Dialog system (INFOMAIR) Group - G03

Dialog system created for the Methods in AI Research (INFOMAIR) course at the Utrecht University.

Teammembers:

• Nathalie Kirch
• Iris Folpmers
• Panagiotis Dallas
• Stan Meyberg

Using the system

A [env.yml] file has been added to this repository in order to easily run the system.

Install locally (using anaconda)

1. conda env create -f environment.yml
2. conda activate infomair_env
3. python -m spacy download en_core_web_sm

Run the dialog system

The following command should be run to interact and use the system:

python main.py

Part 1a: Text classification

To classify the sentences into dialog acts, four different classifiers were developed (2 baselines and 2 based on machine learning).

• Baseline systems
  • The majority class baseline
  • The rule-based system based on keyword matching
• Machine learning systems
  • A Naive Bayes approach
  • A K-nearest neighbours (KNN) approach

In order to run the analysis over all the developed models, the following command can be run:

python ./intent_classification.py

This will train all the models on the train dataset and evaluate the models on the test dataset.

Deliverables

• Python code that implements a majority class baseline and a keyword matching baseline
• Python code that implements two or more machine learning classifiers

Part 1b: Dialog management

In order to implement the dialog management system, the following State Transition Diagram (STD) has been modelled:

This diagram has been implemented with python into a working dialog management system using a state transition function. Furthermore a lookup function has been added to find suitable restaurant suggestions and a algorithm for identifying the preferences of the user based on their utterances.

Deliverables

• The state transition diagram
• A working dialog system interface, implementing a state transition function.
• An algorithm identifying user preference statements in the sentences using pattern matching on variable keywords and value keywords on utterances
• A lookup function that retrieves suitable restaurant suggestions from the CSV database

Part 1c: Reasoning and configurability

Reasoning

In the dialog system a simple reasoning component has been implemented. This uses inference rules to determine additional restaurant properties based on initial properties. The user can choose for the following additional preferences:

• touristic
• untouristic
• assigned seats
• children
• no children
• romantic
• unromantic

Configurability

In order to implement some degree of configurability in the dialog manager the following features can be changed in the .env file:

• levenshtein_distance (0 or higher)
  • Levenshtein edit distance for preference extraction (0 means a 'dynamic' edit distance is used)
• formal (True or False)
  • Use formal or informal phrases in system utterances
• use_caps (True or False)
  • OUTPUT IN ALL CAPS OR NOT
• allow_restart (True or False)
  • Allow dialog restarts
• use_delay (True or False)
  • Introduce a delay before showing system responses
• use_tts (True or False)
  • Use text-to-speech for system utterances

Deliverables

• Implementation of implication rules
• Implementation of configurability for selected features
• Integration of the two deliverables above as part of the dialog management system from Part 1b