Introduction
In the early 1980s, Natural Language Generation (NLG) systems began to demonstrate the ability to answer complex, comparative questions about structured data. Systems like TEXT could explain the differences between entities—for instance, distinguishing a cruiser from a destroyer. However, these systems were limited by a significant bottleneck: they required highly detailed, manually authored knowledge representations (KR) to provide human-like descriptions. Standard database models, such as the Entity-Relationship model, lacked the "saliency" necessary for natural conversation; they could store data but could not identify which attributes were most relevant for a description.
Kathleen McCoy’s ENHANCE system was developed to address this portability problem. By automatically augmenting database schemas with semantic hierarchies, ENHANCE bridges the gap between raw data and the rich knowledge structures required for sophisticated NLG.
The ENHANCE System: Automated Knowledge Augmentation
ENHANCE serves as a pre-processor for the TEXT generation system. Its primary function is to transform a static database into a dynamic generalization hierarchy. It achieves this by applying a tiered set of "World Knowledge Axioms" that guide the discovery of sub-classes and their distinguishing features.
World Knowledge Axioms
The system utilizes three levels of axioms to navigate the data:
- Very Specific Axioms: These handle domain-specific logic not explicitly defined in the schema. For example, an axiom might specify that if a hull number begins with "CV," the entity is an Aircraft Carrier.
- Specific Axioms: These identify "important attributes" within a domain (e.g., displacement or length for naval vessels). These axioms suggest which data paths are likely to yield meaningful comparisons.
- General Axioms: These are domain-independent rules designed to maintain the integrity of the hierarchy. A typical general axiom prevents the system from creating "meaningless" categories, such as a sub-class containing only a single individual.
Methodology: Distinguishing Descriptive Attributes
A core component of the ENHANCE system is the calculation of Distinguishing Descriptive Attributes (DDAs). DDAs represent the minimal set of features that uniquely identify a sub-class relative to its "siblings" in the hierarchy.
Because finding the optimal set of DDAs is a combinatorial challenge, ENHANCE employs a pre-processor to identify "potential-DDAs." This heuristic approach prevents the system from checking every possible attribute combination, which would otherwise lead to exponential processing times. Additionally, the system utilizes a "CLASS-OTHER" strategy to group outliers. By aggregating unique instances that do not fit a broader trend into a single category, ENHANCE maintains a clean and navigable knowledge representation.
Empirical Evaluation
The system was implemented and tested using a subset of an Office of Naval Research (ONR) database. The dataset consisted of 157 records covering vehicles (ships, submarines) and destructive devices (missiles, torpedoes). ENHANCE successfully generated sub-type information for seven primary entities within this scope.
Results and Performance
Qualitative analysis showed that the TEXT system, when powered by ENHANCE-generated data, produced nuanced and accurate natural language descriptions. For example, the system could identify that an aircraft carrier has a "greater length than all other ships" and a "greater displacement than most other ships."
Computationally, the process was intensive. On a VAX 11/780, processing the 157-record dataset required approximately 159,157 CPU seconds. However, because this is a one-time batch process performed before the NLG system goes live, the author argues that the trade-off is acceptable for the benefit of real-time interactive query support.
Discussion and Future Work
The ENHANCE system demonstrates that saliency is not an intrinsic property of an attribute but is contextual. While a database might track fuel capacity, displacement is often more salient for a general description. By deriving these hierarchies from the data itself, ENHANCE significantly improves the portability of NLG systems, allowing them to be moved to new domains without manual KR reconstruction.
Despite its success, the system has limitations. At the time of development, ENHANCE primarily relied on attribute values. Future work suggests incorporating relational data—such as classifying missiles based on the vehicles that carry them—to create even richer hierarchies. Furthermore, scaling the DDA discovery process for massive databases remains a challenge, necessitating more advanced heuristic pruning.
Conclusion
ENHANCE represents an early but pivotal step in automating the semantic enrichment of databases. By transforming raw records into structured, salient knowledge representations, it reduces the manual labor involved in NLG deployment and paves the way for more flexible, data-driven AI communication.
Summary updated