DeepSeek-V3 powers training-free journal recommendation with 70% Top-10 accuracy
No more manual journal matching: LLMs select the right venue with interpretable reasoning.
Journal recommendation has long relied on supervised models, handcrafted features, or historical interaction data—all of which limit generalizability and interpretability. A new paper from Peking University introduces an LLM-powered semantic alignment framework that reframes the problem as a direct semantic matching task between manuscript content and journal scope descriptions. The system feeds article titles, abstracts, keywords, and candidate journal information into DeepSeek-V3, which then infers suitability without any fine-tuning. Tested on a dataset of 23,609 articles from 49 statistics and related-field journals, the framework achieves Top-3, Top-5, and Top-10 accuracies of 40.23%, 53.67%, and 70.05% respectively.
The framework also excels in stability and interpretability. Across repeated runs, Top-5 recommendations show an average Jaccard similarity of 84%, indicating consistent outputs. Incorporating reference information further boosts performance. Perhaps most importantly, the model provides natural language reasoning for each recommendation, giving researchers and editors insight into why a particular journal was suggested. This training-free, scalable approach could transform scholarly decision support—especially for early-career researchers navigating thousands of journals—by eliminating dependency on historical submission data and enabling transparent, context-aware recommendations.
- Uses DeepSeek-V3 on 23,609 articles from 49 journals without any fine-tuning or supervised training.
- Achieves 40.23% Top-3 accuracy, 53.67% Top-5, and 70.05% Top-10 accuracy.
- Generates interpretable reasoning for each recommendation with 84% Jaccard similarity for Top-5 across runs.
Why It Matters
LLMs can replace traditional recommendation systems for journal selection, offering scalable, transparent guidance without historical data.