Accompanying code for participating system at PAN 2021: Profiling Hate Speech Spreaders on Twitter. See this link for our paper with the system description.

Usage

For training or testing the model, run the main.py script with the corresponding parameters. It must be taken into account that we presented different approaches for modeling the users' profiles, for this some parameters differ from one model to another. Also, our architecture is modular, hence every module is trained independently.

Tweets Encoder

This module is based on pre-trained models from HuggingFace Transformer Library and pre-trained based models will be downloaded after the training process. To encode independently each tweet you need to train the Encoders as follows:

    python main.py -l <language> -dp <data-path> -mode tEncoder [-bs <bsize>]

Where:

• language is the data language. For using adapters and build a multilingual model, add a _ at the end of the language (e.g., either ES_ or EN_)
• data-path is the path for the training data
• bsize is the size of the batch for training the models. By default, it is set to 64 and 200 for training and encoding phases respectively.

To Encode the tweets you need to run the same command as before, but setting -mode encode and data-path with the path of data to be encoded.

Profile Modeling

In this step, we are assuming that the isolated tweets have already been encoded by the Tweets Encoder. The profile is modeled by means of an FCNN or a Spectral Graph Neural Net to train these models you must run:

    python main.py -l <language> -mode <model> -epoches <epoch> -phase train -bs 32

Where model is the kind of deep model to be employed (cgnn or tfcnn for convolutional o FCNN respectively) and epoch can be set to -1 to use the default epoch hyperparameter. To encode the profiles just need to set -phase to encode.

Deep Impostor Method

This method needs to have available the modeling of the training profiles for making predictions on test data. For making predictions run:

    python main.py -l <language>  -mode tImpostor -dp <data-path> -rp <proportion> -metric cosine -up <pselection> -ecnImp <encoding source> -dt <data-test> -output <out>

Here again, we are assumed that profile modelings have been computed for training as well as for test data.

• data-path is the path of data for training
• proportion is the proportion of positive and negative classes to be taken as prototypes.
• pselection can be set to random or prototipical to change the approach for selecting the prototypes from random to PCA method exposed in the paper.
• encoding source is the way in which the profiles have been modeled, it can be set to transformer to use the average method exposed in the paper, gcn to employ the Convolutional Graph-based method or fcnn to use the FCNN method.
• data-test is the path of testing data.
• out is the path to save predictions.