Between Reading Time and Syntactic/Semantic Categories

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 404–412, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Between Reading Time and Syntactic/Semantic Categories Masayuki Asahara † and Sachi Kato ‡ National Institute for Japanese Language and Linguistics, Japan †, ‡ masayu-a@ninjal.ac.jp † Abstract This article presents a contrastive anal- ysis between reading time and syn- tactic/semantic categories in Japanese. We overlaid the reading time annota- tion of BCCWJ-EyeTrack and a syntac- tic/semantic category information anno- tation on the ‘Balanced Corpus of Con- temporary Written Japanese’. Statistical analysis based on a mixed linear model showed that verbal phrases tend to have shorter reading times than adjectives, ad- verbial phrases, or nominal phrases. The results suggest that the preceding phrases associated with the presenting phrases promote the reading process to shorten the gazing time. 1 Introduction Most of the studies on sentence processing by humans are based on confirmatory data analy- sis. The methodology involves developing a hy- pothesis, constructing sample sentences, includ- ing the target language phenomena, and perform- ing a psycholinguistic experiment, such as record- ing reading time or event-related potentials. In recent times, the ‘Balanced Corpus of Contem- porary Written Japanese’ (hereafter ‘BCCWJ’) (Maekawa et al., 2014) was compiled and pub- lished. The reading time annotation on BCCWJ: BCCWJ-EyeTrack (Asahara et al., 2016) is avail- able for the linguistic research community. The data in the BCCWJ enable us to perform ex- ploratory data analysis in fair and reproducible en- vironments. We measured the readability of humans. More concretely, we performed a contrast comparison between reading time and syntactic/semantic cat- egories of words. We prepared the annotation of word senses on BCCWJ based on ‘Word List by Semantic Principles’ (国立国語研究所, 1964, 2004). The original WLSP label annotation is on both short unit words and long unit words in the BCCWJ. We then mapped these annotations into Bunsetsu(base phrase)-units. The statistical analysis using a mixed linear model shows that verbal phrases tend to have shorter reading times than adjective/adverbial phrases or nominal phrases. Section 2 presents the related research. Section 3 shows the data and methods. Section 4 presents the results, and Section 5 is the discussion. Sec- tion 6 concludes this article and presents the im- plications of our current work and the future work we plan to conduct. 2 Related Work First, we present related work on eye tracking. The Dundee Eyetracking Corpus (Kennedy and Pynte, 2005) contains reading times for English and French newspaper editorials from 10 native speak- ers of each language that were recorded us- ing eye-tracking equipment. The corpus does not target a specific set of linguistic phenom- ena but instead provides naturally occurring texts for testing diverse hypotheses. For exam- ple, (Demberg and Keller, 2008) used the cor- pus to test Gibson’s dependency locality the- ory (DLT) (Gibson, 2008) and Hale’s surprisal theory (Hale, 2001). The corpus also allows for replications to be conducted; for example, (Roland et al., 2012) concluded that previous anal- yses (Demberg and …

Content truncated to 500 words