2019/09/08 by Che-Ping Tsai, Hung-yi Lee, Tsai, Che-Ping +1 · 1 citation
Computer Science · #Audio and Speech Processing (eess.AS) #Computation and Language (cs.CL) #FOS: Computer and information sciences #FOS: Electrical engineering #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Sound (cs.SD) #Text and Document Classification Technologies #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1909.03434
openalex publication_date 2019/09/08 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Multi-label classification (MLC) assigns multiple labels to each sample.\nPrior studies show that MLC can be transformed to a sequence prediction problem\nwith a recurrent neural network (RNN) decoder to model the label dependency.\nHowever, training a RNN decoder requires a predefined order of labels, which is\nnot directly available in the MLC specification. Besides, RNN thus trained\ntends to overfit the label combinations in the training set and have difficulty\ngenerating unseen label sequences. In this paper, we propose a new framework\nfor MLC which does not rely on a predefined label order and thus alleviates\nexposure bias. The experimental results on three multi-label classification\nbenchmark datasets show that our method outperforms competitive baselines by a\nlarge margin. We also find the proposed approach has a higher probability of\ngenerating label combinations not seen during training than the baseline\nmodels. The result shows that the proposed approach has better generalization\ncapability.\n