We all have gone through it, excited about what we have read and heard on Artificial Intelligence, we have the opportunity to start a conversation with a chatbot that promises to solve our problems quickly, without waiting or endlessly pressing numbers to find the right combination for a customer rep to pick our call. Since it is a company that we know and trust we dive into it with enthusiasm.
And then… not much happens, the chatbot is not as half as intelligent as we expected it to be. After various attempts explaining in different ways what we want to accomplish and a similar number of disappointing responses, we are frustrated and start to strategise on the most efficient way to connect to a human. We can’t blame companies for deploying technology aimed at improving customer service, however, few thought about managing expectations.
Imagine that you need to feed into the AI model a large number of possible ways to ask the same question so that it can identify what the user means. We, humans, are creative species and find innumerable ways to refer to the same idea. Then, what would happen if somebody, using similar words, asks something different? You would then need to train the model again to understand the subtle but relevant difference. And then, what would the model understand the next time a user asks a question using similar words but different structure? Will it give more weight to the first version or the new one?
Training the model and correcting it for new sentences becomes an endless and expensive journey. Costly in terms of engineer hours but more expensive in terms of customer experience and management attention. Some chatbots end up being moved to permanent holidays.
The difference between traditional supervised learning and the more recent pre-trained models for conversational AI
Recent developments in Natural Language Processing bring major advantages to conversational AI by providing context. Rather than identifying words in sentences and comparing them to a long list of fed in examples in an attempt to find similarities, pre-trained models analyse the relevance of each word within a sentence and its relationship to other words, extracting meaning and context from words within sentences and sentences within paragraphs. This, thanks to the use of transformers, deep learning models that apply attention mechanisms weighing the influence of different parts of the input data. This is where things get technical, let’s keep it practical.
Because these models have been trained with billions of data, they have already developed capabilities to understand various angles of a conversation such as hierarchical relations (the relevance one word has versus others), long-term dependency (relations between information presented earlier in the conversation and more recent data) and sentiment.
In practice, this means that pre-trained models are much easier to train for specific applications as they come equipped with a set of core capabilities to interpret language. They still need application specific training to distinguish processes but this is a fraction of the training that traditional supervised methodologies require.
Additional advantages of pre-trained models is that they exhibit a faster response rate and lower operational cost as they require less computational power. The algorithms have been designed in a way that calculations are done in parallel instead of in sequence and this has a relevant impact on performance.
How to train your model?
Despite the significant progress in technology, you still need to train the conversational AI model, just like new human resources need job training. The difference between the old and the more recent, pre-trained models is comparable to hiring toddlers and teaching them the elementary rules of language versus hiring grown ups who already understand and speak a language. Regardless of age, you will need to give training to human employees so that they learn the various use cases of your business and know how to respond accordingly. You also want them to adhere and uphold your company values and brand image. Logically, the same applies to pre-trained conversational AI models, they require training on the specifics of the job to be done. They will just not need hundreds of examples to understand how to respond to your customer questions.
Let’s illustrate this with an example. One training sentence: “I don’t understand my invoice” connected to the answer “Not a problem, I would be happy to help you, which invoice do you have a question about?” is enough for the conversational AI to understand all variations of the customer intent without the need to feed numerous samples to the model. In contrast, the model is also capable of understanding when there is no problem like in the case “I understand my invoice”, or when there is a special request attached to it like “…Can I ask something about that?” or “What is your phone number …”. In this case the response will be tailored.
By drastically improving the level of comprehension, pre-trained AI models are set to become a major interaction point with your customers. In no time they will be outstanding company representatives, able to go beyond customer service, becoming a competitive advantage and an opportunity to increase your sales.
If you are not ready to undergo the evolution into more powerful technologies, our advice is not to throw your chatbot away but to ensure that your customers understand the limitations of the help it can deliver. Traditional, supervised chatbot techniques work well to answer frequently asked and especially distinctive questions but managing expectations, as always, remains key. If, on the other side, your goal is that the nice icon with a catchy name floating on your website makes a difference in your customer service, you ought to consider powering it with the adequate tools.
