ActiveQA — IndicQA Telugu

This repository contains the trained artifacts for an Active Question Answering system.

Architecture

Original Question → Reformulator → Multiple Candidate Questions → QA Environment → Answer/F1 Reward → Selector → Best Reformulation

The reformulator is further optimized using REINFORCE with the QA performance as the reward.

Dataset

ai4bharat/IndicQA

Language: Telugu

Components

Reformulator

Generates alternative formulations of the original question.

QA Environment

Answers the original and reformulated questions using an extractive QA model.

Selector

Selects the candidate question expected to provide the best QA result.

RL

The reformulator receives the improvement in QA F1 as its reward.

Repository

nlpctx/activeqa-indicqa-telugu