This past week I had a chance to work on a conversational UI. It’s a neat way to interact with software and presents an interesting set of problems to implement.
Before even starting to build something I needed to decide on some tools to help with managing the conversations. There are a small number of libraries and services that can help with the natural language processing which would normally present a technical hurdle.
I evaluated 3 options:
wit.ai is a service owned by Facebook that provides some very impressive ways to train a language parser to categorize and extract information from sentences. It operates as a web service that if you send it a message it will return with a confidence measure and any extracted values. It provides a level of fuzziness in it’s matching and as a web service it allows you application to improve over time without deploying new code.
Api.ai is a similar Saas based solution that provides a way to convert a message into a more easily actionable format. It has slightly different semantics than wit.ai in how you describe the message parsing.
RiveScript – is the option I decided to work with. It specifies what is comparable to a markdown type language for defining messages and responses. Unlike wit.ai or api.ai it is more simplistic in the language parsing aspects (closer to regular expression matching). But has the nice property that it doesn’t rely on an external service. Implementations have been written in a bunch of languages which make it easy to get started writing bots.