[D] Machine Learning - WAYR (What Are You Reading) - Week 98

It'd be an overestimation from my side, but their architecture is kinda ingenious and we are trying to do it in our startup, using LM like BERT with position features to embed fields in the documents... It's more like semantically searching a document graph for a query label to detect... Off-the-shelf LMs don't have any mechanism to provide geometry information (that coming from my experience from using sentence-transformer). But, I think it can be done. I am moving in some direction at least. PS: Having tried GNN experiments, they didn't do well either for a highly imbalanced datasets where background nodes overshadow all the major labels.