What is the status, endeavours and prospects for the use of universal/commonsense knowledge bases (KB) for the representation of the meaning of the natural languages in natural language processing / computational linguistics?
Currently distributional (statistical) semantics is very popular in natural language understanding but I think that KB are more rigorous and promising approach, because of:
- KB usually distinguish among concepts (or abstraction entities, classes) and objects (instances of classes). Therefore real world entity can be recognized and tied to the texts (e.g. city Paris, President Obama).
- KB usually contains elaborate definitions of the concepts and instances, therefore it is more easily to reason about concepts and generate new concepts - definitions can be used in this process.
There already are more or less elaborated and formalized KBs like Cyc (OpenCyc), PrimeCog (OpenCog), ConceptNet and WordNet and they should be appropriate for the use in computational linguistics, aren't they?
I guess, the only obstacle to such development is the (time) complexity (and undecidability in some cases) of the reasoning/inference procedures over such KBs, is it so? But this can be solved by adopting approximate reasoning procedures that use limited inference depth (like humans do). What else prohibits adoption of KB and logical approach for NLP/computational linguistics?