license: mit
language:
- en
tags:
- gaze, hallucination
Dataset Details
This repository consists the eye-tracking dataset released as part of EMNLP 2023 paper: Eyes Show the Way: Modelling Gaze Behaviour for Hallucination Detection .
We collect a first-of-its-kind eye-tracking data, IITB-HGC, in which we employ 5 annotators for annotating 500 instances of claim-context pairs, carefully derived from the FactCC dataset (Kryscinski et al., 2020). During the annotation process, we capture the fixation durations of annotators on both the claim and context texts, along with their corresponding labels.
The dataset is formatted as a jsonl file (jsonlines-guide). Each line can be loaded as a json object, and has the following format:
{
'trial_id' : <instance id (can take values from 1-500)>
'text' : <the text presented to the annotator, in a claim:context format>
'participant_id' : <participant id (p01, p02, p03, p04, p05)>
'fixation_seqs' : <sequence ids of fixations>
'fixation_word_ids' : <a sequence containing the ids of the words being fixated in the order in which they were fixated upon>
'fixation_word_texts' : <a sequence containing the words being fixated in the order in which they were fixated upon>
'fixation_durations' : <a sequence containing the fixation durations (in ms) in the order in which the words were fixated upon>
'annotator_labels' : <label given by the annotator (can take a value: hallucinated/non_hallucinated)>
'true_labels' : <true label from the FactCC dataset (can take a value: hallucinated/non_hallucinated)>
}
For the code, check our github repository.
Cite the work
If you make use of the dataset or the code please cite our paper.
@inproceedings{maharaj2023eyes,
title={Eyes Show the Way: Modelling Gaze Behaviour for Hallucination Detection},
author={Maharaj, Kishan and Saxena, Ashita and Kumar, Raja and Mishra, Abhijit and Bhattacharyya, Pushpak},
booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing},
year={2023}
}