Multi-label Text Sentiment Classification: Comparative Experiments of Five Machine Learning Models Based on the GoEmotions Dataset

A course project by a Vietnamese student team, focusing on the multi-label text sentiment classification task. It compares the performance of five algorithms (Logistic Regression, LinearSVC, Random Forest, 1D CNN, and Bi-LSTM) on Google's GoEmotions dataset, and discusses technical challenges and optimization strategies in multi-label scenarios.

In the field of natural language processing, sentiment analysis has evolved from binary classification to multi-dimensional recognition. Google released the GoEmotions dataset in 2021, which contains 58,000 Reddit comments labeled with 28 fine-grained emotion categories. Multi-label classification faces unique challenges: a text may carry multiple emotion labels simultaneously, and label sparsity and co-occurrence patterns make it difficult to directly apply traditional methods.

The project uses an end-to-end pipeline architecture:

1. Preprocessing: Lowercase conversion, special character removal, tokenization, lemmatization;
2. Word vectors: TF-IDF (for traditional models), Word2Vec (for deep learning models);
3. Classification models: Traditional methods (Logistic Regression OVR, LinearSVC OVR, Random Forest), deep learning methods (1D CNN, Bi-LSTM);
4. Threshold optimization: Independently tune thresholds for each label to maximize F1 score.

Quantitative Evaluation: The dataset has extreme class imbalance. High-frequency categories (e.g., amusement, gratitude, love, neutral) perform well, while low-frequency categories (e.g., sadness, pride) and categories with ambiguous semantic boundaries are prone to confusion; Qualitative Comparison:

• Logistic Regression/LinearSVC: Perform well in linear scenarios but are weak in handling non-linear semantic combinations;
• Random Forest: Robust in handling contradictory emotions;
• 1D CNN: Strong at local feature extraction, excellent performance on short texts;
• Bi-LSTM: Strong ability to maintain long-distance semantic dependencies, suitable for complex mixed emotions.

Practical Testing: Four challenging use cases (coexistence of multiple emotions, contradictory semantics, clear features, complex mixing) were designed to simulate real scenarios; Technical Implementation: Google Colab platform was used, with code organized in Jupyter Notebook, supporting GPU acceleration and real-time demonstration. The code is modular for easy reproduction.

Reference value of the project for industrial-grade systems:

• Threshold tuning is a necessary engineering practice in multi-label scenarios;
• Model selection needs to balance scenario requirements (CNN is suitable for short texts, LSTM for long texts, traditional methods have strong interpretability);
• Data quality is more important than model complexity; This project provides a complete learning example of multi-label classification for beginners.