Fusion architectures for automatic subject indexing under concept drift

Abstract

Indexing documents with controlled vocabularies enables a wealth of semantic applications for digital libraries. Due to the rapid growth of scientific publications, machine learning based methods are required that assign subject descriptors automatically. While stability of generative processes behind the underlying data is often assumed tacitly, it is being violated in practice. Addressing this problem, this article studies explicit and implicit concept drift, that is, settings with new descriptor terms and new types of documents, respectively. First, the existence of concept drift in automatic subject indexing is discussed in detail and demonstrated by example. Subsequently, architectures for automatic indexing are analysed in this regard, highlighting individual strengths and weaknesses. The results of the theoretical analysis justify research on fusion of different indexing approaches with special consideration on information sharing among descriptors. Experimental results on titles and author keywords in the domain of economics underline the relevance of the fusion methodology, especially under concept drift. Fusion approaches outperformed non-fusion strategies on the tested data sets, which comprised shifts in priors of descriptors as well as covariates. These findings can help researchers and practitioners in digital libraries to

Annif: DIY automated subject indexing using multiple algorithms

“…In both cases, the result is a list of candidate subjects for the document. In order to determine the final set of suggested subjects for the document, the candidates must then be ranked and only the most promising ones retained (Medelyan, 2009;Toepfer & Seifert, 2018).…”

“…Algorithms for automated subject indexing can generally be divided into lexical and associative approaches (Toepfer & Seifert, 2018). In lexical approaches, frequently occurring or otherwise salient terms in the document are matched with terms in the vocabulary.…”

“…Associative approaches, including machine learning algorithms, instead find correlations between words (or, more generally, short sequences of words called n-grams) in document text and subjects, based on a large amount of training data. These two approaches can be considered complementary, and often the best results are obtained by combining results from both kinds of algorithms using ensembles and/or fusion architectures (Toepfer & Seifert, 2018).…”

“…Fusion methods for automated subject indexing (Toepfer & Seifert, 2018) are ways of combining results from multiple algorithms. The algorithms are combined into an ensemble and the final prediction of subjects is made by using a decision function applied on the predictions of individual algorithms.…”

Representation Learning for Texts and Graphs

“…Figures 6.10c and 6.11c). This resembles a challenge because label annotations suffer from concept drift over time [TS20]. We use the years 2012 and 2013 as test documents for EconBiz and the year 2016 for IREON to obtain a 90:10 train-test ratio, as in the citation recommendation datasets described above.…”