iai-group / moviebot Goto Github PK

Abstract base class with: utterance string, timestamp, and agent/user.
Child classes for AgentUtterance and UserUtterance.

Update dialog history to use this class.

Remove "author" line from all files, as it is not common practice.

Reimplement terminal-based version using Questionary.

Rename BOT to TELEGRAM in config and classes.
Make terminal version the default.

Add init.py to all modules and submodules. For sphinx to treat them as modules instead of namespaces.

Create San class that can be several tokens long. Token is a single word and should be a subclass of Span.
Use Span for semantic annotation.

Extend the base Controller class with necessary method signatures for wanted functionality across all controllers.
Implement a restart method for all controllers.

System revealment:

• User is aware of the capabilities of the system (in case of unsuccessful recommendation, the user understands what went wrong/what information was missing)
• Revealing system capabilities (limitations, what kind of answer is expected)

Proposed solution:

• Create a list of actions that may be desired but are not supported in the system yet. The actions should be recognized in the user intent and handled by the system. The system should inform the user that the given action is recognized, but the system cannot perform it.

Proposed architecture:
Each action should have one of the following labels assigned: {Capability, Limitation, Unknown}. The Capability is the regular action that is well-defined and can be performed by the system in response to user intent. The Limitation is the action that is recognized by the system but is not supported. It requires an explanation provided by the system (system revealment) that includes the type of the recognized action and information on why it cannot be performed. The Unknown action is not recognized by the system. The system should inform the user that the action was not recognized and the intent should be reformulated.

Use sphinx-autoapi to make automatc API documentation (without having to generate .rst files offline).

Add SemanticAnnotation as a parameter to ItemContraint.

This can be useful for evaluation, but also for easier finding of overlapping annotations.

It's Figure 1 in the paper

• Removing preferences doesn't change the current IN (sometimes it doesn't work when using the provided options)
• Using the provided option I want to restart for a new movie was recognized as intent inquire(actors)

The problem

The current implementation does not allow for MovieBot to only run as a RESTful server.
This would make it possible and easier for other services to talk to the bot.

Proposed solution

Have a REST API that uses the response as the bot response utterance. Without any 3rd party integration.

Some particularly problematic parts:

• next_action in dialogue_policy.py
• generate_dact in nlu.py
• update_state_user in dialogue_state_tracker.py

For more robust dialogue modeling, the following solution was proposed. It considers not only the most likely user intent, but a ranked list of predicted user intents.

user_intents = predict_user_intent(user_utterance, conv_history)

for user_intent in user_intents:
  Check if user_intent can be executed	
    Do something
    Break
  Else
    Sorry, I don’t understand

Need to make neccessary modifications to use UserUtterance class instead of raw input string within the application. This will make it easier to keep history and have the preprocessed text (tokens) always available.

Some CI tests are failing because NLTK downloads are not done.

Add a step like that in CI:

- name: Download NLTK dependencies
        run: |
          python -m nltk.downloader punkt wordnet stopwords omw-1.4

Keep all annotation-related classes under nlu.annotators

Currently the recommender model is not a separate module. It is more implicit and based on slots. Factoring out recommender model so that it can be improved in the future is useful.

• Design a RecommenderModel base abstractclass
• Move the current recommender logic under SlotBasedRecommenderModel which is a subclass of RecommenderModel

Make an evaluation folder where diferent annotators can be evaluated.

Hierarchy in the docummentation should be expanded.

It'd be enough to state in the README that we follow the IAI Guidelines

IAI MovieBot doesn’t need to be typeset in italics in the documentation files.

Docstrings should be in google style for better readability.

It would be good to have SlotAnnotator class as abstract so we can inherit different classes from it, i.e rule-based annotation, neural models.

It would be good to have a Contributing.md file so people would know what to consider when making a pull request or a contrinution.

Expand CONTRIBUTING.md with explanation of pull request workflow.

Set up a GitHub page to chat with MovieBot and retire Telegram server.

Add pre-commit hooks to enforce when any file is updated it also has to follow IAI style and formatting guidelines.

TypeError: must be str, not NoneType when user input contains some punctuation marks, like -

In moviebot/nlu/user_intents_checker.py", line 592, in get_annotation_relevance

Move intents and utterance folders to core/shared folder.

Store utterance history in dialogue context.

NLU module is being actively worked on so this is a priority for expanding test coverage.

It would be nice to get >90% coverage for the module.

Below are some important information to add to the documentation:

• State minimum python version
• Add quickstart instructions for the different options (i.e., Telegram, console) to run the demo
• Add user and agent dialogue acts
• Add recommendation dialogue flow