Political Misogynistic Discourse Monitor is a web application and API that detects hate speech against women in Spanish and Portuguese.
This project is part of the 2021 JournalismAI Collab Challenges, a global initiative that brings together media organisations to explore innovative solutions to improve journalism via the use of AI technologies. It was developed as part of the Americas cohort of the Collab Challenges that focused on How might we use AI technologies to innovate newsgathering and investigative reporting techniques? in collaboration between AzMina (Brazil), La Nación (Argentina), CLIP (Colombia) and Data Crítica (México) with the support of the Knight Lab at Northwestern University.
JournalismAI is a project of Polis – the journalism think-tank at the London School of Economics and Political Science – and it’s sponsored by the Google News Initiative. If you want to know more about the Collab Challenges and other JournalismAI activities, sign up for the newsletter or get in touch with the team via hello@journalismai.info
Team members:
- Bárbara Libório
- Marina Gama Cubas
- Helena Bertho Dias
- Gabriela Bouret
- Jose Luis Peñarredonda
- Fernanda Aguirre
We would like to thank Ivan Vladimir for all his help developing the software and the web application. We also want to acknowledge IIMAS for hosting the project.
This collaboration is an attempt to accelerate the development of MonitorA, a project by AzMina with InternetLab and Institute Update, which gathered evidence and insight of systematic misogynistic attacks of women candidates in Brazilian local elections of 2020.
According to the report Violence Against Women in Politics, this kind of violence is a deterrent to the participation of women in the political sphere, where women from marginalized communities are disproportionately affected. UN Women states that in Latin America’s context, women hold barely 30% of parliamentary seats. In addition, UN Women stands out that “gender equality in the highest positions of power will not be reached for another 130 years”. The mentioned facts lead us to analyze how the violence against women is perpetrated and has an impact in their participation. We want to report this kind of disinformation and attacks throughout Latin America in an effort to motivate new narratives where women have a safe space in their involvement in politics.
For the above reason, although this AI model is able to identify violence against women in general, we want to focus on misogyny in political discourse as a case study in Latin America. In our project, we support that the automation of detecting misogynistic discourse is just a tool to help identifying attacks against women among a large volume of data on Twitter, so the system highlights content that can be analyzed by a human moderator afterward.
Since the collaborators are from Latin American countries, this model was trained with Spanish and Portuguese tweets posted from 2020 to 2021. We retrieved 4179 tweets from Twitter in 'csv' format.
There are missing 270 tweets from the database we used to train the model and the database we share in this repository since we couldn't recover the IDs from those tweets. All the amounts from the data analysis belong to the database training.
| database training | database repository |
|---|---|
| 4179 | 3909 |
We created a dictionary on Spanish and one on Portuguese with misogynistic terms and phrases. Along with that, we made a list of usernames for trendy politicians. However, we considered those accounts wouldn't be inclusive enough, so we decided to make a second list exclusively for diverse women (black, indigenous and LGBTQIA+) politicians, journalists and activists from Brazil, Argentina, Colombia and Mexico. Therefore tweets mentioning both lists of usernames were collected from Twitter using Meltwater and filtered by the dictionaries with regular expressions.
The data file includes three columns:
-
ID: As Twitter's policy prevents from sharing tweets messages, we only included the ID from each tweet, considering that IDs are allowed to be downloadable and can be transformed into the original text using available tools.
-
Classification: Tweets are annotated with the label
1if they are misogynistic or0if they are not. Misogynistic discourse were positive in 2637 tweets and negative in 1542 tweets.
- Language: There's a label for the language of the tweet,
esfor Spanish andptfor Portuguese. There are 2087 tweets on Spanish and 2092 on Portuguese.
The annotation for this database to detect misogyny was performed by six human annotators (five women and one man) which first languages are Spanish or Portuguese and that are based on the country of each dataset (Brazil, Argentina, Colombia and Mexico). In order to validate the annotation, all the classification labels had a checker different from the first annotator. If the checker agreed on the label, the classification remained. Otherwise, the tweet was removed from the database.
In order to create the classifier, we made use of five Colaboratory Python Notebooks:
- Data analysis: Basic analysis and statistics of the data.
- Train and evaluate model (2 versions): Trains a model and evaluates it, one for Transformers and another for Adapters.
- Labelling data (2 versions): Labels data from entry form from the notebook or from a file, one for Transformers and another for Adapters.
There are several pre-processing steps on Natural Language Processing that can be applied to the data:
- Lowers: All words are lowered. (e.g., GitHub → github)
- Stop words: Remove words that are very common but don’t provide useful information. (e.g., prepositions)
- Demojize: Change emojis to textual representation. (e.g.,
☺️ → :smiling_face:) - URLs: Replace URLs with
$URL$(e.g., https://github.com/ → $URL$) - Mentions: Replace mentions with
$MENTION$(e.g., @github → $MENTION$) - Hashtags: Replace hashtags with
$HASHTAG$(e.g., #github → $HASHTAG$) - Emojis: Replace emojis with
$EMOJI$(e.g., 😃 → $EMOJI$) - Smileys: Replace smileys with
$SMILEY(e.g., :) → $SMILEY) - Numbers: Replace numbers with
$NUMBER$(e.g., 4 → $NUMBER$) - Escaped characters: Replace escaped characters with
$ESCAPE_CHAR$(e.g., char(2) → $ESCAPE_CHAR$)
It is worth mentioning that we obtained better results lowering the text.
