@voiceflow/natural-language-commander

natural-language-commander

NLC is a tool for connecting natural language commands with callbacks.

To see NLC in action, see the todo-chat live demo, or check out its repo.

It's somewhat inspired by Amazon's Alexa interface - so if you've integrated an app with the Amazon Echo, you should feel right at home.

Use Cases

Say the same thing in different ways

It's targeted to the case where you're building a bot, and want users to be able to ask the same question in a bunch of different ways. Say your bot keeps a todo list for users, and your user needs to buy some milk. You've got some logic for adding items to their list, so maybe you use a regular expression like /put (.+) on my todo list/ to match the command string with your addToList function. That works great, until your first user tells your bot, "add buy milk to my todo list", or "remind me to buy milk", it fails, and they say, "your bot isn't very smart." Ouch.

NLC helps you solve that problem, by connecting addToList to a bunch of potential ways of saying the same command (called utterances), like:

[
  'put {Item} on my todo list',
  'add {Item} to my todo list',
  'remind me to {Item}',
  'put {Item} on my list',
  'add {Item} to my list',
  'put {Item} on the list',
  'add {Item} to the list',
  ...
]

More specific matching

Besides just having a bunch utterances, NLC also lets you only count a slot (like {Item}) as matching if it meets certain criteria, like being in a list of words, looking like a date, matching a regular expression, or matching an arbitrary function. This lets you match "what's today look like" with a weather checker function, and "what's a sloth look like" with an image search function.

Common Misspellings

NLC knows about common misspellings, like definitely/definately, and will handle checking for them - so as long as you spell your utterances correctly, of course!

Gather training data

A system like this, where a human has to think up all possible ways to say the command, is never going to be perfect and catch every case. If building really robust natural language bot is your goal, you're probably going to end up with a machine learning system - which is going to need a large number of commands to train on. If you build your Minimum Viable Product with NLC, you can record every command, and what it matched to (if anything), and use that data to train your machine learning algorithm later. And even if you're planning on sticking with NLC, logging unmatched commands is a good way to find utterances you hadn't thought of.

Installation

Install from npm with

npm install natural-language-commander --save

or

yarn add natural-language-commander

And use in node with:

var NLC = require("natural-language-commander");
var nlc = new NLC();

or with TypeScript:

import NLC = require("natural-language-commander");
const nlc = new NLC();

NLC is written in TypeScript, and comes with its d.ts definition file - so you don't have to worry about a @types package.

Basics

NLC has four basic components - Intents, Utterances, Slots, and Slot Types.

Intents

An intent is a collection of utterances, slots, and a callback that collectively describe a single conceptual command, like "add an item to my todo list."

Utterances

An utterance is a specific way of giving a command, like 'add {Item}to my todo list', or 'remind me to {Item}'. Each intent will have many of these.

Slots

A slot is essentially an argument for an intent, which you expect to show up in a certain place in an utterance. It's often a noun, and is the {Item} in 'add {Item} to my todo list'. Checking for slots part of matching a command to an utterance, and the data in the slot gets passed along to the callback.

Slot Types

A slot type is how you say "for a given slot, only look for these words or patterns". NLC comes with some common defaults, and you can add your own.

Default Slot Types

The default slot types are:

• STRING: Any string of any length.
• WORD: A single word.
• NUMBER: A number, which can include commas. Matches will be returned as numbers.
• DATE: A string that's either a common date format like 1/1/2016, Jan 1, 2016, 2016-01-01, etc., or a word like today or tomorrow.
• SLACK_USER: A single word that starts with an @, like @user.
• SLACK_ROOM: A single word that starts with an # or a @, like #room.

