jeromepatel/cri-comp-2022-text-paraphrase-detection-challenge

Solution and experiments to the challenge CRI competition 2022 - text paraphrase detection

CRI-Comp-2022-Text-Paraphrase-Detection-Challenge - First Rank winner solution

Solution and experiments to the challenge CRI competition 2022 - text paraphrase detection

This repo contains raw code for winning solution of CRI-COMP-2022 Text Paraphrase challenge.

I'll add cleaned code in a separate repo with training instructions, however this repo contains code of all of my experiments, including failed ones.

• This repo includes the baseline approach using sentenceTransformer module.
• Using paraphrase_mining function which calculates sentence embeddings first and uses cosine sim distance to get paraphrase for query sentence.
• Please refer file bertbaseline.py for more details.
• Update: check the jupyter notebook which can directly be ran in colab, the final solution contains two stages, first stage uses fine-tuned bi-encoder embeddings and second stage uses pretrained cross-encoder pairs with custom postprocessing to get the results.
• Final solution scored 0.78 F1-score on private leaderboard (test set) with rank 1.