Deep Transfer Learning for Question Classification Based on Semantic Information Features of Category Labels

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Question classification is crucial for question answering (QA) systems, impacting answer accuracy and type restriction.
• Traditional methods require extensive manually labeled data, which is often unavailable in new domains.
• Data scarcity poses a significant challenge for developing robust QA systems in emerging fields.

Purpose of the Study:

• To investigate the effectiveness of various deep transfer learning methods for cross-domain question classification.
• To address the limitations of manual data labeling in new or specialized domains.
• To enhance the performance of question answering systems through improved question classification.

Main Methods:

• Utilized the ALBERT fine-tuning model as a foundation for deep transfer learning.
• Extracted source domain labels, question text, and target domain predicted labels as input.
• Incorporated semantic information of category labels and employed WordNet for question expansion.
• Compared the performance of different deep transfer learning techniques for cross-domain tasks.

Main Results:

• Deep transfer learning methods, particularly with ALBERT, significantly improve cross-domain question classification.
• Semantic information extraction and WordNet-based question expansion further boosted classification accuracy in target domains.
• The proposed approach effectively mitigates the challenges posed by limited labeled data in new domains.

Conclusions:

• Deep transfer learning offers a viable solution for overcoming data scarcity in cross-domain question classification.
• Combining ALBERT fine-tuning with semantic information and WordNet expansion enhances QA system performance.
• This research provides a practical framework for building accurate question classification models in data-limited environments.