Deep learning and Threshold Probability for Out of Scope Intent Detection in Task Oriented Chatbot
Muhammad Fikri Hasani, Ford Lumban Gaol, Benfano Soewito, Harco Leslie Hendric Spits Warnars · 2022
A lot of research on task-oriented chatbots has been carried out to improve research on intent recognition. However, there are not many studies that discuss out-of-scope. This study introduces the use of threshold for classification results. The use of threshold improves recall performance of OOS detection above baseline for BiLSTM but reduces recall inside scope for all models. Directly applying multiclass classification with combined class of inside scope and out of scope also showed higher recall than baseline.