Exp 2 Language understanding & common sense reasoning Joshua K Hartshorne • Tobias Gerstenberg • Timothy J. O’Donnell • Joshua B. Tenenbaum Massachusetts Institute of Technology Motivation Language is massively ambiguous. Interpretation often requires world knowledge & common sense reasoning. How and why is common sense recruited? Simulate: A beat B and A was strong. If A had not Details & Intuitions Framework message world speaker’s experience speaker’s knowledge speaker’s intentions utterance linguistic knowledge Goal is to recover information about world. Literal meaning is an intermediate stage. Prior beliefs about world plays key role. listener Winograd Schema (Winograd, 1973) The city council denied the protesters a permit because they feared violence. The city council denied the protesters a permit because they advocated violence. • Require common sense • Easy to generate • Easy to score Properties: (cf Levesque, 2011) message P(utterance|message) …because the city council feared violence P(message) …because the protesters feared violence …because the Red Sox feared violence …because the Dutch feared violence … likely likely unlikely unlikely … likely unlikely very unlikely very unlikely … Note: Current focus is on messages achieved from different interpretations of pronoun. Conclusions & Future Directions P(message|utterance) α P(utterance|message) * P(message) Factorization: Find most likely message given utterance need model of world (& speaker) Exp 1 P(message) estimated by Turkers Exp 1a A scared B because he jumped out from behind something. utterance (80 total): Turkers asked: How likely is it that {A/B} jumping out from behind something would cause A to scare B? -2 -1 0 1 2 model -5 -4 -3 -2 -1 0 1 2 3 4 5 log-odds he = A model humans r=.88 Utterance: The city council denied the protesters a permit because they feared violence. Rarely use pronouns to refer to new entities Exp 1b Because A scared B, he screamed. utterance (80 total): Turkers asked: How likely is it that A scaring B would cause {A/ B] to scream? -2 -1 0 1 2 model -5 -4 -3 -2 -1 0 1 2 3 4 5 log-odds he = A model humans r=.73 average weak strong strength B A B’ B’ B’ B’ B’ P(message) calculated via simulation over intuitive theory A B 16 sentences A {almost/-} {beat/lost to} B in tug-of-war {because/although} he is {strong/weak}. A beat B in tug-of-war because he is strong. Note: Assume counterfactual semantics for “because”. p because q = p and q and (without q, no p). Determine probability by simulation over an intuitive model of tug-of-war (top-right). Possible messages: … because A is strong. … because B is strong. etc. average weak strong strength A B A’ A’ A’ A’ A’ log-odds he = A model -3 -2 -1 0 +1 +2 +3 -2 -1 0 +1 +2 +3 humans • • • • • • • • • • • • • • • • r=.92 3 4 5 6 7 B’s strength model humans r=.96 • • • • • • • • • • • • • • • • 3 4 5 6 7 A’s strength model humans r=.98 • • • • • • • • • • • • • • • • Simplifications: Stronger person always wins. Prototype / soft threshold semantics for scalar adjectives. • • Language comprehension as inference over intuitive model of the speaker explains how common sense affects language processing. • Limited need for theory of pronouns per se. • • Can derive pronoun biases (ask me). Productive paradigm for test theories of knowledge (e.g. reasoning about tug-of-war) and semantics (e.g., because=counterfactual). • Note: No notion of informativity (yet). pdf: joshuakhartshorne.org/publications.html