Support will be limited on november 1st and 11th, 2024.

NLU bots: best practices

Congratulations! By using intent detection for your bot, you’ll make your website journey much more intuitive for your users (and much more flexible for you)... and you've also just made your entrance into the world of artificial intelligence! AI is first and foremost based on the correct organization of elements you are already familiar with. Here is some information to guide you. 

Some definitions 

  • Intents categorize the wishes and needs expressed by your visitors, thus enabling a card in a bot's conversation scenario. Example: product information, forgotten password, delivery information...

  • Intents are defined by expressions that all correspond to the same customer need, but express it in different ways.

    For example, the following expressions, which could all be found in various conversations:

    “I have lost my password”, “I can't find my password”, “Forgot password”, “I can't find the code”, “I don't remember my login”, “I don't remember my password!”, “I couldnt fund my password”, etc..., can all be linked to one and the same intent: “forgotten password”, to which the bot will always give the same answer!
  •  
  • Bots with intent detection use artificial intelligence to analyze expressions used by visitors and match them with expressions they know, which are already associated with intents. So to work well, an intent must be associated with a minimum of 10 to 15 different expressions in iAdvize's intent manager. These expressions must be short and vary slightly from each other(using synonyms or typos for example) in order to train the AI successfully.

    For example: “forget pasword”, “I can't remember my login” and “I lost my password” include very different words and even mistakes, but these three expressions are all relevant to the “forgotten password” intent!
  • The entities complement and qualify intents because they correspond to more specific terms. For example, if a visitor types: “I want a ticket for Nantes”, the bot will recognize the “buy a ticket” intent from the expression “I want a ticket”, and will also recognize the entity “Nantes”: so you won't need to list all possible destinations in the expressions associated with the “buy a ticket” intent.

    To detect entities, the bot always relies on inputs, which can be of two types: 
    • value lists: you list reference values (for example, a list of destinations: Nantes, Paris, Venice…) with which you can associate synonyms (for the reference value “Paris”, for example, you can enter: “Paris, France”, “the City of Light”, “the capital”, etc.)

    • regular expressions correspond to certain forms that the bot will definitely recognize: email address, phone number… but also order number or parcel tracking number!

Your intents: where to start? 

To list the relevant intents for your bot, simply start with what you already have: 
  • Your FAQ, which you've probably been building up over the years on your site: this is the best starting point to find common questions as well as questions more specific to your activity, all of which can become relevant intents.

  • Your conversation categories: if you use them, your iAdvize conversation categories will also be a good addition to your FAQ! You can create your answer categories yourself and later find them in your reports, such as the Insights report.

  • Your Insights report: when looking at the word cloud in the “Volume by category” section, you may detect new concerns among your visitors, which you can choose to turn into intents. 

Intents you might not have consideredGBM_-_Les_intentions_auquelles_vous_ne_penseriez_pas.jpg

Sometimes intent detection fails, even if the intents are input properly... because your visitors don't always use your bot as intended! Here is a review of the most common cases:
  • “hello”, “thank you” and “goodbye”: often your visitors greet or thank your bot! Even if the bot is unmoved by their politeness, you can create dedicated intents and associate them with a card that will respond to them before resuming the scenario, which your visitors will appreciate.

  • “joke”: Internet users are used to joking with bots. You can create a dedicated intent and associate it with a “joke” card inspired by your industry, for example, before resuming your scenario. Or you can make your bot answer that it doesn't have a sense of humor yet (which is true).

  • “insult”: bad experiences can also happen. If you want to protect your human respondents from rudeness and trolls, you can, for example, attach the “insult” intent to the action of ending the conversation.

  • “talk to an agent”: some visitors already know how a bot works and make their request directly. You can speed up their transfer to an agent, or, if your bot is also there to collect information, notify them that the transfer will indeed take place once they've told you more about your case.

Your expressions: how do you supply your intents?

As we mentioned above, for the AI to work well, each intent must be associated with a minimum of 10 to 15 different expressions. To train your augmented intelligence faster, iAdvize offers several solutions: 
  • Intent libraries: these include standard intents and expressions that are common to multiple industries (sales, insurance, finance, telecommunication, travel). For each library, you can download all the intents, or select them one by one.

  • Collaborative training: with iAdvize, you can ask your team to help you make your AI more relevant! By enabling collaborative training, you can ask your respondents to help you sort expressions that come from your conversations with your visitors and are suggested directly in their conversation panel by iAdvize when they're not busy talking to visitors. You will soon have a large number of ready-grouped expressions that you can validate in your intent manager.

    Since they come from your own conversations, these expressions will have the advantage of being
    perfectly adapted to the needs of your site. You can also enable collaborative training before you even start using intent detection, to begin gathering relevant expressions and intents!

