About The Project

Master Thesis: Open-domain Conversational Agent based on Pre-Trained Transformers for Elderly-Robot Interaction (ongoing)

Motivation:

• A project that solves a problem and helps others
• Increase knowledge in Deep Learning
• Master Thesis

Built With

This section should list any major frameworks that you built your project using. Leave any add-ons/plugins for the acknowledgements section. Here are a few examples.

• DialoGPT
• Flask
• Streamlit
• Raspberry Pi Zero W

Getting Started

TODO

Prerequisites

• List of things you need:

  pip install -r requirements.txt
  

Mini-Babel

When it comes to evaluating the performance, the ideal would be to have a human-based evaluation. Specially in a system where paraphrasing can lead to inferior metrics, despite still having a good performance. Consequently, to add a score to the translated dataset, a small interface was developed. In this User Interface: http://20.108.185.129:8501/ , from a dataset storing almost 8000 source and target segments pairs, 14 random are displayed to the user. The user has then the option to classify the translation from 1-10.

GrandPal-Chat

To add a score to the performance of the Chatbot, a small interface was developed. In this User Interface: http://20.108.185.129:8052/ , the user has a chance to engage in a conversation with the system. After a six utterance interaction with the system, the user has then the option to classify the conversation from 1-10. One more advantage of this evaluation is allowing to analyze the ability of the model to have context during the dialogue. Since the system, when predicting a new reply, has the conversation history as input. Considering that the model was only trained with 7 sentences for context, the input only stores the last 6 uttered sentences.

GrandPal-Model

• model-app: has the code to build the bot along with a zip contained the fine-tuned model
• model-training: has the code to train and evaluate the model, along with the directory containing the datasets
• mt-api: has the zip containing the mt open source model
• raspberry: has the code to give the bot a body (specifically a Raspberry Pi Zero W). Running the client in the raspberry and the server on a computer, it is possible to build a physical voice assistant. It also comprises a weather and radio API.

Local GrandPal

Spin up GrandPal locally in Docker

• Install Docker
• To build the image that holds the streamlit server, go to GrandPal/GrandPal-Chat/ and run:

  docker build -t grandpal_chat_streamlit .
  

• To spin Grandpal locally:
  1. Go to GrandPal/GrandPal-Model/model-app/
  2. Unzip the GrandPal zip file. (In case this doesn't work, download it from (here)[https://github.com/marianafidalgo/GrandPal/blob/master/GrandPal-Model/model-app/GrandPal-20211024T182801Z-001.zip] and unzip it)
  3. Give the Unzipped directory the name Grandpal
  4. Run:

  docker-compose build
  docker-compose up
  

Connect to GrandPal:

Open the browser in Grandpal Streamlit The first interaction will take a while since it is downloading all the necessary models.

Contact

Mariana Fidalgo

Project Link: https://github.com/marianafidalgo/GrandPal