I need to come up with a topic model, without any labelled dataset, the model should also be multilingual, thinking of using LLM's as they are accurate and awesome but if Im to build one on my own how to approach this problem without any data in hand. Also, the data we are expecting is mostly like customer care or support questions from different domain. Any help?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Yes, you certainly can, but understand that this would be a first step to getting your model off the ground and then later on you can refine the predictions and retrain the model to make it better. One method, that I've used to do this is
Decide whether you want to mark each 'document' or unit of text that you will be predicting with a single label or multiple labels. That matters up front.
Take each 'document' and use 'word embeddings' to convert the entire document into a vector.
From here you can treat the problem with a clustering algorithm. documents which are near each other in the n-dimensional vector space should be related to each other.
Now it's an iterative process of adjusting the number of your clusters and reading the documents that fall into each cluster to see if they make sense when grouped together.
When you're satisfied with the groups, now you can label each group as something sensible. That is now your initial labels for your model.
Now you have a labeled dataset.
Please, refine your model by changing labels on conversations that don't make sense and as time goes by the accuracy of your model will increase into a more acceptable level.
