Use language model (LSTM) to transform an utterance (list of tokens) to an utterance representation
Idea is to combine these utterances with an 'utterance level' rational extractor --> either the whole utterance is kept, or masked.
Then the complete context (with the words in the masked utterances masked) is used to generate next utterance.
Use language model (LSTM) to transform an utterance (list of tokens) to an utterance representation
Idea is to combine these utterances with an 'utterance level' rational extractor --> either the whole utterance is kept, or masked. Then the complete context (with the words in the masked utterances masked) is used to generate next utterance.