AI Literature Review Assistant

The AI Literature Review Assistant helps researchers with literature reviews — a crucial but time-intensive step, especially the initial pass over a large corpus of research. Using keyword selection, database search, and machine learning, it automates classifying papers into “interesting” or “not interesting” based on the researcher’s focus. At its core is a binary-decision-tree model, optimized and deployed through a Jupyter notebook.

Features

• Automated paper classification — a binary-decision-tree model categorizes papers against predefined criteria, filtering out irrelevant ones early. Titles are pre-processed (noise removal, tokenization, feature extraction) and lemmatized so the classifier recognizes different forms of the same word.
• User-friendly GUI — makes reading and classifying titles and abstracts easy, with simple interactive elements to mark papers “interesting” or “not.”
• Scopus database queries — automates search and retrieval from Scopus, returning titles, authors, and key information for quick relevance assessment.
• Python & Jupyter optimization — a collaborative environment for training, tuning, and deploying the model, using metrics like accuracy, ROC/AUC, recall, and F1 — with tuning aimed at minimizing false negatives for “interesting” titles. In practice it reached up to ~85% accuracy and saved up to ~15 hours of manual triage on a review.
• Keyword extraction & query preparation — helps define keywords and build queries for searching Scopus.

At the heart of the classifier is the binary decision tree — trained on the researcher’s own accept/reject decisions, it learns which title features signal a relevant paper:

The corpus itself comes straight from Scopus: the tool runs the prepared queries and pulls back titles, authors, and metadata ready for triage:

Putting it together, the flowchart below shows how a review unfolds with the tool — from keyword definition through automated triage to the final reading list.