huPWKP: A Hungarian Text Simplification Corpus

Proceedings of Recent Advances in Natural Language Processing, pages 898–907 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_097 898 huPWKP: A Hungarian Text Simplification Corpus No´emi Pr´ot´ar E¨otv¨os Lor´and University protarnoemi@student.elte.hu D´avid M´ark Nemeskey E¨otv¨os Lor´and University Department of Digital Humanities nemeskey.david@btk.elte.hu Abstract In this article we introduce huPWKP, the first parallel corpus consisting of Hungarian stan- dard language–simplified sentence pairs. It is the Hungarian translation of PWKP (Zhu et al., 2010), on which we performed some cleaning in order to improve its quality. We evaluated the corpus both with the help of human evalu- ators and by training a seq2seq model on both the Hungarian and the original (cleaned) En- glish corpus. The Hungarian model performed slightly worse in terms of automatic metrics; however, the English model attains a SARI score close to the state of the art on the offi- cial PWKP set. According to the human eval- uation, the corpus performs at around 3 on a scale ranging from 1 to 5 in terms of informa- tion retention and increase in simplification and around 3.7 in terms of grammaticality. 1 Introduction The most important function and goal of hu- man communication is joint meaning construction (Tolcsvai Nagy, 2017): we want every person who participates in the discourse to understand the ref- erential scene (T´atrai, 2017, 2020) – i.e. what we are talking about – exactly (or as similarly as possi- ble) as we intended it to be understood. In order to achieve this, we sometimes need to simplify what we are saying and how we are phrasing it: mean- ing, we need to reduce ”the linguistic complexity of a text, while still retaining the original infor- mation content and meaning” (Siddharthan, 2014). Simplified texts can be of use for several groups of people, e. g. for people with (communicative or other) disabilities (Maaß and Rink, 2020; Maaß and Hernandez Garrido, 2020), non-native speakers (Paetzold, 2015) or children (De Belder and Moens, 2010). However, as text simplification is a fairly time- and resource-consuming task for humans, it seems beneficial to try to automate this task. There have been multiple successful attempts at creating text simplificaton systems: most of them for En- glish, e.g. Zhu et al. (2010) or Xu et al. (2016) or Xu et al. (2015). Less-resourced languages, such as Hungarian, have been largely ignored in the lit- erature. In this paper, we introduce the first (albeit translated) Hungarian parallel corpus consisting of standard language – simplified sentence pairs, as well as a simplification model trained on it. 2 Related work 2.1 Text simplification in NLP Text simplification (TS) is a fairly popular research area in NLP, especially for the English language. Most modern TS systems are capable of abstractive text simplification, meaning they can create new text on the basis of the original, usually on sentence- level units (Paetzold and Specia, 2017). The work of Nisioi et al. (2017) has brought a breakthrough in abstractive text simplification: they used a sequence-based model, originally designed for machine translation, using standard-language …

Content truncated to 500 words