Competition link | Kaggle notebook
The repo contains copy of the Kaggle notebook.
Since research papers are large texts, I first created heuristics to find paragraphs which might contain dataset mentions using spacy's text matching/regex. On those paragraphs, I applied fine-tuned RoBERTa to extract custom dataset entity.
The work involved
- converting the annotations to CoNLL-U format, and
- training the transformer for multiple epochs.
To avoid overfitting only a subset of randomly selected labels was used along with randomly selected negative examples.
The results of this approach were capable of achieving a top 100 rank. The winning solution trained another model instead of applying heuristics for the first step of filtering paragraphs
