How do we segment text? Two-stage chunking operation in reading

Chunking in language comprehension is a process that segments continuous linguistic input into smaller chunks that are in reader’s mental lexicon. Effective chunking during reading facilitates disambiguation and enhances efficiency for comprehension. However, the mechanisms of chunking remain elusive, especially in reading given that information arrives simultaneously yet the written systems may not have explicit cues for labeling boundaries such as Chinese. What are the mechanisms of chunking operation that mediates the reading of the text that normally contains hierarchical information? We investigated this question by manipulating the lexical status of the chunks at distinct levels of grain-size in four-character Chinese strings, including the two-character local chunk and four-character global chunk. Participants were asked to make lexical decision on these strings in a behavioral experiment, followed by a passive reading task when their electroencephalography (EEG) were recorded. The behavioral results showed that the lexical decision time of lexicalized two-character local chunks was influenced by the lexical status of four-character global chunk, but not vice versa, which indicated that the processing of global chunks possessed priority over the local chunks. The EEG results revealed that familiar lexical chunks were detected simultaneously at both levels and further processed in a different temporal order -- the onset of lexical access for the global chunks was earlier than that of local chunks. These consistent behavioral and EEG results suggest that chunking in reading occurs at multiple levels via a two-stage operation -- simultaneous detection and global-first recognition. Significance Statement The learners of a new language often read word by word. But why can proficient readers read multiple words at a time? The current study investigates how we efficiently segment a complicate text into smaller pieces and how we process these pieces. Participants read Chinese strings with different structures while their key-press responses and brain EEG signals were recorded. We found that texts were quickly (about 100 ms from their occurrences) segmented to varied sizes of pieces, and larger pieces were then processed earlier than small pieces. Our results suggest that readers can use existing knowledge to efficiently segment and process written information.

Jinbiao Yang (杨金骉) 1,2,3,4,5 , Qing Cai (蔡清) 1,5 , Xing Tian (田兴) 1,4,5 4 1 NYU-ECNU Institute of Brain and Cognitive Science at NYU Shanghai, Shanghai, 5 China 6 2 Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands 7 3 Centre for Language Studies Nijmegen, Radboud University, Nijmegen, The 8 Netherlands 9  level. The four types of stimuli are listed in Table 1.  We selected and created all stimuli with following steps. We extracted the 1 4 7 GwLw and GwLn stimuli from a database of Sogou Pinyin (a popular product http://chardb.science.ru.nl/). All the GwLw and GwLn stimuli satisfied the 1 5 1 following criteria at the global level:1) noun 1 ; 2) high-frequency 2 ; and 3) no 1 5 2 duplicative characters (e.g., "高高兴兴", translation: happy). In addition, the 1 5 3 GwLw stimuli satisfied the following criteria at the local level: 1) both two-1 5 4 character words were nouns, and 2) high-frequency words. Moreover, the 1 5 5 lexicality of GwLn stimuli at the local level were verified by checking the first In each trial, participants were first asked to focus on a cross presented at the 1  8  0   3   T  h  e  c  h  a  r  a  c  t  e  r  '  s  l  o  g  f  r  e  q  u  e  n  c  y  i  s  d  e  t  e  r  m  i  n  e  d  b  y  t  h  e  S  u  b  t  i  t  l  e  D  a  t  a  b  a  s  e  (  C  a  i  &  B  r  y  s  b  a  e  r  t  ,  2  0  1  0  ) .
center of the screen. After 400 ms the fixation cross disappeared, and a 4- intervals were randomly selected from a range from 800 to 1000 ms. Four stimuli types (GwLw, GwLn, GnLw, GnLn) were fully crossed with task 1 9 2 types (global task vs. local task) and yield 8 conditions. 320 trials were package -Expy (https://github.com/ray306/expy), which is a software for 1 9 7 presenting and controlling psychological experiments. Behavioral data analysis 2 0 0 All participants had response accuracy exceeding 85%, and the average of lexicality, local-level lexicality, and task, followed by planned t-tests for testing 2 0 5 specific hypotheses. The same group of subjects participated in the EEG experiment. The EEG 2 0 9 experiment shared the same stimuli list with the behavioral experiment, but 2 1 0 both the procedure and the task are different. First, the display of each 2 1 1 character string lasted for 300ms. Participants were asked to read the 2 1 2 underlined parts of the stimuli (in order to keep their attention on the stimuli), 2 1 3 but they did not perform any lexical decision task. We used all 320 strings with 2 1 4 80 for each stimuli type in the global task and repeated once in the local task. Moreover, 320 four-symbol strings were included as the visual baseline in the 2 1 6 EEG experiment. The symbols in a symbol string trial were randomly sampled 2 1 7 with replacement from four symbols ("□", "△", " ", and "○"). Underlines were stimuli, we randomly inserted strings of digits for 100 ms and participants 2 2 2 were asked to report the underlined digits by pressing number buttons on a 2 2 3 keyboard. About forty-eight number-report trials were presented to each 2 2 4 participant. were used to monitor vertical eye movements. Electrode impedances were 2 3 0 kept below 10 kΩ. Data were continuously recorded in single DC mode. Data 2 3 1 were sampled at 500 Hz, online referenced to the Cz. shuffled the condition labels on the subjects' ERPs using EasyEEG toolbox  The test of TANOVA involved comparisons on multiple timebins, we corrected Planned post-hoc T-tests were further carried out in each factor to specify the information affect processing at the local level (Fig. 1A). In the local task, the 3 4 0 reaction time in GnLw was significantly longer than that in GwLw the global level (Fig. 1B). In the global task, we didn't find significant information at the local level may leak through to the processing of global 3 5 7 chunks and influence the decision of nonwords. We further test this parallel 3 5 8 processing dynamics in EEG experiment. The reaction times were not different between the trials with underlines either stimuli that were relevant to task did not affect response speed. in the same color but at different time points indicate that they are grouped into the same 3 7 8 cluster --sharing similar features but occurring at different times. The temperature of 3 7 9 colors represents the rank of the cluster distance relative to the Cluster baseline (cluster 3 8 0 defined by the baseline period). About 80 ms after stimulus onset, a novel cluster 3 8 1 (Cluster 2nd) appears at the same time across 5 conditions, followed by another new 3 8 2 cluster. However, in the symbol condition the Cluster 2nd appears earlier with much 3 8 3 shorter duration than 4-character string conditions. 3 8 4 3 8 5 We first carried out the clustering analysis to explore the dynamics of ERP formed a short gap that broke the early processing into two stages. We To test the hypothesis about the lexical detection in the earliest stage, we comparison correction (Fig. 3a). However, the difference topographies and lower on occipital area, highlighted in a red box in Fig. 3a). Therefore, we In the clustering results (Fig. 2), a 'temporal gap' was observed in the early and local levels (Fig. 3). These findings are consistent with the early lexical accessing the semantics. In other words, the familiar lexical chunks are 5 2 0 detected before subsequent processes (e.g. semantic retrieval). This is 100ms and extend to multiple chunk levels. What factor enables this early chunk detection in reading? Top-down 5 2 7 mechanisms have been proposed to account for perceptual and cognitive projection of low spatial frequency information (M. Bar et al., 2006). In perception of these physical attributes vary in a great degree across individuals. Therefore, the factor that leads to the early chunk detection are 5 4 1 likely to be the perceptual consequences --the familiarity of these attributes.

