Machine Learning Architectures for Scalable and Reliable Subject Indexing Fusion, Knowledge Transfer, and Confidence

摘要

Digital libraries desire automatic subject indexing as a scalable provider of high-quality semantic document representations. The task is, however, complex and challenging, thus many issues are still unsolved. For instance, certain concepts are not detected accurately, and confidence estimates are often unreliable. Accurate quality estimates are, however, crucial in practice, for example, to filter results and ensure highest standards before subsequent use. The proposed thesis studies applications of machine learning for automatic subject indexing, which faces considerable challenges like class imbalance, concept drift, and zero-shot learning. Special attention will be paid to architecture design and automatic quality estimation, with experiments on scholarly publications in economics and business studies. First results indicate the importance of knowledge transfer between concepts and point out the value of so-called fusion approaches that carefully combine lexical and associative subsystems. This extended abstract summarizes the main topic and status of the thesis and provides an outlook on future directions.