Text-To-Intents

A library for parsing natural language to a pre-defined set of intents or semantics. The library is meant for flat-structured semantics: intent detection and slot filling with no hierarchies.

Getting Started

Installation

pip install .

Simple example

from texttointent              import TextToIntentSimple
from texttointent.intents.tell import Joke, Weather, WhatYouCan

# Create parser
tti = TextToIntentSimple(intents=[Joke, Weather, WhatYouCan])

# Returns an object of either Joke or Weather or WhatYouCan
# In this case, it should be Weather with a slot of PointInTime type
print(tti.parse("какая завтра будет погода?"))

Intents

Intents are defined using the following information

• A list of keywords. Every intent will be matched against their set of keywords.
• A list of slots. Types of the slots that should be attached to the intent.

How to add your own intent?

### Define custom intent ###
from texttointent.intents import Intent

class CustomIntent(Intent):
    KEYWORD = "custom"

    @staticmethod
    def get_keywords():
        return ["мой", "набор", "ключевых", "слов"]
    
### Create a parser with your intents of interest ###
from texttointent              import TextToIntentSimple
from texttointens.intents.tell import Joke

tti = TextToIntentSimple(intents=[CustomIntent, Joke])

For more examples, check the intents and slots folders.

Supported Solutions

Solution	Num Intents per Request	Num Slots per Intent	Numbers Supported
TextToIntentSimple	1	1	FALSE

Benchmarking

Benchmarking is a crucial part of understanding how good is the system for your particular problem. For this purpose, we provide a simple and convenient interface for benchmarking your datasets against the provided solutions.

How to benchmark a solution for your dataset?

from texttointent              import TextToIntentSimple
from texttointent.intents.tell import Joke, Weather, WhatYouCan
from texttointent.slots        import PointInTime

# Create a benchmarking dataset (usually you would like to parse it from .csv)
# (text, intent type (via an id keyword), slot value)
dataset = [
    ("что с погодой сегодня", Weather.KEYWORD, PointInTime.TODAY),
    ("что завтра с погодой",  Weather.KEYWORD, PointInTime.TOMORROW),
    ("а что ты умеешь делать", WhatYouCan.KEYWORD, None),
    ("никаких ключевых слов тут нет", None, None),
    ("погода обманчива в эти дни", None, None)
]

# Create a parser
tti = TextToIntentSimple(intents=[Joke, Weather, WhatYouCan])

# Obtain a benchmarking report (sci-kit classiciation report)
print(tti.benchmark(dataset))

Fine-tuning

Some solutions are trainable. If you want to fit your particular dataset, just call fit.

from texttointent              import TextToIntentSimple
from texttointent.intents.tell import Joke, Weather, WhatYouCan
from texttointent.slots        import PointInTime

# Create a benchmarking dataset (usually you would like to parse it from .csv)
dataset = [
    ("что с погодой сегодня", Weather.KEYWORD, PointInTime.TODAY),
    ("что завтра с погодой",  Weather.KEYWORD, PointInTime.TOMORROW),
    ("а что ты умеешь делать", WhatYouCan.KEYWORD, None),
    ("никаких ключевых слов тут нет", None, None),
    ("погода обманчива в эти дни", None, None)
]

# Create a parser
tti = TextToIntentSimple(intents=[Joke, Weather, WhatYouCan])

# Fit simple solution for your dataset
# (No gain for this particular dataset is expected)
fit_report = tti.fit(dataset)
print(fit_report)