The role of textual semantic constraints in knowledge-based inference generation during reading comprehension: A computational approach

Abstract

The present research adopted a computational approach to explore the extent to which the semantic content of texts constrains the activation of knowledge-based inferences. Specifically, we examined whether textual semantic constraints (TSC) can explain (1) the activation of predictive inferences, (2) the activation of bridging inferences and (3) the higher prevalence of the activation of bridging inferences compared to predictive inferences. To examine these hypotheses, we computed the strength of semantic associations between texts and probe items as presented to human readers in previous behavioural studies, using the Latent Semantic Analysis (LSA) algorithm. We tested whether stronger semantic associations are observed for inferred items compared to control items. Our results show that in 15 out of 17 planned comparisons, the computed strength of semantic associations successfully simulated the activation of inferences. These findings suggest that TSC play a central role in the activation of knowledge-based inferences.

Keywords:

• Knowledge-based inferences
• Bridging and predictive inferences
• Textual semantic constraints
• Semantic associations
• Associative and strategic processes
• Latent semantic analysis

This research was supported by the European Union [grant number PIEF-GA-2009-254607-IEF] to Menahem Yeari. Facilities for conducting the research were provided by the Brain and Education Lab, the Leiden Institute for Brain and Cognition, and by the Department of Education in Leiden University. We thank Celia M. Klin, Anne E. Cook, Tracy Linderholm, Murray Singer, Connie Shears, and Marc J. Beeman for sharing their data with us.

This research was supported by the European Union [grant number PIEF-GA-2009-254607-IEF] to Menahem Yeari. Facilities for conducting the research were provided by the Brain and Education Lab, the Leiden Institute for Brain and Cognition, and by the Department of Education in Leiden University. We thank Celia M. Klin, Anne E. Cook, Tracy Linderholm, Murray Singer, Connie Shears, and Marc J. Beeman for sharing their data with us.

Notes

¹ Experiments 1a and 1b differed in the duration of the inter stimulus interval (ISI), which did not influence the behavioural results and is beyond the scope of the current simulations.

² All presented means were averaged across ISIs.

³ We counted the two misses found in the simulation of Linderholm’s (2002) data as one miss because the present computational procedure cannot produce different predictions for the low-span and high-span readers.

⁴ One outlier item which diverged from the mean difference by 3.15 SD was excluded from the analysis.

⁵ Shears and Chiarello (2004) found this difference between causal and motive bridging inferences with individuals who had survived brain injuries; Shears et al. (2007) found it under conditions of high cognitive load using a concurrent task.

⁶ Beeman et al. (2000) studied the activation of bridging and predictive inferences separately in each hemisphere (using a divided visual field paradigm). Because the simulation of hemispheric differences is beyond the scope of the present research, the present simulations were not intended to predict Beeman et al.’ s results.

⁷ Being preceded by a larger text segment per se, as in the case of bridging inferences (which were always probed one sentence earlier than the two predictive inferences), is not sufficient to produce greater cosines with the preceding text. To produce greater cosines, these extra text segments should be at least as semantically related to the target inference as the preceding (shared) text segments. This claim receives some support from the finding that the cosine mean at the first predictive test point (M_P1 = 0.092) was higher than the cosine mean at the second predictive test point (M_P2 = 0.083), although the second predictive test point was always preceded by a larger text segment.

⁸ We counted the failure to simulate the results in the one-sentence version of Singer and Halldorson’s (1996) experiment as a failure to simulate the activation of predictive inferences.

⁹ Text length is not the only factor that influences the strength of semantic constraints to the inference activation. For example, the extent to which the text is focused on the event that relates to the probed inference is another factor.