Crowdsourced Corpus of Sentence Simplification with Core Vocabulary

Crowdsourced Corpus of Sentence Simplification with Core Vocabulary Akihiro Katsuta, Kazuhide Yamamoto Nagaoka University of Technology 1603-1, Kamitomioka Nagaoka, Niigata 940-2188, JAPAN {katsuta,yamamoto}@jnlp.org Abstract We present a new Japanese crowdsourced data set of simplified sentences created from more complex ones. Our simplicity standard involves all rewritable words in the simplified sentences being drawn from a core vocabulary of 2,000 words. Our simplified corpus is a collection of complex sentences from Japanese textbooks and reference books together with simplified sentences generated by humans, paired with data on how the complex sentences were paraphrased. The corpus contains a total of 15,000 sentences, in both complex and simple versions. In addition, we investigate the differences in the simplification operations used by each annotator. The aim is to understand whether a crowdsourced complex-simple parallel corpus is an appropriate data source for automated simplification by machine learning. The results, that there was a high level of agreement between the annotators building the data set. So, we believe that this corpus is a good quality data set for machine learning for simplification. We therefore plan to expand the scale of the simplified corpus in the future. Keywords: Corpus, Crowdsourcing, Simplification 1. Introduction Simplification task is the process of rewriting a complex text into a simpler form while preserving its meaning. Sim- plified texts play an important role in providing accessi- ble and easy-to-understand information for a wide range of users who find it difficult to understand texts that have not been simplified due to their linguistic complexity. At- tempts have been made to automate the simplification pro- cess for various languages, including English, Spanish, Brazilian Portuguese, and Portuguese. We have been con- ducted research on simplification since a few years ago (Moku et al., 2012). Recent studies have treated text sim- plification as a monolingual machine translation problem in which simple and synonymous sentences are generated using statistical machine translation (Wubben et al., 2012; Xu et al., 2016). As with statistical machine translation using bilingual par- allel corpora, text simplification therefore requires a mono- lingual parallel corpus for training. In the case of English, there are PWKP (Zhu et al., 2010) and Wikipedia Datasets (Coster and Kauchak, 2011; Kauchak, 2013) made from wiki and Swiki. In German, a simplification corpus with a scale of 7,000 sentences (Caseli et al., 2009), In Italian the PaCCSS-IT (Brunato et al., 2016) with a scale of 63,000 sentences, In Spanish a simplification corpus do exist made by hand by the few rules (Mitkov and ˇStajner, 2014; ˇStajner et al., 2015), and so on. However, only English corpora are publicly available, such as a large-scale simplified English corpus obtained from pairs of Wikipedia and Simple English Wikipedia entries. For Japanese, there is no other than our simplifi- cation corpus (Maruyama and Yamamoto, 2018). We have already conducted the experiment of simplification using corpus (Maruyama and Yamamoto, 2017). Currently, most of the simplification research that has been done for Japanese has emphasized paraphrasing word units (Kajiwara et al., 2013; Kajiwara and Yamamoto, …

Content truncated to 500 words