[21st feb] discussion Step 1: Clean DSTC for template

    Step 2:
        Multiple tasks

    Baselines:
        1. 2017 paper - upper limit.
        2. Lower limit - simple memnn using cleaned up data.
        3. Seq2Seq models are not as good as memnn, so we don't implement them.

    -- Interesting question: LSTMs vs MemNN.

    Template induction:
        1. understand code.
        2. email dipanjan das for code. 

        Data for templates:
            try subreddit for getting moar templates.

            Multilabel NER - template induction would work and provide a single field.
            Even if we want to get templates where each slot can have only one field, 
                where do we get such data???

    Transactional conversations:
        We dont have any.

    Once we have templates, we have to do 2 things:
        JUGAAD:
            - create a model that picks a template given {u_i}
                = can use lower
            - slot filling code.

        - IDEAL: joint inference? slot filling + template picking at the same time.
            = entity extraction
            = intent extraction
            C&C paradigm...
                = microsoft at interspeech conf. ... entity intent joint modeling. (we're doing the exact opposite.)
                = still read about it. get some ideas.

todo:
    HH:
        0. read template induction... get code.
        1. hpc set up. (tomorrow) - TF + tutorials...
        2. cleanup code onto repo...
        3. look at dstc data.
    DR:
        1. start using the repo.