Summarization Corpora of Wikipedia Articles

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6651–6655 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6651 Summarization Corpora of Wikipedia Articles Dominik Frefel University of Applied Sciences Northwestern Switzerland Institute of Data Science, Bahnhofstrasse 6, 5210 Windisch dominik.frefel@fhnw.ch Abstract In this paper we propose a process to extract summarization corpora from Wikipedia articles. Applied to the German language we create a corpus of 240 000 texts. We use ROUGE scores for the extraction and evaluation of our corpus. For this we provide a ROUGE metric implementation adapted to the German language. The extracted corpus is used to train three abstractive summarization models which we compare to different baselines. The resulting summaries sound natural and cover the input text very well. The corpus can be downloaded at https://github.com/domfr/GeWiki. Keywords: Machine Learning, Natural Language Processing, Summarization, Corpus 1. Introduction Advances in deep learning have led to the development of a variety of sequence-to-sequence models (Sutskever et al., 2014; Bahdanau et al., 2014) to solve natural language tasks. Rush et al. (2015) were the first to apply these mod- els to abstractive summarization. More recently, the Trans- former model has reached state of the art performance in various natural language processing tasks (Vaswani et al., 2017) and has successfully been used for abstractive sum- marization (Gehrmann et al., 2018; Hoang et al., 2019). These models require a large number of text samples for the training and evaluation. Large summarization corpora are only sparsely available and most research is focused on the English language. We explore a process to extract summarization data for other languages. Giannakopoulos et al. (2015) use Wikipedia featured articles to evaluate on multiple languages, but their data is not sufficient to train an abstractive model. Based on their idea, we propose a method to extract summarization data from Wikipedia articles and apply it to the German language. This paper is structured as follows: Section 2 describes our method of corpus construction. In section 3 we present our German corpus and in section 4 our experiments and re- sults. Section 5 contains conclusions. 2. Corpus construction 2.1. Source Featured Wikipedia articles are required to have a summa- rizing introduction of the main body (Wikipedia-Featured- articles, 2019). There are only about 2000 featured articles written in German, which is not enough to train deep learn- ing models. However, numerous ordinary articles have a summarizing introduction as well. Our goal is to identify texts with introductions that can be used as a summary of the rest of the article. We hypothesize that useful texts can be selected using only a few metrics, namely: Compression ratio, summary length and ROUGE scores (α = 0)1 of the introduction. 1With α = 0 ROUGE measures the recall scores 2.2. German Rouge Metric We define a ROUGE implementation2 adapted to the Ger- man language following the work of Lin (2004). It applies a number of preprocessing steps: • Split the given text into sentences • …

Content truncated to 500 words