I'm trying to build a Shapley value (marginal contribution) of a text document in terms of information content, given that there are several documents on a given topic.

For example, we have 3 reports describing the ocean:

A: {The ocean is blue.}

B: {The ocean is blue and salty.}

C: {The ocean is salty and a home for fish}

If I'm correct, a first step in the algorithm is identifying a set of features (individual pieces of information): {blue, salty, a home for fish}

Moreover, say each feature has an equal weight.

Then I can compute Shapley values of 1/6 for A, 1/3 for B, 1/2 for C.

I wonder if there are well-established algorithms / papers that deal with this problem. I'm coming from economics myself (I'm familiar with Shapley values abstractly), so apologies if the question is too trivial.

Thank you!


From an NLP perspective, the main issue is to identify and extract the terms/relations of interest from the text.

It may be easy if the sentences always appear like these simple examples: one can use pattern matching "the <entity> is <property1> [and <property2>]" (no need for ML).

However with general text it's a quite complex problem which would probably require training an ad-hoc model from annotated data. This would generally involve the tasks of Named Entity Recognition and Relation Extraction.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.