Poster: Visual-Explainable AI: The Use Case of Language Models

We address the research problem of explainable AI (Artificial Intelligence) in the context of language models using visualization and visual analytics. Our goal is to open the black box of language models to make them more transparent and interpretable for humans, find problems (debugging), and improve accuracy. Different strategies can be chosen to analyze deep learning models. One approach is to explore a model's internal states while performing a prediction. This helps see how they change and contribute to the final prediction. Another method is to explore the actual prediction results that can provide insight into the prediction and provide a quality assessment about the performance of the model. Other approaches, such as exploring the training process, are also possible but are not considered in this work. These strategies can be applied to a variety of data types. Here, we chose the use case of language models. In the area of natural language processing (NLP), AI techniques can create impressive results, but results may still be incorrect. Explainable AI can help understand why certain predictions were made. We investigate NLP as part of data-integrated simulation science: in the form of supporting cognitive aspects in the creation of a digital human model, which is one of the visionary examples of the Cluster of Excellence "SimTech."

PDF Abstract - PDF Poster - Videos/Bilder (DaRUS)

@inproceedings{languagePoster2023,
  author = {Munz-Körner, Tanja and Künzel, Sebastian and Weiskopf, Daniel},
  booktitle = {International Conference on Data-Integrated Simulation Science (SimTech2023)},
  title = {Poster: Visual-explainable AI: The use case of language models},
  year = {2023},
  url = {https://www.simtech2023.uni-stuttgart.de/documents/Theme-2/Munz-Koerner-Tanja.pdf},
}

Download BibTeX