Lecture 11: Mastering TxMM

The first half of this lecture is a short recap of some of the less watched videos. The second half is a guess talk about a master student using text mining in her master thesis project.

Recap Topic Modelling

topic modelling != Clustering

Assume generative model with latent topics.

Topic is distribution over whole vocabulary; but for each topic the top-k words will be different.

How are topics chosen?

Unsupervised: Words are grouped together based on how they co-occur together; you count words, shuffle them around, and maximize probabilities.

Have a look at this for the IR project.

Master Project Myrthe Reuver

Classifying smoking status of patients from free text.

There are some tactics for dealing with little training data:

  1. A rule based regex approach
  2. increasing training data by using weak supervision methods e.g. using SNORKEL
  3. Transfer learning, e.g. using BERT

See the end of the slides; there were some very nice tips for doing a thesis.