A comparison of human attention with computational attention mechanisms
File format: Input.label, Input.text, Answer.Q1Answer, Answer.html_output
Input.label: Original Yelp label for the review, converted to 0 (negative) or 1 (positive). One and two-star reviews are mapped to 0, four and five-star reviews are converted to 1. Input.text: Original review from the Yelp dataset. Answer.Q1Answer: Sentiment label collected from the annotator Answer.html_output: Human attention map (HAM) collected from the annotator.
One exception is "ham_part7.csv" file contains 2 to 4 HAMs per review instead of standard 3 HAMs. This file will be re-uploaded after removing the reviews with 2 HAMs.
You can use the "visualize_single_map>empty-template.html" file for visualizing a single HAM. You should copy "Answer.html_output" field into the relevant part in this file.
You can use the "generate_ham>generate-human-attention.py" file for converting the html format HAM into a binary vector.