Which Intelligence for Human-Machine Dialogue Systems?

Abstract

We present here two experiments carried out to improve the performances of an intelligent human-machine interface, called UKKO. The interface is destined to be implemented in humanoid social robots. From the standpoint of the automatic generation of outgoing messages, the way the interface works is not repetitive but creative. In other words, it does not reproduce talking points extracted from corpus but produces its own utterances by using rules. These rules exploit quality linguistic resources, i.e., formalized descriptions of lexicon that concern its morphological, syntactic and semantic features. The system’s quality of speaking depends on the completeness of descriptions inserted in its database. We discuss an experimentation that uses deep learning techniques to automatically retrieve syntactic and semantic properties of lexical units, which can be inserted into the database. An unsupervised method, based on the Word2Vec algorithm, is used for the semantic properties. The first results obtained are promising and show the interest of this approach for the experimentation. A supervised method, based on a sequential algorithm, is used for the syntactic properties. This method requires the transformation of the original data to match the training data that come from the UKKO system. The initial results achieved need to be improved. A line of research would be the addition of semantic descriptors to syntactic descriptors so as to make the approach achieve a better performance. This is consistent with the linguistic modeling underlying this project.