Meta Learning for Complexity-Aware Optimal Sampling in Few-Shot Intent Classification

Intent classification plays a crucial role in applications such as virtual assistants and chatbots, enabling an accurate determination of the user's intention and providing relevant responses. Minimal supervision models that train from minimal labeled data have become important since they can train models using less amount of data as manually labeling data is time-consuming. To address this problem, the proposed model utilizes a few-shot learning mechanism approach using a metalearning architecture for intent classification. Model-Agnostic Meta Learning model is implemented to enhance generalization given the scenario of constrained labeled samples. Model Agnostic Meta Learning is tested using uniform and complexity aware optimal sampling based methods to assess the impact of data imbalance and robustness of the model. The experimental results on the Banking77 dataset show that the Meta Agnostic Meta Learning architectures in a complexity aware non-uniform sampling setting are able to achieve 100<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> of the results achieved by full train models when integrated with ML models. DL models on the other hand did not report a considerable enhancement when migrated from a uniform to a non-uniform sampling setting.