Junfan Chen, Beihang University
This is the 2 minites promotional video for ?Open-Set Semi-Supervised Text Classification with Latent Outlier Softening?. The pseudo-labeling methods may produce mislabeled texts from the unlabeled-text set. The mislabelling usually lead to false positive inference problem in OSTC, where the biased model are prone to mistakenly predict an OOD example to a ID class. Existing Pipeline OSTC models first filter OOD examples then perform semi-supervised training, which may bring error-labels that may degrade the OSTC performance. Therefore, we propose an unified OSTC model that simultaneously train an STC model and soften the OOD impacts by adaptive weights within a unified training process. Specifically, our Latent Outlier Softening Model (or LOS) adopts the EM algorithm, where the E-step computes the soft weights and M-step update the model parameters. The evaluation results show that LOS significantly outperforms baseline models.