Adding Slot Types

To add a custom slot type, call nlc.addSlot, passing in a SlotType object with the attributes:

• type {string} - The name of the slot type, to be used in intents. Note that you cannot add more than one slot type with the same name.
• matcher {string | string[] | RegExp | (value: string) => string} - This determines if a potential slot value matches.
  • String: The input must match the string exactly (case insensitive).
  • String Array: The input must match one of the strings exactly (case insensitive).
  • Regular Expression: The input must match the expression.
  • Function: The function must return anything but undefined to succeed. The intent will be passed the return value, which lets you normalize slot values.
• baseMatcher? {String} - Behind the scenes, NLC uses regular expressions to match your utterances. By default, it replaces slots with (.+), which is a capture group that will will match a string of any lenght greater than 1. Then it takes the captured string, and checks it against the slot's matcher. This generally works well if there is some text between slots, like {Slot1} then {Slot2}, since the regexp uses the text as natural stopping points. But when two slots are next to each other, like {Slot1} {Slot2}, the first slot may end capturing text that the second slot was looking for, resulting in a failure to match. The baseMatcher lets you fix that by providing an escaped string version of the regexp the utterance matcher should use. So, if the slot should only be a single word, you could use '\\w+', which will only match on a single run of characters, without spaces. Note that, if you use a regular expression for your matcher, it will also be used as the baseMatcher.

Examples

// Add a slot type that matches against the word 'this'
nlc.addSlotType({
  type: "STRING_TYPE",
  matcher: "this"
});

// Add a slot type that matches against 'this' or 'that'
nlc.addSlotType({
  type: "STRING_ARRAY_TYPE",
  matcher: ["this", "that"]
});

// Add a slot type that matches phone numbers with a regexp.
nlc.addSlotType({
  type: "PHONE_TYPE",
  matcher: /\d\d\d-\d\d\d-\d\d\d\d/
});

// Add a slot type with a function that matches strings less than 6 characters long,
// and returns the length.
nlc.addSlotType({
  type: "SMALL_COUNT_TYPE",
  matcher: slot => {
    if (slot.length < 6) {
      return slot.length;
    }
  }
});

// Add a slot type with a baseMatcher that only matches against a single word.
nlc.addSlotType({
  type: "WORD_TYPE",
  matcher: ["this", "that"],
  baseMatcher: "\\w+"
});

To remove a SlotType, you can call:

nlc.removeSlotType("SLOT_TYPE"):

however, you can only do this if no intents rely on the slot type.

Registering Intents

To register an intent, call nlc.registerIntent with an Intent object with the attributes:

• intent {string} - The name of the intent. nlc.handleCallback returns this when it matches an intent.
• slots? {IIntentSlot[]} - Optional. An array of slots included in the intent, with the attributes:
  • name {string} - The name of the slot, to be used in the intent's utterances.
  • type {string} - The name of the associated slot type. If this is a custom type, you should add it BEFORE registering an intent that uses it. If nlc.handleCallback included some data, that will be the first argument, and the slots will come after. This lets you pass along other information about the command, like data about the user who issued it. NLC checks utterances in order - so if you have a more specific utterance, like 'say {Something} to {Username}', list that before a less specific command like 'say {Something}'. Also note that an utterance will match against the start of the user input - so if the utterance is 'tell me a joke', 'tell me a joke!' will match, but 'hey tell me a joke' will not.
• utterances: {string[]} - A list of utterances that will match to the intent.
• callback {Function} - The callback to run when the intent matches a command. Slots (if any) will be passed in as arguments in the order they were listed in the slots array.

Examples

// A simple intent without slots.
nlc.registerIntent({
  intent: "NO_SLOTS",
  utterances: ["this is a test"],
  callback: () => {
    console.log(`it's a match!`);
  }
});

// An intent with a couple slots.
nlc.registerIntent({
  intent: "SLOTS",
  slots: [
    {
      name: "Thing",
      // The default string slot type.
      type: "STRING"
    },
    {
      name: "ThingType",
      // Some custom slot type
      type: "ThingTypes"
    }
  ],
  utterances: ["{Thing} is a {ThingType}"],
  callback: (thing, thingType) => {
    console.log(`${thing} is a ${thingType}!`);
  }
});

// A intent expecting some data.
nlc.registerIntent({
  intent: "NO_SLOTS",
  slots: [
    {
      name: "Thing",
      type: "STRING"
    }
  ],
  utterances: ["{Thing} is a test"],
  // In this case we're expecting nlc.handleCallback to have passed a user object in.
  callback: (user, thing) => {
    console.log(`${user.name} thinks that ${thing} is just a test.`);
  }
});

