Bertopic semi-supervised text classifier

[matrix00100]

Hi everyone,

I have table of text data from customer. I would like to use Bertopic to figure out the topics from the text data and create labels and then train the Bertopic model with these labels. After that, I want to use the model to classify new customer messages and count the message topic label. Is it possible with Bertopic? If you have any other suggestions how to tackle this, I really appreciate this.

[MaartenGr]

After that, I want to use the model to classify new customer messages and count the message topic label. Is it possible with Bertopic?

I believe your use case should be possible with BERTopic but I do get the feeling you might have approached it a bit more complex than necessary. If you want BERTopic to figure out topics and create labels, then you can simply use BERTopic.fit(). Then, when you want to use that model to classify new customer messages, you can use BERTopic.transform(). From your description, I do not think it is necessary to first use BERTopic.fit() to figure out the topics and create labels, and then use those labels for another pass of BERTopic.fit(). The end result will likely be the same.

[matrix00100]

I tried .fit(customer_messages, y=customized_labels). It gave a error "Length of x = 400000, length of y = 33, while it must be equal".

[drob-xx]

So I'm not sure exactly what is going on here, but my guess is that you have a classification rubric that has 33 different classes and 400K documents that you want to place those into? I'm not that familiar with the way BERTopic does guided classification - but you would need to label each of the 400K documents with one your 33 classifications for this to work - in other words the dimension of these two values has to match in n.

It might be helpful to clearly describe your data and what your business case is to give more context.

[matrix00100]

Sorry due to the privacy, I can't post the data. However, the data set looks similar to the IMDB dataset except no labels(positive or negative).

I want to use label these customer messages with customized labels as weak labels supervised learning then do transfer learning later with labeled messages when we have the labeled data.

[drob-xx]

If you are trying to get binary positive/negative labels for the reviews then as far as I know BERTopic isn't the right tool (but @MaartenGr can be more definitive).

If this is what your goal is the reason that BERTopic isn't a good choice is that topic models provide word vocabularies affiliated with an underlying topic representation of text. These won't tell you anything about the kind of speech (positive, negative) used, just a summary of what speech is used. There are a ton of articles about sentiment analysis (determining the positive/negative valence of a collection of words) - here's one article https://medium.com/@pyashpq56/sentiment-analysis-on-imdb-movie-review-d004f3e470bd that seems decent enough - but it is easy to find more.

If you aren't looking to do sentiment analysis then please let me know what I got wrong and why you think BERTopic is a good choice for your project.

[matrix00100]

Oh no I am not looking for sentiment analysis. I use it as an example for a text column of customer data.

I want to create a model to classify these messages with different labels such a ["sale", "product questions, "return questions",]. The list of labels can be like 30 labels.

After that, I want to apply this model toward new customer messages to classify them and count the number of topics for business intelligence.

Hope I make it clearer. I wish I could share the data but that is against company policy.

[drob-xx]

No problem. So when you got the error above:

I tried .fit(customer_messages, y=customized_labels). It gave a error "Length of x = 400000, length of y = 33, while it must be equal".

What was the content of customized_labels, if you want BERTopic to consider your labels when it builds the topics then you would need to supply a label for each document.