Source-linked AI summary
ZuCo 2.0: A Dataset of Physiological Recordings During Natural Reading and Annotation
Nora Hollenstein, Marius Troendle, Ce Zhang, Nicolas Langer
TL;DR
ZuCo 1.0 recorded natural and task-specific reading on different days, so EEG differences could reflect session effects rather than reading paradigm. ZuCo 2.0 provides a validated corpus with both paradigms, enabling their comparison and reuse in NLP research.
Problem
ZuCo 1.0's different-day recording sessions made comparisons between natural and task-specific reading unsuitable because EEG differences could reflect session effects.
Source: §2 · p. 3Method
The authors recorded and preprocessed simultaneous eye-tracking and EEG data from 18 subjects reading 739 sentences across normal-reading and annotation paradigms.
Source: §1 · p. 2, §3.2 · p. 3Results
ZuCo 2.0 is the first dataset that allows comparison between natural reading and reading during annotation.
Source: §5 · p. 7Takeaways & Limitations
Its word- and sentence-level eye-tracking and EEG features can support NLP and machine-learning evaluation, including probing language models and studying annotation quality.
Source: §5 · pp. 7–8
Abstract
from arXiv · showhide
We recorded and preprocessed ZuCo 2.0, a new dataset of simultaneous eye-tracking and electroencephalography during natural reading and during annotation. This corpus contains gaze and brain activity data of 739 sentences, 349 in a normal reading paradigm and 390 in a task-specific paradigm, in which the 18 participants actively search for a semantic relation type in the given sentences as a linguistic annotation task. This new dataset complements ZuCo 1.0 by providing experiments designed to analyze the differences in cognitive processing between natural reading and annotation. The data is freely available here: https://osf.io/2urht/.
1. Introduction
Physiological recordings can illuminate language processing and support NLP, but prior work motivates asking whether reading and annotation activity can be captured directly. ZuCo 2.0 provides simultaneous eye-tracking and EEG recordings across natural reading and annotation, with data and scripts publicly available.
1. Introduction
- Eye movements provide millisecond-accurate gaze records, while EEG measures scalp electrical activity; together they capture word-aligned brain signals during naturalistic sentence reading.Eye-tracking defines word boundaries on the reading timeline, enabling extraction of brain activity signals for each word.Source: §1 · p. 1
- The corpus addresses whether physiological activity recorded during reading can reveal aspects of text understanding and annotation relevant to NLP.The motivation includes using cognitive data to improve or evaluate NLP systems and inform labeling processes.Source: §1 · p. 1
- ZuCo 2.0 is presented as the first dataset for comparing normal reading with task-specific reading during annotation using simultaneous eye-movement and brain-activity recordings.It includes raw and preprocessed eye-tracking and EEG data from 18 subjects reading 739 English sentences, plus recording and preprocessing scripts.Source: §1 · p. 2
- The paper describes corpus construction, experimental design, acquisition, preprocessing, feature extraction, and technical validation of recording quality.The corpus is intended for reuse in NLP training and evaluation, including semantic tasks such as entity and relation discovery.Source: §1 · p. 2
2. Related Work
Earlier work established the value of physiological data for NLP and collected eye-tracking or EEG during reading and annotation. ZuCo 2.0 addresses the lack of data designed to compare natural reading with annotation, while also resolving a comparison issue in ZuCo 1.0.
2. Related Work
- Available reading corpora included eye-tracking datasets such as Dundee, Provo, and GECO, and an EEG corpus from UCL.Prior work also used cognitive processing data to improve and evaluate NLP methods.Source: §2 · p. 2
- Before ZuCo 1.0, no available data combined eye-tracking and EEG during natural reading; prior annotation studies used eye-tracking but not EEG.Prior examples include sentiment, word-sense, named-entity, and coreference annotation, as well as gaze-based active-learning cost modeling.Source: §2 · p. 2
- ZuCo 1.0 recorded normal and task-specific reading on different days, making comparisons unsuitable because session differences could affect EEG activity.ZuCo 2.0 was recorded to address this issue and extend the data with more sentences and subjects.Source: §2 · p. 3
3. Corpus Construction
ZuCo 2.0 documents the construction of a corpus for comparing normal reading with task-specific relation annotation, including its participants, materials, recording design, and processing.
3.1. Participants
- The corpus includes 18 participants and 739 Wikipedia sentences, with 349 read normally and 390 during task-specific relation annotation.The dataset also includes 63 duplicate sentences read under both tasks, enabling comparison across reading goals.Source: §3.2 · p. 3, §3.1 · p. 3
3.3. Experimental design
- Participants read at their own pace in alternating task blocks, with all sentences recorded in one session per participant.Normal-reading blocks involved comprehension questions for a random 12% of sentences; annotation blocks required a relation-present decision, with 17% of sentences serving as controls.Source: §3.3 · pp. 3–4
3.4. Linguistic assessment
- Participants completed LexTALE, and the mean score was 88.54%; comprehension-question and relation-annotation scores were also reported.The sentence materials included seven semantic relation types for the task-specific paradigm.Source: §3.2 · p. 3, §3.4 · p. 4
3.5. Data acquisition
- Eye position and pupil size were recorded at 500 Hz, and high-density EEG was recorded at 500 Hz using a 128-channel system.Eye-tracker calibration was checked before each sentence block; electrode impedances were checked before recording and after every third block.Source: §3.5 · p. 5
3.6. Preprocessing and feature extraction
- The released eye-tracking data include gaze coordinates and word-level features, while EEG processing synchronizes recordings to fixation onsets and derives frequency-band features.Eye features include gaze duration, total reading time, first and single fixation durations, go-past time, fixation count, and pupil size; EEG processing includes artifact rejection and Hilbert-transformed band signals.Source: §3.6 · pp. 5–6
4. Data Validation
Technical validation found task-related differences in eye movements and fixation-related EEG patterns, with the EEG findings consistent with prior work and ZuCo 1.0.
Eye-tracking
- Task-specific reading involved fewer fixations, faster reading, and higher omission rates, with fixations often ending after words decisive for relation detection.The reading material came from the same source and length range, while normal reading continued through sentence endings more often.Source: §4 · p. 6
EEG
- The fixation-related potentials showed components similar to prior natural-reading findings and ZuCo 1.0, and the validation compared results with those studies.The authors describe the aim as checking recording quality and replicating earlier co-registration findings.Source: §4 · p. 6, Figure 8 · p. 7
- Fixation-duration analyses revealed a first positive EEG deflection around 100 ms after fixation onset and a second peak that varied with fixation duration.The duration-dependent second peak was attributed to the time-jittered succeeding fixation.Source: §4 · p. 7, Figure 9 · p. 7
5. Conclusion
ZuCo 2.0 is a freely available corpus of eye-movement and brain-activity recordings during natural reading and annotation, enabling the first comparison of these reading paradigms. Its word- and sentence-level features and semantic-relation annotations support NLP research and studies of annotation quality.
5. Conclusion
- ZuCo 2.0 is the first dataset enabling comparison between normal reading and reading during annotation using eye-movement and brain-activity recordings.The corpus is freely available.Source: §5 · p. 7
- Word-level and sentence-level eye-tracking and EEG features can support NLP and machine-learning evaluation, including assessments of linguistic phenomena in neural models.The paper gives syntactic skills and cognitive plausibility of word representations as examples.Source: §5 · p. 7
- Semantic-relation labels and participant annotations also make the corpus suitable for relation extraction and classification.The two reading paradigms may help improve manual labeling and annotation quality for supervised machine learning.Source: §5 · pp. 7–8