Beyond Domain-Independence: Experience With the Development of a German Language Access System to Highly Diverse Background Systems

Wolfgang Hoeppner, Thomas Christaller, Heinz Marburger, Katharina Morik, Bernhard Nebel, Mike O'Leary, Wolfgang Wahlster

Published in the early 1980s, this paper introduces HAM-ANS (HAMburg Application-oriented Natural language System), a pioneering German-language interface developed at the University of Hamburg. While contemporary research focused on "domain independence"—the ability to discuss different topics—the HAM-ANS team pursued "application independence." This concept refers to a single core linguistic engine's ability to interface with fundamentally different background systems, ranging from real-time computer vision to massive relational databases, while adapting its communicative persona accordingly.

The Challenge of Functional Diversity

In the early era of Natural Language Processing (NLP), interfaces were typically hard-coded for specific tasks. The authors identified that true versatility requires more than just a broad vocabulary; it requires functional diversity. For instance, a hotel reservation system must be persuasive and "interest-based," whereas a database interface must be strictly cooperative and efficient.

HAM-ANS was designed to prove that a modular core could handle these varying communicative roles and technical backends—such as fuzzy logic expert systems or structured SQL-like databases—without requiring a total rebuild for each new application.

Modular Architecture and Semantic Representation

The system's architecture is centered on a dual-layered semantic representation that bridges the gap between human language and machine logic:

  • SURF (Surface-oriented): This layer remains close to natural language structure, capturing linguistic nuances like tense, voice, and initial quantifier ordering. It serves as the primary target for the parser and the source for the generator.
  • DEEP (Deep-level): This layer is optimized for the background system. It resolves ambiguities, such as quantifier scope, and transforms linguistic concepts into specific queries the application can execute.
The modular architecture of HAM-ANS, illustrating the flow from NL-Input through the SURF and DEEP representation layers to NL-Output.

Key Technical Strategies

  1. Generate and Test Strategy: To interpret requests, the system generates potential internal identifiers for objects and "tests" them against the background data to verify if they satisfy the user's description.
  2. ATN Parsing: Augmented Transition Networks (ATNs) are used for syntactic analysis. The system switches between case-driven and surface-grammar strategies depending on the available domain knowledge.
  3. Anticipation Feedback Loop: Before generating a response, the system simulates how its own analysis component would perceive that output to ensure clarity and minimize ambiguity.

Evaluation Across Diverse Application Classes

Rather than using static benchmarks, HAM-ANS was evaluated across three distinct "application classes" that demonstrated its flexibility:

  1. Hotel Reservation (Class I): Interfaced with a LISP/FUZZY expert system. The system acted as a manager in a "competitive" dialogue, using "tendentious descriptions" (hedged adjectives) to present rooms favorably.
  2. Traffic Scene Analysis (Class II): Interfaced with the NAOS/MORIO vision system. It analyzed image sequences of street intersections to answer questions about vehicle movements.
  3. Fishery Database (Class III): Interfaced with a PASCAL/R relational database containing 11MB of expedition data. This required coping with mass data and complex query optimization.
Comparison of the three application classes (Hotel Reservation, Traffic Scene, and Fishery Data) highlighting differences in discourse domain, system behavior, and background architecture.

Key Insights and Results

  • Non-Uniformity as a Strength: Contrary to the trend of seeking a single uniform representation for all processing stages, the authors found that specialized representations for morphology, lexicon, and semantics actually improved performance.
  • Performance: Running on a DECsystem 1070, the system produced responses in 5 to 25 seconds. In 1983, this was considered efficient enough for practical human-computer interaction.
  • Extensibility: Integrating a completely new case system into the core took less than five man-weeks, validating the system's modular design.
  • Handling Vagueness: The system successfully managed vague quantifiers (e.g., "almost every") by tracking "hits" and "misses" during its internal data testing phase.

Conclusion and Future Directions

While HAM-ANS demonstrated significant linguistic breadth—handling anaphora, ellipsis, and complex noun phrases—the authors noted several areas for growth. Future work aimed to support comparative constructions ("the biggest room"), counterfactuals, and more sophisticated "partner models" to better anticipate user knowledge levels. Despite these limitations, HAM-ANS remains a landmark in the transition from brittle, task-specific interfaces to adaptable, multi-application natural language systems.

Summary updated

BibTeX
@inproceedings{hoeppner1983beyond,
  author = {Wolfgang Hoeppner and Thomas Christaller and Heinz Marburger and Katharina Morik and Bernhard Nebel and Mike O'Leary and Wolfgang Wahlster},
  title = {{Beyond Domain-Independence: Experience With the Development of a German Language Access System to Highly Diverse Background Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1983},
}
Figure