Turing's Dream and the Knowledge Challenge

There is a set of clear-cut challenges, all centering around knowledge, that have received insufficient attention in AI, and whose solution could bring the realization of Turing’s dream – the dream of a machine we can talk with just like a person, and which is therefore (at least) our intellectual equal. These challenges have to do with the representation of linguistically expressible knowledge, the role of knowledge in language understanding, the use of knowledge for several sorts of commonsense reasoning, and knowledge accumulation. Concerning the last topic, I briefly present preliminary results of some of our recent efforts to extract “shallow” general knowledge about the world from large text corpora. Hi-fidelity Representation of Linguistically Expressible Knowledge “Language is fundamental to our ability to think; it is ‘more or less synonymous with symbolic thought,’...” – Donald Johanson, Sci. Am. 1998, expounding on Ian Tattersall’s Becoming Human Humans acquire large amounts of knowledge about the world from verbal expressions of such knowledge, and in turn, are able to communicate much of their internalized knowledge in language. It stands to reason, therefore, that our ‘mentalese’ must match the expressive resources of language – and that artificial agents competent in language will also require internal representations no less expressive than language. My collaborators and I have argued for many years (e.g., (Schubert & Hwang 1989; 2000; Schubert 2000)) that a representation capable of supporting natural language and commonsense reasoning in machines must allow not only for predication, logical connectives, and ∀/∃-quantifiers – the devices of classical logic – but also the following: • generalized quantifiers such as most or often; e.g., “Most of the trick-or-treaters who came to our door received several candy bars”; though set-theoretic paraphrases are possible, inference is more straightforward if such quantifiers are directly allowed for; • event/situation reference; e.g., (following the previous sentence) “This nearly emptied our bowl of treats by 7:30 Copyright c © 2006, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. pm”; note the anaphoric reference to the collection of ‘receiving events’, and the temporal modification of the ‘nearly-emptying event’; • modification (of predicates and sentences); examples are “nearly” in the previous sentence, verbs like “become” and “pose as” (with a predicate complement), and sentence adverbials such as “perhaps” and “according to reliable sources”; many such examples require an intensional semantics, i.e., one that does not reduce meaning to reference (extension) in the world; • reification (of predicates and sentences); examples are “Domestic dogs evolved from wolves and African wild dogs” (nominal predicate reification), “Composing music requires problem solving, but is deeply gratifying” (VP predicate reification), “That Turing was brilliant is beyond doubt” (proposition reification), “For Turing to make a mistake was unusual” (event-type reification), and “Who murdered Montague remains a mystery” (question reification); again intensionality is involved in many such examples; • metric/comparative attributes; e.g., “Thousands of men were employed for 20 years to build the Great Pyramid, which rose to a height of 485 feet”; “The frame of the bed is wider than the door is high”; • uncertainty and genericity; e.g., “If John receives a job offer from North Central Positronics, he will probably accept it”; “Elementary school children are usually bused to school”; “Dogs bark”; • metalinguistic capabilities; e.g., “What is the sum of binary numbers 111 and 1, expressed as a binary number?”; “Can you finish the opening sentence of Hamlet’s soliloquy, ‘To be or not to be, ...’?”; It is possible to contrive FOPC paraphrases by hand for some of the above examples, but doing so automatically is deeply problematic, and probably impossible in some cases (e.g., for some examples of modification, reification, genericity, and metalanguage). Our Episodic Logic (EL) representation (Schubert & Hwang 2000) covers many of the above constructs and implements them in the EPILOG system (Schaeffer et al. 1993; Schubert & Hwang 2000). Also (Fox & Lappin 2004; 2005) provide a two-level representation (PTCT – Property Theory with Curry Typing) wherein propositions and other intensional entities are represented as terms, and these are embedded in a first-order language with a predicate True. They provide a tableau proof theory for the language, and their focus is on the logic, more than on its exact connection to various linguistic phenomena, or on computational issues. Much additional work is needed to develop semantics and practical inference methods, including various types of nondeductive inference, for KR languages with the above sort of expressive range. Worries about tractability of such languages are quite misguided and may have hindered progress on NLU. Promoting research into tractable inference is one thing, but inveighing against formalisms that allow intractable inference is like inveighing against programming languages that allow recursion or lambda-abstraction. The Role of Knowledge in Language