Stony Brook SAT reading fixation dataset (SB-SAT)

Additional details about the eyetracking data and models can be found in this paper.

Citation:

@inproceedings{ahn2020towards,
  title={Towards predicting reading comprehension from gaze behavior},
  author={Ahn, Seoyoung and Kelton, Conor and Balasubramanian, Aruna and Zelinsky, Greg},
  booktitle={ACM Symposium on Eye Tracking Research and Applications},
  pages={1--5},
  year={2020}
}

Repo structure

There are three folders in this repo:

• stimuli contains the screenshots of four reading passages and questions used in the task
• model contains the preprocessed fixation dataset used for model training (with different types of data splits, please see the main paper for details). We also include the sample colab code for training models for comprehension prediction.
• fixation contains the raw eye-movemenet data in csv. See below for more detailed description on the reported variables.

More detailed information about the fixation dataset:

raw fixation dataset

18sat_fixfinal.csv contains raw eyemovement measures recorded using Eyelink 1000 (1000hz sampling rate). This file also contains which word of the passage the current gaze is fixated on (CURRENT_FIX_INTEREST_AREA_LABEL column). We exported all eye-movement measures using SR Research Eyelink Data Viewer program. You can see the detailed description about each eye-movement measure in their manual. 18sat_trialfinal.csv contains eye-movment measures are averaged or aggreaged for each trial (each page). Additional details about the other variables:

• RECORDING_SESSION_LABEL: subject id
• INDEX: trial id
• type: whether this trial belongs to reading or question
• book_name: reading passage name id
• book: reading passage integer id
• page: current page number for each reading passage
• total_page: total page number for each reading passage
• RT: total reading time for each page
• answer: subject's key response
• correct_answer: correct response (no correct reponse for reading trials, replaced as -99)
• page_name: the filename used for image stimuli (in stimuli folder)

question response

18sat_labels.csv has subject demographics and preprocessed question responses for model training (see paper for details how we preprocess this data)

• subj: subject id
• language: native langauge
• native: whether native langauge is english (0-no, 1-yes)
• book: reading passage id
• acc: average passage accuracy
• acc_level: average passage accuracy level in quartile (0: lowest 25%, 1: 25-50%, 2: 50-75%, 3: highest 25%)
• subj_acc: individual’s comprehension score averaged over all four passages
• subj_acc_level: subj_acc level in quartile (0: lowest 25%, 1: 25-50%, 2: 50-75%, 3: highest 25%)
• other variables; subject's subjective evaluation on their cognitive and mental states while reading this passage (default is higher the number, higher the state, number range from 0 to 3)
  • confidence: How confident were you when answering to the previous comprehension questions (overall)?
  • difficulty: How do you rate the difficulty level of the passage you read?
  • familiarity: How familiar are you with the topic of this passage?
  • interest: How interesting this passage was?
  • pressured: How much pressure did you feel during reading?
  • sleepiness How sleepy were you during reading this passage
  • sleephours: How long did you sleep last night?(0: less than 3 hours, 1: 3-6 hours, 2: 6-9 hours, 3: more than 9 hours)
  • recognition: Have you read this passage before? (0:no, 1:yes)