Understanding word predictability using Natural Language Processing algorithms

D. FERNANDEZ-SLEZAK; D. E. SHALOM; J. KAMIENKOWSKI; BIANCHI, BRUNO

Wuppertal

Congreso; European Conference on Aye Movements; 2017

Universidad de Wuppertal

During reading our brain predicts upcoming words. Predictability (probability of guessing the next word) is currently estimated by performing cloze experiments, where participants read incomplete statements and have to complete them with one word. During the task, the only information subject can use is the preceding context. To estimate the predictability of one word, it is necessary to ask several participants, and then calculate the proportion of correct answers. Cloze-task is then an expensive experiment, and results are only valid for those words in the analyzed texts. During the last years, different approaches have been taken to automatically estimate this human predictability. Here we analyzed different ways of predicting words, using Natural Language Processing algorithms (LSA, word2vec, n-grams), and explore different aspects of the human predictability (semantic, syntactic, mnemonic). We evaluated the incorporation of these computational measures, both by themselves or combined on Linear Mixed Models with eye movements as dependent variables. Results show that these computational estimations of the word predictability have very good performance and can be used to replace the human predictability in the used models. Further, this is a step forward in understanding and separating the contribution of the different cues we use to predict words.