> Colombo, Pierre. _Learning to Represent and Generate Text Using Information Measures_. # Learning to represent and generate text using information measures 1. How to adapt the MI maximization principle to learn transcript representations? 2. How to use information measures to generate and evaluate generated text? ## Learning a multi-purpose dialog embedding - New loss: utterance-level and dialog-level - Masked utterance modelling - Masked sequence generation with a causal decoder → learn inter-utterance dependencies - Masked language model loss ## Learning to disentangle representations - Transferring style is easier with disentangled representations but it removes information about the content