4 2
In fact, the familiarity has been demonstrated in improving language retrieval 5 4 3 (Bannard & Matthews, 2008;Zheng, Li, & Xiao, 2015). In this study, we Our behavioral results demonstrated that the processing of a global chunk processing of global chunks preceded that of local chunks. The priority of global information has been demonstrated in many cognitive is hierarchical and global processing has priority over local processing; while recognition also implies the activation of high-level information will be faster 5 7 3 than the lower-level information (M. Bar et al., 2006;Moshe Bar, 2003). In level interacts with the letter identification (McClelland & Rumelhart, 1981a).

7 7
This study further demonstrates the influences of phrases on words. Our mechanism can be applied across multiple levels in a hierarchical manner in information and less internal entropy, which can prevent ambiguity.

8 3
Paralleled processing of chunks at both levels 5 8 4 The behavioral results revealed that the judgement of a non-lexicalized 5 8 5 phrase at the global level was harder when the task-unrelated chunks were 5 8 6 familiar words at the local level. This indicated that the local processing may 5 8 7 be initiated before the finish of global processing. The EEG results further 5 8 8 supported that processing at both levels temporally overlapped --the 5 8 9 response patterns of processing global chunks continued after the start of 5 9 0 local processing responses patterns (Fig. 4). This observation of partially 5 9 1 temporal overlap in the processing part-whole hierarchies is consistent with 5 9 2 simultaneous processing mechanisms implemented in the connectionist global level, whereas both hemispheres engaged in processing chunks at the 5 9 8 local level (Fig. 4), suggesting the possible anatomical differences that 5 9 9 mediate the partially temporal paralleled processes at both levels. Based on all results, we tentatively put forward a workflow of processing 6 0 2 multiple-level information in reading (Fig. 5). The segmentation occurs in an processes of chunks at two levels have partially temporal overlap that enables 6 1 0 interaction across levels before final integration. The current study investigated the chunking mechanism in reading.

1 7
Consistent behavioral and EEG results suggested that multiple levels of