Abstract: Computer-based generation of natural language requires consideration of two different types of problems: 1) determining the content and textual shape of what is to be said, and 2) transforming that message into English. A computational solution to the problems of deciding what to say and how to organize it effectively is proposed that relies on an interaction between structural and semantic processes. Schemas, which encode aspects of discourse structure, are used to guide the generation process. A focusing mechanism monitors the use of the schemas, providing constraints on what can be said at any point. These mechanisms have been implemented as part of a generation method within the context of a natural language database system, addressing the specific problem of responding to questions about database structure.
Introduction
In the early 1980s, natural language research primarily focused on the "how" of language—the syntactic and lexical choices needed to turn a pre-defined message into a sentence. Kathleen R. McKeown’s 1982 paper identifies a critical gap in the "what": how a system decides which facts from a massive database are worth mentioning and in what order they should be presented to ensure a cohesive narrative. The TEXT system was developed to address these "global" problems of content selection and paragraph-level organization, moving beyond simple sentence construction to strategic discourse generation.
Methodology
The TEXT system employs a multi-stage pipeline to bridge the gap between structured data and human-readable paragraphs:
- Relevant Knowledge Pool: When a query is received, the system "circumscribes" a subset of the database. It selects entities and attributes closely related to the query object within the hierarchy, ensuring the generated response is focused and relevant.
- Rhetorical Schemas: This core innovation involves encoding common human explanatory patterns into templates. These schemas (e.g., Identification or Constituency) suggest a sequence of "rhetorical predicates" such as analogy, constituency, or attributive descriptions.
- Focusing Mechanism: Based on the work of Candy Sidner, this mechanism maintains discourse coherency. It uses rules to decide which piece of information to discuss next based on the preceding context, prioritizing "legal moves" like staying on the current topic or shifting to a newly introduced item.
- Tactical Component: Finally, the structured message is passed to a component using Martin Kay’s functional grammar to produce the final English text.
Knowledge Representation and Datasets
The system was implemented and tested using a portion of the Office of Naval Research (ONR) database. This dataset contained structured information about vehicles (ships, submarines, aircraft) and destructive devices (torpedoes, missiles), along with attributes like speed indices, travel means, and lethal indices.
As shown in Figure 1, the database is organized hierarchically. The TEXT system uses this structure to determine the "density" of information available for a given topic, which in turn influences schema selection.
Key Insights and Results
McKeown’s research yielded several significant insights into computational discourse:
- Density-Driven Selection: The system chooses a schema based on information availability. If the system possesses extensive knowledge about an object's sub-types, it utilizes a "Constituency Schema"; otherwise, it defaults to an "Identification Schema."
- Efficiency of Focus Shifting: The system posits that human discourse prefers shifting focus to a newly introduced item immediately if more information is available, avoiding the cognitive overhead of re-introducing the topic later.
- Domain Independence: While the specific facts are domain-dependent (e.g., naval vessels), the rhetorical predicates are universal. This suggests the TEXT architecture is portable to other domains, such as medical or financial systems.
Qualitative results demonstrated that the system could successfully define terms, compare entities (e.g., explaining the difference between a "whisky" and a "kitty hawk" submarine/carrier), and describe broad categories. The generated text showed a clear logical progression, using pronouns and transitions that felt significantly more natural than contemporary systems.
Limitations and Future Work
Despite its success, the TEXT system had notable limitations. It assumed a "static, casual, naive user," lacking the ability to adapt detail levels for experts. Furthermore, it was primarily designed for "what is" questions, struggling with "why" questions or process-oriented queries. Future work aimed to address presuppositional failures, where a user might ask a question based on a false premise (e.g., asking about the "wings" of a submarine), requiring the system to detect and correct these errors gracefully.
Summary updated