forked from KEMT/zpwiki
52 lines
2.0 KiB
Markdown
52 lines
2.0 KiB
Markdown
---
|
|
title: Dries Huybens
|
|
published: true
|
|
taxonomy:
|
|
category: [erasmus]
|
|
tag: [nlp, ie, rag, medical]
|
|
author: Daniel Hladek
|
|
---
|
|
|
|
ERASMUS Intern Summer 2026, 20 July - 30 August
|
|
|
|
Topic:
|
|
|
|
Multilingual Knowledge Graphs from medical data
|
|
|
|
Goal:
|
|
|
|
- Construct a knowledge graph from medical package inserts in multiple languages
|
|
- Utilize the graph in an intelligent agent that recommends medication.
|
|
|
|
Tasks:
|
|
|
|
- Continue project of [Bogdan Paul Chis](/interns/bogdan_paul_chis)
|
|
- Study repositories:
|
|
- https://github.com/chis-facultate/erasmus-kosice
|
|
- https://github.com/hladek/mul-me-kg
|
|
- https://github.com/hkuds/lightrag
|
|
|
|
- Learn intelligent agents and generative models - OpenAI API, Agent frameworks, RAG systems.
|
|
- Learn about knowledge graphs and GraphRAG. Read several research papers.
|
|
- Prepare a Python based workflow, use git code repository
|
|
- Visualize the graph
|
|
- Prepare an agent that utilizes the unstructured data and graph-data.
|
|
- Evaluate the agent using DeepEval or RAGAS.
|
|
- Write a report
|
|
- Put all code to GIT
|
|
|
|
|
|
|
|
Project tasks update:
|
|
|
|
- Prepare a multilingual parallel corpus from OPUS data
|
|
- Prepare dataset for visual language model evaluation from foto
|
|
- Prepare dataset for visual language model evaluation from wikipedia. You can use https://huggingface.co/datasets/wikimedia/wit_base .
|
|
- Prepare corpus of text and image data from pravda.sk . For each subdomain, create - raw HTML data with images. Then create corpus of images plus descriptions, subtitles and tags. Then create corpus of extracted text. You can use docling , trafilatura for text extration. Or design your own parser. Give source codes to a git repository with documentation. Put data files to school server. Do not download too fast, so our IP does not reveive a ban.
|
|
|
|
Outupus:
|
|
|
|
- multilingual corpus was analyzed with embedding models. Semantic overlap between languages is low, so multilngual parallel corpus will be small.
|
|
- A [corpus](https://huggingface.co/datasets/driesaster/fotkyzadarmo_slovak) from FotkyZadarmo.sk
|
|
|