Manual Topic Modeling with Pre-Determined Hierarchical Topics

[peterforberg]

I have a set of documents that were hierarchically clustered using another embedding/vectorization strategy, and I was hoping to use BERTopic to retrieve hierarchical topic names to be able to visualize the parent-child relationships for these clusters. Is there a way to perform hierarchical manual topic modeling with a predetermined hierarchy using BERTopic?

[MaartenGr]

Hmmmm... that might be possible if you were to create a manual BERTopic model and then run topic_model.visualize_hierarchy(hierarchical_topics=hierarchical_topics). However, you would need to make sure that the format of your hierarchical topics is the same as in hierarchical_topics, so your best bet would be to follow along here.

[peterforberg]

Thank you for the reply! I am not sure what is meant by "make sure that the format of your hierarchical of your hierarchical topics is the same in hierarchical_topics." If I were to run manual topic modeling in which all of the documents are grouped by the child cluster, then I will have textual representations of just the lowest children of the hierarchy. I don't, at this step, have any information about the parent clusters, other than the knowledge that X and Y child clusters belong to the same parent, and thus that this parent is comprised of X and Y's combined documents. Recreating the dataframe hierarchical_topics absent this knowledge of the parent clusters' representations feels difficult. If I run hierarchical clustering on the manual topic model of the children, the resultant hierarchy is, for many reasons, not the same hierarchy as I have predetermined.

I tried running manual topic modeling on each "level" in the hierarchy (e.g., all the children, then all of their parents, then all of their parents, etc.), which, while not directly comparing the children to the parents, gives textual representations of each level. The question then becomes, when attempting to show what words differentiate the parents from the children, can I just look at the top words of ParentX to ChildX?

Alternatively, one thing I have been thinking about is: I could individually run manual topic modeling on the parent cluster docs and one of the children cluster docs, to see what differentiates the child from the parent. Doing this programmatically for each cluster and its parent would, I think, produce topics that distinguish children from parents.

Update: Third paragraph strategy definitely doesn't work, not really surprised! It differentiates parents and children just fine, but since the children are not being compared to one another, they all start to blend together. I tried a largely manual approach similar to the second paragraph: Group clusters into "levels" by a range of heights, create topic models for each level, then compare the most frequent words of the child clusters to the parent and substitute duplicate words for words more highly ranked in the children. The clusters look good, but the rigor of such a method...may be lacking.

[MaartenGr]

Ah right, I assumed you already had representations. Considering that you don't, I would indeed run manual topic modeling on each level of your hierarchy (assuming you can create levels that way).

Update: Third paragraph strategy definitely doesn't work, not really surprised! It differentiates parents and children just fine, but since the children are not being compared to one another, they all start to blend together. I tried a largely manual approach similar to the second paragraph: Group clusters into "levels" by a range of heights, create topic models for each level, then compare the most frequent words of the child clusters to the parent and substitute duplicate words for words more highly ranked in the children. The clusters look good, but the rigor of such a method...may be lacking.

Glad to hear that the second paragraph looks good to you. I do think that this is the best way to approach it considering you still have to calculate the topic representations. Another trick that you might be able to do is to flatten your hierarchy, duplicate documents, and then calculate the topic representations. I wouldn't be surprised if that actually would work.

[peterforberg]

Thank you again! I had to put down this project for a moment, but I'll try flattening the hierarchy and update this post to see how it changes the results, for the sake of any future readers.