Ruprecht-Karls-Universität Heidelberg

A multi-strategy approach to natural language generation for dialogue systems

Bachelor Thesis by Gavin Gabriel


Dialogue systems have been around for some time now, but only recently have they reached a standard where they are beginning to become a regular part of the digital landscape in form of virtual assistants shipped with most smartphones nowadays, smart home systems and customer service chatbots offered and employed by a growing number of companies. Advances in automatic speech recognition, natural language processing and machine learning techniques are to credit for this resurgence.

Nonetheless, conversations with these systems are still far from natural because system responses often lack variation and are generally inflexible. We believe one reason for this is that, while big advances were made in understanding natural language, generating appropriate responses still relies on inflexible database lookups or non-transparent machine learning approaches. This thesis thus follows the idea that directing more attention towards the natural language generation (NLG) component of a dialogue system could work towards achieving more natural human-machine dialogue. Following insights from research on human dialogue, dialogue system architecture and NLG, we propose a concept for NLG in a dialogue system context, geared towards contextualisation and personalization of system responses. Our prototype implementing this concept shows promising results and we believe it to be expandable into a fully functional system. Beyond that, we gained valuable insights for the design of new dialogue system architectures, equipped for a pervasive computing shaped future and closer to fulfilling the expectations of human dialogue partners.


« back

back to top