How Can You Reduce Your Bot’s Training Time?

When it comes to technologies using artificial intelligence, time is an essential aspect of their algorithm performance. Indeed, the concept  behind these solutions is that they are self-learning, requiring more or less assistance. Getting a conversational robot (chatbot) to perform and train quickly involves several elements.

The Chatbot’s Foundation: NLP Performance 

The natural language processing engine, or NLP, is based on, and integrates, several elements that are more or less complex. The most basic engines only focus on one word. This keyword recognition system is very similar to a search engine.

At dydu, we’ve been developing and optimising our algorithm for over 10 years. When a user asks a question, our engine analyses all the words in the sentence. It corrects spelling mistakes and identifies the root form of a word, as well as its synonyms. This enables the chatbot to understand lots of different ways of expressing the same intent (or question).

The Chatbot’s Framework: Supervised Learning

Most clients create a traditional chatbot from scratch, building a knowledge base according to their business and use cases. A client setting up a chatbot to handle rail transportation issues won’t have the same needs as one looking to provide IT support for employees. Knowledge bases, i.e. all the questions and answers, are specific to different audiences and scenarios, as is the vocabulary. Once the questions and answers have been defined, time needs to be spent on configuring and training the chatbot. This involves thinking about different ways to formulate the same question. And also testing and finetuning the bot repeatedly to optimise its understanding.

Given that conversational robots speak on behalf of brands to potential and existing customers or employees, we decided, at dydu, to opt for a supervised approach. Either way, humans have the last word and decide, not the robot.

Specific tools in the chatbot’s admin console make this calibration easier and quicker. The engine continually suggests matching optimisations and improvements based on its knowledge base:

  • Firstly, by sending alerts in the event of close knowledge articles.
  • Then by highlighting misunderstood sentences and, most importantly, suggesting alternatives (to be confirmed or rejected). The bot manager can add the reworded knowledge articles or complete the matching group in a click.
  • And finally, by performing quality alerts and reporting any answers with errors, such as broken links.

These preventive and auditing features are therefore key to your bot’s ability to be trained effectively.

As far as the process is concerned, we recommend using samples of real testers from the start with different profiles.

Ready-to-Use: Pre-Trained Knowledge Bases

With our Expert chatbots, all of this calibration work, which can take time at the beginning, has already been completed.

We’ve already defined the themes, questions and matching answers. We’ve also grouped the sentences that mean the same thing together. For example, this is what enables our Council chatbot to understand the following terms equally well: canteen, school lunches, cafeteria

For our HR chatbot, dydu has created and tested 110 matching groups associated to 300 knowledge articles.

Creating a bot from an existing model therefore allows you to benefit from all the work that has already been done. This will help your teams quickly adopt the content, by only focusing on the customisation elements specific to your company. Your users will immediately benefit from an excellent experience.  And it will be even easier for you to continue optimising your chatbot in the future!