Advanced NLP major project: Open domain question answering on WikiQA corpus
In this project, we study the problem of Answer Selection, a subproblem in Open Domain Question Answering using the WikiQA dataset. We implement three methods on the problem. Our literature survey, analysis and findings can be found under the reports folder as well as the presentation link shared above.
To try out the code here:
This notebook contains the implementation of Attentive Pooling Networks using Bi-LSTMs and CNNs. This architecture was proposed in this paper for computation of question and answer representations with mutual influence in each.
A simple BERT finetuning beats the baseline established by Attentive Pooling Networks. This notebook can be used to try the BERT-based modeling of the problem.
The Transfer AND Adapt (TANDA) strategy, which is currently a SoTA on the problem, proposed a 2-stage finetuning on top of a pretrained architecture like BERT: