Great question — answering as the UphillTech team:
You described a JSON-like knowledge-graph based structured transformer. It is the neuro-symbolic hybrid approach for modeling language to deal with various types of downstream NLP tasks. The idea is roughly to convert a symbolic-parser parsed hierarchical concept graph into a set of fixed-dimension embeddings through techniques like learned aggregation functions, where nodes at each level of the hierarchy are aggregated via learned functions such as pooling, hyperbolic embedding, or attention. Indeed such hierarchical language models might be much better than the existing popular flattened sequential model in specific use cases and they're still in research stage due to theoretical...
Full answer by cinch: https://ai.stackexchange.com/a/47925 (CC BY-SA 4.0).
Need it done for your business? UphillTech.
