B-T4SA

在过去的几年中，情感分析领域取得了很大进展。研究人员依靠文本数据来完成这项任务，而直到最近他们才开始研究从多媒体内容中预测情绪的方法。随着社交媒体上共享的数据量不断增加，人们对“在野外”工作的方法也越来越感兴趣，即能够处理不受控制的条件。在这项工作中，我们面临着从大量用户生成和未标记的内容开始训练视觉情感分类器的挑战。特别是，我们收集了超过 300 万条包含文本和图像的推文，并利用文本内容的情感极性来训练视觉情感分类器。据我们所知，这是第一次在这种情况下提出和测试跨媒体学习方法。我们通过对视觉情感分析的基准进行比较研究和评估来评估我们模型的有效性。我们的实证研究表明，尽管与每张图像相关的文本通常是嘈杂的，并且与图像内容的相关性很弱，但可以利用它来训练深度卷积神经网络，从而有效地预测以前看不见的图像的情感极性。

Over the past several years, the field of sentiment analysis has witnessed substantial progress. Initially, researchers relied solely on textual data to complete this task; only recently have they begun to explore methods for predicting sentiment from multimedia content. With the continuously increasing volume of shared data on social media, there has been a growing interest in methods that work "in-the-wild", namely approaches capable of handling uncontrolled scenarios. In this study, we tackle the challenge of training visual sentiment classifiers using large-scale user-generated and unlabeled content. Specifically, we collected over 3 million tweets that contain both text and images, and utilized the sentiment polarity of the textual content to train the visual sentiment classifiers. To the best of our knowledge, this is the first time that cross-media learning methods have been proposed and tested under such circumstances. We assess the effectiveness of our model via a comparative study and benchmark evaluation for visual sentiment analysis. Our empirical findings show that, although the text associated with each image is often noisy and only weakly correlated with the image content, it can still be used to train deep convolutional neural networks (CNNs) that can effectively predict the sentiment polarity of previously unseen images.