In the current upsurge of Large Language Models and their application to data2text generation, controllability is an essential topic in the light of hallucinations and toxicity of language. Controllable generation of evaluative markers is a challenge in text generation in general, mainly because the motiviation for an evaluative tone must be grounded in the data.
Empirical studies have been conducted on how evaluative adverbs influence the reader's expectations on certain car features in road test reports. Additionally, we show how to use regression models for approximating the features of cars and delineate boolean borders where data qualifies the generation of evaluative language. We also show how well the reader's expectations in comprehension agree with the input data and therefore validate the use of regression as approximative decision boundaries.
Maurice Langner and Ralf Klabunde (2023) Validating Predictive Models Of Evaluative Language For Controllable Data2Text Generation. In: Proceedings of INLG 2023.
Maurice Langner and Ralf Klabunde (2022) Realizing a Denial of Expectation in Pipelined Neural Data-To-Text Generation. In: Proceedings of the 6th Workshop on Advances in Argumentation in Artificial Intelligence. Udine, Italy, November 2022.