Parlai: Grounding knowledge conversation with blender model

[Saraharas]

Hi,

I would like to use the blender model for knowledge-grounded conversation.

The dataset I use has grounding knowledge facts for each dataset entry. So I created the task following the tutorial and added facts as a bot persona in the InteractiveWorld class.

When I run interactive.py script and ask a question related to the current context it seems like the model does not use context information to generate the answer (see the example below).

[context]: timothy theodore duncan Enter Your Message: do you love basketball ?
[TransformerGenerator]: i do , i love watching it . it ' s one of my favorite sports . what about you ?
Enter Your Message: me too
[TransformerGenerator]: what ' s your favorite team ? i ' m a big fan of the lakers . they ' re my favorite team .
Enter Your Message: do you know tim duncan ?
[TransformerGenerator]: no , i don ' t know much about him . are you a fan of any other players ?
Enter Your Message: no, i would like to know about tim duncan
[TransformerGenerator]: he ' s a great player . he was one of the best centers in the history of basketball .
Enter Your Message: is he nba champion?
[TransformerGenerator]: no , he is not . he is considered to be the greatest center of all time , though .

Could you please help me what script to change to incorporate grounding knowledge correctly?

Thanks in advance

[stephenroller]

BlenderBot isn't trained to incorporate knowledge in context. It was just fine tuned on a dataset that has a lot of factoids, which brings out some knowledge behavior.

If you want to add this, you'll need to do some fine tuning.

[Saraharas]

Thank you @stephenroller for the quick reply.
Could you recommend the best way to add knowledge to the data for fine tuning BlenderBot?

Now my task dictionary looks like this

{"text": 'i like apples', "knowledge": 'an apple is an edible fruit produced by an apple tree',"labels": 'me too'}

Should I concatenate "knowledge" and "text"?

[stephenroller]

Yes, concat is the easiest way.

You can look at our wizard of Wikipedia models too, but it gets hairy fast.

[Saraharas]

@stephenroller Thanks for your answer. I have a question about interactive chat with fine-tuned model: how can I pass the episode grounding knowledge into the the model in interactive mode?

I tried adding facts as the context of conversation using InteractiveWorld in the task. But as far as I have seen during test dialogues information from context is not used in model answers.
Could you please explain how the context is used to generate BlenderBot responses?

[stephenroller]

Easy hack is just to type it into the input with "n", which gets string replaced to a newline.

Look for get_contexts in the interactive worlds for hints on doing it more programmatically

[Saraharas]

Ok, one quick question: am I correct that the first element of get_contexts output goes to the human_agent which just prints it, and the second element is fed to the BlenderBot generator model?

[stephenroller]

Yes, I believe that's correct.