  • Suggestions from your augmented intelligence: if you can't mobilize your teams for collaborative training, your AI will directly suggest clusters of expressions in your intent manager, without having them tested by your respondents.

  • Finally, you can of course manually create your expressions, intents and entities in the intent manager, but also edit and delete them. You can also copy them from your Conversation report, which is full of verbatim quotes and common expressions that you just have to copy in order to supply your intents with expressions.

    In the “intent fallback rate” section of your Automation report, you can export all the expressions your AI was unable to detect... you just need to associate them with the right intents in your manager!

Combine intents and entities

  • As we explained: intents categorize the wishes and needs expressed by your visitors, enabling a card in a conversation scenario - so they lead to an action, whereas entities complement and qualify intents, because they correspond to more specific terms. For example, in the sentence: “I am looking for more information about Futurevac V12”, the AI will recognize both the “product information” intent, which will enable the “transfer to an ibbü expert” card and the “Futurevac V12” entity as a product reference among others, if it appears in a value list you have input.

  • When used properly together, intents and entities will also make your augmented intelligence more exhaustive. With more general intents, you won't need to list specific information such as your product offering or a list of destinations: it's best if the entities take care of this type of information.

  • Because they are more specific, as iAdvize develops, you will be able to use entities to create connections with your API!

    For example, when at your bot's request, your customer gives a reference like johnsmith@jollybank.com, associated with tracking number ZU700001, augmented intelligence will automatically recognize these inputs as Regular Expression (RegEx) entities, and your bot will be able to run a custom API query independently.

A few final tips

GBM_-_Dernie_res_astuces_FR.jpg

To guide you in setting up your intent detection, here are a few principles to keep in mind: 
  • Be sure to have a variety of expressions, while targeting them as much as possible. As we explained, the key to an effective intent lies in the variety of expressions associated with it, at least 10 to 15. The more varied and relevant your expressions are, the better your AI will detect the intent behind them.

    For example, the expressions
    “I would like to buy a ticket for a concert” and “I want a ticket for a show” are short sentences that contain different verbs and complements. So they are two very relevant expressions for the same intent, “buy a ticket”. On the other hand, taken as it stands, the sentence “I would like to book a seat for Classicfest but would like to know who the first violinist is” is not a relevant expression, because it's very long and does not only concern the "buy a ticket” intent. If you make an expression out of it and combine it with that intent, be sure to keep only the first part of the sentence, “I would like to book a seat”, so as not to confuse the intent detection.

  • In their own way, mistakes are also expressions: your AI is based on text and nothing but text, so every typo or spelling mistake is also a variation, just like a synonym. Take all your interlocutors into account, even the most rushed and clumsiest, and include the most common mistakes in your expressions!

  • The word “hello”, a parasite word: many visitors word their query as they would in a store: “hello, I'm looking for a vacuum cleaner”... and of course, this may be the case for all the queries they send you. Therefore, we recommend that you include this term in at least one of the expressions associated with each of your intents, so it doesn't negatively impact the detection. Afterwards, if you spot other “parasite words” that negatively impact detection (failure or misattribution), feel free to associate them with as many intents as possible in order to neutralize them.

  • Supply your intents evenly: the ones you have supplied with the best information (= with the most varied expressions) are more likely to be properly detected... On the other hand, if some are less complete, your AI may decide in favor of the most complete ones, thus making mistakes. This is all the more true if some of your intents can be associated with expressions that are very similar to each other: the more complete your expressions are, the more your AI will be able to focus on what differentiates them!

The life of your bot

GBM_-_La_vie_de_votre_Bot.jpg

Over the months and years, you may have to make changes to your website and bot. The same goes for your intent detection!
  • As we said earlier, the 10 to 15 initial expressions you need to associate with an intent are a minimum - it's in your interest to add new expressions as you come across them!

  • The advantage of intent detection bots is that it's very easy to update them if you identify new needs, questions or uses: just add a new intent and the answer card associated with your existing journey!

  • If, in the same vein, you want to refine your intents, that's certainly possible. For example, it might be helpful to add a specific intent called “featured vacuum cleaner information”, if this product is the subject of 60% of queries currently associated with the “product information” intent.

    In this case, simply make sure you
    “clean up” any expressions associated with the original intent that may relate to the new intent, as they could confuse your augmented intelligence. 
Although intent detection requires an initial investment, there are many aids such as collaborative training and intent libraries to help you speed the process up. But most importantly, have confidence in yourself. Augmented intelligence is based primarily on what you and your team already know: your job and feedback from users of your site!

Soon, they will benefit from an optimal conversation experience, which will be easier for you to edit or enrich according to your wishes and those of your visitors, as well as the latest developments in your industry.