TL;DR
This paper introduces a novel human-centric evaluation platform designed to assess the "helpfulness" of explanations generated by AI systems for Knowledge Graph Completion (KGC) tasks. By moving beyond traditional AI-centric evaluation, this platform aims to understand how well these explanations actually aid human users in judging the correctness of AI predictions, a crucial step for building trustworthy AI.
What Problem Does This Paper Address?
Knowledge Graphs (KGs) are powerful tools for representing structured knowledge, but they are often incomplete. Knowledge Graph Completion (KGC) aims to fill these gaps by predicting missing relationships between entities. However, KGC models, especially those based on embeddings, are often "black boxes," making it hard for users to understand why a prediction was made. This lack of transparency hinders trust, particularly in high-stakes domains like medicine.
Explainable KGC (XKGC) methods aim to provide these explanations. While many XKGC methods exist, there's a significant gap in evaluating how helpful these explanations are to human users. Existing human evaluation paradigms often ask annotators to simulate AI behavior, which is difficult and doesn't reflect real human-AI interaction. This paper addresses this by proposing a platform that directly measures how well explanations help users assess the correctness of KGC predictions, aligning with how humans naturally interact with AI.
Methods
The core contribution is a human-centric evaluation platform for XKGC. The platform offers a two-sided interface:
- Researcher Interface: Allows researchers to upload KGC predictions and their corresponding explanations (in JSON format). They can then filter and select specific predictions and explanations for evaluation, and preview the test setup.
- Tester Interface: Presents predictions and explanations to human testers. Crucially, explanations are shown in two formats: a table and a visual graph representation. Testers are asked to:
- Assess the correctness of the prediction based on the explanations.
- Rate their confidence in their assessment.
- Indicate if the explanation was "Helpful" (via checkboxes or clicking on graph edges).
The platform collects metrics like assessment accuracy, confidence, number of helpful explanations, and time cost. It also includes statistical analysis tools (power analysis, hypothesis testing, mixed-effect analysis, correlation analysis) to derive reliable conclusions from the collected human feedback.
Datasets
The paper uses the kinship dataset (Kok and Domingos, 2007) for their evaluations. This dataset was chosen because it is "easily human understandable" and involves challenges like multiple relations and 1:n relationships, making it suitable for evaluating explanations for lay testers.
Key Insights and Ideas
- Human-AI Collaboration is Key: The platform emphasizes a paradigm where AI facilitates human understanding, rather than expecting humans to perfectly simulate AI. This is a more realistic and scalable approach to human evaluation.
- Visualizations Enhance Understanding: Presenting explanations in both tabular and graphical formats helps users comprehend relationships and reason about predictions more effectively.
- "Path" Explanations are Meaningful: Users tend to look for "paths" that connect the predicted triple's entities to the explanation's entities. Explanations that form meaningful connections (like a "triangle") are more helpful.
- Less is More (Sometimes): Testers often find a small set of explanations (2-3) helpful, and a large number can lead to confusion.
- Beyond Simple Explanations: Existing methods (like Method A) might not provide the depth of information needed. Explanations that link more entities (e.g., four entities) can be more informative.
- Objective Quantification of Helpfulness: The platform's metrics (accuracy, confidence, helpExpl) provide a more objective way to measure the quality of explanations compared to subjective ratings alone.
Key Results
The paper conducted two evaluations:
- User Needs Interviews: Revealed that users seek "paths" connecting predicted triples to explanations and can be overwhelmed by too many explanations.
- XKGC Method Comparison: Compared two XKGC methods (Method A and Method B). Method B significantly outperformed Method A across all four metrics:
- Helpful Explanations: Method B generated more helpful explanations (1.64 vs. 0.55).
- Assessment Accuracy: Method B led to enhanced accuracy in testers' assessments (83% vs. 35%).
- The paper concludes that Method B generates more helpful explanations for human testers in the context of kinship predictions.
Limitations and Future Work
- Dataset Domain Specificity: While the kinship dataset is good for general understanding, the authors note that for domain-specific XKGCs (e.g., medical diagnosis), specialists should be employed as testers.
- Tester Engagement: The paper acknowledges that tester engagement can vary and suggests strategies like random checkpoints and time constraints to mitigate this.
- Future Work: The platform is designed to be a general tool. The authors suggest it can be replicated to design evaluations for other use cases beyond KGC, and that the collected data can inform the development of more human-centered XKGC methods. The platform's architecture is flexible and can integrate with other systems.
Summary updated