Between Reading Time and Clause Boundaries in Japanese - Wrap-up Effect in a Head-Final Language

Between Reading Time and Clause Boundaries in Japanese – Wrap-up Effect in a Head-Final Language Masayuki Asahara National Institute for Japanese Language and Linguistics, Japan Tokyo, Japan masayu-a@ninjal.ac.jp Abstract This paper presents a contrastive analysis be- tween reading time and clause boundary cate- gories in the Japanese language. We overlaid reading time data, made with BCCWJ Eye- Track, and clause boundary categories annota- tion on the Balanced Corpus of Contemporary Written Japanese. Statistical analysis based on the Bayesian linear mixed model shows that the reading time behaviours differ among the clause boundary categories. The result does not support the wrap-up effects of clause-final words. Another result we arrived at is that the predicate-argument relations facilitate the reading speed of native Japanese speakers. 1 Introduction Readability is an important issue for evaluation of sentence generation by both humans and machines. In general, readability should be evaluated by the reading time and the comprehension question for the stimuli. The reading time analysis studies evalu- ated the wrap-up effect (Just and Carpenter, 1980; Rayner et al., 2000) in English, in which readers tend to spend a longer time while reading clause-end phrases than clause-internal phrases. Our motivation is to evaluate whether the wrap-up effect at clause- end phrases occurs in Japanese or not. This paper explains how clause boundaries af- fect Japanese reading time using exploratory data analysis. We use reading time data from BCCWJ- EyeTrack (Asahara et al., 2016), which consists of the ‘Balanced Corpus of Contemporary Writ- ten Japanese’ (hereafter BCCWJ) (Maekawa et al., 2014) as the stimuli for the self-paced reading and eye tracking method. We overlaid the clause bound- ary category annotation data BCCWJ-ToriClause (Matsumoto et al., 2018) on the reading time data. Statistical analysis using the Bayesian linear mixed model (Sorensen et al., 2016) shows that the clause wrap-up effect (Just and Carpenter, 1980; Rayner et al., 2000) does not appear in our exper- iment for Japanese. The analysis also shows the difference among the clause boundary categories. For example, the rela- tive clause end phrases require a shorter reading time than the apposition clause end phrases; the noun clause end phrases have a shorter reading time than the quotation clause end phrases; and there are read- ing time differences between causal clause end and attendant circumstance clause end. Section 2 presents related work. Section 3 de- scribes the data. Section 4 shows statistical analysis method. Section 5 presents results with discussions. Finally, Section 6 presents our concluding remarks. 2 Related Work First, we present models of language analysis in- volving reading time or eye-tracking gaze infor- mation. (Barrett et al., 2016) presented a POS tagging model with gaze patterns. (Klerke et al., 2015) presented a grammaticality detection model for machine-processed sentences. Second, we present related work on eye-tracking corpora. The Dundee Eyetracking Corpus (Kennedy and Pynte, 2005) contains reading times for English and French newspaper editorials collected from 10 native speakers of each language, measured using PACLIC 32 19 32nd Pacific Asia Conference on Language, Information and …

Content truncated to 500 words