Research colloquium: Computational and experimental psycholinguistics (SS 2024)

Title: Can we close the eye-movement gap in reading research by using lower sampling rates?
Abstract: Eye-movement research has revolutionized our understanding of reading, but the use of eye-tracking techniques is still limited to only a few countries in the world. Publication statistics from the last 25 years show that most publications on eye-movements during reading have authors based in Western countries. We argue that eye-tracking is the ideal technique for reading and language research in countries with limited resources, and that it is crucially important to not just study a small subset of languages, but that more needs to be done to make eye-tracking technology accessible for researchers in those countries. This includes evaluating to what extent cognitive processes during reading can be measured with less expensive eye-tracking devices. One such way may be to use devices with a lower sampling rate, which may be much less expensive than high-sampling rate eye-trackers. We present findings from a study that recorded readers’ eye movements during reading at different sampling rates. We show that it is possible to measure the classic effect of word frequency on fixation duration, reflecting ongoing processing during reading, even at sampling rates of 250 Hz and less.
Slides: Click here

Title: A Model of Language Processing as Resource-Rational Sequence Prediction
Abstract: Psycholinguists have long studied humans’ difficulty in comprehending complex sentences as a window into the nature of human language processing. Prominent theoretical accounts assign central roles either to expectations about upcoming content, or to retrieval of past content from short-term memory. Each account is supported by a substantial body of evidence, and unifying them has proven challenging. We propose a unifying model based on resource-rational memory representations, which we scale to the rich statistical structure of language using large-scale text data and contemporary machine learning methods. The model makes fine-grained predictions sharply different from those of existing models, which we confirm in three behavioral experiments. Taken together, our work shows how general cognitive principles, implemented using machine learning, predict fine-grained patterns in human language comprehension that previous theories cannot account for.
Background reading: Hahn, M., Futrell, R., Levy, R., & Gibson, E. (2022). A resource-rational model of human processing of recursive linguistic structure. Proceedings of the National Academy of Sciences, 119(43). http://dx.doi.org/10.1073/pnas.2122602119
Bio: Michael Hahn is an Assistant Professor at Saarland University, Germany. He received his PhD from Stanford University in 2022. His research aims to uncover the computational principles underlying language processing, and their implications for Natural Language Processing. To this end, he combines behavioral experiments with information theory and contemporary machine learning. His research has been published in journals including PNAS, Psychological Review, and Cognition.

Title: Statistical Insensitivity in German Plural Generalization
Abstract: Speakers are highly sensitive to the statistical patterns that characterize their language, and use this distributional knowledge to both predict and produce novel linguistic sequences. This motivates some recent claims that complex patterns in inflectional morphology serve a functional purpose – speakers can use them to reduce uncertainty when producing novel inflected forms. But do speakers consistently use distributional information in this way?

We investigate how adult German speakers use their lexical knowledge when generalizing plural classes to unknown words. We focus on grammatical gender, a highly informative cue to plural class in the German noun lexicon, and sample information-thoretic properties of relevant lexical distributions to formalize three behavioral hypotheses: speakers may Maximize the statistical informativity of grammatical gender (optimal prediction), Mirror its informativity on a phonologically-constrained distribution (suboptimal prediction), or Ignore gender altogether. In three behavioral experiments, we show that nearly all speakers Ignore or at most Mirror gender in plural production, meaning they do not use this distributional knowledge to optimally reduce uncertainty.
Bio: Kate McCurdy is a postdoctoral researcher in the LaCoCo Lab at Saarland University. She received her PhD in 2024 from the University of Edinburgh School of Informatics.

Title: Sounds surprising! How prediction and error can contribute to speech learning
Abstract: When listening to speech in a native or highly proficient language, we barely notice how complex and variable the speech signal is. But when we encounter an unfamiliar language, it is striking how difficult it can be to even grasp the sounds themselves or how we might go about producing such sounds. What learning mechanisms contribute to this change? And how can we account for differences between learning speech sounds in our first language, which seems to come naturally with high accuracy, and learning a second language, which often poses challenges? According to error-driven learning models, learners use available sensory information – cues – to predict important events. Through feedback from prediction error – also called ‘surprise' – learners incrementally learn to discriminate informative from uninformative cues, so as to optimise predictive success. I will present a collection of recent work showing how error-driven learning models can account for a variety of phenomena in speech acquisition.
Background reading: Nixon, J. S. (2020). Of mice and men: Speech sound acquisition as discriminative learning from prediction error, not just statistical tracking. Cognition, 197, 104081. http://dx.doi.org/10.1016/j.cognition.2019.104081