EA Stakeholder Mapping - Interactive Poster

📋 About This Project

This analysis maps the expertise, interests, and geographic connections of attendees to the EAGx Amsterdam conference using a combination of NLP and LLM-based entity extraction.

The goal: to both visualise the main focus of the conference attendees as a representation of the wider EA community; and to create a recommendation system to help attendees find collaborators, mentors, and unexpected connections based on semantic similarity of their profiles.

The latter functionality is available in the associated GitHub repository, while this poster focuses on the data analysis and visualisation.

If you have any questions, feedback, or suggestions, please feel free to get in touch via the contact details in the footer. I'll also be dialling in some time this afternoon, so feel free to pay me a visit.

⚙️ Analysis Pipeline

📊

Data Ingestion

Load & clean attendee profiles

↓

🔍

Entity Extraction

NLP + LLM with majority voting

↓

🧠

Semantic Clustering

Embed & cluster cause areas

↓

📈

Visualization

Interactive charts & network visualization

↓

Recommendations

Find connections & collaborators

📊 Top Cause Areas

Most frequently mentioned topics across all attendee profile, extracted from a combination of NLP and LLM. The profile represents attendees referring to their own areas of experience as well as their interests.

🕸️ Semantic Network

Cause areas clustered by semantic similarity. Nodes are coloured by EA cause area and sized proportional to how often they are mentioned. Take a look at what's grouped close to your favourite cause area: if you haven't explored it, why not?

🗺️ Geographic Distribution

Countries and organizations mentioned by attendees. Note the strong geographical bias due to location of conference. Seems there's still a footprint of EA having had a lot of adoption in the US! Or perhaps it's commenting on the current state of affairs over there...

🧑‍🔬 Expertise vs Interests

What people know vs. what they want to learn, extracted directly from the pre-set keywords. Lots of people interested in academia (and apparently enjoy it!) but AI strategy is getting popular...

💡 Undervalued Areas

High interest but low expertise — opportunity gaps! Note that just because there is a gap, it doesn't mean it's a great thing to look into. The general interest level is indicated by the number of mentions detailed to the right of each bar.

🤖 NLP vs LLM Extraction

Comparing extraction methods. Pretty shocking overlap here! The NLP over-estimates cause areas since it pretty much picks out every proper noun it can find. I'm more unsure as to why you get opposite profiles between the locations and organisations bars. Any ideas?

🛠️ Methodology

Two complementary extraction approaches were used and compared:

🔠 NLP (spaCy)

Fast, deterministic entity recognition using named entity recognition and keyword matching.

🧠 LLM (Llama 3.2)

Local Ollama model with 3-run majority voting to handle stochasticity and filter hallucinations.

Semantic clustering uses sentence-transformers embeddings with predefined EA cause categories for interpretable groupings.

🔑 Key Findings

These results attain to a single conference and were analysed by a stochastic parrot. They may not be entirely accurate and are likely to strongly reflect the nature of this specific conference rather than the wider EA community. Please take them with a beaker of salt.

1 Global health is tailed by AI safety — once people stop talking about themselves (the organisational group), they are first concerned with global health, then AI safety. Note that AI is split between several areas, so as a wider area it's by far the largest. Surprise surprise.

2 Europe-centric network, US-focused associations? — strong clustering in the UK, Netherlands, and France, as is only to be expected given the location of the conference. Little representation from the Global South.

3 Expertise gaps identified — areas like policy implementation and community building show high interest but lower expertise availability. And let's not even get started on 'Space Governance'... Interestingly, there appears to be both high interest and expertise in 'Academic Research'.

4 LLM extraction outperforms NLP for cause area detection, though NLP remains faster for location/organization extraction. I haven't formally quantified the accuracy of the two (only the overlap) but NLP generally extracts very generic and single words rather than reading between the lines to place them within wider cause areas.

🔭 Next steps

Plenty more to do! I don't have the time, but if anyone is interested, get in touch or just fork the repo – see the footer.

4 Formally verify the accuracy of LLM-extracted keywords by manually checking against original data (e.g. for a sample of paragraphs).

5 Analysing the databases from past/future conferences to look for (statistically-significant) shifts in interests, expertise, and general focus areas

6 Ask more specific questions: to what extent is EA conference attendance dominated by people interested in AI? How is this changing with time? Are people predominantly interested in near-term or long-term causes?

🎯 Mapping the EA Community