucial for handling non-local information, such as topicalization (e.g., moving an object to the front of a sentence), by allowing rules to pass "notes" across different parts of the tree.
Evaluation and Key Results
The system was implemented and tested within the Berlin Machine Translation system, focusing on administrative texts from the European Community (EG) for English-German translation. The grammar fragments included approximately 50 to 70 ID and PA rules per language.
Key findings from the implementation include:
- Complex Generation: The system successfully generated passive voice and topicalized sentences while maintaining correct morphological agreement (gender, number, case).
- Performance: On an IBM AT, the system generated complex sentences in approximately 4.7 seconds.
- Modularity: The restricted nature of PA rules ensured the system remained transparent and easy to debug, allowing linguists to swap semantic inputs or target grammars without rewriting the entire generator.
Conclusion and Future Directions
Busemann's work demonstrates that decoupling mapping logic from grammar creates a "tactically" multilingual system. However, challenges remain regarding the manual tuning of rule precedence and the scaling of PA rules to massive, unrestricted natural language grammars. Future work aims to explore bidirectional rules that could function for both parsing and generation.
Summary updated