?? How to explain the prediction of semantic search?

TxtAI is a python package for semantic tasks. It provides an "explain" function that generates an importance score for each word in the sentence.

But how does it work? Read for more details ??

It follows the occlusion-based explanation approach. The steps for score computation are as follows,

• First, several copies of the same sentence are created, each one with one word/token missing.
• Then cosine similarity is computed between the fixed query and each permutation.
• Finally, the similarity score of the permutation is subtracted from the score of the original sentence (with all the words present).

The intuition is as follows,

• if an important word like?"win"?is missing from the sentence, the score of the sentence will be reduced and the difference will be positive
• if an unimportant word like?"man"?is missing from the sentence, the score of the sentence will increase and the difference will be negative

And that's how you can find the importance of words in the sentence that lead to a higher overall score!

More details: https://mohitmayank.com/a_lazy_data_science_guide/natural_language_processing/text_similarity/#txtai

TxtAI: https://github.com/neuml/txtai/blob/master/examples/32_Model_explainability.ipynb