De-confusing Hard Samples for Text Semantic Hashing

Text semantic hashing maps a text to a compact binary code, which is an important part of information retrieval and language processing. There are two main challenges for this task, one is to make the hash codes express the hierarchical category information for improving the retrieval accuracy, and the other is how to deal with the hard samples. In this paper, we adopt the Bernoulli VAE to encode the text semantics and design the parent and child level contrastive losses to learn the hierarchical information of the text. To find the hard samples, for each category, we introduce a latent sphere space to split the majority samples and hard samples, where the center and radius are dynamically calculated based on the semantic distance between samples. For the hard samples, we introduce the de-confusion loss to pull them close to the center. We conduct experiments on three datasets and the results show that the proposed model outperforms the SOTA baselines. The ablation experiments show that the category constraints and the de-confusion loss contribute to the model performance. The results of t-SNE also show that the hash codes learned by our model reflect high category differences.