Incremental learning algorithm for dynamic evolution of domain specific vocabulary with its stability and plasticity...

Area of Science:

• Natural Language Processing
• Information Retrieval
• Machine Learning

Background:

• Domain-specific vocabularies require constant updates for fields like Information Retrieval and Natural Language Processing.
• Traditional methods necessitate retraining from scratch, which is inefficient.
• Incremental Learning offers an alternative by updating existing knowledge without full retraining.

Purpose of the Study:

• To present an incremental learning algorithm for updating domain-specific vocabularies.
• To introduce DocLib, an archive for storing data and vocabulary footprints.
• To evaluate the effectiveness of updated vocabularies in downstream tasks like text classification.

Main Methods:

• Developed an incremental learning algorithm for vocabulary updates.
• Utilized DocLib to archive data and vocabulary terms.
• Employed task-based evaluation, specifically text classification, to measure vocabulary effectiveness.
• Assessed vocabulary stability and plasticity using novel algorithms.
• Tested generalizability across datasets and compared with state-of-the-art techniques.

Main Results:

• The proposed algorithm ensures bounded memory and processing requirements.
• Multiple incremental updates maintained vocabulary relevance and effectiveness.
• Achieved 97.89% accuracy in identifying domain-related data across datasets.
• Demonstrated the ability to assimilate new knowledge while retaining old insights.
• Confirmed effectiveness compared to state-of-the-art techniques on a benchmark dataset.

Conclusions:

• The incremental learning approach effectively updates domain-specific vocabularies.
• The method is efficient, with bounded memory and processing needs.
• The approach shows strong generalizability and potential for various research fields beyond classification.