Symbolic Regression

Recent advances in automatic measurement of semantic similarity, especially in the field of textual information, have been surprising. The truth is that the new solutions based on neural network architectures have achieved what seemed impossible until now: beating all the benchmarks datasets with which the research community worked until now.

However, this fact also presents a great disadvantage. Solutions based on neural networks act like black boxes, so no human operator can really understand how the information that generates the results is processed. This causes not a few problems in a wide range of academic disciplines and business models that need to understand why a specific result is offered.

For these reasons, we have been working on a software solution that is much more interpretable. Our solution uses the concept of symbolic regression to learn a model capable of mapping simple semantic similarity inputs into the desired result provided in the form of background truth. The results we have been able to obtain are quite good. And what is more important, very interpretable.