Multi-view text classification through integrated RNN autoencoder learning of word, sentence, emotion and paragraph...

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Single-view text classification methods limit performance by processing documents through a single lens.
• Capturing the multi-dimensional nature of textual information is crucial for robust text classification.

Purpose of the Study:

• To propose an end-to-end framework, FMV-RNN-AE, that integrates multiple textual views for improved text classification.
• To evaluate the effectiveness of the proposed framework against existing single-view and multi-view methods.

Main Methods:

• The FMV-RNN-AE framework integrates four complementary textual views: word-level embeddings, sentence-level representations, emotion-based features, and paragraph-level semantics.
• Standard RNN autoencoders are employed to learn compressed view-specific representations.
• A learnable fusion module and joint optimization are used for classification, focusing on principled integration.

Main Results:

• FMV-RNN-AE demonstrated consistent improvements of 4.7% over single-view approaches and 2.2-4.0% over existing multi-view methods across seven benchmark datasets.
• The framework achieved high accuracy on sentiment-oriented tasks, including 93.5% on Hate Speech and 92.7% on IMDb.
• Compared to BERT, FMV-RNN-AE uses significantly fewer parameters and less memory, with comparable average accuracy.

Conclusions:

• The proposed FMV-RNN-AE framework offers a memory-efficient and task-sensitive alternative for text classification, particularly in latency-tolerant scenarios.
• Carefully designed multi-view autoencoder integration can enhance text classification robustness across diverse domains under constrained memory budgets.