Avoiding Unwanted Conversational Implicatures in Text and Graphics

Joseph Marks, Ehud ReiterAAAI 1990

Abstract

This paper addresses the challenge of preventing automated systems from generating misleading text or graphics through "conversational implicatures." By applying Gricean maxims to natural language and graphical design, the authors introduce the FN and ANDD systems. These systems utilize preference-based heuristics to ensure that generated outputs do not lead users to infer false relationships or meanings, such as non-existent rankings or groupings.

Introduction

When humans communicate, we follow unwritten rules known as Gricean maxims. If a speaker provides more detail than necessary, the listener often infers that the extra information is significant. This phenomenon, known as a conversational implicature, can lead to misunderstandings when computer systems generate overly specific text or inconsistent graphics. This classic 1990 paper from the AAAI proceedings explores how to prevent AI systems from accidentally "lying" by saying—or showing—too much.

The Problem of Spurious Implicatures

The authors argue that computer systems often fail to account for how users interpret "Manner" and "Quantity." In text, an automated system might generate a perfectly accurate but overly detailed description (e.g., "a 10Mbit/sec packet-switched computer network" instead of "Ethernet"), leading a user to mistakenly think the object is not a standard Ethernet.

In graphics, a similar problem occurs. A system might accidentally group icons together or use varying line thicknesses for no reason, leading a viewer to hunt for a non-existent "hidden meaning" in the layout. Because humans naturally look for patterns—proximity, symmetry, and alignment—a graphic designer must actively manage these "Gestalt" effects.

Methods: The FN and ANDD Systems

The paper introduces two systems designed to avoid these pitfalls:

1. FN (Natural Language Generation)

This system generates "attributive descriptions" of objects using a preference function. It starts with a basic description and incrementally improves it by applying three key rules:

  • No Unnecessary Components: Modifiers are excluded unless required for identification.
  • Lexical Preference: "Basic-level" terms (e.g., "shirt") are preferred over overly specific or general ones.
  • Local Brevity: Concise terms are prioritized over long strings of attributes.

2. ANDD (Automated Network-Diagram Designer)

The ANDD system applies pragmatic logic to graphics in two stages:

  • Expressive Mapping: A rule-based system maps data attributes to graphical properties (shape, color, size). It specifically avoids mapping simple categories to "ordered" properties like size to prevent users from inferring a ranking where none exists.
  • Diagram Instantiation: This stage handles layout, using heuristics to avoid "spurious" perceptual groupings that might imply non-existent relationships between nodes.
Figure 6: A legend and diagram demonstrating the ANDD system's approach to mapping data attributes (vertex type, emphasis) to graphical variables (shape, intensity).

Key Insights

  • Pragmatics is Universal: Grice’s Maxims apply just as much to pixels as they do to words. A "spurious" line thickness in a diagram is essentially a "lie" of Manner.
  • The "Basic Level" in Graphics: Just as "dog" is a more natural category than "mammal," there are "preferred" graphical values. Using a non-preferred value (like a blinking icon) implies a specific reason for that choice.
  • Perceptual Grouping is Mandatory: You cannot "turn off" human vision. Designers must ensure Gestalt principles reinforce, rather than contradict, the data.
Figure 3: A variation of a disk subsystem diagram where node size is used to distinguish component types, which can trigger false implicatures about the relative importance or capacity of the components.

Key Results and Limitations

The authors formally prove that finding the absolute shortest description is NP-Hard, justifying their use of incremental, heuristic-based algorithms. The FN and ANDD systems successfully demonstrated that incorporating "negative constraints" (rules about what not to do) produced outputs that were significantly less confusing than naive generation methods.

However, limitations remain. The ANDD system's heuristic layout might still occasionally produce accidental alignments. Furthermore, the systems rely on an accurate model of user knowledge; if the system's assumptions about what the user knows are incorrect, the "optimized" output may still be misleading. Future work remains in the area of "media coordination"—deciding when to use text versus diagrams.

Summary updated

BibTeX
@inproceedings{marks1990avoiding,
  author = {Joseph Marks and Ehud Reiter},
  title = {{Avoiding Unwanted Conversational Implicatures in Text and Graphics}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure