# Beyond Labels: Empowering Human with Natural Language Explanations through a Novel Active-Learning Architecture

**Authors:** Bingsheng Yao, Ishan Jindal, Lucian Popa, Yannis Katsis, Sayan Ghosh, Lihong He, Yuxuan Lu, Shashank Srivastava, James Hendler, Dakuo Wang  
**arXiv:** [2305.12710](https://arxiv.org/abs/2305.12710)  
**Date:** 2023-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.12710v1/

## Key points

- Data annotation is a costly task in machine learning
- Existing Active-Learning (AL) techniques focus on labeling data points but overlook the importance of providing natural language explanations alongside the labels
- Proposed AL architecture generates both labels and explanations to reduce the need for human annotations
- Explanation-generation model assists humans in decision-making process in real-world applications
- AL data selection strategy leverages explanation annotations and outperforms traditional strategies
- Human evaluation shows preference for explanations generated by proposed system over state-of-the-art system
- Lack of explainability in language models may lead to mistrust of predictions
- Proposed framework has limitations including applicability to other NLP tasks beyond classification datasets and reliance on data diversity-based AL selector design

## Summary

Data annotation is a costly task in machine learning, and researchers have proposed Active-Learning (AL) techniques to support human annotators. However, existing AL works only focus on labeling data points and overlook the importance of providing natural language explanations alongside the labels. In real-world scenarios, such as medical diagnoses, humans often require both labels and corresponding explanations simultaneously. To address this gap, this work proposes a novel AL architecture that reduces the need for human annotations by generating both labels and explanations. The proposed AL architecture incorporates an explanation-generation model that can explicitly generate natural language explanations for the prediction model. These explanations assist humans in their decision-making process in real-world applications. The framework also includes a data diversity-based AL data selection strategy that leverages the explanation annotations. To evaluate the effectiveness of the proposed approach, automated AL simulation evaluations were conducted. The results consistently showed that the data selection strategy outperformed traditional data diversity-based strategies. Additionally, a human evaluation was conducted to compare the generated explanations with those from a state-of-the-art (SOTA) explanation-generation system. The results revealed that humans preferred the explanations generated by the proposed system. The introduction highlights recent advancements in Natural Language Processing (NLP) and emphasizes the need for explainability within language models. While these models demonstrate impressive performance on various NLP tasks, their lack of faithful explainability may lead to mistrust of their predictions. Humans typically develop intermediate information as rationales to aid decision making which serves as faithful explanations. The paper also discusses limitations of the proposed framework including its applicability to other NLP tasks beyond classification datasets like e-SNLI and its reliance on a data diversity based AL selector design. Overall, this work presents an innovative approach to reduce human annotations by incorporating an explanation generation model into an AL architecture. The results demonstrate improved performance compared to traditional strategies and highlight human preference for generated explanations over existing systems.