Interaction-Enhanced Textual Similarity Matching Distance Measurement – Exploration on Chinese

Textual similarity matching is one of the most valuable tasks in natural language processing. Meanwhile, due to the wide development of applications and increase in requirements, various subtasks are proposed and expected to be resolved. In addition, Chinese, which owns the largest population of native speakers, introduced more challenges because of its different grammar and characteristics. In this project, we explore the best design of representation, interaction, and prediction layers for Chinese textual similarity matching by extensive experiments on two public datasets. Results prove that the precise interaction paradigm and unlearnable distance metric is more suitable for matching task considering effectiveness and efficiency. And for the option of representation structure, we should simultaneously consider textual information distribution within datasets and computational cost.