Related Experiment Videos
A temporal adaptive dictionary-constrained LDA and Bi-calibrated dual granularity DTM framework for dynamic topic
1School of International Studies, Tianjin University of Commerce, Tianjin, 300134, China. yinxq@tjcu.edu.cn.
Scientific Reports
|May 29, 2026
Summary
This study introduces a new AI topic model to track research trends, identifying generative AI and large language models as explosive growth areas. The model improves accuracy and dynamic tracking for better research and policy decisions.
Area of Science:
- Artificial Intelligence (AI)
- Data Science
- Information Science
Background:
- Accurate tracking of artificial intelligence (AI) research evolution is vital for topic selection, policy, and innovation.
- Traditional topic models face limitations in balancing static topic accuracy and dynamic evolution tracking.
- Existing methods struggle to capture the nuanced, temporal shifts in AI research landscapes.
Purpose of the Study:
- To propose a unified framework for accurately capturing dynamic topic evolution in AI research.
- To enhance static topic recognition and dynamic evolution tracking capabilities beyond traditional models.
- To provide decision support for researchers and management departments in navigating AI field frontiers.
Main Methods:
- Developed a Temporal Adaptive Dictionary-Constrained Latent Dirichlet Allocation (LDA) combined with a Bi-calibrated Dual Granularity Dynamic Topic Model (DTM).
- Optimized LDA topic initialization using a domain-specific dictionary for enhanced static topic clustering.
- Employed an 'annual + quarterly' dual granularity in DTM to track topic strength, keyword composition, and cross-topic associations over time (2014-2023).
Main Results:
- Identified six core AI topic clusters, with 'generative AI', 'large language models', and 'multimodal learning' emerging as explosive topics post-2021.
- Observed a declining trend in traditional machine learning, shifting focus to 'few-shot learning' and 'edge computing adaptation'.
- Top topic couplings include 'generative AI-large language models', 'computer vision-multimodal learning', and 'reinforcement learning-robotics'.
Conclusions:
- The proposed combined model significantly outperforms single models, reducing perplexity by 17.3% and topic drift by 23.5%.
- The framework effectively captures both static topic accuracy and dynamic evolution in AI research.
- Findings offer valuable insights for understanding AI research frontiers and optimizing resource allocation.