What matters in a transferable neural network model for relation\n classification in the biomedical domain?

Abstract

Lack of sufficient labeled data often limits the applicability of advanced\nmachine learning algorithms to real life problems. However efficient use of\nTransfer Learning (TL) has been shown to be very useful across domains. TL\nutilizes valuable knowledge learned in one task (source task), where sufficient\ndata is available, to the task of interest (target task). In biomedical and\nclinical domain, it is quite common that lack of sufficient training data do\nnot allow to fully exploit machine learning models. In this work, we present\ntwo unified recurrent neural models leading to three transfer learning\nframeworks for relation classification tasks. We systematically investigate\neffectiveness of the proposed frameworks in transferring the knowledge under\nmultiple aspects related to source and target tasks, such as, similarity or\nrelatedness between source and target tasks, and size of training data for\nsource task. Our empirical results show that the proposed frameworks in general\nimprove the model performance, however these improvements do depend on aspects\nrelated to source and target tasks. This dependence then finally determine the\nchoice of a particular TL framework.\n

Ähnliche Arbeiten