Reproduce “TIGTEC : Token Importance Guided TExt Counterfactuals” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Milan Bhan, Jean-Noel Vittaut, Nicolas Chesneau, Marie-Jeanne Lesot
TIGTEC generates counterfactual explanations for text classification models by integrating token importance methods with masked language models and a guided search strategy. The approach produces sparse, plausible, diverse, and semantically relevant explanations, achieving over 98% success rate on the IMDB sentiment analysis dataset while producing more fluent text compared to competing methods.
Key insights
- Guiding counterfactual generation with token importance methods leads to more focused and efficient explanations.
- Leveraging a cost function that balances target class probability and semantic distance helps achieve high-quality counterfactuals that are sparse, plausible, and semantically relevant.
- Attention-based token importance, derived directly from transformer models, effectively identifies influential tokens, outperforming random selection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.