Multimodal Sentiment Analysis of Songs Using Ensemble Classifiers

Gomez-Saavedra, Esteban

Permalink

https://hdl.handle.net/2142/79038 Copy

Description

Title

Multimodal Sentiment Analysis of Songs Using Ensemble Classifiers

Author(s)

Gomez-Saavedra, Esteban

Contributor(s)

Do, Minh N.

Issue Date

2015-05

Keyword(s)

• Music Information Retrieval
• Sentiment Analysis
• Multimodal Classification
• Classification Algorithms
• Multimodal Fusion

Date of Ingest

2015-08-13T16:08:52Z

Abstract

We consider the problem of performing sentiment analysis on songs by combining audio and lyrics in a large and varied dataset, using the Million Song Dataset for audio features and the MusicXMatch dataset for lyric information. The algorithms presented on this thesis utilize ensemble classifiers as a method of fusing data vectors from different feature spaces. We find that multimodal classification outperforms using only audio or only lyrics. This thesis argues that utilizing signals from different spaces can account for interclass inconsistencies and leverages class-specific performance. The experimental results show that multimodal classification not only improves overall classification, but is also more consistent across different classes.

Type of Resource

text

Genre of Resource

Other

Language

en

Permalink

http://hdl.handle.net/2142/79038

Owning Collections