With Blinkers on: Robust Prediction of Eye Movements across Readers

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 803–807, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics With blinkers on: robust prediction of eye movements across readers Franz Matties and Anders Søgaard University of Copenhagen Njalsgade 142 DK-2300 Copenhagen S Email: soegaard@hum.ku.dk Abstract Nilsson and Nivre (2009) introduced a tree- based model of persons’ eye movements in reading. The individual variation between readers reportedly made application across readers impossible. While a tree-based model seems plausible for eye movements, we show that competitive results can be obtained with a linear CRF model. Increasing the inductive bias also makes learning across readers pos- sible. In fact we observe next-to-no perfor- mance drop when evaluating models trained on gaze records of multiple readers on new readers. 1 Introduction When we read a text, our gaze does not move smoothly and continuously along its lines. Rather, our eyes fixate at a word, then skip a few words, to jump to a new fixation point. Such rapid eye movements are called saccades. Sometimes we even jump backwards. Backward saccades are called re- gressions. Gaze can be recorded using eye track- ing devices (Starr and Rayner, 2001). Since eye movements in reading give us important information about what readers find complicated in a text, and what readers find completely predictable, predicting eye movements on new texts has many practical ap- plications in text-to-text generation and human com- puter interaction, for example. The problem of predicting eye movements in reading is, for a reader ri and a given sequence of word tokens w1 . . . wn, to predict a set of fixation points F ⊆ {w1, . . . , wn}, i.e., the fixation points of ri’s gaze. For each token wj, the reader ri may skip wj or fixate at wj. Models are evaluated on record- ings of human reading obtained using eye tracking devices. The supervised prediction problem that we consider in this paper, also uses eye tracking data for learning models of eye movement. Nilsson and Nivre (2009) first introduced this su- pervised learning task and used the Dundee corpus to train and evaluate a tree-based model, essentially treating the problem of predicting eye movements in reading as transition-based dependency parsing. We follow Hara et al. (2012) in modeling only forward saccades and not regressions and refix- ations. While Nilsson and Nivre (2009) try to model a subset of regressions and refixations, they do not evaluate this part of their model focusing only on fixation accuracy and distribution accuracy, i.e., they evaluate how well they predict a set of fixation points rather than a sequence of points in order. This enables us to model eye movements in reading as a sequential problem of determining the length of forward saccades, increasing the inductive bias of our learning algorithm in a motivated way. Note that because we work with visual input, we do not tokenize our input in our experiments, i.e., punctuation does not count as input …

Content truncated to 500 words