后验正则化综述

Survey on posterior regularization

在模型训练过程中,训练语料通常包含许多针对特定问题的边信息,而模型往往无法直接利用这些边信息。后验正则化(posterior regularization,PR)由于其框架的灵活性和简单性,在分类任务、自然语言处理以及远程监督系统等领域获得广泛应用。首先系统性地对后验正则化问题进行描述;然后详细介绍了三类后验正则化的通用框架,并指出了各个后验正则化框架被提出的原因以及其模型的具体形式、所具有的优缺点、适宜解决的问题等;进而又指出了近年来几类后验正则化框架的典型应用,并指明了后验正则化框架未来可能的发展方向;最后对全文内容进行概括总结。

During the model training,training corpora usually contains a lot of external problem-specific information that cannot be used by the model directly.Due to the flexibility and simplicity of its framework,Posterior regularization has been widely applied in the fields of natural language processing and sample selection of classification tasks.The aim of the posteriori regularization framework is to limit the capacity of the model posteriori on unmarked data,thus guiding the model to learn and achieve the desired performance.Firstly,this paper systematically described the problem of PR.Then it introduced the three general frameworks of PR,pointed the motivation of every PR and the specific form of every PR model,and the advantages of every PR model,the problems that every PR model could solve and the typical application scenarios that PR could be used.Lastly,this paper pointed out the research directions of the posterior regularization models in the future and summarized the content of the total paper.