We present our experiments on RuATD Shared Task Dataset that containts sentences generated by humans and different generative models.
Our pipeline was constructed in a following way:
We experimented with models for binary classification (baselines are marked as bold):
| encoding | classifier | binary: accuracy |
|---|---|---|
| BoW+TF-IDF | FNN | 0.73 |
| word2vec | LSTM | 0.59 |
| character | CNN | 0.66 |
| RuBERT-tiny | fine-tuning | 0.81 |
| text features | FNN | 0.6 |
| text features | Logreg | 0.61 |
| text features | RandomForest | 0.64 |
| BERT | 0.79622 | |
| TF-DF | Logreg | 0.63562 |
We chose RuBERT-tiny as the best model.
Then we tried different models on classification of texts that RuBERT-tiny recognised as "non-human".
| encoding | classifier | accuracy | macro f1-score |
|---|---|---|---|
| TF-IDF | FNN | 0.3 | 0.29 |
| word2vec | LSTM | 0.29 | 0.25 |
| character | CNN | 0.37 | 0.32 |
| text features | FNN | 0.26 | 0.17 |
| text features | Logreg | 0.22 | 0.1 |
| text features | RandomForest | 0.27 | 0.24 |
| Combo Model | 0.39 | 0.35 |
We chose Combo model that includes parts of other models' acrhitectures in the following way:
Combined with RuBERT-tiny, the Combo model shows 0.5 accuracy.
For details see our presentation (in Russian).
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent