Assembly Minutes Summary

Purpose

For the development of our society, the National Assembly conducts numerous meetings. These meetings are recorded in detail and can be viewed through the ‘National Assembly Minutes Big Data’ site currently operated by the government.

However, in order to check the speaker’s intention or opinion on the site, you need to read the entire meeting, which takes a lot of time and effort. We developed this service to improve this inconvenience and make it easier for citizens to understand and engage with law and politics.

Process

The Bart model used in this project is a self-supervised learning model, and it is a pre-trained model. Bart has a Bidirectional Encode structure, and the original data is encoded with noise, Autoregressive Decoder learns by restoring (predicting) the original data from the noisy encoded data.

Based on this structure, we fine-tuned it for this project to train the model to summarize. In this case, the encoder contains the original text and the decoder contains the human labeled summary. As noise is added to the original text input to the encoder and restored by the decoder, the model is fine-tuned for summarization. the model is fine-tuned on the summaries.

After the fine-tuning was completed, the summarization was performed on the real-world data, the National Assembly Minutes dataset. We used rouge score, a matrix often used in summarization tasks, as a performance evaluation metric for the model, 80% accuracy was achieved.

We then built a DB using Mysql to store the data.

The project was completed with the website built using Javascript and React on the front end and Express on the back end.

Expectations

Better accessibility to citizens seeking easy review through summarized parliamentary proceedings. It can also be used for parliamentary or law-based projects.

It can be used to gain the sustained interest and participation of members of society in relation to politics, and can be used to educate younger generations, for whom comprehension skills are required.

If introduced to a national website, many people can obtain good-quality information on a single site.