Columbia MVSO Image Sentiment Dataset

TL;DR

This paper introduces the Columbia MVSO Image Sentiment Dataset, a large multilingual image sentiment dataset with human annotations, designed to benchmark automatic sentiment prediction systems in images.

Contribution

It provides a new, annotated benchmark dataset for evaluating image sentiment analysis systems, focusing on multilingual and adjective-noun pair concepts.

Findings

Dataset includes 15,600 concepts in 12 languages.

Human judgments collected for 3,911 English ANPs.

Dataset aims to serve as a benchmark for sentiment prediction systems.

Abstract

The Multilingual Visual Sentiment Ontology (MVSO) consists of 15,600 concepts in 12 different languages that are strongly related to emotions and sentiments expressed in images. These concepts are defined in the form of Adjective-Noun Pair (ANP), which are crawled and discovered from online image forum Flickr. In this work, we used Amazon Mechanical Turk as a crowd-sourcing platform to collect human judgments on sentiments expressed in images that are uniformly sampled over 3,911 English ANPs extracted from a tag-restricted subset of MVSO. Our goal is to use the dataset as a benchmark for the evaluation of systems that automatically predict sentiments in images or ANPs.