Text extraction in the field of computational linguistics often encounters a variety of problems. Real readers provide the basis for natural language processing. Reader's cognitive processing is influenced by text structures. The current paper controls the text structure and combines the method of eye tracking and machine learning to investigate readers’ cognitive mechanism of text comprehension with two studies(four experiments). The first study examined how text structures influenced text comprehension processing. The second study examined the relationship between text structures and eye movement measures to further explore the text reading cognitive processes and their computational modeling.
The first study contains two experiments for revealing the structure effect during text reading comprehension and its cognitive processes.
Experiment 1 controlled the thematic representative structure of texts and investigated the interactive effect between text structure and repeat task. It is found that the position of topic sentence influenced readers’ eye movement behavior, moderated by repeat task. Readers processed texts strategically during repeat period. The result supported situation model theory, a high hierarchical situation model constructed and facilitated subsequent processing during text comprehension. The repeat effect also supported context-dependent representation model.
Experiment 2 manipulated the thematic and refutation structures of the texts, recording the eye-movement trajectories of 68 readers on reading the 32 structure-manipulated texts. The results showed that the refutation structure improved reading efficiency by reducing the total reading time and looking back at areas of misconception and increasing the rereading time of scientific concepts, implying that conceptual change altered the readers' online processing strategies. The effect of the refutation structure was also moderated by the position of the topic sentence, with the effect significant only in the condition where the topic sentence was in its initial position and not in the condition where the topic sentence was in its final position, suggesting that topic sentences placed in the initial position helped readers to construct a model of the textual context. Readers conceptual change with a complete situation model in mind reduced their cognitive load and reading time.
Experiment 2 manipulated the thematic and refutation structures of the texts, recording the eye-movement trajectories of 68 readers on reading the 32 structure-manipulated texts. The results showed that the refutation structure improved reading efficiency by reducing the total reading time and looking back at areas of misconception and increasing the rereading time of scientific concepts, implying that conceptual change altered the readers' online processing strategies. The effect of the refutation structure was also moderated by the position of the topic sentence, with the effect significant only in the condition where the topic sentence was in its initial position and not in the condition where the topic sentence was in its final position, suggesting that topic sentences placed in the initial position helped readers to construct a model of the textual context. Readers conceptual change with a complete situation model in mind reduced their cognitive load and reading time.
Experiment 3 used the eye-movement data of readers reading four- structure discourses in Experiment 2, compressed the eye-movement features with the Lasso algorithm and used support vector machine to test the prediction effect, showing that the eye-movement data can predict the text structure, supporting the findings of Experiment 2 in reverse. It shows that reading processing is moderated by discourse structure, and the eye-movement measures that are most sensitive to distinguishing discourse structures were successfully screened.
Experiment 4 collected eye-movement data from 1200 discourses in the weibo corpus and used methods of cognitive modelling to construct the cognitive process of readers when reading texts. High similarity score refers to high eye movement trajectory similarity. Then we used Hidden Markov chain to model eye movement trajectory within claused and then calculate the transfer matrix. The results show that both the structure and position of the clauses and sentences in a natural discourse can influence the reader's processing.
From the studies, we examined that the process of reading a text is moderated by text structures, which can enable readers to read the text more strategically and improve teading efficiency. This study uses machine learning methods to improve the reliability and validity of this study, and also provides new inspiration for future psycholinguistic analysis methods.
修改评论