Extraction Of Citation Data From Websites Based On Visual Cues

In this master’s thesis a system for extracting meta-information, specifically citation
data, from webpages is proposed. Machine Learning models like Artificial Neural
Networks and Random Forests are trained to classify elements on a given webpage
based on visual cues. Visual properties of elements are analysed in detail in order
to derive meaningful numerical features for classification. After applying sensible
post-processing filters, the system is able to recall up to 80% of the desired data at
a precision of up to 90%. Relying purely on visual cues however has it’s limitations
for robust extraction of some of the citation data. Possible approaches to facilitate
that are discussed at the end.