Group mention non-economic attributes classifier

A multi-label classifier for detecting non-economic attribute categories referred to in a social group mention, trained with setfit based on the light-weight sentence-transformers/all-mpnet-base-v2 sentence embedding model.

The non-economic attributes classified are:

attribute	definition
age	People referred to based on or categorized according to their age, generation, or cohort such as children, young people, old people, future generations.
family	People referred to based on or categorized according to their familial role such as fathers, mothers, parents.
gender/sexuality	People referred to based on or categorized according to their gender or sexuality such as men, women, or LGBTQI+ people.
place/location	People referred to based on or categorized according to their place or location such as peolple from rural areas, urban center, the global south, or global north.
nationality	People referred to based on or categorized according to their nationality such as natives or immigrants.
ethnicity	People referred to based on or categorized according to heir ethnicity such as people of color or ethnic minorities.
religion	People referred to based on or categorized according to their religion of belief such as christians, jews, muslims, etc.
health	People referred to based on or categorized according to their health condition or relation to aspects of health such as disabled/handicapped people or chronically sick people.
crime	People referred to based on or categorized according to their relation to crime such as offenders/criminals or victims.
shared values/mentalities	People referred to based on or categorized according to their shared values or mentalities such as people with a growth mindset, meritocratic values, environmental or peace mentalities or a more equal society.

Model Details

Model Description

Group mention non-economic attributes classifier

Developed by: Hauke Licht
Model type: mpnet
Language(s) (NLP): ['en']
License: apache-2.0
Finetuned from model: sentence-transformers/all-mpnet-base-v2
Funded by: The Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy – EXC 2126/1 – 390838866

Model Sources

Repository: tba
Paper: tba
Demo: [More Information Needed]

Uses

Bias, Risks, and Limitations

Evaluation of the classifier in held-out data shows that it makes mistakes.
The model has been finetuned only on human-annotated labeled social group mentions recorded in sentences sampled from party manifestos of European parties (mostly far-right and Green parties). Applying the classifier in other domains can lead to higher error rates.
The data used to finetune the model come from human annotators. Human annotators can be biased and factors like gender and social background can impact their annotations judgments. This may lead to bias in the detection of specific social groups.

Recommendations

Users who want to apply the model outside its training data domain should evaluate its performance in the target data.
Users who want to apply the model outside its training data domain should contuninue to finetune this model on labeled data.

How to Get Started with the Model

Use the code below to get started with the model.

Usage

You can use the model with the setfit python library (>=1.1.0):

Note: It is recommended to use transformers version >=4.5.5,<=5.0.0 and sentence-transformers version >=4.0.1,<=5.1.0 for compatibility.

Classification

import torch
from setfit import SetFitModel

model_name = "haukelicht/all-mpnet-base-v2_noneconomic-attributes-classifier"
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
classifier = SetFitModel.from_pretrained(model_name)
classifier.to(device);

# Example mentions
mentions = ["working class people", "highly-educated professionals", "people without a stable job"]

# Get predictions
with torch.no_grad():
    predictions = classifier.predict(mentions)
print(predictions)

# Map predictions to labels
[
    [
        classifier.id2label[l]
        for l, p in enumerate(pred) if p==1
    ]
    for pred in predictions
]

Mention embedding

import torch
from sentence_transformers import SentenceTransformer

model_name = "haukelicht/all-mpnet-base-v2_noneconomic-attributes-classifier"
device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"

# Load the sentence transformer component of the pre-trained classifier
model = SentenceTransformer(model_name, device=device)

# Example mentions
mentions = ["working class people", "highly-educated professionals", "people without a stable job"]

# Compute mention embeddings
with torch.no_grad():
    embeddings = model.encode(mentions)

Training Details

Training Data

The train, dev, and test splits used for model finetuning and evaluation will be made available on Github upon publication of the associated research paper.

Training Procedure

Training Hyperparameters

num epochs: (1, 4)
train batch sizes: (32, 4)
body train max teps: 75
head learning rate: 0.010
L2 weight: 0.01
warmup proportion: 0.15

Evaluation

Testing Data, Factors & Metrics

Testing Data

The train, dev, and test splits used for model finetuning and evaluation will be made available on Github upon publication of the associated research paper.