Our team participated in the CupIT competition in which we had to develop a language model for comment ranking in a social network. We were given a dataset with posts, where each post contains a post's text and texts of four comments to it. The comments were sorted by their rating. We conducted an EDA and decided to train DistilBERT for this task.
We used Pyhton3 and Google Collab for this task. We used such modules as:
- scikit-learn
- nltk
- numpy
- pandas
- transformers
- pytorch
- matplotlib
- json
- textblob
- textstat
- spellchecker
After training the model acheived NDCG metric value of 0.88.
Correlation between comment length and its rating (with 0 being the highest and 4 the lowest).
Distribution of positive/negative comments among all.
Correlation between the percentage of spelling mistakes in a commentary and its rating (with 0 being the highest and 4 the lowest).