Fast entity desambiguation in large-scale databases

要评价一个研究人员的工作、一个学术机构的影响或一项研究政策的有效性，就需要对科学出版物的相应产出进行分析。全球科学成果代表了大量信息，可以通过几个数据库访问，但模糊信息的存在和数据库的庞大规模使得难以在合理的时间内提取准确的结果。到目前为止，处理的一个重要部分，称为消歧，仍然必须手动执行，以达到令人满意的结果。在过去的几个月里，在NSERC ENGAGE项目的框架内，与工业合作伙伴Science-Meetings合作，对自动实体消歧的新方法的属性进行了初步调查。这种新的方法是基于一种技术，用于在图像序列中的人脸检测的图像处理。它的一个显著特点是使用级联分类算法，其中分类特征的选择和排序对于快速准确算法的推导至关重要。 由于初步研究产生了非常有希望的结果，本项目的总体目标是开发和测试一个成熟的级联消歧方法，能够处理实际规模的书目数据库，具有更好的准确性和召回率比现有的商业或公开可用的技术。为了达到这一目标，该方法的所有主要组成部分，即，数据预处理、训练方法、分类方法、可用性和计算效率将与科学-气象中心密切合作，重新审视和改进。由此产生的方法将为Science-Meetings提供显著的技术和商业优势，这反过来又将有利于实体消歧，大规模数据分析和更普遍的高级信息处理领域的经济和学术活动。

Evaluating the work of a researcher, the impact of an academic institution, or the effectiveness of a research policy requires the analysis of the corresponding output in terms of scientific publications. The global scientific production represents a huge amount of information which can be accessed through several databases, but the presence of ambiguous information and the large size of the databases make it difficult to extract accurate results in a reasonable lapse of time. So far, an important part of the processing, known as disambiguation, must still be performed manually in order to achieve satisfactory results. In the past few months, a preliminary investigation of the properties of a new approach to automatic entity disambiguation was carried out within the framework of a NSERC ENGAGE project, in collaboration with industrial partner Science-Metrix. This new approach is based on a technique used in image processing for face detection in image sequences. One of its salient characteristic is the use of a cascaded classification algorithm in which the choice and ordering of classification features are critical to the derivation of a fast and accurate algorithm. Since the preliminary study produced very promising results, the general goal of the present project is to develop and test a full-fledged cascaded disambiguation method capable of processing real-size bibliographic databases with better accuracy and recall rates than existing commercial or publicly available techniques. In order to reach this goal, all major components of the method, i.e., preprocessing of the data, training approach, classification methodology, usability and computational efficiency, will be revisited and improved in close collaboration with Science-Metrix. The resulting method will give Science-Metrix a significant technical and commercial edge, which in turn will be beneficial to the economic and academic activities in the areas of entity disambiguation, large-scale data analysis and more generally high level information processing.