Text by: Vincenzo Scotti – POLIMI
In this blog post, we want to discuss an important issue of modern AI tools based on machine learning. These powerful techniques have become a standard for dealing with many complex problems. The key concept behind machine learning techniques is that the AI learns to do its task from examples, but how does it behave when there are too few examples to learn from?
We want to report our experience with this kind of issue and tell how we overcame this problem on voice analysis. In fact, we worked on extracting users’ emotions from their voices using machine learning tools and Natural Language Processing (NLP).
Usually, the examples to train machine learning tools for NLP are available only in English. At least the collections with a significant number of samples, which are crucial to learning correctly, are available only in English. In the case of WorkingAge, we had to work with Greek and Spanish, which do not have many collections of examples for emotion recognition.
Fortunately, machine learning offers approaches to deal also with these problems: there are AI algorithms based on machine learning that allow extracting new mathematical representations, even for the human voice, which makes more straightforward the learning step, consequently relaxing the requirements on the number of examples. Although it is not always possible to take advantage of these valuable representations, having access to such solutions made it easier for us to develop the voice analysis tool.
