Semantic analysis of answers written in natural language : talk4’s difference
Talk4 platform integrates a COBOT – COllaborative roBOT -, which is under the control of a human: its supervisor, analyzes and sorts out, in a super-fast way, the answers answered by the participants to the open questions asked in a talk.
This function allows you to output in a few minutes, to all participants, their answers sorted by coherent ideas, to help them still thinking together.
How does it work and why this choice ?
The COBOT talk4 is a support tool for verbatim analysis, which works in collusion with its ‘human’: the supervisor, from whom it learns how to classify the answers, in the specific context of the discussion in progress.
The COBOT, which has been trained to understand the language used, learns during the classification plan, the logic that its supervisor uses to sort the first verbatim answered by the participants to the open questions asked.
Quickly, the COBOT proposes to the supervisor only to intervene to sort out the answers that call for different ideas, and thus probably require the creation of a new subgroup.
Why this choice ?
Because at talk4, we know that there is no single way to sort answers to an open-ended question because there is no consensus on which we might want to stretch.
The level of knowledge of the environment, or the particular view of the person who reads them, related to his / her role or experience, will greatly influence the reading grid of the exchanges.
It is the learning of this particular grid of reading, which allows to output a relevant classification in the supervisor’s point of view, and which will allow to capture more nuances than what a generic algorithm could do.
talk4’s point of view to assume this observation and to take advantage of it to go further in the design of COBOT talk4 by integrating a Machine Learning which learns the particular reading grid of the pilot.
The machine learning integrated in COBOT talk4 has the particularity to learn in 2 stages: upstream it was trained to understand the language used by the participants, it continues to learn then, in supervised mode, how to classify the supervisor.
This approach could be defined as « dynamic machine learning », as the learning and adjustment of the hyper-parameters of the model is done simultaneously with the arrival of new data.
The other particularity of COBOT talk4 is that it works in a super-fast way, to allow a restitution in a few minutes of answers classified by ideas, to feed the dynamics of the group of participants and its reflection.
To facilitate this complex process, it is desirable to ask participants to produce one idea per contribution, the number of contributions per participant being obviously unlimited.
From the groupings obtained, the user sees emerging trends as well as weak signals (the few contributions that are attached to any group).