Repository navigation
Replies: 1 comment
|
With the default HDBSCAN setup, the reviewer is right: |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi all,
I'm using BERTopic for a bibliometric study and have a core question about interpreting the probs output.
I ran: topics, probs = topic_model.fit_transform(docs, embeddings)
In my paper, my method is to calculate a topic's trend (e.g., for "Topic A") by summing the probabilities of all documents for that specific "Topic A." I justified this by treating probs as a "soft-clustering" distribution (like LDA), allowing one document to contribute to multiple topics.
A reviewer challenged this, arguing that:
The probs value returned by fit_transform is not a distribution across all topics.
It is merely the membership confidence score for the single topic assigned in the parallel topics array.
Therefore, my method is just a "weighted count" (summing confidences only for documents assigned to "Topic A"), not a true soft assignment (summing all documents' contributions to "Topic A").
My Question: Is the reviewer correct? Is the probs value from fit_transform strictly the confidence for the single assigned topic?
If so, is there another way (perhaps calculate_probabilities=True?) to get a full probability matrix, allowing me to validly sum all documents' probability scores for a single, specific topic?
Thanks for your help!
All reactions