Martin Körner

Martin Körner

The extraction of individual reference strings from the reference section of scientific publications is an important step in the citation extraction pipeline. Current approaches divide this task into two steps by first detecting the reference section areas and then grouping the text lines in such areas into reference strings. We propose a classification model that considers every line in a publication as a potential part of a reference string. By applying line-based conditional random fields rather than constructing the graphical model based on the individual words, dependencies and patterns that are typical in reference sections provide strong features while the overall complexity of the model is reduced.

Martin Körner, Tatiana Sennikova, Florian Windhäuser et al.

Wikipedia articles about the same topic in different language editions are built around different sources of information. For example, one can find very different news articles linked as references in the English Wikipedia article titled "Annexation of Crimea by the Russian Federation" than in its German counterpart (determined via Wikipedia's language links). Some of this difference can of course be attributed to the different language proficiencies of readers and editors in separate language editions, yet, although including English-language news sources seems to be no issue in the German edition, English references that are listed do not overlap highly with the ones in the article's English version. Such patterns could be an indicator of bias towards certain national contexts when referencing facts and statements in Wikipedia. However, determining for each reference which national context it can be traced back to, and comparing the link distributions to each other is infeasible for casual readers or scientists with non-technical backgrounds. Wikiwhere answers the question where Web references stem from by analyzing and visualizing the geographic location of external reference links that are included in a given Wikipedia article. Instead of relying solely on the IP location of a given URL our machine learning models consider several features.

Rene Pickhardt, Thomas Gottron, Martin Körner et al.

We introduce a novel approach for building language models based on a systematic, recursive exploration of skip n-gram models which are interpolated using modified Kneser-Ney smoothing. Our approach generalizes language models as it contains the classical interpolation with lower order models as a special case. In this paper we motivate, formalize and present our approach. In an extensive empirical experiment over English text corpora we demonstrate that our generalized language models lead to a substantial reduction of perplexity between 3.1% and 12.7% in comparison to traditional language models using modified Kneser-Ney smoothing. Furthermore, we investigate the behaviour over three other languages and a domain specific corpus where we observed consistent improvements. Finally, we also show that the strength of our approach lies in its ability to cope in particular with sparse training data. Using a very small training data set of only 736 KB text we yield improvements of even 25.7% reduction of perplexity.