文本分类中的主动多域学习

现有主动学习主要着眼于对单个域训练方法的研究,不同域有不同的特征,同时也存在一些隐含的共性.如何从多个域中选择合适数据样本成为多域学习中减少人工标注工作量的关键.本文提出了一个新颖的主动多域学习框架,该框架充分考虑了重复信息,并可从多个域中选择合适的数据样本.该框架首先找到一个包含不同域间隐含共性的共享子空间,然后将所有数据样本分解为公共域部分和个性域部分,其中公共域部分可视为域间的重复信息,该部分在查询时需要被考虑到.最后,将主动多域学习方法与最新的主动学习方法的性能进行了比对,实验结果表明,本文提出的主动多域学习方法在减少人工标注工作量方面有显著作用.

Multi-Domain Active Learning in Text Classification

T he existing active learning methods are mainly focus on training a single domain .Different do-mains have different characteristics ,but there are some implied commonalities .T herefore ,how to choose the right data samples from multiple domains becomes the key to reduce the workload of manual tagging in multi-domain learning .This paper presents a novel multi-domain active learning framework .The frame-work fully considered the duplicate information and selected the appropriate data samples from multiple do-mains .Firstly ,in this framework ,a sharing subspace containing implicit commonalities between different domains is found ;T hen ,all the data samples are broken dow n into the individual domain portions and the public domain portions ,and the public domain portions can be considered as the duplicate information be-tween domains which needs to be considered in the query .Finally ,the multi-domain active learning meth-ods and the latest active learning methods are compared in terms of performance .The experimental results show that the proposed multi-domain active learning methods are more marked effect in reducing the work-load of manual tagging .