Automatically evaluating the conceptual complexity of German texts

Automatically evaluating the conceptual complexity of German texts Freya Hewett 12 1 Humboldt Institute for Internet and Society, Berlin, Germany freya.hewett@hiig.de Manfred Stede 2 2 Applied Computational Linguistics University of Potsdam, Germany stede@uni-potsdam.de Abstract Conceptual complexity is concerned with the background knowledge needed to understand concepts within a text and their implicit con- nections (Hulpus, et al., 2019). In the present study, a recently proposed framework from Hulpus, et al. (2019), which assesses the con- ceptual complexity of English newspaper arti- cles, is replicated and adapted to German lex- ica entries aimed at three different age groups. The final results on the corpus of 885 German texts improve upon the original study in both a pairwise classification task and a ranking task, showing that the framework transfers well to a different language and a different genre. We release the dataset used, as well as an extended version with a total of ca. 3000 texts. 1 Introduction Text simplification aims to reduce the complexity of a text whilst retaining the main informational content. Conceptual complexity is concerned with the background knowledge needed to understand concepts within a text, and the implicit connections between the concepts that contribute to understand- ing a text (Hulpus, et al., 2019). The present study aims to evaluate the conceptual complexity of Ger- man texts, by recreating a recent study from Hulpus, et al. (2019) in which they assess the conceptual complexity of English language newspaper articles from the Newsela corpus (Xu et al., 2015), which contains articles at five different levels of complex- ity. To do this, they develop a framework which is based on psycholinguistic theories on reading com- prehension, in particular priming, which states that words are recognised faster if preceded by words related in meaning (Collins and Loftus, 1975). In the present study, this framework is directly applied to German texts from three lexica designed for beginner readers, children and adults. The framework is then slightly adapted to account for nuances specific to the German language, such as compound words. The results show that the model adapts well to German texts and works well across domains. We also release the lexica dataset to fos- ter research on German text simplification, and a script to build the dataset as the lexica grow.1 2 Background The main hypothesis in Hulpus, et al.’s (2019) study is that the more priming in a text, the lower the con- ceptual complexity. A spreading activation (SA) framework (Quillian, 1962, 1967; Collins and Lof- tus, 1975) is used to illustrate the priming pro- cess. The framework compares concepts to nodes in a network, with the properties of concepts rep- resented as labelled relational links from the node to other concept nodes. Whenever a concept is mentioned in a text, it activates other neighbouring concepts in the graph (Collins and Loftus, 1975). The amounts of activation generated by this pro- cess are used to symbolise the amount of priming in the text. The rest of this section provides a summary of …

Content truncated to 500 words