https://github.com/mihirkudale/megatron-nlp-project
https://github.com/mihirkudale/megatron-nlp-project
megatron natural-language-processing
Last synced: 7 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/mihirkudale/megatron-nlp-project
- Owner: mihirkudale
- Created: 2023-06-21T13:05:18.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2024-06-22T10:45:19.000Z (over 1 year ago)
- Last Synced: 2025-01-25T18:11:16.735Z (9 months ago)
- Topics: megatron, natural-language-processing
- Language: JavaScript
- Homepage:
- Size: 16.3 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# MegatronBot - Let's Chat
MegatronBot is a fully fleged chatbot with easy to update, integrate with website, easy to deploy in any cloud services like AWS, GCP and azure with a capibility to work in production enviorment.Megatron accepts various formats of inputs you can give a text input, you can also give a Speech as a input.
## What is the need of ChatBot?
Megatron was build by the team of ineuron.ai where I have worked upon a task of Intent Classification and State Tracking.The need of the MegatronBot came as ineuron.ai has a overwhemling responses and queries asked by the student over their newly released courses, as their is a Skype Support team to clarify their queries always but they cant be available 24x7 hours due to this ineuron.ai wants to build a such a great solution where the user can clarify their queries anytime without waiting in a queue for hours.
## How I have apporached to such Solution?
Building a Megatron like ChatBot requires a huge amount of data and various state of the art components.So, the first task is to get a large data,The dataset was creatd by scaraping the queries and answers asked by students over a year to their Skype support team, the dataset include of 20k sentences which were transformed and added to CSV and json format.
I have tried many State of the art language models from BERT-large to DialoGPT to RoBERTa but got an awesome results over Distilled BERT Models with ELMO embeddings.The model then trained over the 20k queries after preprocessing then adding tokens like [SEP], [START], [END] and [EOS].
The Architecture of MegatronBot: -
Below are some results:
