Generic semantics-based task-oriented dialogue system framework for human-machine interaction in industrial scenarios

In Industry 5.0, in which automatization has an increasingly important role, human workers and their well-being are placed at the centre of the production process. In this context, task-oriented dialogue systems allow workers to delegate simple tasks to industrial assets while working on other, more complex ones. Also, the possibility of naturally interacting with these systems reduces the cognitive demand to use them and triggers acceptation.

However, most current solutions do not allow a natural communication, and modern techniques to obtain such systems require large amounts of data to be trained, which is scarce in these scenarios. This causes industrial task-oriented systems to be highly specific, which limits their capacity to be modified or reused in other use cases, which is bound to high development time and costs.

To overcome these challenges, this thesis leverages Semantic Web Technologies and NLP techniques to develop KIDE4I, a semantics-based task-oriented dialogue system for industrial scenarios that allows a natural communication between human workers and industrial systems. The modules in KIDE4I have been designed from a generic perspective, with the objective of simplifying the process of adaptation to new use cases. In this line, different methodologies and resources have been proposed to promote reuse and encourage the continuous improvement of the dialogue system through new interactions.

Among these resources, the TODO (Task-Oriented Dialogue management Ontology) is the core of KIDE4I. This modular ontology is in charge not only of modelling the domain, but also the dialogue process and the storage of traces. It has been developed by following a well-known methodology, LOT, which is reflected in its high quality.

KIDE4I has been implemented and adapted for four industrial use cases, proving that the adaptation process is not complex and it benefits from the reuse of resources. Three of these have been evaluated through user studies, and the results obtained are reported.