基于深度迁移学习的跨领域文本情感分类方法研究

Exploring the implicit opinions and sentiment in subjective texts in social media is critical to the applications such as product recommendation, e-commerce, and social sentiment monitoring. The existing deep learning model faces challenges of weak supervision, specific task needs, distribution difference gap, and data multi-source when solving text sentiment mining tasks in specific domains. Cross-domain sentiment classification is the process of transferring the sentiment polarity classification task of unlabeled data in the target domain by transferring the sentiment information of the source domain task to the target domain. Based on the characteristics of the domain diversity, sentiment domain dependence, and the consistency of semantic structure, this project develops cross-domain sentiment classification methods under single and multiple source domains based on the deep transfer learning mechanism. Specifically, the main contents include: (1) Exploring the internal mechanism of cross-domain sentiment transfer, and giving the basic strategies and algorithms for sentiment deep transfer learning; (2) Studying cross-domain knowledge representation and distance metrics, and introducing semantic structure learning and attention transfer mechanism; (3) Using multi-source domain joint learning mechanism to solve multi-source cross-domain sentiment classification tasks; (4) Developing a product recommendation system based on cross-domain sentiment analysis technology to provide services for researchers and enterprises. The results of the project can not only expand the applications of deep transfer learning in textual social media, but also effectively solve the problem of insufficient data marking in specific domains, which can provide new research clues and directions for social media sentiment analysis.

挖掘社会媒体主观性文本中隐含的观点和情感，对产品推荐、电子商务和社情舆情监控等应用至关重要。深度学习模型在解决特定领域的文本情感分类任务时面临弱监督、特定任务需要、分布差异鸿沟和数据多源性等挑战。跨领域情感分类是通过迁移源领域任务的情感信息到目标领域，实现对目标领域不带标签数据的情感极性分类任务的过程。本项目针对文本领域多样性、情感领域依赖性和语义结构一致性特点，开展社会媒体跨领域文本情感分类研究。主要研究内容包括：（1）探求跨领域情感迁移机理，给出深度情感迁移学习基本策略和算法；（2）研究跨领域知识表示和距离度量，引入句法结构一致性与注意力迁移机制；（3）运用多源领域联合学习机制，解决多源跨领域情感分类任务；（4）开发基于跨领域情感分析技术的产品推荐系统。项目成果可拓展深度迁移学习在文本特性社会媒体中的应用场景，有效解决特定领域数据标记不足问题，还可以为社会媒体情感分析提供新的研究思路。

作为社会媒体文本情感分析的重要研究课题之一，跨领域文本情感分类旨在利用源领域资源或模型迁移地服务于目标领域的文本情感分类任务，其可以有效缓解目标领域中带标签数据不足问题。本项目开展了以下四点研究工作：（1）跨领域文本情感分类研究进展总结，总结了目前跨领域情感迁移面临的主要挑战，并对其未来的研究方向进行了展望；（2）基于参数迁移和注意力共享机制的跨领域文本情感分类方法，设计了两个分层注意力网络：源领域网络和目标领域网络。考虑了跨领域句子级别和单词级别的注意力对应关系，并设计了一种有效的跨领域注意力共享机制；（3）基于多源领域适应联合学习的跨领域文本情感分类方法，构建了包含情感分类损失、参数共享损失、领域融合损失和正则项在内的联合损失函数，设计了多源领域适应联合学习训练算法，联合训练多个源领域和目标领域的带标签数据；（4）基于句法结构迁移和领域融合的跨领域情感分类方法，可以解决特定领域对带标签数据依赖问题，句法结构迁移方面，将依存语法特征加入到递归神经网络中，设计一种可迁移的依存句法递归神经网络模型。并基于跨领域情感分析技术构建了智慧旅游平台应用。此平台利用跨领域情感迁移理论与方法开发了一个基于评论大数据分析的智慧旅游平台，可以为研究者和企业提供在线情感分析服务。受到项目资助以第一作者发表论文5篇，授权发明专利2项，实质审查中发明专利2项，软件著作权10项，获得奖励2项。