This title slide introduces "Multi-Modal Sentiment Analysis," a Capstone Project for MCA. It explores the fusion of text, audio, and visual data for advanced sentiment detection.

Multi-Modal Sentiment Analysis

A Capstone Project for MCA

Exploring Text, Audio, and Visual Fusion for Advanced Sentiment Detection

Source: Capstone Project for MCA

The presentation agenda outlines four main sections: Introduction & Problem Statement on sentiment analysis challenges in multi-modal data, Objectives & Methodology covering project goals, datasets, and fusion methods. It continues with Results & Comparative Analysis of experimental outcomes and performance insights, followed by Conclusion & Future Work summarizing findings and research directions.

Presentation Agenda

1. Introduction & Problem Statement

Overview of sentiment analysis challenges in multi-modal data.

2. Objectives & Methodology

Project goals, dataset details, and proposed fusion methods.

3. Results & Comparative Analysis

Experimental outcomes, comparisons, and key performance insights.

4. Conclusion & Future Work

Summary of findings and directions for further research. Source: Multi-Modal Sentiment Analysis — A Capstone Project for MCA

This section header slide introduces the "Introduction & Problem Statement" (Section 01) of the Multi-Modal Sentiment Analysis capstone project for MCA. It highlights the subtitle on challenges in fusing text, audio, and video for accurate emotion detection.

Multi-Modal Sentiment Analysis — A Capstone Project for MCA

01

Introduction & Problem Statement

Challenges in Text, Audio, and Video Fusion for Accurate Emotion Detection

Source: Introduction & Problem Statement

The slide outlines objectives for developing a multi-modal sentiment analysis model that fuses text, audio, and visual features. It aims to achieve superior accuracy over unimodal baselines and evaluate performance on benchmark datasets.

Objectives

• Develop multi-modal sentiment analysis model
• Fuse text, audio, and visual features
• Achieve superior accuracy over unimodal baselines
• Evaluate performance on benchmark datasets

Source: Multi-Modal Sentiment Analysis — A Capstone Project for MCA

The slide details multi-modal datasets CMU-MOSEI and IEMOCAP, each with over 1000 samples including text, audio, and video. These datasets feature positive/neutral/negative sentiment labels and are preprocessed for feature extraction.

Dataset Details

• Multi-modal datasets: CMU-MOSEI, IEMOCAP
• 1000+ samples with text, audio, video
• Labels: positive/neutral/negative sentiment
• Preprocessed for feature extraction

Source: Multi-modal datasets for sentiment analysis

The methodology workflow processes text with BERT for 768-dim contextual embeddings, audio with MFCC + VGGish for spectral and learned features, and video with ResNet-50 for 2048-dim frame-level spatial features. It then applies early/late fusion via concatenation or decision-level methods, followed by LSTM for temporal modeling and sentiment classification output.

Methodology

Source: Multi-Modal Sentiment Analysis — A Capstone Project for MCA

The slide presents a comparative analysis table of model performances, with metrics for Accuracy and F1-Score. The multi-modal "Ours" model leads with 85% accuracy and 0.83 F1-Score, outperforming Unimodal Text (72%, 0.70) and Audio-Only (65%, 0.63).

Comparative Analysis

Source: Multi-Modal Sentiment Analysis — A Capstone Project for MCA

The slide presents experimental results with 85% accuracy (+13% over baseline) and an F1 score of 0.83 (+15% gain from unimodal models). It also highlights a +13% accuracy gain relative to state-of-the-art and +15% F1 improvement from multi-modal fusion.

Experimental Results

• 85%: Accuracy

+13% improvement over baseline

• 0.83: F1 Score

+15% gain from unimodal models

• +13%: Accuracy Gain

Relative to state-of-the-art

• +15%: F1 Improvement

Multi-modal fusion benefit Source: Multi-Modal Sentiment Analysis Capstone