Deep Learning Based Text Mining for the Analysis of Sales Contracts in Hungarian (BSc thesis)
- Period
- 2021 – Jan 2022
- Organisation
Budapest University of Technology and Economics- Role
- BSc thesis, supervised by Dr. Gyula Max
- Location
- BME Faculty of Electrical Engineering and Informatics (buildings Q and I)Magyar tudósok körútja 2, 1117 Budapest, Hungary
- Posted
A client–server application that reads Hungarian-language sales contracts and extracts their key entities with a transformer-based named-entity recognition model. C# server and Windows Forms client, SQL Server storage, and a Python NLP component.
- Transformer-based NER for Hungarian legal text
- C#/.NET server and client with user accounts and SQL Server storage
Problem
Sales contracts ("adásvételi szerződés") follow a familiar structure, but each one is written differently. Extracting the key information from them by hand is slow and error-prone, and most NLP tooling targets English rather than Hungarian.
What I built
A two-part application:
- Server (C#/.NET, MS SQL Server). Handles user registration and login, stores contracts and extracted data, and calls into Python to run the NER model.
- Client (Windows Forms). The user interface for loading contracts and reviewing the extracted entities.
- NLP component (Python, Hugging Face
transformers). A transformer-based named-entity recognition model applied to Hungarian contract text.
What it taught me
It was my first project where a machine-learning model was only one part of a working system. Most of the effort went into plumbing: process boundaries between .NET and Python, configuration, and a database schema that could hold uncertain model output next to user-confirmed data.
References
- D. Frey. 2022. Deep learning based text mining for the analysis of sales contracts in Hungarian. BSc thesis, Budapest University of Technology and Economics. diplomaterv.vik.bme.hu/en/theses/deep-learning-alapu-text-mining-magyar-nyelvu
- Hugging Face Transformers library. github.com/huggingface/transformers