The aim of this project is to create Abstractive opinion summarization using the power of transformers.
The study focuses on training different Transformers used in this project that is:
Stacking the transformers encoder and decoder with pre-trained models.
Using pre-trained transformers model which are trained as a whole on data ex. T5
It also compares the performance of different Transformer models based on the above training techniques. The models are Bert2Bert ,T5and Roberta2Roberta .
The pipeline is based on Opinion Digest Framework
| Model | Rouge-1 | Rouge-2 | Rouge-L |
|---|---|---|---|
| Bert2Bert | 25.03239 | 4.05246 | 16.35603 |
| Roberta2Roberta | 25.96 | 4.08 | 16.66 |
| T5 | 27.97 | 4.6 | 17.1 |