Utilizing Language Models to Expand Vision-Based Commonsense Knowledge Graphs

The introduction and ever-growing size of the transformer deep-learning architecture have had a tremendous impact not only in field natural language processing but also other fields. transformer-based models contributed to renewed interest commonsense knowledge due abilities deep learning models. Recent literature has focused on analyzing embedded within pre-trained parameters these embedding missing using graphs fine-tuning. We base our current work empirically proven understanding very large expand limited graph, initially generated visual data. few-shot-prompted can learn context an initial graph with less bias than fine-tuned corpus. It is shown that offer new concepts are added vision-based graph. This two-step approach vision mining model prompts results auto-generation well equipped physical commonsense, which human gained by interacting world. To prompt models, we adapted chain-of-thought method prompting. best knowledge, it novel contribution domain generation result five-fold cost reduction compared state-of-the-art. Another assigning fuzzy linguistic terms triples. process end graphs. means triples verbalized language, after being processed, converted back