Stefan TrockelonFeb 23, 2021 makes training AI chatbots easy

We expect chatbots and voice assistants such as Google Now, Alexa and Amazon Echo to understand us correctly in all situations, and to recognize our intentions and needs. These assistants are based on Natural Language Understanding methods that extract meaning from our input in order to respond accordingly. For example, in response to an input "I would like to buy a Spanish wine", we expect suggestions from the bot that further narrow down the options: "Would you prefer red wine, rosé or white wine?". To the question "When will my delivery arrive?", the bot ideally responds with details about the delivery date: "The package should be there on Monday. The tracking number is JJD1410302937991."

Chatbots are designed to recognize a number of relevant user intentions, called intents. The intents in our examples above would be "search for product" or "track shipment." These intents can be expressed in very different ways, such as "When will my package arrive?", "When will my shipment be delivered?" and "Will my shipment arrive on time?". It is important that chatbots are robust in the sense that they can recognize these different expressions and match them to the correct intent. To achieve this robustness, bots need to be trained using machine learning techniques. This is done by giving the bot examples of utterances with the corresponding intent as learning input; usually 8 to 10 examples suffice. The bot learns to understand other similar utterances based on these examples. The understanding is then not only limited to the given examples; the bot recognizes regularities and recurring patterns and can transfer these to completely new utterances. For example, the bot then understands "When will my shipment arrive?", even though this sentence may not have been included in the training, based on a generalization of the sentences "When will my package arrive?" and "When will my shipment be delivered?". The platform incorporates several machine learning approaches, allowing it to automatically adapt to the scope of the existing training sentences. All approaches are pre-configured to achieve optimal learning success without the need to configure parameters yourself.

In addition, offers pre-configured e-commerce bots that can be used directly to support marketing automation, sales and after-sales communication automation, on WhatsApp, Facebook Messenger, and other channels. These bots are pre-trained; this reduces the effort required to release a first bot to practically zero. They support product searches, inquiries about the status of an order, returns processing, and a whole host of FAQs.

However, no bot that goes live can already perfectly understand everything. For e-commerce store providers who want to reduce response time and thus increase customer satisfaction, it is essential that their bots have a high level of robustness, so that customers do not bail out in frustration. High robustness can be achieved by retraining a bot over and over again in live operation, so that it improves over time. 

The trick is to optimize the process of training a bot in live mode in such a way that examples are provided where they are most helpful, that is, where the learning effect is highest. In this way, resources are used optimally and no time is wasted in providing examples that the bot already understands. The key technology for such efficient training is called "active learning". Active learning is implemented by default in all bots. The bots are able to estimate the degree of their own understanding by calculating a confidence degree. If this confidence degree is too low, the example is transferred to the "homework list" of the bot project in the platform. In the next training interaction, the bot can proactively ask a human for the correct interpretation. This leads to effective training and works completely without technical knowledge. The bots explain themselves what they did not understand and you can continuously improve the bots with a few clicks.

Thus, through live training, high-quality bots can be created that have a high recognition rate in understanding user utterances. With the platform, the quality of the interaction can be monitored at any time. This ensures that customers do not abandon the interaction prematurely, increases conversion rates and leads to more satisfied customers whose concerns can be addressed quickly and competently. For example, less than a second after expressing a desire to return an item, the customer could have the return label sent to them by the bot, quickly and conveniently. As a result, everyone wins: e-commerce store operators can run their store more efficiently and cost-effectively, and customers feel that they are being served quickly and around the clock.

Stefan Trockel