Dialogue PL, part 1 #805
Draft
Dialogue PL, part 1 #805
+8,865
−5,824
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
With most of the interesting logic stubbed out
Or whether it should throw an exception. This includes all the logic to dispatch to multiple dialogues in parallel.
Extract the dialogue generation code into a new module, and wire up the interfaces used by DialogueInterface in a new PartialDialogue class. The IO of DialogueInterface is rewritten in terms of templates, so we can integrate with the existing synthesis. These templates are added dynamically to the user/agent generator, which are then expanded normally.
Nothing actually works yet
Separate the "dialogue runtime" which tracks conversation histories/APIs/logs and is independent of ThingTalk/dialogue PL, from the "thingtalk dialogues" (dialogue PL)
Move the responsabilities of the old DialoguePolicy class to InferenceSentenceGenerator, which initializes the SentenceGenerator on-demand based on the skills currently in the context. Introduce a InferenceTimeDialogue, which is a DialogueHandler for ThingTalk dialogues but also bridges to DialogueInterface (dialogue PL). Refactor non-standard dialogue code (interactive-annotate, simulate-dialogs) to also use InferenceTimeDialogue as a dialogue handler. Notifications are still WIP.
Refactor the executor classes so they interact with the user through the DialogueInterface instead of directly through DialogueLoop. Special questions asked prior to execution are now represented as regular states in the state machine, with dedicated dialogue acts. Notifications are still messy.
With this commit, we have "hello" and "cancel" commands working, and the agent can start and stop the dialogue as needed.
Templates written in Genie language will remain as "building blocks", but the dialogue logic will be written in imperative js code. In preparation for that, separate lib/templates containing the single-sentence templates, and lib/transaction-dialogues containing the logic of the transaction policy.
The module has grown to contain all sorts of unrelated things. Split into a module containing the ContextInfo class, a module containing the code to create context phrases, and a module with actual state manipulation. Later, the state manipulation will stop using the context info and will move to utils/thingtalk
Except for a few helpers that I need to get rid of. As part of this change, state manipulation code is now updated to return target states instead of full states, because that makes more sense.
And stop "tagging" the context to know what the agent should do Also fix entity handling so replies have the correct content.
gcampax
added
enhancement
State tracking in the dialogue interface can get messed up if the calls are nested.
To do that, we need to track what the agent actually said in the last turn, after all the templates are expanded and the eithers are resolved.
This commit contains all the small fixes needed to get the agent tests to pass, and pass without too many changes.
gcampax
force-pushed
the
wip/dialogue-pl
branch
from
90a993d to
c1112b6
Compare
yeah, let's do this in person. maybe next week? still have a bit of coughing, probably a bad idea to meet this week |
Map notifications to a command with a special dialogue act, and then dispatch as usual.
It doesn't yet do anything useful, but at least it goes through
We need to include the last turn of the user when the agent terminates the dialogue, otherwise we don't include any "cancel" turn (which is all we generate at the moment)
So we can see that we generate some real dialogues
Which might cause problems
And in the process catch various places where we use context tags that were removed.
Concatenate them together into a single dynamic template
gcampax
force-pushed
the
wip/dialogue-pl
branch
from
6e23764 to
2964d1b
Compare
Should be fixed in ThingTalk but it is not.
gcampax
force-pushed
the
wip/dialogue-pl
branch
from
187c449 to
eaa14ec
Compare
|
Alright, synthesis is not done yet but at least tests pass now. @rayslxu how do you want to review this? Do we want to merge this first PR somewhere? |
Ugh... this is a lot to read. I guess I will just skim through it to get a sense of what's going on... (unless you will split the PR into smaller ones) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
All the way to running the agent and synthesizing dialogues, without regressions, and hooking up the existing transaction model, but while keeping most of the logic in the form of templates.