Question Answer System

Install

conda env create -f ODQA.yml

activate environment

conda activate ODQA

Pipeline

Information Retrieval - (Re-ranking passage)

Use Quesntion encoder and Passage encoder model of Dense Passage Retrieval which is trained on multi data. Then, use dot-product to perform similarity of embeddings.

Download pretrained model DSP

python Retrieval/download_pretrained.py

Reader (Reading Comprehension)

Download SQuAD v2 dataset

python download_data.py

Fine tune ELECTRA on SQuAD v2

Download pretrained ELECTRA model

python Reader/download_pretrained.py

Training

python Reader/train.py

We trained ELECTRA model on SQuADv2. You can download weights model here

Result

Epoch	Loss	Exact Match
​1	1.1731	0.83
2	0.7471	0.84
3	0.6016	0.85

Demo

python app.py

UI