NLU systems improve over time
Measured by accuracy of response, the systems improve when they are repeatedly trained. The larger the date set that is fed to train them the better is the accuracy of the model since it has covered all test cases that exists in the real world. A learning system can never be perfect since it will always have to keep on learning based on the cases it encounters from the real world and make changes to its neural algorithms and patterns.
Intents in the bot world are typically classified as General intent and Business intent. General intent are small talk items, like, Greeting types, Hi, Hello, Hola, Who are you? What is the weather? How are you today? What can you do? What’s your name? Will you be my friend? Etc.
Trained for casual intents
Such intents work for Affirmative and Negative intents. E.g. Yeah you are right is an affirmation intent for the bot. vs. No I did NOT mean that- is a negative affirmation intent. Your bot should be able to handle small talk and discern between casal, business and the varieties of affirmative and negative intents. The nature of general intents are casual conversational elements that are important to human interactions but not relevant to an objective oriented bot. The best bots try to don a human mask for small talk elements to keep the style conversational and humorous while engaging the users with smart-alecky responses.
Business intent is more transactional in natures, look up, transaction, tell me, show me or get it for me kind which involves retrieval of stored data or information that is the primary purpose of the bot. The bot can be any kind of bot, so a school bot will be trained to have information of holiday lists, class schedules, special notes, emergency close information and many others that are specific to the business of why the bot was created in the first place. Business intent bots or the interaction patterns that classify as business interactions, typically involve the retrieval of information from a stored data source to be trained later.
The more the examples for every intent that the model is tracking the better the model gets. The examples fed to the model for intent classification should be similar in meaning so that the vectors between them are close to each other.
Imagine this intent to be a cloud and intents around this cloud to be their own cloud. When the model encounters a new sentence, it measures the word vector, adds up to the sentence vector and determines based on the score which intent cloud is it closest to.
You can also use vector science to associate sentence intents by traits or other factors. As long as the measuring and classification methods are consistent, you can develop a learning model. It will be for the machine to score and categorize any new information that comes in way. At the bottom of it is statistical science, some algebra and vector science.
There are challenges here to solve for after the base vector models are in place.
Colloquial and local vocabs
The common ones like r u ok? are still easy to factor into a learning model. GoT rules over TWD on NF any day. ;-). How do you classify GoT rules as being from the Game of Thrones rather than the usual got?
How do you solve for negation? E.g. a NOT in a sentence changes the meaning to be the opposite.
Is a bull market closer to a cattle market? What is the intent of the word bull if not cattle? How does the machine learn to categorize a bull market as being something related to the stock market and not someplace where live stock is traded for the abattoir.
Stay tuned for more information.