Query: Inference on new data

[dwlmt]

I came across your paper, and it seems like a promising alternative to BERTopic. Running your code on a small custom dataset with a custom sentence encoder produced good topics. One thing is, I couldn't see a way of inferring topics on new documents in the code without rebuilding all the topics. I assume for zero-shot classification, I could just find the nearest neighbours against vec_t for each sentence and resolve it to the topic word list? But not sure if there is a better way of approaching this.

[JohnTailor]

Thanks for the comment! Sorry for replying late. You are right, the library so far does not support (incremental) inference. Your approach seems reasonable.

I might add that feature in the future. Currently, the paper is under review at a conference and I will work further on it, once the paper gets accepted.