ThisIsNotARealTalkshow

TINART is a student project from the 2020 fall semester at CODE University of Applied Sciences. It allows the user to simulate an interactive political talk show with intelligent agents representing the rethorical style of selected German politicians.

Getting started

Prerequisites

There should be a recent Python version (3.x) installed on your computer.

Moreover, we recommend using a common web browser like Chrome or Firefox.

Setting up an environment

First, navigate to the local directory you would like to place the project code in. Then, clone the TINART repository.

cd <PATH_TO_DIRECTORY>
git clone https://github.com/JokusPokus/TINART_app.git

Create a virtual environment. The first command is only required if the virtualenv package is not yet installed on your machine.

pip install virtualenv
virtualenv venv

Activate the virtual environment.

For Linux users:

source venv/bin/activate

For Windows users:

.\venv\Scripts\activate

Alternatively, you can use your preferred Python IDE and select the venv there. This project was created using PyCharm.

Next, install all required packages.

pip install -r requirements.txt

To run the application locally, you need to execute the application file:

python application.py

Note that iOS users need to replace the pip and python commands with pip3 and python3, respectively.

You may also wish to install the app in your virtual environment. Make sure to navigate to the Nim_AI root directory and execute:

pip install -e .

Language Models

A language model must be provided for each politician that can be selected in the talk show. A politician's language model must be saved in the following directory:

\language_model\gpt2-{}

where {} is a placeholder for the politician's name. For example:

\language_model\gpt2-merkel

All models must be consistent with the AutoModelForCausalLM requirements from the transformers library. We recommend using this repository for model fine-tuning.

Built with

• Python 3.7.4
• Flask 1.1.2
• Werkzeug 1.0.1

Authors

• Jakob Schmitt (NLP, Backend)
• Igor Lapinski (Design, Frontend)