sebastiaanram / mlp-coleridge

Let US show you the data

This repository houses notebooks belonging to the "Coleridge Initiative - Show US the data" competition, hosted by Kaggle. The notebooks have been set up by three students of team 2 of the Radboud University Masters course "Machine Learning in Practice".

Dataset

The dataset consists of 14.000 training publications with 130 unique labels. The hidden test set has roughly ~8.000 publications, many times the size of the public test set.

Goal

The goal of this competition is to predict the occurance of data set labels within the text of scientific publication using natural language processing techniques.

Methods

We attempted a named entity recognition (NER) approach using the spaCy library. Our best model originated from a SciBERT pretrained transformer and resulted in a public score of 0.028.

Specifications

The notebooks present in this repository were directly downloaded from Kaggle. The notebooks were written in Python and ran on a Kaggle GPU accelerator. Additionally, libraries like spaCy, PyTorch, NLTK and Huggingface Transformers were used in combination with our own generated data sets.

Notebooks referenced

Prashans Dixit (2021, April). 📝Coleridge Initiative-EDA📚 & Baseline Model🎯. https://www.kaggle.com/prashansdixit/coleridge-initiative-eda-baseline-model