search

PrithivirajDamodaran / Parrot_Paraphraser

A practical and feature-rich paraphrasing framework to augment human intents in text form to build robust NLU models for conversational engines. Created by Prithiviraj Damodaran. Open to pull requests and other forms of collaboration.
Apache License 2.0
867 stars 143 forks source link

Parrot returns very similar description without paraphrasing for some sentences #4

Closed ElisonSherton closed 3 years ago

ElisonSherton commented 3 years ago

Hi Prithviraj,

Good Day!

Awesome work on building this library! I tried to use it for a personal project from the fashion domain and here's what I observed:

image

image

Have a look at the two sentences above. I have provided the input sentence and the paraphrased sentence obtained using parrot. Except some punctuation and contractions, there's not much that the model is able to do.

Such is the case even for most of the descriptions that I have scraped through fashion retailers. Could you advise how can I use parrot to obtain better paraphrased suggestions please?

Thanks & Regards, Vinayak Nayak.

liuzzi commented 3 years ago

i have also noticed this.

PrithivirajDamodaran commented 3 years ago

There is no universal model similarly there is no universal paraphraser which will work across all text nuances from across domains. Performance and variety is a function of the data it has seen.

For your needs and domain text feel free to fine tune your own paraphraser.