If you have trained a chatbot yourself or you have read our previous posts about the topic for Dialogflow (former API.ai) and LUIS, you already know that creating a functioning chatbot is a long and tedious process. On top of the complexity, even if you fulfill an “appropriate” training, this does not ensure that you will obtain a chatbot able to understand various inputs from users.
A sentence structure you have not thought of, a slightly more complex syntax, or the mention of a variable you have not come up with, can render your chatbot useless. Focusing on the chatbot engine Amazon Lex, I will explain in detail the two main ways Bitext proposes to help you build a robust chatbot.
We train your bot:
Imagine you have a big data set that you consider has a high coverage. We could be talking about hundreds of millions of sentences the users have written. Now think of the amount of time and effort it would take to tag all those sentences manually to train your bot. Alongside this challenge, you can add human-made mistakes on the tagging (tagging is a very mechanic and tedious task).
Training a bot involves several steps, as seen in this blog the past weeks:
- Create your chatbot
- Add all the ‘intents’ or types of requests a user can make to your bot, e.g: ‘purchase pizza’, ‘purchase drink’, or ‘cancel order’.
- Fill in what Amazon Lex calls ‘slots’. All the items related to your intents, for example, for the intent ‘purchase pizza’ logical slot types would be ‘type of pizza’, ‘ingredients’, ‘type of crust’, or ‘size’.
- Create and manually tag ‘Sample Utterances’. These are all the conceivable sentence structures a user could utilize to fulfill an intent (for example, to order a pizza)
As you can see, the tasks above, particularly 3 and 4, are endless.
How can Bitext improve the training process?
With our system based on linguistic knowledge, we can create and tag your training data set fast and accurately.
This reduces your training time drastically, and it radically increases the quality of your bot, allowing you to have a bigger more accurate and cheaper to create data set than a manually generated one. By tagging a large set accurately, you improve UX avoiding interactions like the following:
This happens if you forget a possible slot, or if the user mentions a pizza that is not on the menu:
This happens if the user asks in a way you have not trained your bot to understand.
The larger the dataset is, and the better tagged it is, the more accurate your chatbot will be, and the better business outcome will achieve. Data equals more sales for e-commerce bots and greater savings for customer care ones.
If you want to see how Bitext can help you to train your bot faster skipping all this process and achieve better accuracy schedule a demo with us!