A general question about pretraining and training of dialogue systems

[helehh]

Hi!

I am rather new to open-domain dialogue systems and while reading the papers (for example Recipes for building an open-domain chatbot) about the topic I noticed that there's a detail about the training that is never explained (not only in this paper but in many papers that I checked). More precisely, it is always said that pretraining of a generative model is done by generating a comment conditioned on the full thread leading up to the comment. However, it is never said whether one thread counts as only one training example or is it used to construct multiple (N-1) training examples.

To illustrate my question, let's say we have a following thread:
A: Hello!
B: Hi!
A: How are you?
B: Good. How are you?

Would it be a single training example:

1. Context: Hello! Hi! How are you?, Label: Good. How are you?

or is it used to create multiple examples:

1. Context: Hello!, Label: Hi!
2. Context: Hello! Hi!, Label: How are you?
3. Context: Hello! Hi! How are you?, Label: Good. How are you?

In addition, is the same (either one) approach applied during fine-tuning on multi-turn dataset?

Thank you in advance!

[stephenroller]

We train to predict every non-root node in the tree.

    A
   / \
  B   C
 / \
E   F

Then we generate 4 (N-1) examples:

Context -> Label
A -> B
A -> C
A, B -> E
A, B -> F

[helehh]

Thank you for this clarification, Stephen!

[helehh]

Hi again, @stephenroller!

Am I correct that in terms of ParlAI Dialog Format each example would count as one episode, e.g. the training data for pre-training would look like:

text:A    label: B    episode_done:True
text:A    label: C    episode_done:True
text:A B    label: E    episode_done:True
text:A B    label: F    episode_done:True

[stephenroller]

Correct