小样本学习的跨模态情感识别方法研究

通过少量样本快速学习到目标的本质信息，是小样本学习研究领域的重要内容，具有广泛的适用场景与应用领域。如何准确地理解、分析与处理跨模态小样本数据，特别在情感分析方法研究领域，仍然是一个待解决的问题。本项目重点展开基于小样本学习的跨模态情感智能分析的研究，具体包括了一下三个方面的内容：1). 基于宽度学习技术，结合孪生网络极强的度量能力，探索适用于情感数据的小样本学习模型，充分学习小样本数据的情感特征；2). 基于跨模态表征学习及迁移学习方法，研究跨模态数据的属性共享机制，挖掘出能够指导语义表达的情感智能分析共同属性；3). 利用对抗网络建立易扩展和强泛化能力的优化模型，解决小样本学习模态泛化能力弱及可扩展性差等问题。本项目对增强神经网络框架理解，提升跨模态信息的语义挖掘能力，以及推进情感分析方法和相关应用，具有十分重要的意义。

Quickly learning the essential information of things through a small number of samples is an important content in the field of few-shot learning research. It has a wide range of applications. How to understand, analyze and process the small cross-modal sample data, especially in the field of affective computing, is still a problem to be solved. This project focuses on the cross-modal affective computing based on few-shot learning, specifically including the following three aspects: (1)based on the Broad Learning System and Siamese Network, exploring some few-shot learning models that are suitable for affective computing to fully learn the emotion features of small sample data; (2) Based on cross-modal representation learning and transfer learning methods, studying the attribute sharing mechanism and mine the common characteristics of emotional cross-modal data that can guide semantic representation. (3) The adversarial network can be used to establish a model that is easy to expand and has a strong generalization ability. It can solve the problems of bad generalization and poor scalability on few-shot learning models. This project is of great significance for enhancing the understanding ability of the neural network framework, improving the ability of semantic mining for cross-modal information, and promoting the research of affective computing and its related applications.