Surprise! How we help your chatbot learn over time

A conversational agent, or chatbot, can learn over time in different ways. One important way is through human intervention when we resolve the most ambiguous cases and respond to emerging topics.

Your users will always surprise you.

After you set up your chatbot — flow, and a great FAQ dataset with all possible knowledge that your users, job seekers and candidates, would like to know about your company — you publish it and then you realise that chatbot users do want to know how the chatbot is doing, say hello or quickly access the main menu. Or something else entirely! The good news is — once our team at jobpal notices this kind of user behavior across all chatbots, we can help the chatbot respond to contextual questions better, without creating additional work for you.

How many small talk queries chatbot encounters and what are the main types?

Behind every great new Machine Learning technology, there is a human factor — people that dedicate their time to support the learning process of the particular language model, allowing it to respond to questions like “How are you?”, “What is your purpose?” or “Who are you?” without creating a need for supporting these topics in the company-specific FAQs.

Small talk and Job discovery are maintained by jobpal, FAQ responses come from out clients’ accounts and are maintained/trained by the client

All incoming messages go through intent recognition and entity extraction. Some, which are recognized as frequently asked questions (FAQs) go to a dataset specific to a particular company. You can find out more about what topics FAQs cover in a previous article.

Training an FAQ dataset is different than training the Intents dataset: Intents are more open-ended and affect all parts of the bot flow. It takes more than one person to assess the most ambiguous, troublesome commands. For example, users might join multiple intents when they ask questions like “Hello, how are you?”(should the bot respond to Hello, or to How are you?) or “Thank you, goodbye” (should it respond to Bye, or to Thanks intent?). To resolve these ambiguities, we created an internal Slack channel to maintain data quality.

Other topics we discussed in this channel involved misspellings (when to keep and when to delete?); using synonyms (United Kingdom — UK); training common welcome phrases for foreign language models (should “Hi” remain in a German or Chinese dataset); or training single words into the categories (should “Graduates” trigger a job search?). This internal channel serves as a resolution to each particular case, and is a great learning experience for the team in the process, especially when it comes to replicating learnings in all supported languages.

Your users will always surprise you! Whether maintaining the overall chatbot behavior through training intents, or company-specific FAQ sets, the key is to approach data quality as a team. With hundreds or thousands of daily incoming new messages, most are handled correctly — and for those ambiguous cases that need attention, it takes the team to assess, correct the direction, and maintain high data quality.