Chatbots Magazine Become a Writer Intro to Chatbots

What have been the experiences with the NLP platforms such as api and wit.ai?

We have talked to a number of people that have had various levels of success with the services. Some have created their own NLP platforms while others are looking for ways to improve the training to get better results. What is your mileage?

We started with wit.ai and we ended up bringing it all in house only because there are some great NLP libraries out there that do the same thing. Plus the latency was an issue with wit.ai, we can now store all our training in memory and access it with little delay.

I should say we still have some bots in the wild using wit.ai because the training was already done and still services the needs but going forward we won’t be using anything but our own.

1 Like

We used wit.ai at CartSkill for the first bot we developed but abandoned it quickly. Training takes time and the results are difficult to predict and often times irrelevant, so it was difficult to build something of high value for customers. As we have all the necessary AI talent in-house with our own team (RNN, CNN, LSTMs) we put together our own framework and are now building all of our bots on top of that. Provides much better accuracy and relevancy + we get to deploy good quality bots quicker.

It’s easier for “english” chatbots to use your own NLP.
For other languages, it’s hard to find a good corpus/libraries that handle the especific language characteristics, so, we need to run to watson or luis.ai

We used Luis. AI for one of the bots but they have limitations on number of intents (20) that can be trained in a model. You can workaround it by training multiple models with no overlap In intents.

Which NLP libraries do you use? What the best for a java application?

I was debating among Watson and Api.ai, finally I decided for Api.ai is more easy to use and some API Rest habilities useful for me but seriously Watson is pretty powerful but is quite modulated.

We decided to build quickly on API.AI since we didn’t have a lot of capacity in the dev team. Yes, it will bring you only so far, but for proof of concept and quick tryouts, it is easy and saves time.

Microsoft’s LUIS. Initially we started out with botkit + wit.ai (before FB), but when news came out of Microsoft’s BotBuilder, and its integration with its LUIS, we migrated to it. However, there were problems with BotBuilder and the direction they had wasn’t what we want for now. So we came back to botkit and retained the LUIS model.

As @rajesh.bindal has said, there are limitations, and during our time developing and training, they were changing the responses and you suddenly wonder why your bot behaves differently. Otherwise, LUIS has been sufficient for our needs, but its slowly showing that we will be diverting away from it for a different solution.

1 Like

Ive been using Wit. It definitely takes a bit of ‘fiddling’ around to get it to do what you want reliably. The thing that helped me most with Wit was when I started trying to make every story as modular as possible. I’ve also had to rely quite heavily on creating context markers in order to guide it. My ultimate aim is to go in house, so i’m piggy backing off wit for now and comparing results with those that I get in house in an attempt to derisk the switch