Infobip Creates Conversational AI Chatbots Using High Quality Datasets

Appen’s custom datasets help Infobip to create conversational AI chatbots which improves satisfaction and lessens the cost of customer service

“In our mission to create world-class artificial intelligence chatbots at record speed, high-quality data sets are essential. Appen is a very important partner in this process because we can rely on them for exactly that: speedy and high-quality datasets that we use to train our AI engine and provide conversational experience. Their global coverage allows us to provide a premium service for our clients, all over the world, in any language.”

— Ante Stjepanovic, Product Marketing Manager at Infobip

The Company

Infobip is a cloud communications platform that specializes in creating tools for customer communications across a variety of channels, including SMS, email, voice, WhatsApp business, Messenger, and more. They enable businesses to have the most efficient and accessible communication with their customers.

One of Infobip’s most popular products, Answers, is an omnichannel chatbot building platform, which businesses can use in the cloud to create AI chatbots for meeting customers on their preferred channels, including WhatsApp, SMS, Facebook Messenger, Google Business Messages Creation of such chatbots on customer’s preferred channel helps Infobip’s clients in lowering customer service costs, providing customer service around the clock, and improving customer satisfaction.

Customers want to interact with businesses on the channel that gets them the fastest response and is most convenient for them. For many customers, this means using a chat app, such as WhatsApp or Messenger, to interact with businesses and find solutions to their problems. For most businesses, Answers acts as a first line of defense for solving customer problems. If the AI chatbot can’t help with the customer’s issue, then the customer is connected to a human agent, which is part of Infobip’s Conversations product.

Infobip’s goal for their customers is chatbot containment and customer satisfaction while reducing the need for human agents, which can be costly.

The Challenge

Some of Infobip’s clients use their help in building the best possible version of chatbots and to meet customer demands, Infobip needs a ton of data. The best data for training this type of machine learning model is crowdsourced data that’s got global coverage and a wide variety of intents. Infobip’s challenge with Answers was receiving quality datasets in a short time frame. They needed fast delivery of quality datasets and assurance the datasets had been properly validated for quality.

Infobip wanted to make sure that when they are building an AI chatbot, it’s trained with a variety of different customer intents, so no matter how the customer phrased a question, the chatbot would be able to understand the intent behind the message.

Infobip estimated that they need a large number of representative phrases per intent to make sure that the chatbot is properly trained on phrase variances. Each phrase would need to be unique enough to cover every potential phrase a customer might use. Infobip needed high-quality data quickly, without sacrificing accuracy.

Infobip has customers around the world who work in a variety of different industries. To get the vast range of data they need in a number of different languages and dialects, they needed a data partner with as global a reach as they have.

The Solution

Infobip reached out to a few different data providers, trying to find a suitable partner to meet their needs Appen provided them with better results and higher-quality datasets, so was chosen as their data partner.

To ensure success and make sure the chatbots were properly trained, Infobip and Appen started with just a few different languages, English, Spanish, and Hindi including some regional variations of those. For each language, Infobip requested hundreds of utterances that fit a particular intent. The main requirements were to:

  • Be in the target language
  • Be relevant to the provided intent (a 95% threshold)
  • Fulfill specific uniqueness criteria to avoid duplicates

To ensure high-quality data annotation, Appen:

  • Screened contributors for language proficiency
  • Collected10% more utterances than needed to allow for discards
  • Conducted a manual, human QA check of annotations from a random sample
  • Checked word count of utterances to ensure they were not too long or short
  • Used Machine Learning assisted smart validators to check for duplicate utterances and target language

If 95% relevance was achieved, the data passed the QA check and was sent to Infobip for use in training its AI chatbot model.

The Results

When Infobip was looking to prepare chatbots for their clients, they knew they needed a lot of data. For smaller projects, they had done data collection and annotation in-house, but with only one team member focused on data, it was a slow process.

By working with a data partner like Appen, Infobip has been able to reduce their time to deployment. They’re able to have more data and higher-quality datasets to train their model and deploy AI chatbots.

Infobip shares that another benefit of working with Appen is the Appen Managed Services Team. When a customer like Infobip works with Appen, they have a single point of contact that they can reach out to with any problems or to check on a project’s timeline, and they loved the timely responses they received, which is imperative when working with tight deadlines.

Website for deploying AI with world class training data