Optimization of Few-Shot Learning NER Models Through Grammatical Conditioning of Training Data

Nowadays, the increased demand for information is constantly posing new challenges for the technical possibilities of AI. Information that has to be collected in the form of digital data and evaluated by AI is increasingly coming from topic-specific peripheral areas with a high degree of formalization. The evaluation of highly formalized information, such as that which can result from textual image analysis, can cause AI models great difficulties, especially in the field of natural language processing (NLP). Texts with a high degree of formalization usually only have a low grammatical richness and are therefore only suitable for training NLP models to a limited extent, as the recognition performance is also impaired by the sentence structures to be examined. In addition, highly formalized texts can also have a low variance in content, so that only a small amount of training data is available. In the following, an approach for optimizing NER models generated from highly formalized training data is presented. The aim is to investigate how raw training data can contribute to improving the recognition performance of NER models with the help of grammatical conditioning.