Chatbot Training

How to train chatbots

By ian tomlin | Jun 09, 2021

How to train chatbots Cover Image

The promise of bots

Customer Service Experience is the most frequently targeted area for digital transformations in the corporate world today. This is understandable. In a world that has come closer thanks to mobile telecommunications and the always-on web, buyers want to serve themselves with products, services and knowledge at any time of day. To achieve that level of service delivery capacity, firms are turning to bots.

No question, your buying audience is ready to use ‘good bots.’ According to a report by MINDSHARE titled ‘Humanity in the machine,’ even when it comes to muddling through with the poor quality bots of today’s generation, 63% of people would consider messaging an online chatbot to communicate with a business or brand.

To deliver on their promises, bots need to work. In the same way that customer service agents can only really do a great job when they know products and services inside and out–and more importantly, know the industry and the nature of enquiries customers might have–that means they need to be well trained.

The essentials of bot training

Traditional bot training methods have required IT programmers to trudge through thousands of rows of conversation data and manually pick up interpretive errors–and it’s been a trial and error process, offering no guarantee of success.

Futuristic robot assistant with artificial intelligence in public place

Modern AI-driven alternative training solutions, combine a knowledge of linguistics (behind the scenes this work has been about Natural Language Processing backed up by linguistic research methodologies) with the ability of machines to progressively learn by themselves and recommend better outcomes.

There are two main problems to fix:

1. False positives

Most training agendas focus on ‘false positives.’ These are occasions where the bot misinterprets a request and suggests the wrong outcomes. The obvious consequence of operating bots that serve up false positives is that customer service drops, transfers to human call centre agents grow, and outcome results flat line. Fixing false positives for chat and voice bots has become the bain of IT teams. Whilst everyone knows what promises bots offer in terms of reducing back-office costs, improving customer service quality and availability, and generating more new business opportunities, achieving this holy grail outcome has been the torment of IT leaders since bots came on the scene.

2. Missed Intents

Whilst there are a number of training solutions to overcome false positives, some of which do clever things with artificial intelligence (like INTNT.ENGINE) to speed the process of correcting answer data banks, overcoming missed intents has become the holy grail of the bot training industry. The challenge with overcoming missed intents is the fact that bots ‘don’t know what they don’t know.’

There are so many potential requests that can be made to bots, and a near-infinite number of ways for humans to raise their questions. Training bots to seemingly know the answer to every question inevitably becomes a task for artificial intelligence-based systems that can learn on the job–because the notion of paying human programmers to code them soon becomes unrealistic.

Even with AI, to resolve the challenge of missed intents, bots need to be great conversationalists and grasp not just ‘what is said’ but what the human is hoping to learn. This is why expert suppliers like INTNT.AI have built tech platforms whose core intellectual property comes from Natural Language Processing research.

Robot's arm working with Virtual Reality touchscreen

Using patent-pending NLP research understanding INTNT.ENGINE is able to understand aspects of conversation such as intent, anger, frustration, humour and annoyance. Blending these characteristics of human conversation with ‘the words spoken’ helps INTNT.ENGINE to predict more accurately the response the enquirer is looking for.

The learning process with INTNT.AI

Adopting a training regime supported by INTNT.ENGINE is relatively painless. It goes something like this:

1. The IT team produces a chatlog, strips it of any personal data, and uploads it to the INTNT.ENGINE ‘inbox’ on the cloud. The dataset includes the question posed, the answer given, and the ‘intent code’ used to identify the conversation thread in the answer knowledge bank. As this data is only a simple text string, the file isn’t very large, so uploads are easy to post.

2. INTNT.ENGINE processes the file and steps through each of the conversation threads, making recommendations on smarter answers that yield more positive results. Any suggested changes are made to bot managers in the form of recommendations.

3. Recommendations are reviewed by the bot manager. These can be accepted with a simple mouse click, and this updates the answer knowledge bank.

Guaranteed, predictable results

Businesses today have the opportunity to transform their online presence and offer 24/7 support to customers via all form-factor smart devices by transforming their chatbot army into an effective vehicle for above and beyond customer experience.

According to a 2020 study by DRIFT, the usage of chatbots as a brand communication channel has seen a 92% increase since 2019

Training bots manually through trial and error efforts to guess better answers is not only soul-destroying work for IT programmers, it’s costly and time-consuming. One of the great things about adopting a training regime that’s based on a proven NLP formula is that results are guaranteed to improve over time.

As Andy Leong, Senior Product Manager for FWD Insurance puts it, ‘INTNT-ENGINE has turned our bot around. Our escalation to live-chat has dropped from 21% to 6% while customer satisfaction has increased from 1 to 4 out of 5 star.’

Live case stories of deployments in the financial services, pharmaceuticals and manufacturing industries show that using AI-driven bot training with INTNT.ENGINE cuts call transfers with 70% of calls deflected, boosts experiences, with an 85% reduction in missed intents and grows sales results, achieving an 80% growth in outcomes.