文本过滤已成为文档分析不可缺少的技术手段。这项技术对于大数据时代下的各类应用中的信息检索、商品推荐等服务提供了技术层面的支撑。现有研究工作主要基于监督学习技术，需要借助大量人力去构建类别的训练文档。然而，在大数据时代的背景下，文档可能覆盖的类别数量不断增长，构建所有类别的训练文档是不现实的。本项目旨在通过类别的语义关键词（种子词），研究基于种子词的零样本/小样本学习的文本过滤技术，克服上述的文本种类不断增长和标注文档样本代价昂贵的难题。项目将采用深度学习与信息检索相关技术理论，重点研究：1）基于词向量表示的文本类别关联特征表示学习技术；2）面向零样本的文本类别语义关联信息的提取机制；3）面向小样本的文本语义匹配技术；4）基于零样本文本过滤模型的小样本文本过滤方法。本项研究对于推动文本处理技术的进一步发展以及降低大数据时代构建训练数据成本的需求具有重要的科学意义和应用价值。

Document filtering has become an indispensable technical process for document analysis. This class of techniques provides technical support for information retrieval and recommendation related applications in the era of big data. Existing works mainly develop the filtering solutions in the paradigm of supervised learning, which requires a lot of manually identified documents for each category in their training phase. In the era of big data, the number of possible categories covered by documents becomes constantly larger. It is unrealistic to manually identify a lot of positive examples for each possible category. In this project, our endeavor is devoted to the development of zero/few-shot document filtering techniques based solely on the provided semantic keywords (called seed words) of each category, aiming at reducing the cost for building the expensive training data. Following the deep learning and information retrieval methodologies, we mainly focus on the following research topics: (1) the category-specific relevance representation learning for documents based on word embeddings; (2) zero-shot driven category-specific relevance signal extraction for documents; (3) few-shot driven semantic matching for documents; (4) few-shot document filtering on the basis of zero-shot document filtering model. This project will make significant contributions to the development of text processing techniques and fulfill the requirement on the cost reduction for building training data.