Session: 4

PS4-18 | Metrics of Reading in Short Stories

Lena Valeska Peroni

Laboratorio de Inteligencia Artificial Aplicada, Instituto de Ciencias de la Computacion, FCEyN, UBA-CONICET

When reading, our brain constantly infers the structure of the sentence as it develops. Predicting upcoming words allows us to integrate information and to guide eye movements. Cloze-task experiments are used to estimate these predictions by asking a group of participants to complete the most likely word that follows a given context. Thus, collecting several responses per word.
A variety of measures could describe different aspects of this response distribution. The Predictability of a given word in a sentence, is estimated as the proportion of times that word is guessed over all guesses. Surprisal is the negative log of Predictability and assesses the unexpectedness of a word in its context. Entropy is a function of the probabilities of the different responses for a given context in the experiment.
Here, we present an analysis of a corpus of sentences drawn from short stories. An online cloze-task was conducted using isolated or contextualized sentences. Our previous work showed that Predictability is highly dependent on large contexts and is a strong predictor of eye movements and brain activity. Now, we explore how these metrics interact with other text variables (such as word-position, word-frequency, etc) and eye movement variables (fixation duration, regressions). Cloze-task-derived information-theoretical complexity metrics that connect theories of parsing and grammar to reading times and brain signals can help us understand various aspects of eye-movement control.