基于众包标注的多标记学习研究

Multi-label learning deals with examples simultaneously associated with multiple labels. While multi-label tasks are ubiquitous in practical applications, it's often difficult to collect the true labels for training data. By distributing the tasks to multiple easy to access non-experts, crowdsourcing provides a fast and flexible alternative to collecting labeling information. This project proposes to study on multi-label crowdsourcing learning problem, and learn from multi-label training data labeled by crowds. The research contents include: 1) proposing methods to make robust use of label correlations, which tackle the label correlation estimation difficulties in the crowdsourcing scenario; 2) proposing methods to learn from low quality annotations, which complement existing learning methods that require all labels annotated for the tagged instances; 3) proposing method to actively collect annotations and learning from them, which address the inefficiency and high cost of existing annotation collection strategies. Our project provides novel directions for multi-label crowdsourcing learning. It is expected to publish 4-6 papers on high quality journals or conferences and apply 1-2 patents.

多标记学习研究一个样本同时与多个标记相关联的学习问题。在实际应用中，多标记学习的任务普遍存在，而收集训练数据的真实标记往往很困难。通过将任务分配给多个易于访问但可能犯错的非专家，众包提供了一种快速灵活的标记收集方式。本项目针对多标记众包学习问题进行研究，从众包标注的多标记训练数据中学习分类模型。主要研究内容包括：（1）针对众包设定下标记关系利用困难的问题，提出面向鲁棒关系利用的多标记众包方法；（2）针对现有学习方法过于单一、要求被标注样本的所有标记的问题，提出适用于低质标注数据的多标记众包方法；（3）针对现有标注收集策略低效、浪费成本的问题，提出基于主动众包标注的多标记众包方法。本项目为多标记众包学习开辟了新的思路和方向，成果有望在国内外重要学术期刊和会议上发表高质量论文4-6篇，并申请专利1-2项。

多标记学习研究一个样本同时与多个标记相关联的学习问题。在实际应用中，多标记学习的任务普遍存在，而收集训练数据的真实标记往往很困难。通过将任务分配给多个易于访问但可能犯错的非专家，众包提供了一种快速灵活的标记收集方式。本项目针对这一问题进行研究，提出面向鲁棒关系利用的多标记众包方法；提出适用于低质标注数据的多标记众包方法；提出基于主动众包标注的多标记众包方法。相关成果发表高水平论文5篇，申请专利4项，培养研究生3名。