An Initial Study on Sentiment Recognition of Social Network Messages Based on Texts, Images, and Videos
Date Issued
2015
Date
2015
Author(s)
Yang, Yu-Chung
Abstract
Nowadays, Internet users prefer to share texts, images, and videos on social networks rather than share them via E-mail or blogs, thus creating many research topics related social networks. One of the research problems is to identify the connoted sentiment of an user’s post automatically. Image, text, and video sentiment recognition have been studied extensively over recent years independently. In this thesis we would like to propose a comprehensive system which incorporates all three types of an user’s posts. Here we proposes a procedural system that utilizes Sentibank visual sentiment ontology for image sentiment recognition, one-hot and Word2vec representation for text sentiment analysis, and the combination of visual and audio features for video sentiment recognition. Lastly we use a combination of these three types of model to predict whether the sentiment of an user’s post is positive or negative. We collect 1113 Facebook posts with text and image as our testing dataset for evaluating the performance for image and text input, and 1101 emotion video dataset for video input. Although the proposed system does not perform satisfactorily in terms of computation time and recognition accuracy with video input, it achieves a comparable recognition accuracy with image input and outperforms Sentibank with text input to existing methods proposed by Sentibank.
Subjects
social networks
image sentiment recognition
text sentiment recognition
video sentiment recognition
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-104-R01922056-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):869df38cdd261a5417bd2e08608786d1