24.10.18 - Text Simplification Experiments

Find Text Simplification Corpora $D$ ;
1. Split in train / test;
(Optional) Use a instruction-tuned model $m$ to do zero-shot prompt-based simplification;
1. Evaluate the results using SARI, READ-IT and other classic text-simplification / readability evaluation metrics;
Fine-tune the pre-trained model for text-simplification $m$ using $D$ , obtaining $m_{D}$ ;
1. Evaluate the results by using SARI, READ-IT and other classic text-simplification evaluation metrics;
Use a collection of texts (possibly from multiple domains) and use $m$ to simplify its sentences:
1. Sample at least 2 simplified version of each sentence;
2. Rank the options, the best is the one tagged as the preferred the other as the rejected;
Use DPO to further fine-tune $m_{D}$ and evaluate it using the same evaluation metrics as the previous steps.

Notes

Check how DPO models are evaluated.

Admin-it (OP) = Simpitiki PA

Totale:
1212 test
11152 train