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

TL;DR

This paper introduces HAM-ANS, a German natural language access system designed to be domain-independent, meaning it can interface with diverse background systems like expert systems, vision systems, and databases. The authors demonstrate that achieving this ambitious goal is feasible by building a core system with a set of diverse applications from the outset, focusing on common knowledge sources and processing units.

What Problem Does This Paper Address?

The paper tackles the challenge of creating truly domain-independent natural language (NL) interfaces. While previous efforts focused on domain independence, they often struggled to adapt to vastly different types of background systems (e.g., CAD, simulation, databases) and varying dialog types, user profiles, and system behaviors. The authors aim to move beyond this by developing a core NL system that can serve as a foundation for interfaces to a wide range of applications, differing not only in their domain but also in their interaction characteristics.

Methods

The core of HAM-ANS is a robust NL dialog system with both depth and breadth, designed for practical real-time interaction. Key methodological aspects include:

  • Core System Design: Building a unified core system with diverse applications integrated from the beginning. This allows for testing modifications and identifying common components, interfaces, and tools.
  • Representation Formalisms: Utilizing logic-oriented semantic representation languages, SURF and DEEP. SURF is designed to be close to natural language utterances, while DEEP is optimized for evaluating utterances against domain-specific knowledge. This dual approach facilitates both parsing and reasoning.
  • Multi-Strategy Parsing: Employing ATN (Augmented Transition Network) parsing with two strategies: case-driven analysis (using specific restrictions for case slot fillers) and surface grammar (handling word-order variations).
  • Ellipsis Recognition: A mechanism to save complete substructures of utterances as SURF representations, enabling the ellipsis recognition component to handle contextual ellipsis by inserting partial semantic representations.
  • Generate-and-Test Interpretation: A core strategy for interpreting DEEP expressions, where "generate" procedures assign object identifiers to variables and "test" procedures evaluate the truth value of instantiated formulas. This iterative process allows for handling complex quantifiers and spatial relations.
  • Cooperative Response Generation: The system is designed to provide extended and cooperative answers, going beyond simple yes/no responses and incorporating only new information.

Datasets

Not explicitly discussed in the paper. The system interacts with background systems like an expert system, a vision system (image sequence analysis system NAOS/MORIO), and a database (fishery database).

Key Insights and Ideas

  • Domain Independence is Achievable (with caveats): The paper demonstrates that building a core NL system adaptable to diverse applications is feasible, though full generality might be a distant goal.
  • Application-Driven Core System Development: Integrating diverse applications from the start is crucial for identifying and refining domain-independent components.
  • Dual Semantic Representation (SURF/DEEP): The distinction between a surface-oriented (SURF) and a deep, logic-oriented (DEEP) representation is a powerful approach for handling both linguistic nuances and semantic reasoning.
  • Cooperative Dialog Behavior: The system's ability to provide extended, informative, and cooperative responses (e.g., modifying quantifiers instead of a flat "no") significantly enhances user experience.
  • Non-Uniformity as a Strength: While uniformity is often a goal in AI, the authors found that a strictly uniform representation schema is unfeasible for diverse NL systems. Instead, they advocate for a core set of uniform formalisms alongside specialized ones where necessary.

Key Results

  • Operational System: HAM-ANS is operational as an interface to an expert system, a vision system, and a database system.
  • Real-time Performance: The system produces typed German responses quickly enough (5-25 seconds on an unloaded POP-10) for practical real-time interaction.
  • Handling Complex Language: The system demonstrates the ability to interpret and generate responses involving complex linguistic phenomena like anaphora resolution, user modeling, extended responses, focus shifting, vague quantifiers, and hedged predicates.
  • Successful Interpretation Example: The paper provides a detailed walkthrough of interpreting a user's question about lamps and beds, showcasing the generate-and-test process and the system's ability to handle spatial relations and quantifiers, ultimately producing a nuanced answer like "Yes, next to almost every one."

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that full domain independence is an ambitious goal and may not be attainable in its complete generality with the current state of the art.
  • While the core system is operational, it is still in the system development stage and not ready for distribution.
  • The paper focuses on the interpretation unit and only briefly sketches analysis and generation.
  • Future Work:
    • Broadening the system's capabilities with a more elaborate partner model.
  • Improving the handling of NL database updates.
  • Developing facilities for explaining the system's speech acts.
  • Enhancing the treatment of word formation.
  • Implementing NL knowledge acquisition.

Summary updated

BibTeX
@inproceedings{1983beyond,
  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