Scrib-AI is a student project that aims to summarize texts in an abstract way. You can test this algorithm directly on our site: http://vps528736.ovh.net/ and the code is in opensource on github https://github.com/AdriBenben/Scrib-AI-model-pytorch (for the python code) and https://github.com/AdriBenben/Scrib-AI (for the website).
This project is coded in pytorch version 0.3.0 and runs on GPUs. Depency of nltk.tokenize
In the Tutorial-projet-ScribAI-English.ipynb, you can find:
1- To reproduce the Seq2seq model of the SaleForces publication, without the reinforcement part.
2- To create a template called "Reward" to score the summary from the feedbacks.
3- To subtrain these two models independently. The first on CNN-Daily and the second with the red metric on the summaries generated by Seq2Seq.
4- Connect the two models and continue learning with the left and right human labeling by connecting it to the interface.
5- To be able to evaluate it
6- To explain our approach.
At the end of this project, our model will be able to: put a note to articles and summarize articles according to this standard
WORK IN PROGRESS ! We have to train our model !
This work is a 8 months (hard-) work. Human Feedbacks are very welcome.