How to Teach Your Chatbot With Training Data
The chatbot represents a booming trend in online interaction, helping to provide information quickly to customers. They help users get straight to the point without the need to hang on to a hold message or scour a FAQ page for information.
Update: A more recent version with examples and training data resources can be found here.
The Bots They Are A Learning
While chatbots may be a boom area for companies, they don’t just pop into existence fully-formed, although it is possible that designers can build a rough example in a few hours.
Instead, the chatbot needs to be created and trained based on sets of data. Most businesses, from early adopters like the hospitality sector or marketing to health, government and B2Bs will already have much of that data, but it is likely to stored in various unstructured forms.
Different types of chatbots handle data in different ways, scripted bots can only offer a limited set of functions or questions and will only accept a narrow range of responses. Machine learning lets bots develop a growing set of knowledge and understanding by studying the data from previous chats or a firm’s current FAQs and other reference material.
They can also watch live conversations (be they on the phone through the type of voice understanding in Alexa and Siri, and by text in IM chats or email) and learn from these over time.
Data can come from:
- The typical queries that your receptionists, help email account or support line fields on a daily basis.
- Information stored in FAQ pages, support chat scripts, call logs, email trails and other written resources.
- Personal knowledge, especially support and customer-facing staff.
- Public sources, such as government or open source data sets.
Trending Articles on Chatbot Marketing:
1. How to Get to 1 Million Users for your Chatbot
2. How to Get Users for Free using a Viral Loop
3. How I grew JokeBot from 26k subscribers to 117k subscribers
How Bots Learn
Whatever the source, data may need to be converted into a structured form that a chatbot can learn from. For a scripted bot, designers can use it to focus on the most commonly asked questions first and highlight the most common answers, working down the list of information by priority.
Writing a decision or process tree down may sound archaic, but it is often the simplest way to start developing a bot. However, things are a lot easier when the bot can train itself using Natural Language Processing (NLP), Natural Language Understanding (NLU) and other Machine Learning skills. These advanced technologies sound may sound scary to any non-technical person but are available as a service, just like your cloud email or office suite to help businesses provide services without fear of the details.
Machine learning bots, such as those you can build on SnatchBot’s platform, can be fed a set of data to get them warmed up before being let loose on the actual job. SnatchBot’s NLP uses a declarative approach to intent and entity recognition, where large numbers of example sentences show the bot what terms are important in the conversation and what users want to achieve.
Once up and running, analytics from the conversations allow bot builders to spot where the bot has difficulties correctly analyzing sentences and, in addition to the automated machine learning that takes place for their bots, SnatchBot recommends adopting a ‘supervised machine learning’ approach, where builders add critical new sentence examples manually.
Benefits of the Machine Learning Age
Advanced training includes sentiment analysis where the bot looks at the language used using NLP. It can discern if a user is upset (and can learn what to do that will make them happy), or confused (and layout a simpler set of options.)
A final set of data can come from customer satisfaction scores at the end of each chat. A simple “please rate your experience from one to five stars…” option will let your team know if the chatbot is meeting customer expectations. If they are unhappy, asking why can help find the flaws in your bot.
As smart bots gain more knowledge over time, they can expand the range of features they offer the client or customer, making them of more value to the business by saving time. That frees people up for more important tasks, and saves revenue. And, as AI becomes smarter and prevalent in chatbots, they will be able to be left more to their own devices to continuously learn and improve the service they offer.
Even so, any company looking to build or evolve their chatbot service should ensure they test and check through user interactions on a regular basis to ensure it meets the needs of the business and is as simple (and fun, businesses should ensure their bots have some mote of personality) to use for the customer.
