Reading Time and Vocabulary Rating in the Japanese Language: Large-Scale Japanese Reading Time Data Collection Using Crowdsourcing

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5178–5187 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5178 Reading Time and Vocabulary Rating in the Japanese Language: Large-Scale Reading Time Data Collection Using Crowdsourcing Masayuki Asahara National Institute for Japanese Language and Linguistics, Japan Tokyo University of Foreign Studies masayu-a@ninjal.ac.jp Abstract This study examined the effect of the differences in human vocabulary on reading time. This study conducted a word familiarity survey and applied a generalised linear mixed model to the participant ratings, assuming vocabulary to be a random effect of the participants. Following this, the participants took part in a self-paced reading task, and their reading times were recorded. The results clarified the effect of vocabulary differences on reading time. Keywords: Japanese, Reading Time, Vocabulary Rating 1. Introduction This study collected large-scale reading time data for the Japanese language using crowdsourcing. The En- glish language, Natural Story Corpus (NSC) (Futrell et al., 2018) was built by recruiting people to partic- ipate in Amazon Mechanical Turk to collect reading times. In a similar experiment, this study recruited par- ticipants through Yahoo! Japan’s crowdsourcing and utilised ibex1 to collect large-scale reading time data using a self-paced reading method via a web browser. In addition, this study conducted a word familiarity rat- ing experiment (Asahara, 2019) to evaluate word fa- miliarity and the subject’s vocabulary rating. The cor- relation between reading time and subject vocabulary size was estimated based on the collected reading time and vocabulary rating data. This study demonstrated a method for collecting reading times and presented the results of statistical analyses. The result shows that the vocabulary size are inversely correlated to the reading time. 2. Literature Review This study surveyed datasets of reading time data from the corpus using eye-tracking and self-paced reading methods. The Potsdam Sentence Corpus (Kliegl et al., 2004; Kliegl et al., 2006) is a reading time corpus of 144 German sentences (1138 words) read by 222 par- ticipants. The Dundee Eye-Tracking Corpus (Kennedy and Pynte, 2005) (Kennedy, 2003) contains reading times for English and French newspaper editorials from 10 native speakers for each language, recorded using eye-tracking equipment. The English version of the Dundee Eye-Tracking Corpus is composed of 20 edito- rial articles with 51,501 words. The Dutch DEMONIC database (Kuperman et al., 2010) consists of 224 Dutch sentences with reading times of 55 participants. The 1https://github.com/addrummond/ ibexfarm UCL Corpus (Frank et al., 2013) consists of 361 En- glish sentences drawn from amateur novels with self- paced reading and eye-tracking data. The sentences are presented without contexts. The Potsdam-Allahabad Hindi Eyetracking Corpus (Husain et al., 2015) con- sists of 153 Hindi-Urdu treebank sentences with eye- tracking data from 30 participants. The Ghent Eye- Tracking Corpus corpus (Cop et al., 2017) (Cop et al., 2007) contains monolingual and bilingual (L1 and L2) corpuses with eye-tracking data of participants reading a complete novel The Mysterious Affair at Styles by Agatha Christie. The Natural Stories corpus (Futrell …

Content truncated to 500 words