Mitigating Data Imbalance Problem in Transformer-Based Intent Detection

There are two major problems when deploying a practical intent detection system for new customer. First, domain-specific data from the customer could be limited and imbalanced. Additionally, despite different customers might share same domain, their categories each other. Thus, it difficult to combine datasets collected into single larger one. In this paper, we use class weights in loss computation alleviate imbalance problem. The defined inversely proportional frequency of training set order give more influence less observed classes. We also employ two-pass fine-tuning procedure utilize information in-domain datasets. Experimental results show that performance is improved significantly weighted function used together with transfer learning procedure. absolute improvement percent accuracy approximately 2% over transformer-based baseline.