PolicyBERTa-7d

This model is a fine-tuned version of roberta-base on data from the Manifesto Project. It was inspired by the model from Laurer (2020).

It achieves the following results on the evaluation set:

• Loss: 0.8549
• Accuracy: 0.7059
• F1-micro: 0.7059
• F1-macro: 0.6683
• F1-weighted: 0.7033
• Precision: 0.7059
• Recall: 0.7059

Model description

This model was trained on 115,943 manually annotated sentences to classify text into one of seven political categories: "external relations", "freedom and democracy", "political system", "economy", "welfare and quality of life", "fabric of society" and "social groups".

Intended uses & limitations

The model output reproduces the limitations of the dataset in terms of country coverage, time span, domain definitions and potential biases of the annotators - as any supervised machine learning model would. Applying the model to other types of data (other types of texts, countries etc.) will reduce performance.

from transformers import pipeline
import pandas as pd

classifier = pipeline(
    task="text-classification",
    model="niksmer/PolicyBERTa-7d")

# Load text data you want to classify
text = pd.read_csv("example.csv")["text_you_want_to_classify"].to_list()

# Inference
output = classifier(text)

# Print output
pd.DataFrame(output).head()

Training and evaluation data

PolicyBERTa-7d was trained on the English-speaking subset of the Manifesto Project Dataset (MPDS2021a). The model was trained on 115,943 sentences from 163 political manifestos in 7 English-speaking countries (Australia, Canada, Ireland, New Zealand, South Africa, United Kingdom, United States). The manifestos were published between 1992 - 2020.

Canadian manifestos between 2004 and 2008 are used as test data.

The Manifesto Project mannually annotates individual sentences from political party manifestos in 7 main political domains: 'Economy', 'External Relations', 'Fabric of Society', 'Freedom and Democracy', 'Political System', 'Welfare and Quality of Life' or 'Social Groups' - see the codebook for the exact definitions of each domain.

Tain data

Train data was higly imbalanced.

Overall count: 115,943

Validation data

The validation was created by chance.

Overall count: 20,461

Test data

The test dataset contains ten canadian manifestos between 2004 and 2008.

Overall count: 8,330

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

training_args = TrainingArguments(
    warmup_steps=0,
    weight_decay=0.1, 
    learning_rate=1e-05,
    fp16 = True,
    evaluation_strategy="epoch",
    num_train_epochs=5,
    per_device_train_batch_size=16,
    overwrite_output_dir=True,
    per_device_eval_batch_size=16,
    save_strategy="no",
    logging_dir='logs',   
    logging_strategy= 'steps',     
    logging_steps=10,
    push_to_hub=True,
    hub_strategy="end")

Training results

Validation evaluation

Test evaluation

Evaluation per category

Evaluation based on saliency theory

Saliency theory is a theory to analyse politial text data. In sum, parties tend to write about policies in which they think that they are seen as competent. Voters tend to assign advantages in policy competence in line to the assumed ideology of parties. Therefore you can analyze the share of policies parties tend to write about in their manifestos to analyze the party ideology.

The Manifesto Project presented for such an analysis the rile-index. For a quick overview, check this. But PolicyBERTa isn't fine-tuned to predict the rile-index, if you're interested in that, check ManiBERT or RoBERTa-RILE.

In the following table, the predicted and original share of the individual policy domains are shown per manifesto in the test dataset. Overall the pearson correlation between the predicted and original shares is 0.965.

Framework versions

• Transformers 4.16.2
• Pytorch 1.9.0+cu102
• Datasets 1.8.0
• Tokenizers 0.10.3