AI : How to make virtual assistants intelligent

"It is generally said that Artificial Intelligence or "AI" starts where the computer program stops", explains Guillaume Mallen, Artificial Intelligence Developer.

The abundance of data is leading companies to integrate Artificial Intelligence (AI) into their system/business model for better performance, innovation and competitiveness. In common parlance, AI is often perceived by the general public as a way to replace humans. But what is it really?

Here is Guillaume's view on this complex notion:

Guillaume, can you introduce yourself in a few words?

I work at Ubilink, in the Research & Development department. More specifically, I am in charge of work on Artificial Intelligence.

As a novice in this field, can you explain to me what Artificial Intelligence is?

It is generally said that Artificial Intelligence or "AI", starts where the computer program stops. AI includes different technological concepts such as machine learning, deep learning, ...

At Ubilink, we are more interested in deep learning, which is a system based on a neural network. In other words, it is a computer system that evolves over time thanks to its learning capabilities. AI is then designed to reason like humans, which gives a machine the ability to perform functions that would normally be associated with human intelligence.

How does it work and what kind of technologies do you use?

Artificial Intelligence is based on two main functionalities which are training and prediction.

At Ubilink, we train our Artificial Intelligence every day. This step consists of creating a model that recognises a certain type of data and makes good predictions.

To do this, we use the Python programming language, which is the most popular language in the field of Artificial Intelligence. We will also use the Tensorflow library. This is a machine learning tool developed by Google, one of the most widely used in AI. As well as Keras, which is a high-level API of Tensorflow allowing the creation and training of Artificial Intelligence models.

How is AI implemented in iBIM?

It is important to know that in the solution we propose, iBIM, the aim is to understand and answer the questions asked by a user. For this, we use Natural Language Processing (NLP) principles. NLP is a technology frequently used in the field of AI. It allows machines to better understand human language. It consists of analysing textual information and converting it into an object that the computer can understand.

After processing the data, we will be able to train our model. The neural network used is a bidirectional LSTM (Long Short Term Memory). It is a recurrent neural network (RNN).

The RNNs have the particularity of having an internal memory that allows them to keep the data they receive in memory. When they make a prediction, they therefore take into account the current input, but also what they have learned from previous inputs. This feature makes them a good choice for textual data, which is exactly what Ubilink needs.

What role does AI play in your solution?

We will find AI in iBIM through intelligent avatars, with their own knowledge base. For example, in a hotel, we could have an avatar like the SPA manager who focuses on information about the SPA, and another avatar, the waiter, who has information about his restaurant. This knowledge base is made up of intentions that reflect the "demand" or "will" of the user. The goal of our AI will be to recognise the user's intention through his request in order to return the appropriate answer.

We will also use AI in our product to guide, orient, animate, and even train users in the virtual world of iBIM.

Are there any limitations to using this technology?

Artificial Intelligence does have its limitations, and we have had to develop algorithms to overcome them.

First of all, during the training phase, it is common for a model to lose quality when new data is added. This is a phenomenon that really must be avoided. To be sure to obtain the best performing model possible, we constantly test its performance using the cross-validation technique.

Secondly, it is possible that when a user types in a query, he or she may make a spelling or grammar mistake. To deal with this, we will apply a stemmer to each word of the question asked in order to keep only its radical.

As a further improvement, we have given our avatars the ability to doubt the user's intention when their query is not precise enough. To do this, we define intentions with two keywords, a verb and a complement, thus making it possible to doubt on predictions belonging to the same keyword family.

We then provided context to influence the avatar's responses based on the prediction that was made earlier. For example, if you ask the avatar: "where is the restaurant?" and then ask: "what are the opening hours?", the avatar will suspect that you are talking about the opening hours of the restaurant.

Finally, each avatar has its own area of competence. It was therefore necessary to deal with the case where the user's query was outside the avatar's domain of competence. For this, we use the cosine similarity technique which looks at the similarity between the question asked and the questions known in the predicted intent.

However, this out-of-domain question may be known by another avatar in iBIM and in this case we will offer the client to redirect to that avatar.