Klexikon: A German Dataset for Joint Summarization and Simplification

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2693–2701 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2693 Klexikon: A German Dataset for Joint Summarization and Simplification Dennis Aumiller and Michael Gertz Institute of Computer Science, Heidelberg University Im Neuenheimer Feld 205, Heidelberg, Germany {aumiller, gertz}@informatik.uni-heidelberg.de Abstract Traditionally, Text Simplification is treated as a monolingual translation task where sentences between source texts and their simplified counterparts are aligned for training. However, especially for longer input documents, summarizing the text (or dropping less relevant content altogether) plays an important role in the simplification process, which is currently not reflected in existing datasets. Simultaneously, resources for non-English languages are scarce in general and prohibitive for training new solutions. To tackle this problem, we pose core requirements for a system that can jointly summarize and simplify long source documents. We further describe the creation of a new dataset for joint Text Simplification and Summarization based on German Wikipedia and the German children’s encyclopedia ”Klexikon”, consisting of almost 2,900 documents. We release a document-aligned version that particularly highlights the summarization aspect, and provide statistical evidence that this resource is well suited to simplification as well. Code and data are available on Github: https://github.com/dennlinger/klexikon Keywords: Summarization, Text Simplification, German 1. Introduction The goal of Text Simplification (TS) is to produce easily understandable texts that benefit disadvantaged readers such as children, dyslexic, or language learn- ers. Simplifications are often generated by adapting a source text written for adult/native readers. However, recent work in simplification has mostly addressed TS as a monolingual translation task, where individ- ual sentences are ”translated” into a simplified version (Zhu et al., 2010; Coster and Kauchak, 2011; Hwang et al., 2015). The main focus is put on either lexi- cographic replacements, paraphrasing, sentence split- ting, or the dropping of words within a single sentence (Amancio and Specia, 2014), which implies that the simplification of any input document will consist of roughly the same number of sentences. While this ap- proach is appropriate for sufficiently short source doc- uments, longer articles become strenuous for disadvan- taged readers. As can be seen in Table 1, articles in different corpora come with varying lengths of their re- spective source texts. When simplifications are gener- ated via manual sentence-by-sentence translations, the simplified texts tend to have more sentences than the source documents. When alignments are constructed from a source and simplification text on the same topic instead, they exhibit a drastic length disparity. Current simplification systems are, however, inherently limited in their ability to address the problem of joint simplification and summarization from much longer in- put documents. Sentence-level alignments were tradi- tionally seen as one way to circumvent certain prob- lems in TS, namely: 1. Human feedback for judging simplification qual- ity is more consistent for sentences, compared to longer samples, such as entire documents. Aligned Avg. #Sentences Resource Articles Source Simple Klexikon (Ours) 2,898 242.09 32.51 (Hewett and Stede, 2021) 978 10.12 43.54 (Battisti …

Content truncated to 500 words