braindatalab / gecobench

NLP Benchmark for XAI methods
BSD 3-Clause "New" or "Revised" License
3 stars 1 forks source link

We have a normalization strategy for computing word-based most common feature attribution #47

Closed rickwg closed 6 months ago

rickwg commented 7 months ago

Currently, we visualize a word-based most common feature attribution, where we normalize the cumulated or mean attribution by word frequency to avoid high-frequency characters showing up in a top-5 ranking. However, this approach tends to emphasize rare words like names. To circumvent this behavior other normalization strategies like tf-idf might be better suited.