The problem you are describing is not a classic NLP problem.
There is a similar classic NLP problem called "topic modelling", which consists of discovering topics in a collection of text documents. Topics are defined by a list of words relevant to the topic itself. The most paradigmatic approach to this problem may be Latent Dirichlet Allocation (LDA). It is an unsupervised learning approach.
Your problem, nevertheless, has somewhat also been approached from a machine learning perspective, at least partially. I can refer you to the article Unsupervised Topic Segmentation of Meetings with BERT Embeddings by Meta. This is its abstract:
Topic segmentation of meetings is the task of dividing multi-person meeting transcripts into topic blocks. Supervised approaches to the problem have proven intractable due to the difficulties in collecting and accurately annotating large datasets. In this paper we show how previous unsupervised topic segmentation methods can be improved using pre-trained neural architectures. We introduce an unsupervised approach based on BERT embeddings that achieves a 15.5% reduction in error rate over existing unsupervised approaches applied to two popular datasets for meeting transcripts.
The authors released their source code at github.
To understand its contents, you will need to have some background on BERT, an NLP neural network based on the Transformer architecture's encoder part. On https://datascience.stackexchange.com/ you can find plenty of specific questions and answers about it (and you can ask more if you don't find your specific doubts).