The current unsupervised GML solution is limited by inaccurate and insufficient knowledge conveyance. In this paper we propose a weakly supervised solution(W-GML) based on the paradigm of GML for the task of SLSA which can effectively exploit a few labeled samples to facilitate gradual knowledge conveyance. W-GML performs labeling by gradual phases, each of which is supposed to label only a proportion of target instances. In each phase, it selects target instances in the decreasing order of evidential certainty such that the distribution difference between the selected instances and the labeled instances is minimal. Then it labels the selected instances in a self-training way with GML as the base model.
As the proposed framework gradually leverages newly generated labels to fine-tune feature representations, it can effectively compensate for the scarcity of labeled training data.