Quantum neural networks successfully calibrate language models
摘要
Recently, quantum neural networks have been applied to different problems, like in computer vision and natural language processing. In this paper, we show an application of quantum models as probability calibrators to language models in Question and Answering (Q &A) tests. Our experiments show comparable results relative to the baseline Q &A models on all metrics. Here, we show another application of this class of models as confidence calibrators, where the parametrized quantum circuit receives a set of features coming from the language model and has to output a corrected confidence measure. The idea is to have a quantum model tell us when the language model is generating good answers or not.