Inappropriate Comment Detection in 360-Degree Feedback with Learning Data Augmentation using Text Generation Models

Abstract

Many companies have implemented "360-degree feedback" to assess employee skills and job performances, providing objective insights into their strengths and weaknesses. This feedback often includes additional comments. However, some comments include inappropriate or too-aggressive content that undermines the effectiveness of the skill development of target employees. Because of this issue, HR personnel must spend significant time manually reviewing comments. Thus, this study aims to reduce this workload by developing models to detect inappropriate comments automatically. We proposed a BERT model fine-tuned with actual annotated data, which we confirmed to achieve a high recall rate of 80%. Additionally, we examined the effect of data augmentation using text generation models, demonstrating that data augmentation with a human-made word dictionary improved the accuracy of the BERT model.