ChatGPT has had a lot of buzz around it recently, and for good reason. It has shown some amazing capabilities in responding to new information, as well as in generalizing new information that has been given to it.
However, ChatGPT also regularly makes logical and mathematical errors. Some have commented that it really tries its best based on the data that was given to it to "sound correct" (almost like a freshman "bullshitting" a term paper), but it not actually great in general at making logical inferences.
One of the ways this issue could be solved is, rather than using a pure large language model (as essentially I understand ChatGPT is doing), we can instead use a semantic parser to parse natural language sentences into a logical form, such as first-order logic, higher-order logic, various forms of modal logic -- or what have you. This approach can also be supplemented with neural methods -- for instance, either by directly training a semantic parser as a sequence-to-sequence problem, or by learning "weights" for the likelihood of different parses.
I know that before the "first AI winter", the more symbolic/logical approach dominated, and a lot of work was done in Prolog, so I could probably find some examples of what I am looking for by looking at said historical approaches.
However, what I am looking for is a state-of-the-art chat bot based at least partially on symbolic techniques (specifically, using semantic parsing + a logical inference engine).
Does such a thing exist? Has much progress been made on that front since the first AI winter? I'd especially be interested in approach that combine neural/deep learning with symbolic approaches in some way.
References would be appreciated, but I'm mainly looking for concrete open-source implementation so I can play around for myself with the "state-of-the-art" and see what the current best offerings are, and their limitations.