Adaptive selection: External knowledge and internal context representation for few-shot intention recognition

作     者：Tang, Jingfan Li, Pengfei Zhang, Xuefeng Xia, Jianmei 

作者机构：Hangzhou Dianzi Univ Coll Comp Sci & Technol Hangzhou 310018 Peoples R China Zhoukou Normal Univ Sch Network Engn Zhoukou 466000 Peoples R China Ningbo Univ Coll Sci & Technol Sch Informat Engn Ningbo 315300 Peoples R China Westlake Univ Affiliated Hangzhou Peoples Hosp 1 Sch Med Dept Gynecol & Obstet Hangzhou 310006 Peoples R China 

出 版 物：《ALEXANDRIA ENGINEERING JOURNAL》 (Alexandria Engineering Journal)

年 卷 期：2025年第117卷

页      面：301-310页

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 08[工学] 

基　　金：Zhejiang Provincial Technical Plan Project [2020C03105] Fundamental Public Welfare Research Program of Zhejiang Province [LGF22F020036] Ningbo "Innovation Yongjiang 2035" Key RD Programme [2024Z202] LGG21F020006 

主　　题：Few-shot intention recognition Adaptive selection Knowledge graph Capsule network Dynamic routing algorithm 

摘      要：Intention recognition models usually need to be trained on a large number of data, and when new intentions appear in the discriminant field, only a small amount of data is available. The method based on few-shot learning can well deal with this problem. In existing methods, the intent representation is often obtained by summing or averaging the sample representations. This will make the distance between the intents too close and cause the classification to fail. In this paper, we propose a few-shot intent recognition method based on adaptive selection of external knowledge and internal context representation. This method combines external knowledge representation extracted from a knowledge graph with internal context representation within sentences. It achieves hierarchical modeling of sentences through a dynamic routing algorithm and adaptively selects entity semantic representation information to enhance the semantic representation of entities in sentences. Experiments conducted on the Banking 77 dataset demonstrate that this method outperforms existing methods in both crossdomain and same-domain scenarios, particularly achieving an accuracy rate of 80.73 % under the 3-way 30shot setting.