Model name problem

[rzhao1]

Could you please clarify the meaning of each model in the code of paper "Decoupling Strategy and Generation in Negotiation Dialogues"?

1. "rulebased"
2. "hybrid"
3. "cmd"
4. "fb-neural"
5. "pt-neural"

Thank you very much!

[hhexiy]

rulebased: rule parser + rule policy + retrieval generator
hybrid: rule parser + learned policy + retrieval generator
cmd: human-controlled bot used in a command line interface
fb-neural: obsolete
pt-neural: seq2seq

[rzhao1]

Thank you for the clarification! It is really helpful! Basically, I want to run use hybrid model to generate the simulated dialogue. However, when I run the following command, I got the error "AttributeError: 'HybridSystem' object has no attribute 'env'".

PYTHONPATH=. python reinforce.py --schema-path data/bookhatball-schema.json
--scenarios-path data/train-scenarios.json
--valid-scenarios-path data/val-scenarios.json
--agent-checkpoints checkpoint/lf2lf/model_best.pt checkpoint/lf2lf/model_best.pt
--model-path checkpoint/lf2lf-margin
--optim adagrad --learning-rate 0.001
--agents hybrid pt-neural
--report-every 500 --max-turns 20 --num-dialogues 5000
--sample --temperature 0.5 --max-length 20 --reward margin
-- templates templates.pkl
--policy model.pkl

[rzhao1]

I noticed that HybridSystem class does not have function of loading trained policy?

[hhexiy]

The parameters are loaded through the manager object of HybridSystem, which is the learned policy. Also, I don't think the command would run, because you cannot back-propagate through hybrid.

[rzhao1]

Thank you! If I want to create an agent that use neural dialogue model as manager and rule-based template as generator (a hybrid system) to talk with human user, which kind of command I should use? The default commands you provided in README only output dialogue act instead of the utterance?

[rzhao1]

One more extra question: what does it mean lf2lf vs lflm? Thank you!

[hhexiy]

You can use https://github.com/stanfordnlp/cocoa/blob/master/scripts/generate_dataset.py to generate bot-bot/human chat by setting one agent to be hybrid and the other to be cmd (human). lflm means learning a LM as the action predictor instead of a seq2seq model, this is obsolete though.

[rzhao1]

Thank you! What does "LF" stand for?

[hhexiy]

logical form