We present KnowledgeVIS, a visual analytics tool in your browser for exploring and browsing relationships that language models (LMs) have learned by identifying, comparing, and summarizing LM predictions. LMs are seeing increased use in everyday artificial intelligence (AI) applications; e.g., in classrooms, LMs are learning to read students' online forum posts and predict potential connections between students, towards fostering collaboration and fighting social isolation. LMs can answer questions about facts and concepts they have learned (e.g., a student's interests) by predicting missing words from sentences (i.e., fill-in-the-blank). To visualize this information, we present interactive views that identify the strength and uniqueness of predictions, compare sets of predictions between entities, and summarize patterns in predictions across multiple entities. Understanding what LMs predict could reveal harmful biases or stereotypes learned that can be used to improve the model's performance, as well as foster trust in everyday users of AI technologies powered by LMs.
KnowledgeVIS: Visualizing What Language Models Have Learned
Our goal is to help people make sense of data. We research and develop interactive visualizations that couple machine learning with visual interfaces of data for exploration and sensemaking.