Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Explainable agency: human preferences for simple or complex explanations

Michelle Blom, Ronal Singh, Tim Miller, Liz Sonenberg, Kerry Trentelman, Adam Saulwick

TL;DR

This paper addresses explaining automated causal reasoning by converting proof traces into layered explanations. It introduces three abstraction rules—Flatten_Logic, Flatten_Rules, and Filter_Knowledge—that transform an explanation graph $\

Abstract

Research in cognitive psychology has established that whether people prefer simpler explanations to complex ones is context dependent, but the question of `simple vs. complex' becomes critical when an artificial agent seeks to explain its decisions or predictions to humans. We present a model for abstracting causal reasoning chains for the purpose of explanation. This model uses a set of rules to progressively abstract different types of causal information in causal proof traces. We perform online studies using 123 Amazon MTurk participants and with five industry experts over two domains: maritime patrol and weather prediction. We found participants' satisfaction with generated explanations was based on the consistency of relationships among the causes (coherence) that explain an event; and that the important question is not whether people prefer simple or complex explanations, but what types of causal information are relevant to individuals in specific contexts.

Explainable agency: human preferences for simple or complex explanations

TL;DR

This paper addresses explaining automated causal reasoning by converting proof traces into layered explanations. It introduces three abstraction rules—Flatten_Logic, Flatten_Rules, and Filter_Knowledge—that transform an explanation graph $\

Abstract

Research in cognitive psychology has established that whether people prefer simpler explanations to complex ones is context dependent, but the question of `simple vs. complex' becomes critical when an artificial agent seeks to explain its decisions or predictions to humans. We present a model for abstracting causal reasoning chains for the purpose of explanation. This model uses a set of rules to progressively abstract different types of causal information in causal proof traces. We perform online studies using 123 Amazon MTurk participants and with five industry experts over two domains: maritime patrol and weather prediction. We found participants' satisfaction with generated explanations was based on the consistency of relationships among the causes (coherence) that explain an event; and that the important question is not whether people prefer simple or complex explanations, but what types of causal information are relevant to individuals in specific contexts.
Paper Structure (24 sections, 5 figures, 2 tables)

This paper contains 24 sections, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Explaining Proofs and Proof-based Reasoning
  4. Explanation and the Semantic Web
  5. User perception of complexity in explanations
  6. Problem Definition
  7. Explanation System
  8. Knowledge Types
  9. Explanation Graphs
  10. Simplifying Rules
  11. Study Design
  12. Procedure
  13. Results
  14. Data Analysis
  15. Study 1: MTurk Participants
  16. ...and 9 more sections

Figures (5)

  • Figure 1: Process employed by our explanation system to produce a layered explanation from a proof trace. The proof trace is parsed to extract different kinds of knowledge (Section \ref{['sec:KnowledgeTypes']}), which is then used to form an explanation graph (Section \ref{['sec:FormProofGraph']}). The graph is simplified via the application of a number of rules (Section \ref{['sec:SimplifyingRules']}) to form a layered explanation.
  • Figure 2: Average rating by question for rule combination comparison pairs. The bars on the left (dark blue) are ratings for Explanation 1. The bars on the right (yellow) are ratings for Explanation 2. The title shows Set number and pair type: Set 1 (FR), Set 2 (non-FR), and Set 3 (FR vs non-FR). The y-axis shows the mean rating for each question, with question numbers on the x-axis. Error bars denote one standard deviation. The five questions are outlined in Table \ref{['tab:5Qs']}.
  • Figure 3: FL-FK vs FL. The first graph shows the rating for each of the five questions outlined in Table \ref{['tab:5Qs']}. The centre graph shows whether participants found Explanation 2 to contain more information than Explanation 1. The right graph shows the difference between the ratings given to the two explanations.
  • Figure 4: FL vs no abstraction. The first graph shows the rating for each of the five questions outlined in Table \ref{['tab:5Qs']}. The second graph shows whether participants found Explanation 2 to contain more information than Explanation 1. The last graph shows the difference between the ratings given to the two explanations.
  • Figure 5: FL-FR vs FL-FK. The first graph shows the rating for each of the five questions outlined in Table \ref{['tab:5Qs']}. The centre graph shows whether participants found Explanation 2 to contain more information than Explanation 1. The right graph shows the difference between the ratings given to the two explanations.