USCDataScience/parser-indexer-py

Support cross-sentence relation extraction with jSRE

Opened this issue · 1 comments

wkiri commented

We discussed trying a simple approach to enable jSRE to propose cross-sentence relations. Given a window parameterized by k, this approach will concatenate the current sentence +/- k sentences into a pseudo-sentence by removing terminal punctuation and adding a semi-colon, or the word "and", to join sentences together. (May also need to lower-case the first word in the concatenated sentence?)

Then, entities that are candidates for a jSRE relation can be gathered from the entire pseudo-sentence instead of just the current sentence. We are interested in comparing this approach to Yuan's cross-sentence relation extraction methods.

wkiri commented

One idea is to create a superclass like RelationExtractionParser to handle the "k" window argument for simple concatenation.