Dos and Don’ts of a High-Grade Chatbot Building
A short recommendation for avoiding the most common mistakes in Conversational AI
Okay, I’m skipping the part telling what a chatbot is and how to apply it. This article is for those who already decided to build-up their first (second or third) chatbot. Below you may find handy tips that will keep you from making common mistakes.
Always keep in mind the reason you’ve decided to build a chatbot, never forget WHERE and HOW it will be used
1. Avoid long phrases. No story-telling when un-called for. Average length phrases are the best. Remember, this is a dialogue you are building, not a monologue.
Bad “Hi! I’m Symphony, an audiophile bot. I can’t live a day without music, it is my passion! Also, I got a very good taste in music and thousands of songs are in my base. I’d love to share it with you!”
Good “Hi! I’m Symphony, an audiophile bot. I can recommend an artist judging from your taste in music. What’s your favorite music genre?”
Top 4 Most Popular Bot Design Articles:
1. What 10 Billion Messages can Teach us About Making Chatbots
2. How to Quickly Improve your Chatbot’s Retention & Engagement
3. Zero to 45k users: how we created a chatbot to track airfares
4. Our journey of building a scalable knowledge management chatbot
2. Avoid open-ended questions and statements, it’s better to move users toward action.
Bad “Hi! I’m Symphony, an audiophile bot. I’d love to recommend an album and tell you everything about it.”
Good “Hi! I’m Symphony, an audiophile bot. I’d love to recommend an album and tell you everything about it. Do you want to know something about a song of the day?”
3. Avoid bureaucratic style. Make your text simple and easy-to-understand.
Bad “It is crucially important to notice that following albums this brilliant singer recorded have become platinum, which makes it possible to draw a conclusion about the successfulness of the debut album as a means of an effective entrance to the international arena.”
Good “Debut album drew the world’s attention to the artist. No wonder his next records became platinum!”
4. Add personification.
“Symphony, how was this album rated?”
“Well, critics have poured some cold water on it — the rating is 3.14”
5. Remember the user profile.
“Symphony, how was this album rated?”
“It was rated as outstanding — 9.6!”
“Wow, I loved it too”
“By the way, you rated the previous album much lower — only 4 out of 10!”
6. Don’t teach your users how to talk, but at the same time don’t leave them alone in the dark.
Bad “I can find top critic reviews or I can tell what your friends say of the artists and bands. In case you want to know the rating, say “rating”. In case you are interested in reviews, say “reviews”.
Good “I can find top critic reviews or I can tell what your friends say of the artists and bands. They actually praised the new Foals album. Would you like to hear it?”
7. Stay ethical — don’t offend or discriminate anyone.
Bad “Symphony, how was this album rated?”
“Well I guess girls really can’t rap — the album is rated 5.4”
Good “Symphony, how was this album rated?”
“Unfortunately, critics didn’t appreciate the efforts — the album is rated 5.4”
8. Process correctly all the communication breakdowns.
“Symphony, Benedict Cumberbatch”
“Sorry, I do not understand”
“Symphony, Benedict Cumberbatch”
“Sorry, I do not understand”
“Symphony, Benedict Cumberbatch”
“Sorry, I do not understand. Could we maybe just talk about the song of the day?”
Think of your users good and bad at the same time
Bad — because they may talk nonsense and you have to provide a nice answer for it.
Good — because we do want to solve their problem, find smth for them, help them, and make them happy as effectively as possible.
Conversational UI + platform restrictions = 😭
System constraints and how to circumvent them:
- Timeouts. Couldn’t keep it under N seconds? Your problems.
What to do:
- Try asynchronous calls
- Use workaround or bandaid solution
2. A platform can’t initiate the dialog. A bot can answer questions, but it can’t start the dialog itself.
What to do:
- Grin and bear
- Try to build up a dialogue, partly closing the cases where it may be required
3. Multiple intents. When using FAQ bot user may ask two questions in one sentence.
What to do:
- Deal with it at the NLU side
- Do nothing (when user figures out they talk to a bot, they might change the formulation of the question)
4. ASR recognition errors. Benedict Cumberbatch = beneficial cabbage-patch 🤦♀️
What to do:
- Grin and bear
- Build a Catch-All for such requests
5. The user is breaking in before the bot has finished its phrase.
“Hi! This is Justin Timberlake, do you have a mi-… ”
“YES!”
What to do:
- Monitor the signal level in a user’s speech path
Summary
- Keep in mind the bot’s mission, never forget where and how it will be used
- Try to make people happy but remember that they are not obliged to make YOU happy
- Develop your bot proceeding from the device it would run on
- Be honest and ethical
- Read aloud what you’re putting into your bot
Please, feel free to contact us if you need help with the conversation design or chatbot/smart assistant development — we at Just AI are always happy to help 😌
Don’t forget to give us your 👏 !