You can also add an utterance to an existing intent (if you're generating them from some learning algorithm or something) by calling:

nlc.addUtterance('INTENT_NAME', 'New utterance');`

or remove one by calling:

nlc.removeUtterance('INTENT_NAME', 'New utterance');`

You can remove an entire intent by calling:

nlc.deregisterIntent('INTENT_NAME')`

Handling Commands

Once your intents are set up, you can start handling commands from users. To do that, call nlc.handleCommand, optionally passing in some arbitrary data to pass along to the matching intent, and then the text of the user's input.

Examples

// Not passing data.
nlc
  .handleCommand(userInput)
  .catch(() => {
    console.log(`${userInput} didn't match :-(`);
  })
  .then(intentName => {
    console.log(`${userInput} matched with ${intentName}!`);
  });

// Passing data.
nlc
  .handleCommand(user, userInput)
  .catch(() => {
    console.log(`${userInput} didn't match :-(`);
  })
  .then(intentName => {
    console.log(`${userInput} matched with ${intentName}!`);
  });

Examples

// Register a question for future use.
nlc.registerQuestion({
  name: "KLONDIKE_QUESTION",
  slotType: "STRING",
  utterances: [`I'd {Slot}`, `I would {Slot}`, `{Slot}`],
  questionCallback: () => console.log(`What would you do for a klondike bar?`),
  successCallback: () => console.log(`Wow, I wouldn't do that!`),
  cancelCallback: () => console.log(`Sorry I don't know what you mean.`),
  failCallback: () => console.log(`Fine don't tell me then.`)
});

// Ask the klondike question.
nlc.ask({
  userId: "12345",
  question: "KLONDIKE_QUESTION"
});

// This would print `Wow, I wouldn't do that!`
nlc.handleCommand({
  userId: "12345",
  command: `I'd wrestle a bear.`
});

To de-register a question, call:

nlc.deregisterQuestion("QUESTION_NAME");

Handling commands that don't match

To register a function to be called when a command doesn't match any intents, call nlc.registerNotFound(callback: (data?: any) => void), passing in the callback. Note that this will not be called when an answer command doesn't match, since that has its own seperate callback.

You can also use nlc.handleCommand(command).catch() to catch bad commands - but that will also be called when an answer doesn't match, which may cause your bot to respond from both the catch callback and the question's failCallback.

Full Example

Here's a full example of using NLC to guess a favorite color.

const NLC = require("natural-language-commander");

const nlc = new NLC();

const favoriteColor = "blue";

// Add a custom color slot type.
nlc.addSlotType({
  type: "Color",
  matcher: ["red", "orange", "yellow", "green", "blue", "purple"]
});

// Register an intent for guessing if the bot likes a color.
nlc.registerIntent({
  intent: "FAVORITE_COLOR_GUESS",
  slots: [
    {
      name: "Color",
      type: "Color"
    }
  ],
  utterances: [
    "is your favorite color {Color}",
    "is {Color} the best color",
    "do you like {Color}",
    "do you love {Color}"
  ],
  callback: color => {
    if (color.toLowerCase() === favoriteColor) {
      console.log(`Correct! ${color} is my favorite color.`);
    } else {
      console.log(`Sorry, I don't really like ${color}.`);
    }
  }
});

// Register a question for asking the user their favorite color.
nlc.registerQuestion({
  name: "USER_FAVORITE_COLOR",
  slotType: "Color",
  questionCallback: () => console.log(`What is your favorite color?`),
  successCallback: color => {
    if (color.toLowerCase() === favoriteColor) {
      console.log("Mine too!");
    } else {
      console.log(`meh.`);
    }
  },
  cancelCallback: () => console.log(`Fine don't tell me then.`),
  failCallback: () => console.log(`That's not even a color!`)
});

nlc.registerNotFound(() => console.log(`Sorry I'm not sure what you mean.`));

