Fine Tuning Transformers for Political Manifesto Summarization and Quantification
Open Access
- Author:
- Elisii, Patrick
- Area of Honors:
- Computer Science
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Amulya Yadav, Thesis Supervisor
John Joseph Hannan, Thesis Honors Advisor - Keywords:
- Machine Learning
Natural Language Processing
Political Documents
Summarization
Classification
BERT - Abstract:
- This thesis presents a method to summarize political manifestos in an unbiased manor. Biased reporting of political candidates is prevalent in the media, which influences viewers to adopt that bias. To overcome this issue, the proposed method extracts specific policy standards from the manifestos to provide an impartial summary. The research involves various model experiments that fine-tune pre-trained language transformers to classify each sentence of political manifestos. The dataset used in the study is comprised of hand-annotated labelled provided by the Manifesto Project. The results of the experiments show that language models can identify political stances, even when there are many categories. However, the model's performance declines on categories with a small number of examples in the dataset. This study suggests and experiments a few ways to improve accuracy on these categories. The study makes significant contributions to the field of natural language processing and political science by providing a new approach for summarizing political manifestos. This approach could be helpful to political analysts, journalists, and policymakers in summarizing complex political documents. Furthermore, the proposed method can be extended to other languages, making it more widely applicable. In summary, this study suggests that language models are powerful tools in summarizing political manifestos and can contribute significantly to promoting unbiased reporting of political candidates.