Content-Based Quality Estimation for Automatic Subject Indexing of Short Texts Under Precision and Recall Constraints

摘要

Digital libraries strive for integration of automatic subject indexing methods into operative information retrieval systems, yet integration is prevented by misleading and incomplete semantic annotations. For this reason, we investigate approaches to detect documents where quality criteria are met. In contrast to mainstream methods, our approach, named Qualle, estimates quality at the document-level rather than the concept-level. Qualle is implemented as a combination of different machine learning models into a deep, multi-layered regression architecture that comprises a variety of content-based indicators, in particular label set size calibration. We evaluated the approach on very short texts from law and economics, investigating the impact of different feature groups on recall estimation. Our results show that Qualle effectively determined subsets of previously unseen data where considerable gains in document-level recall can be achieved, while upholding precision at the same time. Such filtering can therefore be used to control compliance with data quality standards in practice. Qualle allows to make trade-offs between indexing quality and collection coverage, and it can complement semi-automatic indexing to process large datasets more efficiently.