# On Explaining Your Explanations of BERT: An Empirical Study with Sequence Classification

**Authors:** Zhengxuan Wu, Desmond C. Ong  
**arXiv:** [2101.00196](https://arxiv.org/abs/2101.00196)  
**Date:** 2021-01-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2101.00196v1/

## Key points

- BERT has gained attention for its ability to create new benchmarks in natural language processing tasks through fine-tuning
- Various attribution techniques have been proposed to explain BERT models, but they are often limited to sequence-to-sequence tasks
- The authors adapt existing attribution methods to explain the decision-making process of BERT in sequence classification tasks
- Extensive analyses using four different datasets in sentiment analysis are conducted, applying four existing attribution methods
- Reliability and robustness of each method are compared through various ablation studies
- Investigation is done on whether these attribution methods can explain generalized semantics across semantically similar tasks
- Findings provide valuable guidance for utilizing attribution methods to explain the decision-making process of BERT in downstream classification tasks
- Explanations can enhance transparency and interpretability in natural language processing applications by shedding light on the inner workings of BERT.

## Summary

BERT has gained significant attention in recent years for its ability to create new benchmarks in natural language processing tasks through fine-tuning. Various attribution techniques have been proposed to explain BERT models, but they are often limited to sequence-to-sequence tasks. In this study, the authors adapt existing attribution methods to explain the decision-making process of BERT in sequence classification tasks. The authors conduct extensive analyses using four different datasets in sentiment analysis and apply four existing attribution methods. They compare the reliability and robustness of each method through various ablation studies. Additionally, they investigate whether these attribution methods can explain generalized semantics across semantically similar tasks. The findings of this study provide valuable guidance for utilizing attribution methods to explain the decision-making process of BERT in downstream classification tasks. By shedding light on the inner workings of BERT, these explanations can enhance transparency and interpretability in natural language processing applications.