Few-Shot Learning for Domain-Specific Intent Detection in Task-Oriented Dialog Systems

The success of task-oriented dialog systems depends on the effective intent detection. Nevertheless, annotated data are not easily found in novel or specialized fields, and this is a difficult task. This paper examines how few-shot learning can be used to speed up domain-specific intent detection and aims to enable dialog systems to adapt quickly and use few-shot labeled examples of tasks at low costs. In this direction we present a meta-learning model based on prototypical networks which is trained on a small number of annotated utterances per intent. It uses contextual and semantic representations to make someone else understand the intents that were already understood in the source domains that have adequate training data to a new domain that have very few training examples. Increased adaptability and reduced overfitting The method relies on contextualized language embeddings and fine-tuning to ensure more adaptability. Our widespread benchmark dataset experiments and domain-specific experiments demonstrate that our approach can be significantly more effective than other supervised and transfer learning baselines and can be used with severe data constraints. The experiments carried out show that few-shot learning does not only decrease the time of developing a dialog system but can also be used to obtain high intent classification accuracy (including rare intents). This paper shows that meta-learning is effective in achieving the objectives of scalable, flexible and efficient deployment of intent detection modules in future domain-adaptive task-oriented dialogue systems.

Ҳали таржима қилинмаган