--- annotations_creators: - found - expert-generated language_creators: - found - machine-generated language: - de license: - cc-by-sa-4.0 multilinguality: - monolingual size_categories: - 1K` tags, which excludes lists, captions and images. * `klexikon_text` (`List[str]`): List of sentences of the Klexikon article. We apply the same processing as for the Wikipedia texts. ### Data Splits We provide a stratified split of the dataset, based on the length of the respective Wiki article/Klexikon article pair (according to number of sentences). The x-axis represents the length of the Wikipedia article, and the y-axis the length of the Klexikon article. We segment the coordinate systems into rectangles of shape `(100, 10)`, and randomly sample a split of 80/10/10 for training/validation/test from each rectangle to ensure stratification. In case of rectangles with less than 10 entries, we put all samples into training. The final splits have the following size: * 2350 samples for training * 274 samples for validation * 274 samples for testing ## Dataset Creation ### Curation Rationale As previously described, the Klexikon resource was created as an attempt to bridge the two fields of text summarization and text simplification. Previous datasets suffer from either one or more of the following shortcomings: * They primarily focus on input/output pairs of similar lengths, which does not reflect longer-form texts. * Data exists primarily for English, and other languages are notoriously understudied. * Alignments exist for sentence-level, but not document-level. This dataset serves as a starting point to investigate the feasibility of end-to-end simplification systems for longer input documents. ### Source Data #### Initial Data Collection and Normalization Data was collected from [Klexikon](klexikon.zum.de), and afterwards aligned with corresponding texts from [German Wikipedia](de.wikipedia.org). Specifically, the collection process was performed in April 2021, and 3145 articles could be extracted from Klexikon back then. Afterwards, we semi-automatically align the articles with Wikipedia, by looking up articles with the same title. For articles that do not exactly match, we manually review their content, and decide to match to an appropriate substitute if the content can be matched by at least 66% of the Klexikon paragraphs. Similarly, we proceed to manually review disambiguation pages on Wikipedia. We extract only full-text content, excluding figures, captions, and list elements from the final text corpus, and only retain articles for which the respective Wikipedia document consists of at least 15 paragraphs after pre-processing. #### Who are the source language producers? The language producers are contributors to Klexikon and Wikipedia. No demographic information was available from the data sources. ### Annotations #### Annotation process Annotations were performed by manually reviewing the URLs of the ambiguous article pairs. No annotation platforms or existing tools were used in the process. Otherwise, articles were matched based on the exact title. #### Who are the annotators? The manually aligned articles were reviewed by the dataset author (Dennis Aumiller). ### Personal and Sensitive Information Since Klexikon and Wikipedia are public encyclopedias, no further personal or sensitive information is included. We did not investigate to what extent information about public figures is included in the dataset. ## Considerations for Using the Data ### Social Impact of Dataset Accessibility on the web is still a big issue, particularly for disadvantaged readers. This dataset has the potential to strengthen text simplification systems, which can improve the situation. In terms of language coverage, this dataset also has a beneficial impact on the availability of German data. Potential negative biases include the problems of automatically aligned articles. The alignments may never be 100% perfect, and can therefore cause mis-aligned articles (or associations), despite the best of our intentions. ### Discussion of Biases We have not tested whether any particular bias towards a specific article *type* (i.e., "person", "city", etc.) exists. Similarly, we attempted to present an unbiased (stratified) split for validation and test set, but given that we only cover around 2900 articles, it is possible that these articles represent a particular focal lense on the overall distribution of lexical content. ### Other Known Limitations Since the articles were written independently of each other, it is not guaranteed that there exists an exact coverage of each sentence in the simplified article, which could also stem from the fact that sometimes Wikipedia pages have separate article pages for aspects (e.g., the city of "Aarhus" has a separate page for its art museum (ARoS). However, Klexikon lists content and description for ARoS on the page of the city itself. ## Additional Information ### Dataset Curators The dataset was curated only by the author of this dataset, Dennis Aumiller. ### Licensing Information Klexikon and Wikipedia make their textual contents available under the CC BY-SA license, which will be inherited for this dataset. ### Citation Information If you use our dataset or associated code, please cite our paper: ``` @inproceedings{aumiller-gertz-2022-klexikon, title = "Klexikon: A {G}erman Dataset for Joint Summarization and Simplification", author = "Aumiller, Dennis and Gertz, Michael", booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference", month = jun, year = "2022", address = "Marseille, France", publisher = "European Language Resources Association", url = "https://aclanthology.org/2022.lrec-1.288", pages = "2693--2701" } ```