# Answer ranking in Community Question Answering: a deep learning approach

**Authors:** Lucas Valentin  
**arXiv:** [2212.01218](https://arxiv.org/abs/2212.01218)  
**Date:** 2022-10-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.01218v1/

## Key points

- Community Question Answering (CQA) is a subfield of computational linguistics that deals with the analysis of questions and answers posted on community-based websites such as Quora or Stack Overflow.
- Automated methods are required to organize, categorize, and quantify the information due to the sheer volume of posts on these sites.
- One common problem in CQA is ranking multiple answers to a question by how informative they are in solving the original question.
- The authors aimed to advance the state-of-the-art in answer ranking for CQA using deep learning models.
- They created a large dataset of questions and answers from Stack Overflow and used natural language processing techniques such as dense embeddings and Long Short-Term Memory (LSTM) networks to predict the accepted answer attribute.
- Numerical features extracted from metadata found in Stack Overflow posts or derived from text data were used to assist with answer ranking tasks.
- The performance of their deep learning models was compared against ensemble methods such as forest and boosted trees but could not improve upon baseline results.
- New methods need to be developed that can handle out-of-vocabulary words present in programming code snippets more effectively.
- Further research is needed to overcome existing challenges.

## Summary

Community Question Answering (CQA) is a subfield of computational linguistics that deals with the analysis of questions and answers posted on community-based websites such as Quora or Stack Overflow. The sheer volume of posts on these sites requires automated methods to organize, categorize, and quantify the information. One common problem in CQA is ranking multiple answers to a question by how informative they are in solving the original question. In this study, the authors aimed to advance the state-of-the-art in answer ranking for CQA using deep learning models. The authors created a large dataset of questions and answers from Stack Overflow and used natural language processing techniques such as dense embeddings and Long Short-Term Memory (LSTM) networks to predict the accepted answer attribute. They also produced numerical features extracted from metadata found in Stack Overflow posts or derived from text data. These features were used to assist with answer ranking tasks. The performance of their deep learning models was compared against ensemble methods such as forest and boosted trees but could not improve upon baseline results. The authors speculate that this lack of improvement may be due to the large number of out-of-vocabulary words present in programming code snippets found in questions and answers text. To address this issue, new methods need to be developed that can handle out-of-vocabulary words present in programming code snippets more effectively. In conclusion, while deep learning approaches may be beneficial for answer ranking problems in CQA, further research is needed to overcome existing challenges. This study highlights the difficulties faced by researchers working on CQA problems and suggests potential avenues for future research.