Live at:- https://huggingface.co/spaces/Shiva-gsk/Text-Summarizer-with-T5
Text Summarization on Product Reviews using Pretrained Transformer So, We basically need to use transformers library from huggingface and do text-summarization task on Amazon Product Reviews
Firstly, I searched Internet to get a good dataset Containing Product Reviews and summaries. Then I explored Dataset using Kaggle and found various insights regarding data. Next, I started to experiment summarization with existing pipeline using T5-small Transformer from Hugging-face and came to know that model's performance on the data set is not good (ROUGE ~= 0.08) So, I started Experiments on Finetuining model on my Local Machine. With only training it for few hours the ROUGE score increased. (ROUGE ~= 0.14) So, I confirmed an increase in model performance on finetuning and Did some training on dataset using Colab. After training i tested it any test dataset and got a ROUGE score of 0.137 which is not so bad on t5-small. Results might be better using t5-large but a lot of computational power is required for that. (I think start of art at the time of T5 is around 0.38)
On Final Training the model for about 2 hours, it achieved a ROUGE score of 0.1370
Word-cloud on summaries
Word-cloud on review-text
We can see that Summaries contains more adjectives (e.g. great, good) and some product related words
- As reviews has a long difference of length (ranging from few words to length paragraphs) it was hard to set the min_length and max_length for output tokens by module.
- However after test trying models many times i came to conclusion on what to be done.
- The model size of so large that i thought it's hard to host it
- However doing it with huggingface spaces using gradio made things easier