Explaining and Arguing With Examples

Edwina L. Rissland, Eduardo M. Valcarce, Kevin D. AshleyAAAI 1984

This 1984 research from the University of Massachusetts explores a fundamental principle of human cognition: examples—both real and hypothetical—are the backbone of effective explanation and argumentation. The authors present two pioneering AI systems that move beyond abstract rules to provide concrete illustrations: an intelligent on-line HELP facility for software users and the HYPO system, which uses "dimensional analysis" to generate legal hypotheticals in trade secret law.

The Problem: The Rigidity of Abstract Systems

The authors address the limitations of early automated systems in two distinct domains:

  1. On-line Help: Most help systems of the era (such as VAX/VMS) provided abstract syntactic descriptions that were difficult for novices to parse. They lacked concrete illustrations and failed to account for the user's specific context or expertise level.
  2. Legal Argumentation: Legal reasoning is rarely a simple application of rules; it involves comparing cases. However, the pool of litigated cases is finite. To test the "rule-like" propositions of law, attorneys need to generate hypotheticals ("hypos") that probe the boundaries of legal concepts.

The core challenge identified is how a system can intelligently retrieve, modify, and present examples to satisfy a user's need for clarity or a lawyer's need for a stronger argument.

Methodology: Retrieval and Modification

The researchers proposed a "retrieval + modification" approach powered by an Example-Knowledge-Base (EKB).

TEXPLATES for On-line HELP

For software assistance, they introduced "TEXPLATES"—script-like templates that separate the control of a help session from its content. When a user requests help, the system retrieves a template, fills it with text and examples from the EKB, and "personalizes" them. For instance, if a user struggles with a PRINT command, the system might use the user's actual filenames in the example. It also employs a taxonomy of examples (start-up, reference, model, counter-example) to order information based on the user's expertise level.

Dimensional Analysis in HYPO

For legal reasoning, the authors developed HYPO, which represents legal cases as frames. The key innovation is Dimensional Analysis. The system identifies "dimensions"—legal factors that strengthen or weaken a party's position, such as whether a secret was generally known in the industry or the amount of money expended on development.

Figure 1: HYPO's dimensional analysis process, showing the modification of a base case (HYPO-1) into a hypothetical (HYPO-1A).

As shown in Figure 1, HYPO can take a base case (like RCAVICTIM v. SWIPEINC) and "slide" it along these dimensions to create a hypothetical that is more or less favorable to the plaintiff or defendant. In the transition from HYPO-1 to HYPO-1A, the system modifies variables like development time (D-T-E) and money expended (D-M-E) to test how these changes affect the legal standing of the parties, often using existing precedents like JCN Corp. v. Terex as a guide.

Key Insights and Ideas

  • Examples as Cognitive "Hooks": Examples serve as a memorable entry point into abstract information, particularly for beginners who struggle with formal syntax.
  • Taxonomic Ordering: Not all examples are equal. The system uses a hierarchy to ensure a "start-up" example (simple) precedes an "anomalous" one (complex), tailoring the flow of information to the user's progress.
  • The "Slippery Slope" Algorithm: In legal reasoning, a "slippery slope" argument can be modeled by iteratively modifying a case along a dimension. For example, if a secret is shared with one person, it remains a secret; if shared with thousands, the legal protection likely vanishes. HYPO automates this probing of limits.
  • Contextual Intelligence: If a user asks about "PRINT" and then asks for a definition of a "queue," the system provides an example of a print queue rather than a generic data structure definition.

Results and Impact

The authors demonstrated a help system capable of providing pragmatic advice that standard manuals lacked, such as "undo" assistance (e.g., showing how to use DELETE/ENTRY to stop a PRINT job).

In the legal domain, HYPO successfully generated variations of trade secret cases. By modifying the "expenditure" or "knowledge" slots in the case frame, it allowed attorneys to cite specific precedents more effectively or distinguish their current case from unfavorable ones. This demonstrated that AI could assist in the strategic, adversarial nature of legal argument rather than just performing simple database lookups.

Limitations and Future Work

Despite its successes, the research identified several areas for growth:

  • User Modeling: The authors noted that their user modeling was relatively crude (based on directory size) and required more sophistication to accurately gauge expertise.
  • Natural Language Processing: The systems assumed requests were already parsed into keywords rather than handling raw natural language.
  • Dimension Collisions: Modifying a case along one dimension might logically contradict another (e.g., making a secret "generally known" while claiming it is "not learnable elsewhere"). Future work would require better mediation between these conflicting constraints.

Summary updated

BibTeX
@inproceedings{rissland1984explaining,
  author = {Edwina L. Rissland and Eduardo M. Valcarce and Kevin D. Ashley},
  title = {{Explaining and Arguing With Examples}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1984},
}
Figure