Analyzing Reaction Time and Error Sequences in Lexical Decision Experiments

We discuss the role of entropy of the set of unfolding word candidates in the context of DIANA, a computational model of human auditory speech comprehension. DIANA consists of three major interacting components: Activation, Decision and Execution. The Activation component computes activations of word candidates that change over time as a function of the unfolding audio input. The resulting set of word candidate activations can be associated with an entropy that is related to difficulty of the decision when one of these candidates must be selected at time T. The paper presents the close relation between entropy measures and the between-word competition during the unfolding of the auditory stimuli, and at the end of the stimuli if no decision could be made before stimulus offset. We present a way for computing the entropy that takes into account linguistic-phonetic constraints that play a role in speech comprehension and in lexical decision experiments. Using the BALDEY data set and linear mixed effects regression models for RT, we show that entropy measures explain differences between RTs of words with different morphological structure.

Section: Regression Analysis 41 Removing Local Speed Effects From The...mentioning

confidence: 99%

Section: Set-upmentioning

confidence: 99%

Word Competition: An Entropy-Based Approach in the DIANA Model of Human Word Comprehension

Bosch

Boves

2021

“…When modeling RTs, it is necessary to take into account the 'local speed effect', according to which a substantial part of the observed variation in the RT sequence is due to local trends [14]. To remove those effects it is necessary to regard the RTs as a sequence.…”

Section: Removing Local Speed Effectsmentioning

confidence: 99%

“…the concept 'cognate' was not included explicitly in the experimental design, we marked a subset of the word stimuli as cognates for this study to enable a comparison with the data from [8]. Before computing log(RT ) the RT sequences were filtered using the methods presented in [14]. In computing log(RToffset) the stimuli for which the response was given before the offset of the stimulus were discarded.…”

Section: An Experiments With Dutch Learners Of Frenchmentioning

confidence: 99%

Models of Reaction Times in Auditory Lexical Decision: RTonset versus RToffset

Brand¹,

Mulder²,

Bosch³

et al. 2021

We investigate how the role of predictors in models of reaction times in auditory lexical decision experiments depends on the operational definition of RT: whether the time is measured from stimulus onset or from stimulus offset. In a large body of literature, RTs are measured from the onset of the stimulus to the start of the response (often a button press or an oral response). The rationale behind this choice is that information about the stimulus becomes available to the listener starting at onset. Alternatively, the RT from offset is less dependent on stimulus duration and is assumed to focus on those cognitive processes that play a role late(r) in the word and after word offset, when all information is available.The paper presents RT-onset and RT-offset-based linear mixed effects models for three different lexical decision-based data sets and explains the significant differences between these models, showing to what extent both definitions of reaction time reveal different roles for predictors and how early and later contributions to the overall RT can be differentiated.

“…Clever manipulation of those features should make it possible to draw conclusions about cognitive processes that drive the reactions. For over a decade, RTs obtained in lexical decision experiments have been modeled using linear mixed-effect models (LMM) (e.g., [1,2,3,4]). The most important advantage of LMMs over traditional linear models in many fields of science is well known (e.g., [5]): the random structure in mixed-effects models makes it possible to estimate more conservative statistical models and to consider e.g.…”

Section: Introductionmentioning

confidence: 99%

Time-to-Event Models for Analyzing Reaction Time Sequences

Bosch¹,

Boves²

2021

We investigate reaction time (RT) sequences obtained from lexical decision experiments by applying Time-to-Event modelling (Survival Analysis). This is a branch of statistics for analyzing the expected duration until one or more events happen, associated with a set of potential 'causes' (in our case the decision for a 'word' judgment as a function of conventional predictors such as lexical frequency, stimulus duration, reduction, etc.). In this analysis, RTs are considered a by-product of an (unobservable) cumulative incidence function that results in a decision when it exceeds a certain threshold.We show that Survival Analysis can be effectively used to narrow the gap between data-oriented models and processoriented models for RT data from lexical decision experiments. Results of this analysis technique are presented for two different RT data sets. The analysis reveals time-varying patterns of predictors that reflect the differences in cognitive processes during the presentation of auditory stimuli.