Between Reading Time and Information Structure

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 15–24 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Masayuki Asahara Between Reading Time and Information Structure Masayuki Asahara National Institute for Japanese Language and Linguistics, National Institutes for the Humanities, Japan masayu-a@ninjal.ac.jp Abstract This paper presents a contrastive analysis be- tween the reading time and information struc- ture in Japanese. We overlaid the reading time annotation BCCWJ-EyeTrack and an infor- mation structure annotation on the Balanced Corpus of Contemporary Written Japanese. Statistical analysis based on a mixed linear model showed that the “specificity,” “sen- tience,” and “commonness” of the Japanese information structure affect the reading time. These three characteristics produce different patterns of delay in the reading time. Espe- cially, the reading time patterns differ depend- ing on the commonness such as new informa- tion or bridging. The results suggest that new information and bridging can be classified by the reading time pattern. 1 Introduction Information structures1 such as definiteness and specificity affect the selection of articles. However, some languages such as Japanese and Russian do not have articles for noun phrases (NPs). The definite- ness and specificity in such languages are not overtly marked in their surface form. Nagata et al. (Nagata et al., 2005) proposed a sta- tistical model to detect article errors made by En- glish learners. They constructed the model by using an enormous text produced by English native speak- ers. However, when we consider the choice of arti- cles in some languages by native speakers of a lan- guage with no articles, careful attention must be paid 1We focus on the information status which is a fundamental notion to define the theme and rheme in the information struc- ture. to the information structures in their NPs. Moreover, when texts in language without articles are trans- lated by humans or machine translation systems into a target language with articles, we should consider the information structure of the source language. Other aspects of information structures are the information status and commonness. The infor- mation status indicates whether the co-referred mention appears in the preceding discourse (discourse-old) or not (discourse-new). Commonness is defined as whether the speaker assumes that the hearer already knows the informa- tion (hearer-old), can estimate the information (bridging), or does not know the information hearer-new. Bridging reference cannot be resolved purely from the surface forms of written texts. Previous research (Hou et al., 2013) tried to resolve bridging references by using annotated corpora and world knowledge. However, bridging is an information structure for the language recipient (hearer). It should incorporate recipient features such as the reading time to estimate whether an NP is bridging or not (i.e., hearer-new). Moreover, identifying if readers can resolve a bridging reference with their own knowledge is important for user-oriented information extraction and document summarization. These information structures are correlated with animacy, sentience, and agentivity. The features of information structures may affect the reading time; there are various ways to monitor the reading time, such as …

Content truncated to 500 words