/*
 * Test some commands
 */
nlc.handleCommand("is your favorite color Blue?"); // 'Correct! Blue is my favorite color.'
nlc.handleCommand("do you like blue"); // 'Correct! blue is my favorite color.'
nlc.handleCommand("is red the best color?"); // 'Sorry, I don't really like red.'
nlc.handleCommand("do you love Green"); // 'Sorry, I don't really like Green.'
nlc.handleCommand("do you love tacos"); // Sorry I'm not sure what you mean.
nlc.handleCommand("do you think blue is pretty?"); // Sorry I'm not sure what you mean.
nlc.handleCommand("what is the meaning of life?"); // Sorry I'm not sure what you mean.
nlc.ask("USER_FAVORITE_COLOR"); // What is your favorite color?
nlc.handleCommand("blue"); // Mine too!
nlc.ask("USER_FAVORITE_COLOR"); // What is your favorite color?
nlc.handleCommand("tacos"); // That's not even a color!
nlc.ask("USER_FAVORITE_COLOR"); // What is your favorite color?
nlc.handleCommand("nevermind"); // Fine don't tell me then.
nlc.ask("USER_FAVORITE_COLOR"); // What is your favorite color?
nlc.handleCommand("do you like blue"); // 'Correct! blue is my favorite color.'

/*
 * Logging matches
 */
function logMatch(command) {
  nlc
    .handleCommand(command)
    .then(intentName => {
      console.log(`Matched ${intentName}`);
    })
    .catch(() => {
      console.log(`No match`);
    });
}

logMatch("is your favorite color green?"); // Matched FAVORITE_COLOR_GUESS
logMatch("do you think blue is pretty?"); // No match

Hubot

NLC pairs particularly well with Hubot, GitHub's bot framework. It's designed to use regular expressions to match callbacks to commands, but it's pretty simple to use NLC for better matching, while still having access to the Hubot res object for replying and getting data about users and rooms.

In your hubot script files, instead of using:

robot.respond(/is your favorite color (.+)/i, (res) => { ... })

Register an intent like:

nlc.registerIntent({
  intent: 'SOME_COMMAND',
  slots: [
    name: 'Color',
    type: 'Color'
  ],
  utterances: [
    'is your favorite color {Color}',
    'is {Color} the best color',
    'do you like {Color}',
    'do you love {Color}',
  ],
  // The callback is expecting the Hubot res object, which is the same object
  // the original callback was getting.
  callback: (res, color) => { ... })
});

Then, somewhere in your scripts directory, add a catch-all that passes any uncaught messages to NLC for matching:

robot.catchAll((res: hubot.Response) => {
  // This returns early if the text didn't start with the robot's name or one of its
  // aliases, so you're not running every message in the chat through NLC.
  if (!text.match(robot.respondPattern(""))) {
    return;
  }

  // Pass the message to nlc. Any matching callback should probably respond with
  // res.send() or something.
  nlc.handleCommand(res, message).catch(() => {
    // Send a failure message if the command didn't match.
    res.send(`Sorry, I can't do that ${res.message.user.name}`);
  });
});

Sanitizing Data

NLC is case insensitive, and handles things like removing extra spaces and common misspellings, so you shouldn't have to do much processing on an input before passing it in. However, you should be careful to sanitize any slot values before putting them in a database, displaying them to users, or otherwise evaluating them, since those are still user-generated strings. It's probably also a good idea to truncate strings to a reasonable character count, to stop users from passing in a gig of data and slowing down your server.

Since NLC doesn't know how you're going to be using it (maybe you want users to be able to tell your bot long stories, who knows), it doesn't handle sanitization for you.

Known Bugs

• Can't use curly braces in a Slot Type baseMatcher.

Change Log

[0.2.0]

• Add deregister methods for Intents, Questions, SlotTypes, and Utterances

[0.1.4]

• Fixed bug with multiple commas in a NUMBER Slot Type.
• Added the CURRENCY Slot Type.

[0.1.3]

• Fixed bug with nested questions.

[0.1.2]

• Added registerNotFound()

[0.1.1]

• Added questions
• Added addUtterance()

[0.0.12]

• Fixed inconsistent slot type name in the demo full example.