Radziwill, Emelie; Ockert, Max; Leubner, Eric. Customisable Voice Interaction: Feature Model Configuration by Means of Deep Learning. Research Project and Complex Practical Task. 2023.
Abstract and Task Description
The project goal is the development of an adaptable concept for speech interaction for the target group of older people, based on the configuration of feature models and configuration selection through a recommender system, with the following sub-goals:
Familiarization with the existing project results:
The results of this project should make speech interaction adaptable, so the architecture of a speech assistant is required. For this purpose, the Mycroft implementation (provided) may be used, or Amazon’s Alexa (possibly in combination with RASA). At the same time, adaptability should be enabled through a configuration of feature models, which will be selected by a recommender system. The results of the project on the deep learning approach (also provided) should be used for this purpose.
Development of a feature model to map language variations for a skill:
With knowledge of the required input feature model, a feature model should be designed that can represent variable speech interaction suitable for automated configuration. The development of the feature model must be done in a suitable scientific manner, primarily based on scientific literature work. This part should be presented in a comprehensible manner in the project documentation. The result can be two models: a user model for describing users and a variability model for describing speech interaction.
Development of suitable training data:
Initial training data for the recommender system is needed to solve the so-called cold-start problem. There are different permissible approaches for this. Ideally, such an approach results in collecting as many plausible data as possible. For example, suitable personas can be researched or created, from whose perspective several suitable configurations can be designed. The personas themselves can be described using the user model (so-called user profile).
Exemplary application of the data using the recommender and a speech assistant:
The existing project results serve as a basis to test and demonstrate the designed model and prepared training data. For this purpose, a skill for the speech assistant must be developed that will use a proposed configuration in interaction with the user. To make a suggestion suitable for the user, the system must support the consideration of user profiles.