The holy grail of AI. Any thoughts on how this kind of integration can be achieved?

I’d suspect whoever made this slide must have some idea?

The slide was from a graph database company (AllegroGraph). It didn't go into much detail but they linked to this pre-LLM paper where two out of those three aspects were integrated: https://sci-hub.se/https://doi.org/10.1016/j.fmre.2021.08.013

I'm working on this stuff in my FoC project and in my day job. More on the knowledge extraction and symbolic tool use, than on having models work directly on the graphs. That seems a little... black boxy for my purposes. The basic idea is give natural language to the LLM, have it output a symbolic representation of what it learns, and have it integrate that symbolic representation with what it has already learned elsewhere. Then have a question-answering agent with the ability to browse & query the resulting knowledge graph, using a combination of symbolic reasoning, graph analysis, and semantic similarity to extract the most valuable context for the question. So far, we've done each of these pieces in isolation, trying to find a decent way to combine them and see if it improves performance on certain tasks.

There is a summer school later this week that I expect might answer some of these questions (?) (I know nothing about this field, just going by the name) https://neurosymbolic.github.io/nsss2024/schedule.html

I was hoping to attend virtually, but circumstances may not allow.

You might like: https://queue.acm.org/detail.cfm?id=3676287

I heard about ILASP at the summer school, yesterday. Very interested to see if an LLM could extract enough data from natural language cases to use ILASP to derive symbolic expressions of common law principles.

@Jason Morris ask the paper authors!