Descriptor-Invariant Fusion Architectures for Automatic Subject Indexing

Abstract

Documents indexed with controlled vocabularies enable users of libraries to discover relevant documents, even across language barriers. Due to the rapid growth of scienti c publications, digital libraries require automatic methods that index documents accurately, especially with regard to explicit or implicit concept dri , that is, with respect to new descriptor terms and new types of documents, respectively. is paper rst analyzes architectures of related approaches on automatic indexing. We show that their design determines individual strengths and weaknesses and justify research on their fusion. In particular, systems bene t from statistical associative components as well as from lexical components applying dictionary matching, ranking and binary classi cation. e analysis emphasizes the importance of descriptor-invariant learning, that is, learning based on features, which can be transferred between di erent descriptors. eoretic and experimental results on economic titles and author keywords underline the relevance of the fusion methodology in terms of overall accuracy, and adaptability to dynamic domains. Experiments show, that fusion strategies combining a binary relevance approach and a thesaurusbased system outperform all other strategies on the tested data set. Our ndings can help researchers and practitioners in digital libraries to choose appropriate methods for automatic indexing. CCS CONCEPTS •Computing methodologies →Supervised learning; Machine learning; Natural language processing; •Information systems →Digital libraries and archives;

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

“…As a consequence, the performance of these methods largely depends on the availability of appropriate training examples and the stability of term and concept distributions, whereas lexical methods require vocabularies that exhaustively cover the domain. When concept drift occurs, that is, if observed terms and the set of relevant concepts differ between training data and new data, both types of indexing approaches considerably decrease in performance [8]. Interestingly, since these algorithms merely learn to assign recognized subjects of the controlled vocabulary, they will silently miss to assign relevant subjects not covered by the controlled vocabulary, and moreover they are unable to recognize and represent the loss in document-level content representation.…”

“…Regarding economics, we use three datasets, which comprise roughly 20,000 (T20k), 60,000 (T60k), and 400,000 documents (T400k), respectively. Each document is associated with several descriptors, for instance 5.89 on average for T400k, from the STW Thesaurus for Economics (STW) 8 . Both, the STW and EUROVOC, comprise thousands of concepts, yielding challenging multi-label classification tasks.…”

“…The relevance of multi-label text classification engendered research in several disciplines. Although considerable progress has been made over the last decades [2,3,4,5,6,7,8,9], several challenges remain. Just to give an example, precision@5 = 52% [9] has recently been reported for a dataset in the legal domain (EURLEX [5]), which means that on average per document only half of the five top-ranked subjects matched human annotations.…”

Multi-Modal Adversarial Autoencoders for Recommendations of Citations and Subject Labels

“…We also conduct the split between training set and test set on the time axis (see Figure 5). This is challenging because label annotations suffer from concept drift over time [39]. We use the years 2012 and 2013 as test documents to obtain a train-test ratio similar to the scenario in Section 5.1.…”

Fusion architectures for automatic subject indexing under concept drift

“…Then, the fusion layer (below) is responsible for combining these predictions. The most interesting property of this layer is the descriptor-invariant decision function [32], i.e., a function that allows to perform predictions for all (also unseen) descriptors. Optionally, the fusion module may additionally consult the knowledge base or the professionally in- dexed documents for its decisions and use a descriptorspecific fusion component.…”