Detecção de intenção do usuário utilizando modelos de aprendizagem profunda com uso de hashing semântico

Abstract

A user intent recognition module can be considered the main compoenent of any conversation system. Intentions are purposes or goals expressed in a user input through an application, search engine, chat, etc. Several techniques are used and work satisfactorily according to the application scenario. Currently, machine learning techniques are considered state of the art for this type of task and have been applied in many works. Such models have high representational capacity and can easily learn the relationships between the terms of a training set. However, for some scenarios there are not a large number of samples and, as a result, proper learning by the method is compromised. Another important point concerns the disposition of terms within a sentence, a difficulty that many researches do not take into account. This paper implements a user intent recognition module based on hybrid deep learning models, taking into account the disposition of the terms of a sentence generating an additional value (semantic hash) and text embedding algorithms. We propose using this extra value only for the words considered most important within a sentence creating a more targeted representation. As a result, the developed model reached an average accuracy of 93.95%, exceeding by more than 2 percentage points the other works evaluated showing the possibility of gain with the use of additional values referring to some og the sentence.