Along with that, we followed a machine learning methodology in which we used part of the labelled data to train a model which then is tested on another part of the data. During training we validated the progress of the model using a third part of the data.
| Split | Percentage | Tweets |
|---|---|---|
| Train | 80% | 3,343 (1673 pt, 1669 es) |
| Test | 10% | 418 (210 pt, 209 es) |
| Validation | 10% | 418 (209 pt, 209 es) |
This section shows some statistics and graphics of the labelled data.
| Frequency | Description | |
|---|---|---|
| count | 19063 | Number of different words |
| mean | 3.444841 | The average number words appear |
| std | 13.935922 | The standard deviation associated to the words |
| min | 1 | The minimum number that a word appears |
| 25% | 1 | Up to 25% of the words appear |
| 50% | 1 | Up to 50% of the words appear |
| 75% | 2 | Up to 75% of the words appear |
| max | 1062 | The maximum number that a word appears |
This graph shows the full vocabulary of the data:
This graph shows the fifty most common words in the data:
These graphs show the number of tweets with a certain length:
This is a wordcloud with the most common words:
We tested several Transformers and Adapters models. Nevertheless, cardiffnlp/twitter-xlm-roberta-base was the one with the better performance on F1 score:
| Model | Type | both | es | pt |
|---|---|---|---|---|
| cardiffnlp/twitter-xlm-roberta-base | Multilingual | 0.8728 | 0.9191 | 0.8235 |
| neuralmind/bert-base-portuguese-cased | Portuguese | - | - | 0.875 |
| dccuchile/bert-base-spanish-wwm-uncased | Spanish | - | 0.8985 | - |
| mudes/multilingual-base | Multilingual | 0.8641 | 0.8929 | 0.8339 |
| neuralmind/bert-base-portuguese-cased | Portuguese | - | - | 0.8496 |
| PlanTL-GOB-ES/roberta-base-bne | Spanish | - | 0.9027 | - |
For more information about all the model performances, checkout this technical report.
This is the workflow structure we follow for the project:
To enable communication with the API, we need a HTTP library to make a request-response. There are a few libraries to make HTTP requests in Python. However, we'll make use of requests due to it is well-documented and simple.
Installing package with conda:
conda install requestsInstalling package with pip:
pip install requestsThe POST method is used when we want to submit data to be processed to the server. Here's an example of the syntax:
requests.post(url, headers={key: value}, json={key: value}, data={key: value})
For more information about HTTP request methods, checkout this guide.
| Parameter | Description |
|---|---|
| url | A string with the endpoint |
| headers | A dict to send to the url |
| json | A dict to send to the url |
| files | A dict of files to send to the url |
| data | A dict or list of tuples to send to the url |
The status code method shows the result when a request is sent. Responses can be grouped in five categories:
- Informational
100-199 - Succesful
200-299 - Redirection
300-399 - Client error
400-499 - Server error
500-599
For more information about HTTP response status codes, checkout this guide.
import requests
url = 'https://turing.iimas.unam.mx/pmdm/api/classify'
headers = {'access-token': 'token'}
tweet = {'tweet': 'text to classify'}
response = requests.post(url, headers=headers, json=tweet)
print(response.status_code)
response.json()Default tweet arguments:
{
'tweet': 'string',
'use_lower': 'false',
'demojize': 'true',
'process_urls': 'true',
'process_mentions': 'true',
'process_hashtags': 'true',
'process_emojis': 'false',
'process_smileys': 'false',
'process_numbers': 'false',
'process_escaped_chars': 'false'
}import requests
url = 'https://turing.iimas.unam.mx/pmdm/api/classify_file'
headers = {'access-token': 'token'}
files = {'uploaded_file': open('filename', 'rb')}
# Tweet arguments required
data = {
'model': 'es',
'use_lower': 'false',
'demojize': 'true',
'process_urls':'true',
'process_mentions': 'true',
'process_hashtags': 'true',
'process_emojis': 'false',
'process_smileys': 'false',
'process_numbers': 'false',
'process_escaped_chars': 'false'}
response = requests.post(url, headers=headers, files=files, data=data)
print(response.status_code)
response.json()For more examples, see this Jupyter Notebook
For future work we would like to create datasets from Latin American countries not included at this point in the interest of keeping the model's training. Furthermore, we will use the API to streamline the detection and to analyze instances of misogynistic discourse on social media.
Since we are aware that the management of an API is still not very accessible for a lot of newsrooms in the region due to technical requirements, we want to document and methodize use applications that hopefully inspire and help other organizations to work with this tool.
If you want to collaborate or just to know more about the project, please reach out to us:
violentometro-online -> Documentation
- Datasheets for Datasets
- Ethical and technical challenges of AI in tackling hate speech
- Detección de Discurso de Odio en Redes Sociales mediante Transformers y Natural Language Processing
- Violência Política de Gênero: as diferenças entre os ataques recebidos por mulheres e seus oponentes
- Tackling Online Abuse and Disinformation Targeting Women in Politics
- #ShePersisted: why gendered disinformation