A lightweight semantic‐enhanced interactive network for efficient short‐text matching

Knowledge‐enhanced short‐text matching has been a significant task attracting much attention in recent years. However, the existing approaches cannot effectively balance effect and efficiency. Effective models usually consist of complex network structures leading to slow inference speed and the difficulties of applications in actual practice. In addition, most knowledge‐enhanced models try to link the mentions in the text to the entities of the knowledge graphs—the difficulties of entity linking decrease the generalizability among different datasets. To address these problems, we propose a lightweight Semantic‐Enhanced Interactive Network (SEIN) model for efficient short‐text matching. Unlike most current research, SEIN employs an unsupervised method to select WordNet's most appropriate paraphrase description as the external semantic knowledge. It focuses on integrating semantic information and interactive information of text while simplifying the structure of other modules. We conduct intensive experiments on four real‐world datasets, that is, Quora, Twitter‐URL, SciTail, and SICK‐E. Compared with state‐of‐the‐art methods, SEIN achieves the best performance on most datasets. The experimental results proved that introducing external knowledge could effectively improve the performance of the short‐text matching models. The research sheds light on the role of lightweight models in leveraging external knowledge to improve the effect of short‐text matching.