# Question Answering Survey: Directions, Challenges, Datasets, Evaluation Matrices

**Authors:** Hariom A. Pandya, Brijesh S. Bhatt  
**arXiv:** [2112.03572](https://arxiv.org/abs/2112.03572)  
**Date:** 2021-12-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.03572v1/

## Key points

- The internet has led to an increase in available information, requiring automated answering systems
- Question-Answering (QA) is used to provide relevant answers using Natural Language Understanding (NLU)
- QA involves mapping user questions, retrieving relevant information, and finding the best answer
- Deep learning models have shown significant improvements in QA tasks
- Open challenges include automatic question generation, similarity detection, and low resource availability for language processing
- State-of-the-art models on QA datasets are evaluated based on performance metrics such as F1 score and EM score
- Approaches used by researchers include pre-training BERT or GPT models without feature extraction or using Bidirectional Long Short Term Memory (BLSTM) networks with embedding layers for end-to-end training
- Hybrid Question Answering requires multiple semantic clues to constrain the answer set for complex questions
- QA can be divided into Raw Text-Based QA and Knowledge-Based QA (KBQA)
- Deep learning techniques are gaining popularity for resource-rich languages in QA research but still far off for low-resource languages where rule-based or machine learning approaches prevail.

## Summary

The past decade has seen a significant increase in the usage and amount of information available on the internet, leading to the need for automated answering systems that can extract useful information from redundant and transitional knowledge sources. These systems rely heavily on the field of Question-Answering (QA) to provide users with relevant answers using Natural Language Understanding (NLU). QA involves several steps, including mapping user questions to pertinent queries, retrieving relevant information, and finding the best suitable answer from retrieved data. Recent improvements in deep learning models have shown compelling performance improvement in all these tasks. This review work analyzes research directions in QA based on question type, answer type, source of evidence-answer, and modeling approach. The paper also highlights open challenges such as automatic question generation, similarity detection, and low resource availability for language processing. The authors present a survey of available datasets and evaluation measures. State-of-the-art models on QA datasets are evaluated based on their performance metrics such as F1 score and EM score. The authors discuss various approaches used by researchers such as pre-training BERT or GPT models without feature extraction or using Bidirectional Long Short Term Memory (BLSTM) networks with embedding layers for end-to-end training. Hybrid Question Answering is another area where multiple semantic clues are required to constrain the answer set for complex questions. Researchers have proposed various methods like generating multiple query graphs for a given question or dividing the model into two parts: question interpretation and answer inference. Based on storage where we look for an answer, we can divide QA further into Raw Text-Based QA and Knowledge-Based QA (KBQA). While humans can easily detect an answer paragraph or sentence from a given passage, machines struggle with this task due to reasoning disparities between humans and machines. Overall, while deep learning techniques are gaining popularity for resource-rich languages in QA research; it is far off for low-resource languages where rule-based or machine learning approaches still prevail.