InstructNet: A novel approach for multi-label instruction classification through advanced deep learning.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Search engines are primary information resources, with "How To" queries being prevalent for task-oriented learning.
• Categorizing instructional text is crucial for building effective knowledge bases and facilitating task completion.
• Existing methods require robust approaches for accurately classifying multi-label instructional content.

Purpose of the Study:

• To develop and evaluate a multi-label classification system for instructional text, specifically "How To" articles.
• To determine the effectiveness of transformer-based deep neural architectures for this task.
• To propose an approach named 'InstructNet' for multi-label instruction categorization.

Main Methods:

• Utilized a dataset of 11,121 wikiHow "How To" articles, each with multiple categories.
• Employed transformer-based deep neural architectures, including Generalized Autoregressive Pretraining for Language Understanding (XLNet) and Bidirectional Encoder Representation from Transformers (BERT).
• Evaluated model performance using accuracy and macro F1-score metrics.

Main Results:

• The XLNet architecture within the InstructNet approach achieved a high accuracy of 97.30%.
• Micro and macro average scores reached 89.02% and 93%, respectively, demonstrating strong multi-label classification performance.
• The evaluation provided insights into the strengths and weaknesses of the proposed architectures.

Conclusions:

• The XLNet-based InstructNet approach is highly effective for multi-label instruction classification.
• Transformer architectures show significant promise for organizing and understanding instructional content.
• Further refinements can build upon this successful multi-level evaluation strategy.