Automatic generation of speeches, FHNW School of Computer Science

Automatic generation of a speech for the opening of the Swiss Digital Days 2022

Natural Language Processing (NLP) methods have advanced enormously in recent years. Transformer-based language models have made it possible to take the context of each occurrence of a particular word into account. These models have significantly improved the quality of language-specific NLP applications such as translation, text summarization, and text generation.

As part of an experiment, the Federal Department of Foreign Affairs (FDFA) sought to investigate the possibilities and limitations of modern language models for the automatic generation of speeches and to test them using the opening speech delivered by President of the Swiss Confederation Ignazio Cassis at the Swiss Digital Days 2022 as an example.

For this project, the FHNW Institute for Data Science investigated state-of-the-art models, namely GPT-2 and GPT-3. An attempt to fine-tune GPT-2 using around 50 publicly available speeches by President of the Swiss Confederation Ignazio Cassis produced significantly poorer results than the experiments with GPT-3 Davinci.

The speech published here is the result of systematic experiments with GPT-3 Davinci. Two manual steps were required. First, the FHNW project team formulated instructions in the form of “prompts”, which produced longer passages of machine-generated text in German and English. Second, the Federal Department of Foreign Affairs (FDFA) selected suitable passages from several dozen speeches generated with GPT-3 Davinci, namely the “story of the farmer and the stranger” and the “goals of digitalisation”.