Abstract
The research described here is aimed at unresolved problems in both natural language generation and natural language interfaces to database systems. How relevant information is selected and then organized for the generation of responses to questions about database structure is examined. Due to limited space, this paper reports on only one method of explanation, called "compare and contrast". In particular, it describes a specific constraint on relevancy and organization that can be used for this response type.
Introduction
In the early development of natural language interfaces (NLIs) to databases, systems were primarily designed to retrieve specific data points—such as listing employees in a particular department. However, these systems often failed when users asked "meta-questions" about the database structure itself. Without an understanding of the underlying schema, users could not formulate effective queries. Kathleen McKeown’s 1980 paper addresses this gap by focusing on the "strategic" component of Natural Language Generation (NLG): the process of deciding what information is relevant to a user's query and how that information should be structured to provide a coherent explanation.
The Strategic vs. Tactical Divide
McKeown distinguishes between two primary components of NLG:
- The Tactical Component: This involves the linguistic mapping of formal representations into grammatically correct sentences. Most research at the time focused here.
- The Strategic Component: This involves content selection and text organization. It determines the "shape" of the message before a single word is chosen.
This paper argues that generating multi-sentence responses requires a computational strategy that goes beyond simple template filling, particularly when responding to comparative questions like, "What is the difference between a part-time and a full-time student?"
Methodology: Conceptual Closeness
The proposed system utilizes an augmented database schema containing a generalization hierarchy (a taxonomy of "is-a" relationships). The strategic component operates in two distinct phases based on a metric called conceptual closeness.
1. Selection of Relevant Information
The system determines how closely related two entities are within the hierarchy.
- Close Entities: If two items share a common immediate parent (e.g., two types of graduate students), the system identifies shared attributes and specific differences.
- Distant Entities: If items are in fundamentally different branches of the hierarchy (e.g., a raven and a writing desk), the system identifies the high-level class differences (animate vs. inanimate) rather than attempting to compare irrelevant attributes like "number of legs."
2. Text Organization
Once the information is selected, the system chooses an organizational outline:
- Similarities-then-Differences: For close items, the system first validates the user’s likely intuition that the items are related by stating their commonalities, followed by their distinctions.
- Separate Descriptions: For distant items, the system provides independent descriptions that emphasize their membership in incompatible classes, clarifying why a direct comparison of attributes would be nonsensical.
Key Insights and Qualitative Results
A central insight of the paper is that a good explainer must account for user assumptions. When a user asks for a difference, they often assume a similarity exists. By first addressing these shared traits, the system provides a helpful context for the subsequent differences.
The paper demonstrates this through qualitative examples:
- Close Comparison: When comparing "part-time" and "full-time" students, the system identifies both as graduate students before contrasting their course loads and employment status.
- Distant Comparison: Addressing the Lewis Carroll riddle of the "raven vs. writing desk," the system avoids a literal attribute comparison. Instead, it explains that one is a bird and the other is furniture, concluding they are fundamentally different types of things.
Limitations and Future Work
As an early-stage implementation, the 1980 system had several limitations. It did not address the "tactical" issues of lexical choice or complex syntax, nor did it account for the user's existing knowledge or the preceding discourse. McKeown noted that future iterations would need to incorporate context awareness to avoid explaining concepts the user already understands. Furthermore, while this paper focuses on "compare and contrast," other expository methods—such as analogy or illustration by example—remain areas for further research.
Conclusion
McKeown’s work laid the groundwork for modern NLG by framing text generation as a structural and strategic problem. By using conceptual closeness to dictate both content and organization, the system moved beyond simple data retrieval toward true automated explanation.
Summary updated