Fountain is a natural language data augmentation tool that helps developers create and expand domain specific chatbot training datasets for machine learning algorithms.

In order to build better AI assistants, we need more data, better models aren’t enough.

Most of NLU system requires entering thousands of possible queries that future users would -- most possibly -- use, and annotate every sentence segment that can identiy user's intentions. It is generally a hectic and tedious manual process. Fountain aims to help developers smooth away this process and generate a volume of training examples to make it easier to train and build a robust chatbot systems.

The tool intends to make it easy to build the same dataset for different intent engines (Amazon's Alexa, Google's API.ai, Facebook's Wit, Microsoft's Luis). At the moment, the tool generates training datasets compatible with the RasaNLU format.

You can install the package via:

$ pip install git+git://github.com/tzano/fountain.git

Install the dependencies:

$ pip install -r requirements.txt

Fountain uses a structured YAML template, developers can determine the scope of intentions through the template with the grammar definitions. Every intent should include at least one sample utterances that triggers an action. The query includes attributes that identify user's intention. These key information called slots. We include different samples to be able to generate datasets.

We use three operations:

• Argument ({slot_name:slot_type}): used to declare slot pattern.
• Argument (( first_word | second_word )): used to provide a set of keywords, these words could be synonymes (e.g: happy, joyful) or the same name with different spellings (e.g: colors|colours)

A simple example of an intent would be like the following

book_cab:
  - utterance: Book a (cab|taxi) to {location:place}
    slots:
      location:place:
        - airport
        - city center

This will generate the following intent json file using to_json.

[
    {
        "entities": [
            {
                "end": 21, 
                "entity": "location", 
                "start": 14, 
                "value": "airport"
            }
        ], 
        "intent": "book_cab", 
        "text": "book a cab to airport"
    }, 
    {
        "entities": [
            {
                "end": 25, 
                "entity": "location", 
                "start": 14, 
                "value": "city center"
            }
        ], 
        "intent": "book_cab", 
        "text": "book a cab to city center"
    }, 
    {
        "entities": [
            {
                "end": 22, 
                "entity": "location", 
                "start": 15, 
                "value": "airport"
            }
        ], 
        "intent": "book_cab", 
        "text": "book a taxi to airport"
    }, 
    {
        "entities": [
            {
                "end": 26, 
                "entity": "location", 
                "start": 15, 
                "value": "city center"
            }
        ], 
        "intent": "book_cab", 
        "text": "book a taxi to city center"
    }
]

The same file would generate the following csv file using to_csv.

intent	utterance
book_cab	book a cab to airport
book_cab	book a cab to city center
book_cab	book a taxi to airport
book_cab	book a taxi to city center

The library supports several pre-defined slot types (entities) to simplify and standardize how data in the slot is recognized.

These entities have been collected from different open-source data sources.

• Dates, and Times

  • FOUNTAIN:DATE
  • FOUNTAIN:WEEKDAYS
  • FOUNTAIN:MONTH_DAYS
  • FOUNTAIN:MONTHS
  • FOUNTAIN:HOLIDAYS
  • FOUNTAIN:TIME
  • FOUNTAIN:NUMBER

• Location

  • FOUNTAIN:COUNTRY
  • FOUNTAIN:CITY

• People

  • FOUNTAIN:FAMOUSPEOPLE

In order to build Fountain's builtin datatypes, we processed data from the following data sources:

• Wikidata WikiData
• GeoNames Geographical database
• Pantheon dataset

You can easily load and parse DSL template and export the generated dataset (Rasa format).

You can find this sample under the directory \labs

# DataGenerator
data_generator = DataGenerator()
# load template
template_fname = '<file>.yaml'
# parse the DSL template
results = data_generator.parse(template_fname)

# export to csv file
data_generator.to_csv('results.csv')
# export to csv file
data_generator.to_json('results.json')

pytest

You can find examples on how to use the library in labs folder. You can enrich the builtin datasets by adding more files under data/<language>/*files*.csv. Make sure to index the files that you insert in resources/builtin.py.

For more information about Chatbots and Natural Language Understanding, visit one of the following links:

• RASA NLU
• Voice Design Guide - A great resource for learning conversational and voice user interface design.

• RASA NLU (Supported)

• Wren - News Chatbot to discover & explore daily news stories. We used Fountain to generate more than 20,000 samples. The Yaml file is available here.

If you are having issues, please let us know or submit a pull request.

The project is licensed under the MIT License.