Given the following linguistics (and some non-linguistic) constructions:

entailment, implicature, Strawson entailment, semantic underspecification, discourse representation theory, rhetorical structure theory and description logic.

My question is: Is there a system trying to integrate all the above cited constructions and designing a powerful Machine Comprehesion algorithm?

I've googled for a kind of linguistic "theory of everything" enabling a machine to read like a man but I haven't find an answer.

2 Answers 2


Adam is correct, however, there are several companies who are working on large scale NLU / machine comprehension systems. Most companies have not yet released their offerings, or they've only released portions. Companies include:

Each solution will vary with the focus of their offering and the specific features. However, the GLUE benchmark serves as an interesting set of requirements that most find necessary: https://gluebenchmark.com/diagnostics

I'll also note that at least half of the companies are attempting to do deep NLU without applying linguistic techniques. Instead, they use deep learning routines that are fine tuned for a final application such as question & answering. Others are embracing the concepts you mention. This is an exciting space, and with Bill Gates mentioning just last week that if he were to do a startup, it would be to "teach a computer to read".

  • Thanks. So how would a non-deep-learning (presumably rules-based) approach deal with the high dimensionality? (It seems like it must be even higher than at the phrase-level.) Jul 3, 2019 at 10:25
  • 1
    Non-DL approaches wouldn't use vector spaces as their primary data structure. For example, they might dump discourse relations into a knowledge graph, where the discourse tag was the tag on the edge. Going full circle, they might have a DL model then subsume the knowledge graph, etc.
    – user21043
    Jul 3, 2019 at 13:26

Basically no, full AI doesn't exist, and there is not much work on representing language at the discourse level.

The bleeding edge would be the Winograd schema challenge and anaphora / co-reference resolution, where the scope is nearby phrases or sentences.

Other tasks that may interest you are summarisation and question answering, eg SQuAD.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.