schema-classification 0.1.4

Perform Intent Classification using an External Schema

schema-classification

This microservice performs the classification of parse results

Usage

The input format looks like this

input_tokens = [
    {
        "normal": "my",
    },
    {
        "normal": "late",
    },
    {
        "normal": "transport",
    },
    {
        "normal": "late_transport",
        "swaps": {
            "canon": "late_transport",
            "type": "chitchat"
        }
    },
]

Calling the service looks like this

from schema_classification import classify

absolute_path = os.path.normpath(
    os.path.join(os.getcwd(), 'resources/testing',
                    'test-intents-0.1.0.yaml'))

svcresult = classify(
    absolute_path=absolute_path,
    input_tokens=input_tokens)

The output from this call looks like

{
    'result': [{
        'classification': 'Late_Transport',
        'confidence': 99 }],
    'tokens': {
        'late': '',
        'late_transport': 'chitchat',
        'my': '',
        'transport': ''}
}

Classification via Mapping

Classification of Unstructured Text is a mapping exercise

The mapping is composed of these elements

1. Include One Of
2. Include All Of
3. Exclude One Of
4. Exclude All Of

The classifier will map extracted entities from unstructured text using the listed elements.

for example,

TEST_INTENT
  - include_one_of:
    - alpha
    - apple
  - include_all_of:
    - beta
    - gamma
  - exclude_one_of:
    - delta
  - exclude_all_of:
    - epsilon
    - digamma

This intent will be selected if the set of extracted entities has either alpha or apple and has both (beta, gamma). The intent will be discarded if delta occurs or if both (epsilon, digamma) occur.

In python, everything can be loaded into a native set structure and use native operations like difference, intersection, union, and symmetric difference.

Because all set operations are native (underlying C modules), it's extremely fast to find an accurate classification.

The system adds more smarts by figuring out what to do if the rule states delta is excluded, and a descendant of delta is present.

Or if alpha should be included and a sibling or child of alpha is present, etc.

In this case, I usually rely on a heuristic to boost or lower confidence and tweak that overtime to get a good result.