TextConvoNet: a convolutional neural network based architecture for text classification.

Area of Science:

• Natural Language Processing
• Machine Learning
• Deep Learning

Background:

• Existing Convolutional Neural Network (CNN) models for text classification primarily extract intra-sentence n-gram features using one-dimensional filters.
• These models represent text data as a Sentence Matrix, limiting their ability to capture broader contextual information.

Purpose of the Study:

• To introduce TextConvoNet, a novel CNN architecture designed for binary and multi-class text classification.
• To enable the extraction of both intra-sentence and inter-sentence n-gram features for improved text representation.
• To evaluate the performance of TextConvoNet against established machine learning, deep learning, and attention-based models.

Main Methods:

• Developed TextConvoNet, a CNN architecture utilizing a novel input matrix representation.
• Applied a two-dimensional multi-scale convolutional operation to capture both intra- and inter-sentence n-gram features.
• Conducted experiments on five diverse binary and multi-class text classification datasets.

Main Results:

• TextConvoNet demonstrated superior performance compared to existing models across all tested datasets.
• The model achieved high scores in accuracy, precision, recall, f1-score, specificity, gmean1, gmean2, and Mathews correlation coefficient (MCC).
• Experimental results confirmed the effectiveness of TextConvoNet in capturing richer n-gram features.

Conclusions:

• TextConvoNet offers a significant advancement in CNN-based text classification by incorporating inter-sentence feature extraction.
• The proposed architecture provides a more comprehensive understanding of text data, leading to enhanced classification performance.
• TextConvoNet represents a promising new direction for deep learning approaches in natural language processing tasks.