About the BB2 paper

[tungnt55]

I'm reading the paper "Beyond Goldfish Memory∗: Long-Term Open-Domain Conversation", I have two questions regarding the model and the evaluation baseline.

1. Let's say we don't use internet search. If there is no memory existing at the current dialog turn i.e. at the beginning of the conversation or we set knowledge_access_method to None, then how does the model generate the response?
2. If I understand correctly, the baseline "Dialogue History" in Table 4 is a standard encoder-decoder that generate response from the whole dialog history. What's the correct way to replicate this model with parlai?

[mojtaba-komeili]

1. In that case we have a simple encoder-decoder model. Very similar to what we had in BB1. The context is going to be the conversation history only.
2. Don't know the answer to this one. But @jxmsML should know.

[tungnt55]

Thank you for the reply. In that case, can we get the baseline by setting knowledge_access_method to None?
Or I need to train a baseline model by setting knowledge_access_method to None?

[mojtaba-komeili]

You don't need to train a new model.

[tungnt55]

ParlAI/parlai/agents/rag/modules.py

Line 190 in 4bc8f17

# Retrieve, get expanded input

I actually found the function that does the encoding of context + retrieved memory here.
Do you recommend using input instead of expanded_input for encoding to replicate the encoder-decoder baseline or it's better to set knowledge_access_method to None?