Neural Architecture Search with Heterogeneous Representation Learning for Zero-Shot Multi-Label Text Classification.

Zero-shot multi-label text classification has become a hot topic recently and has a wide range of applications, including assigning legal concepts to legislation, category information to goods, and disease information to patient records. Existing approaches employ a series of neural networks to represent the text and labels separately by relying on artificial experience, which can not achieve good performance. To effectively represent text data and multi-label data together, it is critical to aggregate the neighboring information with graph structure information in zero-shot multi-label text classification. To solve this problem, we propose a neural architecture search (NAS) approach with heterogeneous representation learning for the representation of text data and labels data together. We split the original search space of NAS into two heterogeneous search spaces and reformulated NAS with heterogeneous representation learning to aggregate neighboring information better and reduce unnecessary search. Besides, we design an alternating search strategy to search for suitable neural architectures for zero-shot multi-label text classification. We conduct neural architecture search and retraining experiments on ERULEX57K dataset. The results demonstrate that our method outperforms previous zero-capable methods and improves the normalized discounted cumulative gain at the top 5 predicted labels (nDCG@5) by 3.0%, 2.4%, 18.9% and 1.0% for overall, frequent, few-shot, and zeroshot labels, respectively.