Semantic Matching Using an Efficient Multi-task Method

Abstract:
Semantic matching is one of the most important tasks in natural language processing and plays a vital role in question answering systems, dialogue systems, modern search engines, and business systems. However, semantic matching tasks usually face higher labeling costs and greater modeling difficulty. In this paper, we propose a multi-task learning method that can easily introduce external annotated data while efficiently modeling the correlations and differences between different tasks. In more detail, we propose a task-aware attention mechanism to learn differential representations of a sentence under various tasks. At the meantime, we propose a multi-expert network to learn the commonalities and differences among different tasks. Experimental results on the benchmark dataset Chinese-Glue demonstrate that our proposed method outperforms existing methods.