Authors: Mohammad N.S. Jahromi, Satya. M. Muddamsetty, Asta Sofie Stage Jarlner, Anna Murphy Høgenhaug, Thomas Gammeltoft-Hansen, Thomas B. Moeslund

Published on: February 05, 2024

Impact Score: 8.2

Arxiv code: Arxiv:2402.03043

Summary

• What is new: The extension of the SIDU XAI method to the text domain (SIDU-TXT) for better understanding of ‘black-box’ models using heatmaps at a word-based level.
• Why this is important: The challenge of making AI’s decision-making process transparent, especially in the text domain.
• What the research proposes: SIDU-TXT employs feature activation maps to elucidate how textual elements contribute to model predictions.
• Results: In sentiment analysis of movie reviews, SIDU-TXT showed superior performance over existing benchmarks. In legal text analysis, it showed promise, though it didn’t fully meet expert expectations.

Technical Details

Technological frameworks used: SIDU-TXT, a comprehensive evaluation framework consisting of Functionally-Grounded, Human-Grounded, and Application-Grounded assessments.

Models used: Comparison models include Grad-CAM and LIME.

Data used: Movie review dataset for sentiment analysis; legal text dataset for asylum decision-making.

Potential Impact

Legal tech firms, sentiment analysis services, AI-powered content moderation platforms, and any sector requiring transparent AI decision-making.

Want to implement this idea in a business?

We have generated a startup concept here: ClearText Insights.