dracoblue.net

Some days ago I was participating in a bluesky thread and a Bluesky post by Yves Venedey reminded me of a project idea I wanted to give a try: Fact Checking on bsky! You can actually extend the atproto ecosystem with a labeler, which makes these labels available to everyone who wants to see them.

Fast forward: Today I publish atproto-fact-labeler and have a live instance at @facts.kiesel.app as a BETA for you to give a try!

It's a matter of pressing "subscribe":

Similar to TopicWatchdog (which tried to extract claims from public youtube videos) I do this in my spare time and for my personal learning journey using my machine learning and federated social networks exploration.

Currently, I am running Rancher Desktop for Docker CLI needs locally. But I've found a viable solution I wanted to give a try.

It's called container and it is released by Apple at https://github.com/apple/container for Apple Silicon Mac devices.

In Week 1 (extraction), Week 2 (embeddings + KMeans), and Week 3 (stable topics with BERTopic) I built the foundations. This week applies the same idea to claims — using BERTopic to cluster claim snippets and keep stable claim_ids via a registry + dim table.

This week we explore BERTopic + stable claim IDs:

• Use pre-computed embeddings from BigQuery (same pipeline as before).
• Fit/Load a BERTopic model (UMAP + HDBSCAN) in Python.
• Assign internal cluster IDs per batch, then map them to stable claim_ids.
• Persist to video_claims, claim_registry, and dim_claims tables for analysis.
• Inspect behavior in Looker Studio and reflect on limitations.

In Week 1 (extraction) and Week 2 (embeddings + KMeans in BigQuery ML) we laid the groundwork. This week I built a Python BERTopic stage whose IDs stay stable across runs by mapping BERTopic’s internal clusters to stable topic IDs in BigQuery. I use Google Gemini again to generate nice labels for the extracted topic clusters.

This week we explore BERTopic + stable topic IDs (via an ID registry):

• Train a BERTopic model in Python (UMAP + HDBSCAN).
• Map BERTopic’s internal clusters (model_version, internal_topic_id)
• Ensure topic IDs remain consistent across retraining (no more ID jumps).
• Join human-readable labels and persist results into video_topics for analysis.
• Inspect results in Looker Studio and reflect on limitations.

This post documents Week 2 of the TopicWatchdog project.
Last week we successfully extracted topics and claims from German political short videos and persisted them in BigQuery.
However, topics often appeared under slightly different names — making aggregation unreliable.

This week we explore embeddings + clustering:

• Generate embeddings of canonical topics and claims with BigQuery ML.
• Train a KMeans model on those embeddings to group semantically similar entries.
• Assign clusters back to each topic/claim.
• Inspect first results in Looker Studio and reflect on limitations.