Workshop Schedule
Held September 30, 2026 · Tilburg University
The workshop ran from 10:00 to 16:00, following the same day rhythm as the main conference, with a break in each half of the day. The programme as it ran:
| Time | Session | What happens |
|---|---|---|
| 09:45-10:00 | Arrival and coffee | |
| 10:00-10:15 | Welcome and setup | Introductions, environment check, plan for the day |
| 10:15-11:00 | Bulk paper retrieval | Talk: finding and downloading papers at scale |
| 11:00-11:15 | Coffee break | |
| 11:15-12:00 | Bulk paper retrieval: hands-on | Build a corpus from a search query of your own |
| 12:00-13:00 | Lunch | |
| 13:00-14:15 | From PDFs to causal claims | Cleaning paper text into sentences, then extracting cause and effect |
| 14:15-14:30 | Coffee break | |
| 14:30-15:45 | Harmonizing constructs and building the map | Grouping constructs that mean the same thing, then assembling the causal map |
| 15:45-16:00 | Wrap-up and questions | Open questions and how to apply this to your own research |
What We Cover
Morning: building a corpus
Erik Tjong Kim Sang, Netherlands eScience Center
A talk followed by a hands-on session on getting the papers in the first place. How bulk retrieval works, what the eScience Center download package does, and how to get from a search query to a folder of papers and metadata on your own machine.
By the end of the morning you have your own corpus on a topic you choose.
Afternoon: from papers to a causal map
Rasoul Norouzi, Tilburg University
The rest of the pipeline, in four steps:
- PDF to sentences. Extracting text from PDFs and cleaning it: dropping running heads and page furniture, joining broken paragraphs, splitting into sentences, stripping inline citations and statistics. We also look at what got thrown away and why, because that is where most mistakes hide.
- Sentences to causal claims. Running SocioCausaNet over the sentences to find the ones that state a cause and an effect, and to mark the two spans. Every relation keeps the sentence, page and document it came from.
- Harmonizing constructs. The same construct is worded differently in every paper. Three ways to deal with that: map the spans onto a published taxonomy such as ELSST or the APA thesaurus, group them by similarity, or cluster them and name the clusters. We compare what each route gives you and where each one breaks.
- The causal map. Assembling the harmonized pairs into a graph, looking at which constructs connect the most, following causal chains, and tracing any edge back to the sentence it came from.
By the end of the day you have a causal map built from the papers you downloaded in the morning, and a pipeline you can run again on your own literature.
The slides and the notebook cover both halves, so you can follow the same steps on your own.
Nothing in the pipeline is generated or paraphrased. Every relation in the final map points back to the exact sentence, page and paper it came from.
Working Through It on Your Own
You can work on your own machine or in Google Colab. Either is fine.
On your own machine
- Python 3.11 or newer.
- A working notebook environment: either Jupyter, or VS Code with the Jupyter extension (setup guide).
- Open a notebook and run a cell first, so you know it works.
- Then install the two packages and download the notebook: the Materials page has the commands.
In Google Colab
- Nothing to install. You need a Google account (create one).
- Open the notebook from the Materials page and save your own copy.
- Create a free account at huggingface.co so you can load models from the Hub.
- Skim the model card for rasoultilburg/SocioCausaNet before you start.