augmentedsocialscientist 3.0.0

A Package to Easily Train Bert-Like Models for Text Classification

The Augmented Social Scientist

This package makes it extremely easy to train BERT-like models for text classifications.

For more information on the method and for some use cases from social sciences, see "The Augmented Social Scientist: Using Sequential Transfer Learning to Annotate Millions of Texts with Human-Level Accuracy" published on Sociological Methods & Research by Salomé Do, Étienne Ollion and Rubing Shen.

To install the package

pip install augmentedsocialscientist

To train a BERT model

Import

from augmentedsocialscientist.models import Bert

bert = Bert()  #instanciation

Functions

The class Bert contains 3 main methods:

• encode() to preprocess the data;
• run_training() to train, validate and save a model;
• predict_with_model() to make predictions with a saved model.

Example

import pandas as pd
import numpy as np

from augmentedsocialscientist.models import Bert

bert = Bert() #instanciation


# 1. Training 
## Load training and validation data
train_data = pd.read_csv('https://raw.githubusercontent.com/rubingshen/augmented_tutorial/main/clickbait/clickbait_train.csv')
test_data = pd.read_csv('https://raw.githubusercontent.com/rubingshen/augmented_tutorial/main/clickbait/clickbait_test.csv')

## Preprocess the training and validation data
train_loader = bert.encode(
    train.text.values,      #list of texts
    train.label.values      #list of labels
    )    
test_loader = bert.encode(
    test.text.values,       #list of texts
    test.label.values       #list of labels
    )      

## Train, validate and save a model
scores = bert.run_training(
    train_loader,             #training dataloader
    test_loader,              #test dataloader
    lr=5e-5,                  #learning rate
    n_epochs=3,               #number of epochs
    random_state=42,          #random state (for replicability)
    save_model_as='clickbait' #name of model to save as
)
# this trained model will be saved at ./models/clickbait
# the output "scores" contains precision, recall, f1-score and support for each classification category, assessed against the provided test set

# 2. Prediction on unlabeled data

## Load prediction data
pred_data = pd.read_csv('https://raw.githubusercontent.com/rubingshen/augmented_tutorial/main/clickbait/clickbait_pred.csv')

## Preprocess the prediction data
pred_loader = bert.encode(pred_data.text.values) #input a list of unlabeld texts

## Prediction with the saved trained model
pred = bert.predict_with_model(
    pred_loader, 
    model_path='./models/clickbait'
    )
# the output "pred" is a ndarray containing probabilities for each text (row) of belonging to each category (column)

## Compute the predicted label as the one with the highest probability

pred_data['pred_label'] = np.argmax(pred, axis=1)
pred_data['pred_proba'] = np.max(pred, axis=1)

Tutorial

Check here for an interactive tutorial on Google Colab.

Languages supported

Bert is a pre-trained language model for the English language. The module augmentedsocialscientist.models also contains models for other languages:

• ArabicBert for Arabic;
• Camembert or Flaubert for French;
• ChineseBert for Chinese;
• GermanBert for German;
• HindiBert for Hindi;
• ItalianBert for Italian;
• PortugueseBert for Portuguese;
• RussianBert for Russian;
• SpanishBert for Spanish;
• SwedishBert for Swedish;
• XLMRoberta which is a multi-lingual model supporting 100 languages.

To use them, just import the corresponding model and instanciate it as in the previous example.

For example, to use the French language model Camembert:

from augmentedsocialscientist.models import Camembert

bert = Camembert()  #instanciation

You can then use the functions bert.encode(), bert.run_training(), bert.predict_with_model() as in the previous example.

To use a custom model from Hugging Face

The package also allows you to use other BERT-like models from Hugging Face, by changing the argument model_name to the desired model name when instanciating the class Bert.

For example, to use the Danish BERT model DJSammy/bert-base-danish-uncased_BotXO-ai from Hugging Face:

from augmentedsocialscientist.models import Bert

bert = Bert(model_name="DJSammy/bert-base-danish-cased_BotXO-ai")

To use a custom torch.Device

By default, the package automatically detects the presence of a GPU and uses it to accelerate computation. You can also set your own device, by providing a torch.Device object to the parameter device when instanciating Bert.

from augmentedsocialscientist.models import Bert

bert = Bert(device=...)  #set your own device