Publication Classifier

About Publication Classifier

Publication Classifier is a hybrid NLP system that analyzes research papers to determine their publishability and recommend the most suitable conference. It combines modern transformer models like SciBERT and Sentence-BERT with classic machine learning and self-training strategies to generate interpretable, high-quality predictions.

The classifier uses text embeddings from SciBERT to evaluate whether a paper should be published. If deemed publishable, a semantic similarity approach using Sentence-BERT matches the paper with top-tier conferences such as CVPR, NeurIPS, EMNLP, KDD, and TMLR.

• SciBERT for Content Understanding
  Generates deep contextual embeddings from full paper content for classification tasks.

• Self-Training Publishability Classifier
  Uses a small set of labeled papers to iteratively train on larger unlabeled datasets with pseudo-labels.

• Sentence-BERT for Conference Matching
  Identifies the most semantically relevant conference based on similarity with topic prototypes.

• Plug-and-Play with CSV Input
  Accepts labeled and unlabeled CSVs, and outputs results to results/output.csv—ready for evaluation or submission.

• Expandable & Modular Codebase
  Easy to fine-tune, extend with new conferences, or upgrade with better models like Longformer or GPT.

Tech Stack

• Embeddings & Transformers: SciBERT, Sentence-BERT, Hugging Face Transformers
• Modeling: scikit-learn, self-training classifier
• Data Handling: pandas, numpy
• Evaluation: precision, recall, F1, confusion matrix
• Execution: Python 3.8+, CLI-compatible scripts

Credits

• SciBERT: For domain-specific contextual embeddings of scientific content.
• Sentence-BERT: For high-quality semantic similarity computation.
• scikit-learn: For baseline classification pipelines and metrics.
• pandas/numpy: For efficient CSV handling and data preprocessing.

Author

Developed by Ayush Sharma & Rishabh Kothari.
Check out the full project on Here.
Check out more projects on GitHub or reach out via LinkedIn.