Degree Type

Creative Component

Semester of Graduation

Spring 2019

Department

Computer Science

First Major Professor

Adisak Sukul

Second Major Professor

Wallapak Tavanapong

Degree(s)

Master of Science (MS)

Major(s)

Computer Science

Abstract

Ad tone defines the aim of a political video advertisement, which can be either to promote a specific candidate, to attack the candidates or to contrast the candidates. Depending upon the aim, a political video advertisement can be classified into either promote, attack or contrast class. Analysis of ad tone in political video advertisements can provide more insights about the political campaign to political science researchers. Political campaigns are investing more and more on online platforms, which creates a large amount of political video advertisements. Manual classification of ad tones in political video advertisement is time-consuming, labor intensive and not scalable. Hence, there is a need for an efficient and effective classification model for automatic classification of the ad tones in political video advertisements. The available labeled dataset is very small in size and suffers from class imbalance. Due to this reason, the performance of the minority class is poor compared to the majority class. Moreover, due to the way the different classes are defined, all three classes decompose into sub-parts and suffer from class overlapping problem. There has been an attempt in automatic classification of political ad tones, but it does not take class imbalance into account. We investigate a couple of data augmentation techniques to overcome the class imbalance problem and the effectiveness of deep learning models on ad tone classification using text-based features. In our experiments, the best deep learning model offers a better F1 score of 0.570 on the minority class compared to the F1 score of previous work, which is 0.527. However, the performance is still unsatisfactory. We design hand-crafted features specific for ad tone classification using Support Vector Machine as the classifier. Our proposed approach gives the best weighted average F1 score of 0.860 on the entire test set and F1 score of 0.657 on the minority contrast class.

Copyright Owner

Baskota, Mohit

File Format

PDF

Share

COinS