Language comprehension models at different levels of explanation

Abstract

This paper reviews two recent models of language comprehension, the RI‐Val model (Cook & O'Brien, 2014; O'Brien & Cook, 2016b) and the dynamic generative framework (Kuperberg, 2016). Both of these models have been heavily informed by research on the processing of semantic anomalies, but they have arisen in different literatures. The RI‐Val model has roots in the text processing literature and focuses primarily on bottom–up aspects of comprehension; the dynamic generative framework has roots in the event‐related brain potential (ERP)/sentence processing literature and focuses more heavily on top–down aspects of comprehension. Relating these models to each other and considering the findings in the semantic anomaly literature that support each one pushes us to consider the possibility that these models might be describing the comprehension system at different levels of explanation (Marr, 1982). This view has a few important consequences. First, it provides new ways of thinking about, and consequently increases the importance of, the small literature about anomaly detection in fictional contexts. It also opens the door for beginning to speculate about potential mappings between these two models of language comprehension. Our hope is that by bringing together and highlighting potential connections between models from different traditions, we will provoke bigger picture points of view that raise new questions and spur new ideas for research.

The cheese was green with… envy: An EEG study on minimal fictional descriptions

“…Both the RI-Val model and the dynamic generative framework can accommodate the current results, furthering the idea that they offer complementary views on language comprehension, i.e., that a swap in the source of information makes readers update their beliefs accordingly, respectively. Colvin and Warren (2020) further suggest that these proposals tap into different levels of explanation (Marr, 1982). The RI-Val outlines part of the mechanisms (i.e., competition among memory sources and corresponding activation levels) through which the goal of the reader identified by the dynamic generative framework (i.e., building a generative model that best accounts for a given event) is executed.…”

“…Even though the dynamic generative framework also offers some insight on mechanisms (e.g., preactivation of semantic features), the authors argue that it is mostly concerned with the computational level, whereas the RI-Val model describes comprehension processes at the algorithmic level. Because our aim is to underline this complementarity, we will discuss the models together and in line with Colvin and Warren (2020) proposal.…”

“…Finally, as a complementary discussion, it is noteworthy to point out that recent language comprehension models have offered proposals as to how information that is inconsistent with general knowledge affects processing, and, importantly, on how a perspective switch (Colvin & Warren, 2020) can occur. For the Resonance-Integration-Validation Model (RI-Val;Cook & O'Brien, 2014), which has been chiefly investigated using eye tracker (e.g., Cook & O'Brien, 2014;Walsh et al, 2018), inconsistencies can affect processing through a swap between sources of information.…”

“…The dynamic generative framework draws on ERP literature on the effects of semantic anomalies and argues that the N400 reflects the ease to which the semantic features associated with an upcoming word were accessed, i.e., their degree of preactivation (e.g., Kuperberg, 2007;2016). Importantly, these models are complementary (e.g., Colvin & Warren, 2020), as changes in the degree of activation of different sources of information can inform readers' updating of preactivated features. However, the models do not specify how much information is required for readers to switch perspective.…”

The use of linguistic and world knowledge in language processing

“…One body of empirical work that looks at interactions between linguistic and world knowledge in language comprehension has focused on whether contextual manipulations can override processing difficulty when a comprehender encounters a highly unlikely word sequence like: the peanut danced or teach the bread or He lifted the lorry (see Colvin & Warren, 2020, for a recent review). This work lends itself naturally to a Venhuizen et al style model.…”