hideaki-j
commented
3 years ago - Q. How did they select mentions
- A. Use FFNN with the input of the first and last token representation (see Sect. 3.3)
- Q. Why does their model require so many computational resources?
- (1) The number of questions might be really large
- Also check "Possible for Training on multiple GPUs?" in https://github.com/ShannonAI/CorefQA/issues
- (2) They calculate bi-directional coreference relation (not only P(ei|ej) but also P(ej|ei)) (see Sect. 3.4)
- Q. How did they propagate the gradient?
CorefQA: Coreference Resolution as Query-based Span Prediction
Contribution summary
[Author] did [Method] to solve [Motivation] and found [Insight]
Authors
Wei Wu, Fei Wang, Arianna Yuan, Fei Wu, Jiwei Li (Zhejiang University, Stanford, ShannonAI)
Motivation
Existing method is/has/uses ...
Method
Proposed method is/has/uses ...
Results / Insight