Dealing with hallucination and omission in neural Natural Language Generation: A use case on meteorology

Abstract

Hallucinations and omissions need to be carefully handled when using neural models for performing Natural Language Generation tasks. In the particular case of data to text applications, neural models are usually trained on large-scale datasets and sometimes generate text including divergences with respect to the input data. In this paper, we show the impact of the lack of domain knowledge in the generation of texts containing input-output divergences through a use case on meteorology. To analyze these phenomena we adapt a Transformer-based model to our specific domain, i.e., meteorology, and train it with a new dataset and corpus curated by meteorologists. Then, we perform a divergences’ detection step with a simple detector in order to identify the clearest divergences, especially those involving hallucinations. Finally, these hallucinations are analyzed by an expert in the meteorology domain, with the aim of classifying them by severity, taking into account the domain knowledge.