# Unsupervised multiple choices question answering via universal corpus

**Authors:** Qin Zhang, Hao Ge, Xiaojun Chen, Meng Fang  
**arXiv:** [2402.17333](https://arxiv.org/abs/2402.17333)  
**Date:** 2024-02-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.17333v1/

## Key points

- The paper addresses unsupervised question answering by eliminating the need for large-scale annotated data in new domains.
- Focuses on unsupervised multiple-choice question answering (MCQA) and proposes a framework using synthetic MCQA data from the universal domain.
- Method involves extracting potential answers from context to create related questions and incorporating named entities (NE) and knowledge graphs for plausible distractors.
- Demonstrates effectiveness in generating accurate responses without relying on annotated data through experiments on various MCQA datasets.
- Offers a promising solution for unsupervised question answering tasks by utilizing synthetic data generation techniques and leveraging existing knowledge resources like NE and knowledge graphs.

## Summary

The paper "Unsupervised Multiple Choice Question Answering via Universal Corpus" by Qin Zhang, Hao Ge, Xiaojun Chen, and Meng Fang addresses the challenging task of unsupervised question answering. This approach is crucial as it eliminates the need for large-scale annotated data in new domains. The authors focus on studying the unsupervised multiple-choice question answering (MCQA) problem and propose a novel framework that utilizes synthetic MCQA data generated from contexts in the universal domain without manual annotation. Their method involves extracting potential answers from the given context and using them to create related questions. They also incorporate named entities (NE) and knowledge graphs to identify plausible distractors, resulting in complete synthetic samples for MCQA. Through experiments on various MCQA datasets, the authors demonstrate the effectiveness of their approach in generating accurate responses without relying on annotated data. Overall, this innovative framework offers a promising solution for addressing challenges in unsupervised question answering tasks by utilizing synthetic data generation techniques and leveraging existing knowledge resources such as NE and knowledge graphs. The results highlight its potential in advancing research in MCQA and other related domains.