Abstract
Temporarily ambiguous sentences such as (1) are read more slowly than their unambiguous counterparts such as (2) with the same meaning. Why is this the case?
- The little girl fed the lamb remained relatively calm.
- The little girl who was fed the lamb remained relatively calm.
One account of sentence processing, Surprisal Theory, argues that the amount of time taken to read a word depends on how
predictable the word is in a given sentence: the word “
remained” is more predictable in (2) because of the disambiguating words “who was”. To test this, previous work has measured the predictability of words using neural network language models (NLMs) and used this to predict human reading times in sentences like (1) and (2). This work found that the estimates from current NLMs underestimated processing difficulty.
The goal of this project is to modify the training data of these models in a targeted manner, and test if any of these targeted changes result in estimates that better align with human reading times. Such an investigation can shed light on
why humans generate the predictions they do when reading sentences. Students working on this project will:
- Read background literature to generate hypotheses about why the data that these models are trained on might cause them to underestimate the processing difficulty in sentences like 1.
- Use these hypotheses to design alternative training datasets that might alleviate some of the issues in the current data.
- Train existing models on the new data sets.
- Compare predictability estimates from the newly trained models to existing human data.
As part of these tasks, students will learn to use and modify existing code to generate sentences, train neural network models, and generate plots.