Using Text Segmentation Algorithms for the Automatic Generation of E-Learning Courses

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 132–140, Dublin, Ireland, August 23-24 2014. Using Text Segmentation Algorithms for the Automatic Generation of E-Learning Courses Can Beck, Alexander Streicher and Andrea Zielinski Fraunhofer IOSB Karlsruhe, Germany {can.beck, alexander.streicher, andrea.zielinski}@iosb.fraunhofer.de Abstract With the advent of e-learning, there is a strong demand for tools that help to cre- ate e-learning courses in an automatic or semi-automatic way. While resources for new courses are often freely available, they are generally not properly structured into easy to handle units. In this paper, we investigate how state of the art text segmentation algorithms can be applied to automatically transform unstructured text into coherent pieces appropriate for e-learning courses. The feasibility to course generation is validated on a test corpus specifically tailored to this scenar- io. We also introduce a more generic training and testing method for text seg- mentation algorithms based on a Latent Dirichlet Allocation (LDA) topic model. In addition we introduce a scalable ran- dom text segmentation algorithm, in or- der to establish lower and upper bounds to be able to evaluate segmentation re- sults on a common basis. 1 Introduction The creation of e-learning courses is generally a time consuming effort. However, separating text into topically cohesive segments can help to re- duce this effort whenever textual content is al- ready available but not properly structured ac- cording to e-learning standards. Since these seg- ments textually describe the content of learning units, automatic pedagogical annotation algo- rithms could be applied to categorize them into introductions, descriptions, explanations, exam- ples and other pedagogical meaningful concepts (K.Sathiyamurthy & T.V.Geetha, 2011). Course designers generally assume that learn- ing content is composed of small inseparable learning objects at the micro level which in turn are wrapped into Concept Containers (CCs) at the macro level. This approach is followed, e.g., in the Web-Didactic approach by Swertz et al. (2013) where CCs correspond to chapters in a book and Knowledge Objects (KOs) correspond to course pages. To automate the partition of an unstructured text source into appropriate seg- ments for the macro and micro level we applied different text segmentation algorithms (segment- ers) on each level. To evaluate the segmenters in the described scenario, we created a test corpus based on fea- tured Wikipedia articles. For the macro level we exploit sections from different articles and the corresponding micro structure consists of subse- quent paragraphs from these sections. On the macro level the segmenter TopicTiling (TT) by Riedl and Biemann (2012) is used. It is based on a LDA topic model which we train based on the articles from Wikipedia to extract a predefined number of different topics. On the micro level, the segmenter BayesSeg (BS) is applied (Eisenstein & Barzilay, 2008). We achieved overall good results measured in three different metrics over a baseline approach, i.e., a scalable random segmenter, that indicate This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added …

Content truncated to 500 words