FLC: Empower Large-Scale Multi-Label Text Classification via Free Label Correlations Derived from Massive Raw Texts

Abstract

Large-scale multi-label text classification (LMTC) aims at tagging texts with its most relevant subset of labels from a large candidate label set. State-of-the-art LMTC methods are mostly based on deep neural networks trained using a cross-entropy loss. These methods require massive human labeled training data to capture the underlying label correlations implicitly, which could be a burden in real-world applications. We observe that many correlated labels are missed by human annotators, but their label names in fact co-occur often in document contexts. In this paper, we propose a novel text classification framework FLC, incorporating “free” label correlations derived from massive raw text corpora. Specifically, we design a principled similarity measurement to evaluate the label correlation based on the co-occurrence statistics of their corresponding label names. Based on the derived label correlations, we add a complementary loss to make the overall loss more robust when correlated labels were predicted by the classifier but missed by human annotators. Experiments on two real-world datasets demonstrate the superiority of FLC over state-of-the-art methods, especially when the human labels are not perfect and the training data is limited. Moreover, the computational overhead of employing FLC is negligible.