论文标题
神经对比聚类：完全无监督的偏见减少情感分类
Neural Contrastive Clustering: Fully Unsupervised Bias Reduction for Sentiment Classification
论文作者
论文摘要
背景：神经网络由于相关偏差而产生偏见的分类结果（即使这些相关性不代表因果关系，他们学习输入和输出之间的相关性以对样本进行分类）。 目的：这项研究引入了一种完全无监督的减轻相关性偏差的方法，并通过对COVID-19社交媒体数据的情感分类证明了这一点。 方法：情感分类中的相关性偏差通常在有关有争议的主题的对话中产生。因此，本研究使用对抗性学习来对比基于情感分类标签的群集，以及由无监督的主题建模产生的簇。这使神经网络不受学习与主题相关的特征的影响，从而产生偏见的分类结果。 结果：与基线分类器相比，神经对比聚类的聚类大约可以使人标记的Covid-19社交媒体数据的偏置句子的准确性翻了一番，而不会对分类器的整体F1分数产生不利影响。尽管是一种完全无监督的方法，但神经对比的聚类与偏见句子的准确性相比，与监督的掩盖方法相比，准确性的准确性更大。 结论：神经对比聚类减少了情感文本分类中的相关性偏差。需要进一步的研究来探索将此技术推广到其他神经网络架构和应用领域。
Background: Neural networks produce biased classification results due to correlation bias (they learn correlations between their inputs and outputs to classify samples, even when those correlations do not represent cause-and-effect relationships). Objective: This study introduces a fully unsupervised method of mitigating correlation bias, demonstrated with sentiment classification on COVID-19 social media data. Methods: Correlation bias in sentiment classification often arises in conversations about controversial topics. Therefore, this study uses adversarial learning to contrast clusters based on sentiment classification labels, with clusters produced by unsupervised topic modeling. This discourages the neural network from learning topic-related features that produce biased classification results. Results: Compared to a baseline classifier, neural contrastive clustering approximately doubles accuracy on bias-prone sentences for human-labeled COVID-19 social media data, without adversely affecting the classifier's overall F1 score. Despite being a fully unsupervised approach, neural contrastive clustering achieves a larger improvement in accuracy on bias-prone sentences than a supervised masking approach. Conclusions: Neural contrastive clustering reduces correlation bias in sentiment text classification. Further research is needed to explore generalizing this technique to other neural network architectures and application domains.
