# Linear Classifier: An Often-Forgotten Baseline for Text Classification

**Authors:** Yu-Chen Lin, Si-An Chen, Jie-Jyun Liu, Chih-Jen Lin  
**arXiv:** [2306.07111](https://arxiv.org/abs/2306.07111)  
**Date:** 2023-06-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.07111v1/

## Key points

- Authors emphasize the importance of using simple baselines like linear classifiers in conjunction with advanced methods like BERT for text classification
- Linear methods demonstrate competitive performance, high efficiency, and robustness for many text data sets
- Running linear classifiers alongside advanced models allows researchers to compare results and determine if the advanced model's performance is acceptable or if further improvements are needed
- Simple baselines act as a benchmark to confirm whether the outcomes of advanced models align with expectations
- Experimental results presented by the authors validate their argument and reinforce the significance of incorporating simple baselines into text classification workflows
- Using linear classifiers as a baseline enhances understanding of model performance and helps make more informed decisions regarding deployment strategies.

## Summary

In their paper titled "Linear Classifier: An Often-Forgotten Baseline for Text Classification," authors Yu-Chen Lin, Si-An Chen, Jie-Jyun Liu, and Chih-Jen Lin discuss the importance of using simple baselines like linear classifiers in conjunction with advanced methods such as large-scale pre-trained language models like BERT for text classification. While these advanced models have gained popularity due to their superior performance, the authors argue that directly training them for a few epochs and deploying the obtained model may not always yield satisfactory results. The authors highlight two key reasons why incorporating simple baselines is crucial. Firstly, linear methods demonstrate competitive performance, high efficiency, and robustness for many text data sets. By running linear classifiers on bag-of-words features alongside advanced models, researchers can compare the results and determine if the advanced model's performance is acceptable or if further improvements are needed. Secondly, the authors emphasize that even advanced models like BERT may only achieve optimal results when properly applied. Simple baselines act as a benchmark to confirm whether the outcomes of advanced models align with expectations. This validation step helps researchers assess if they have effectively utilized the capabilities of these sophisticated methods. To support their argument, the authors present experimental results that validate their points. These findings reinforce the significance of incorporating simple baselines into text classification workflows to ensure accurate and reliable outcomes. Overall, this opinion paper sheds light on an often overlooked aspect of text classification by emphasizing the importance of using linear classifiers as a baseline alongside advanced models like BERT. By doing so, researchers can enhance their understanding of model performance and make more informed decisions regarding deployment strategies.