Semantic Similarity

The field of semantic similarity measurement has become very popular in recent years. In part, this is due to the great advances made in relation to word embeddings such as word2vec. That is, techniques that use deep neural networks to convert what used to be text into a set of numerical features that can be used as input into neural networks and other soft computings techniques.

However, and despite that with word embeddings, the problem of accuracy has been solved. But there is still another problem, that of interpretability. That is, through the use of neural computing solutions it is possible to achieve good results, but the cost is that models are obtained that are not easily interpreted by human operators.

For this and other reasons, we have proposed the so-called semantic similarity controllers that are artifacts based on fuzzy logics that allow us to learn knowledge models that can be put into exploitation with certain guarantees of performance and yet are easily understood by the human operators who make use of them. We hope that in the next few years, more such solutions will appear and be gradually used in both academia and industry.