ADPT-BE

Overview

ADPT-BE stands for Automatic Detection of Persuation Techniques - Backend.

The repository contains a small Flask application that loads the model and exposes an endpoint to perform model inference.

To have a better understanding of the model and how it was created, you can check ADPT-AI which contains the code and research papers produced during the work.

Currently, the model created only supports inference for textual data. There are plans in the future to add image classification.

Endpoint

Method	Resource	Query Parameters	Body
POST	predict	-	{"text": $sentence, "probs": true|false}

Bellow there is an example of a response setting probs to true, and after that, an example of setting it to false.

{
    "Appeal to authority": 0.08427731692790985,
    "Appeal to fear/prejudice": 0.3441292643547058,
    "Bandwagon": 0.1411733627319336,
    "Black-and-white Fallacy/Dictatorship": 0.11429373919963837,
    "Causal Oversimplification": 0.14675235748291016,
    "Doubt": 0.04427878186106682,
    "Exaggeration/Minimisation": 0.5388496518135071,
    "Flag-waving": 0.3002663254737854,
    "Glittering generalities (Virtue)": 0.1867496371269226,
    "Loaded Language": 0.9543203115463257,
    "Misrepresentation of Someone's Position (Straw Man)": 0.29040059447288513,
    "Name calling/Labeling": 0.10895266383886337,
    "Obfuscation, Intentional vagueness, Confusion": 0.2316785603761673,
    "Presenting Irrelevant Data (Red Herring)": 0.03351626917719841,
    "Reductio ad hitlerum": 0.07988007366657257,
    "Repetition": 0.11493965983390808,
    "Slogans": 0.5039657950401306,
    "Smears": 0.030777540057897568,
    "Thought-terminating cliché": 0.07756602019071579,
    "Whataboutism": 0.026369892060756683
}

[
    "Exaggeration/Minimisation",
    "Loaded Language",
    "Slogans"
]

Deployment

As the time, the service is not hosted anywhere due to free cloud providers resource limitations.

It can be deployed locally using the provided Dockerfile.