How Artificial Intelligence will transform Marketing Influence
After several months of implementation and adjustment, Data Science at Kolsquare already has its share of victories. After many iterations, our teams can now ensure that this technology accurately recognizes the languages of social network posts by more than 99%. This allows us to qualify very precisely the linguistic profile of influencers.
Beyond this qualitative victory, the Kolsquare platform is also capable of analysing more than 3 billion texts in a few hours, thanks to its efforts and perseverance.
The current focus is now on the automatic detection of influence themes. A complex and complete subject that only bodes well for the future of the Kolsquare platform, especially on extremely advanced identification features.
In accordance with the values we defend, we have chosen authenticity and the search for excellence. Marc Caillet, Team Lead Data Science, explains the challenges and successes of Data Science at Kolsquare.
Marc, can you remind us of your role at Kolsquare by Brand and Celebrities?
Marc :I arrived in Kolsquare a little over a year ago to take up a position as Team Lead Data. My first focus was on data collection. I then set up the Data Science activity to meet the specific needs of the Product team.
To form the starting team, I suggested to Koji Grandpierre, then an intern in the Community team, to accompany me in this beautiful adventure based only on his motivation. On that occasion, I became Team Lead Data Science. In parallel with these activities, I am also actively contributing to the complete overhaul of our Data Engineering architecture.
If you had to explain what Artificial Intelligence is in a few words, what would you say?
Marc : In a few words, I would say that Artificial Intelligence is a computer program that has the particularity of simulating human reasoning after having learned by itself the rules that allow it to perform the task for which it was designed. Be careful, I don’t say « the » human reasoning in its entirety.
To give you an example, the first conceived Artificial Intelligence recognizes the languages in which a text from a social network is written. During its development, it developed by itself the rules that allow it to distinguish one language from another, among 11 languages.
You submit a text to her, she tells you if it is written in French, English or Italian. It also handles multilingual texts perfectly: it can, for example, identify that a text is 60% written in French, 25% in Japanese and 15% in English.
Finally, she has the ability to answer that she does not know if the text presented to her is not written in any of the languages for which she was designed.
Can you tell us more about the use of Artificial Intelligence at Kolsquare? What were the objectives of this implementation?
Marc : Since its creation, the Data Science team has been at the service of the Product team. All its efforts are dedicated to improving the experience of Kolsquare platform users in their search for influencers that best match the message they want to communicate and the target they want to reach.
Through Artificial Intelligence, we determine what influencers talk about on social networks and how they do it. The work done on languages falls into the second category.
In parallel to the development of this Artificial Intelligence, we have designed two others.
The first one learned how to automatically classify influencers into influence themes. The second, which has been brought to a lower level of maturity to date, has learned to distinguish positive tone texts from negative tone texts.
For all these subjects, we have opted for a Deep Learning approach which consists in modelling Artificial Intelligence in the form of an artificial neural network. This approach is based on an analogy with the topology and functioning of a biological neural network.
How does the creation of Artificial Intelligence work?
Marc : In general, we rely on the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology, which we have developed somewhat.
After the « product need » has been clearly identified and after the development of Artificial Intelligence has appeared to be the most appropriate response to this need, we go through a phase of analysis of social network posts. This phase has a twofold objective.
The first is to determine which modalities we will use to develop this A.I. For influencers’ languages, the modality to use was very simple to determine: the texts of the posts. With regard to the classification of influencers by influence themes, we have opted for photos.
The second objective is to understand the diversity of data with which Artificial Intelligence will be confronted. Let me take the example of language recognition again. Texts on social networks cover a very wide spectrum: it can range from sustained to vulgar and from perfectly academic spelling to more personal spelling.
To be effective once it is implemented, Artificial Intelligence must have been trained to recognize the language of texts in all this diversity. We then build three sets of data, datasets, annotated. That is, for each of the texts in these sets, we indicate in which language it is written.
If I tell you that we have used several hundred thousand texts for each of the 11 languages recognized today by our Artificial Intelligence, it can give you, I think, a good idea of the scope of this essential task.
By the way, I would like to pay tribute to the titanic work accomplished by Koji who, among many other things, has consistently selected and annotated all these texts.
And then you also have to be creative. Our combined language skills have allowed us to easily select texts with all the desired diversity written in French, English and Japanese. However, we are, like me, completely ignorant of the eight other languages we have learned at our Artificial Intelligence.
We then completely revised our datasets strategy based on a key idea that allowed us to bypass our linguistic boundaries. However, this idea will remain a trade secret.
Once the datasets have been compiled and annotated, we use them to teach Artificial Intelligence to perform the task for which it is intended and to evaluate its performance.
Until we are satisfied with the performance, we adjust the A.I. parameters or modify our datasets, depending on the deficiencies revealed.
After many iterations, we have achieved an accurate prediction rate of over 99%. It’s really very high. This level of requirement is a characteristic feature of the Data Science team. It is at the heart of all the projects we carry out.
What were the benefits noted internally?
Marc : Data teams are composed of specialists in their field on the one hand, and talents with multidisciplinary skills on the other. This gives us an agility thanks to which we can easily modulate our organization in order to best meet the specificities of each project.
We work closely with Kolsquare’s most experienced specialists in product, community and marketing influence. Their enthusiasm is always palpable. Their contribution greatly contributes to the quality of the results we produce.
In addition, we regularly communicate our results internally. Each time, we generate enthusiasm, which is very satisfying. I was also able to perceive the birth of a sense of pride in each Kolsquare employee.
This is particularly striking in a context where a recent survey revealed that few companies actually use Artificial Intelligence, while many claim to do so. Not only do we use them, but we also create them ourselves, we can make them evolve at any time, we are not dependent on any third party!
What do you think of the results of this investigation?
Marc :Everyone chooses the strategy that best suits them to stand out from the competition. At Kolsquare, in accordance with the values we defend, we have chosen authenticity and the search for excellence.
Finally, where are you in your projects?
Marc : Artificial Intelligence for language recognition is fully operational on the Kolsquare platform: for several months now, our users have been able to filter influencers according to their preferred languages of expression on social networks.
To achieve this state, we had to upgrade our data engineering architecture to be able to process several billion texts in a few hours! It was a great challenge.
Artificial Intelligence for classifying influencers by influence themes has reached maturity. Its integration into the data processing architecture has begun.
We are also working on topics that do not require the development of Artificial Intelligence. We might talk about it in a future interview.
Would you like to know more about Kolsquare and its technologies? Do you want to set up a marketing influence campaign? Contact us here.