BERTopic: Is it okay to ignore the first two topics?

Question

I used BERTopic to generate a topic model over a large dataset of texts. The result is very appealing and the modeled topics are mostly perfectly interpretable for a human, especially compared to other topic modeling approaches.

According to the documentation (e.g. https://maartengr.github.io/BERTopic/getting_started/quickstart/quickstart.html) the topic with number -1 refers to outliers and should be ignored. Topic -1 is in my case indeed a topic consisting of unrelated common words without a possible interpretation.

However, in my topic model, as well the second topic with number 0 is just a mixture of unrelated words and not a "nice topic" in terms of human sense. All the following topics are very nice topics with a clear meaning.

My question: Is it okay to ignore the first two topics (-1 & 0) modeled by BERTopic and only start using the topics from topic 1 on? Or is this problematic and indicating an issue with the model? Is there a parameter that can be changed to change this behavior in order to ensure that only topic -1 will be an uninterpretable topic?

Answer 1

I received a answer for this question on github:

It might be that topic 0 groups documents together that have little meaning and as such can be considered to be a topic. From that perspective, they could indeed be considered outliers. For example, if you pass the model a couple of hundred of empty documents, it will generate a perfectly valid cluster with no labels. That cluster, could then be ignored. You could, however, merge -1 and 0 together with .merge_topics or .update_topics to make sure they are indeed -1.