Paraphrase Detection on Microsoft Paraphrase Corpus
89.5AccuracyRoBERTa
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RoBERTaModel Size=large2021.07 | 89.5 | 92.2 | |
| ParaBLEUModel Size=large2021.07 | 88.8 | 91.5 | |
| ALBERTModel Size=large2021.07 | 88.2 | 91.3 | |
| ParaBLEUModel Size=base2021.07 | 85.2 | 88.9 | |
| TF-KLDTraining Methodology=Supervised, Transfer Protocol=No Transfer2017.05 | 80.4 | 85.9 | |
| BiLSTM-Max (on AllNLI)Training Methodology=Supervised, Source Dataset=AllNLI2017.05 | 76.2 | 83.1 | |
| SCD-BERT_baseBackbone=BERT-base2022.03 | 75.71 | — | |
| IS-BERT_baseBackbone=BERT-base2022.03 | 74.24 | — | |
| SCD-RoBERTa_baseBackbone=RoBERTa-base2022.03 | 73.71 | — | |
| Unigram-TFIDFTraining Methodology=Unsupervised2017.05 | 73.6 | 81.7 | |
| SimCSE-BERT_baseBackbone=BERT-base2022.03 | 73.51 | — | |
| Skip-thoughtModel=Skip-thought2022.03 | 73 | — | |
| GloVe embeddings (avg.)Model=GloVe (avg.)2022.03 | 72.87 | — | |
| BERT [CLS]-embeddingBackbone=BERT-base, Pooling=CLS2022.03 | 72.29 | — | |
| SimCSE-RoBERTa_baseBackbone=RoBERTa-base2022.03 | 72.23 | — | |
| RoBERTa [CLS]-embeddingBackbone=RoBERTa-base, Pooling=CLS2022.03 | 72.17 | — | |
| Avg. BERT embeddingsBackbone=BERT-base, Pooling=Avg2022.03 | 69.54 | — | |
| MTMETRICS2017.04 | 0.774 | 0.841 | |
| SDAE2017.04 | 0.764 | 0.834 | |
| BYTE MLSTM2017.04 | 0.75 | 0.828 | |
| SKIPTHOUGHT2017.04 | 0.73 | 0.82 | |
| BERTbackbone=BERT2019.11 | — | 88 | |
| BERT+DLbackbone=BERT, loss=Dice Loss2019.11 | — | 88.71 | |
| BERT+DSCbackbone=BERT, loss=DSC2019.11 | — | 88.92 | |
| BERT+FLbackbone=BERT, loss=Focal Loss2019.11 | — | 88.43 | |
| XLNetbackbone=XLNet2019.11 | — | 89.2 | |
| XLNet+DLbackbone=XLNet, loss=Dice Loss2019.11 | — | 89.33 | |
| XLNet+DSCbackbone=XLNet, loss=DSC2019.11 | — | 89.78 | |
| XLNet+FLbackbone=XLNet, loss=Focal Loss2019.11 | — | 89.25 |