Synonymy and Near-Synonymy in Deep Lexical Semantics Niloofar Montazeri and Jerry R. Hobbs Information Sciences Institute University of Southern California Marina del Rey, CA Deep Lexical Semantics: Methodology words link with axioms core theories Construct core theories of abstract phenomena in various domains Express meanings of word senses as logical axioms in terms of predicates in core theories Core Theories and Lexical Periphery Define (Characterize) words in terms supplied by the core theories. range(x,y,z) <--> scale(s) &subscale(s1,s) & bottom(y,s1) & top(z,s1) & in(u1,x) & at(u1,y) & in(u2,x) & at(u2,z) & (u x)( v s1) at(u,v) s y s1 v z x = {u1 . . . . . . . u . . . . . u2} word: “range” linking axiom core theory Axiomatize core theories with richly explicated core predicates: Core Theory of Scales: scale, <, subscale, top, bottom, at Abstract Words in Context By specializing “at” and the scale in various ways, we can get a whole range of possible meanings for “range”: The scores on the test ranged from 33 to 96. The timber wolf ranges from Mexico to the Arctic. His behavior ranged from cheerful to sullen. Deep Lexical Semantics of Event-Related Words (forall (x y z) (iff (give x y z) (exist (e1 e2 e3) (and (cause x e1)(change’ e1 e2 e3)(have’ e2 x y) (have’ e3 z y))))) x gives y to z = x causes a change from x having y to z having y Builds on old work by Gruber, Lakoff, Schank, Jackendoff, and many others on lexical decomposition, but .... “primative” predicates are explicated in core theories: restricts possible interpretations of the predicates enables reasoning within theory logic, not syntax Use in Textual Inference T: Russia is blocking oil from entering Ukraine. H: Oil cannot be delivered to Ukraine. not’(n2,c2) & can’(c2,x2,d2) & deliver’(d2,x2,o2,u2) Lexical decomposition axioms cause’(d2,x2,c3) & changeTo’(c3,h2) & have’(h2,u2,o2) possible’(p1,c4) & cause’(c4,x3,e1) in’(h2,o2,u2) cause’(c1,x1,n1) & not’(n1,p1) & possible’(p1,e1) block’(b1,x1,e1) & enter’(e1,o1,u1) Defeasible inferences from core theories Core Theories: Change change(e1,e2): state e1 changes into state e2 e1 and e2 involve a common entity; change of state of something Transitive if the something is the same e1 and e2 are different unless there is an intermediate state (cyclic change) “move” is change of state of “at” relation change(e1,e2) = changeFrom(e1) = changeTo(e2) Theory of Causality: Causal Complex causal complex s e1 causal-complex(s,e) e2 e3 e4 e e1 s, .... .... effect When every event or state in s happens or holds, then e happens or holds. All eventualities in s are relevant to the effect. A rigorous, nondefeasible notion, but can’t specify everything. Theory of Causality: Cause In a causal complex, some eventualities are distinguished as causes. presumable power on cause finger in socket Causes are the focus of planning, prediction, explanation, interpreting discourse (but not diagnosis) shock What is presumable depends on task, context, knowledge base, .... “Cause” is a useful but defeasible notion. Methodology Having axiomatized these core theories, .... Focus on most common 450 word senses involving change of state and causality Determine radial structure of set of WordNet senses of word, and characterize by incremental differences in associated axioms Encode axioms for the most abstract or general senses or “supersenses” Evaluate on a textual entailment task “Enter” At each hop, we specialize a predicate or constrain an argument enter-S1: x enters p = changeTo(p(x)) enter-V2: enter a race enter-V4: enter into calculations enter-V9: enter into career p = at/in enter-S11: x enters y = changeTo(in(x,y)) enter-V1: enter a room enter-V6: enter from stage left + cause enter-S2: x enters y in z = cause(x,enter-S11(y,z)) Logically, and sometimes chronologically enter-V5: enter in ledger enter-V8: enter picture into text “Hit” Supersenses give topological structure. Specific senses specialize general predicates or put constraints on arguments hit-S1: x hits y = changeTo(at(x,y)) We hit Detroit by noon. The temperature hit -20. + sudden impact hit-S11: x hits y = changeTo’(e,in(x,y)) & sudden(e) & impact(x,y) The car hit a tree. + cause hit-S2: x hits z with y = cause(x,hit-S11(y,z)) He hit the ball. Synonymy and Near-Synonymy Synonymy: The axiomatic lexical decompositions are the same Near-synonymy: The axiomatic lexical decompositions differ only incrementally (similar to word senses in radial structure) “Receive” and “Get” receive-S1: x receives y from z = change(have(z,y), have(x,y)) -> changeTo(have(x,y)) = FrameNet sense 1 subsumes all of WordNet’s senses where “have” is specialized to owning, having a property, perceiving, hosting, etc. get-S11: x gets y = cause(x, changeTo(have(x,y))) He always gets what he wants. I’ll get the book at the library. get-S12: x gets y = changeTo(have(x,y)) He got the flu. I got a call from Sue. You’ll get your results tomorrow. Synonyms or near-synonyms “Go”, “Hit”, and “Reach” go-V1: x goes from e1 to e2 = change(e1,e2) & arg*(x,e1) & arg*(x,e2) 3rd FrameNet sense go-FV3: specialize e1 and e2 to “at” relations x goes from y to z = change(at(x,y), at(x,z)) hit-S1: x hits z = changeTo(at(x,z)) reach-S1: x reaches z = changeTo(at(x,z)) x is an argument of ei or a participant in ei Not synonymous at more general levels of “go” Not synonymous at more specific levels of “hit” “Deliver”, “Give”, and “Provide” deliver-S1: x delivers y from w to z = cause(x, change(rel(y,w), rel(y,z))) deliver-S11: specializes rel to have stipulate that x = w = cause(x, change(have(x,y), have(z,y))) give-S0: cause(x, exist(y)) synonyms give-S1: cause(x, change(have(x,y), have(z,y))) provide-S1: x provides z with y = cause(x, changeTo(possible(e))) & arg(y,e) & need(z,e) provide for medical emergencies provide-S11: possible(e) specializes to have(z,y) = cause(x, changeTo(have(z,y))) & need(z,e) near synonyms An Aside on “Need” (Gordon& Hobbs) In a core theory of cognition (based on beliefs and goals), Define badFor(e,x) = e causes a goal of x’s not to happen need(x,e) = cause(not(e), e1) & badFor(e1,x) Overlapping Radial Structures deliver-S1 give-S0 give-S1 deliver-S11 provide-S11 provide-S1 A Problem Synonymy is a relation between word senses. Carving the uses of words into word senses is highly arbitrary; e.g., WordNet is very fine-grained; VerbNet less so. word-1, sense-i word-2, sense-j Why not just stipulate that these are three different senses, so that in the middle the senses are perfectly synonymous? Context-Dependent Synonymy give = cause to have provide = cause to have + need She gave food to the hungry man. She provided food for the hungry man. The sentences are equivalent even though “give” and “provide” are only near synonyms. Synonymous Specific Senses So far the examples have been at an abstract level, but intersection of radial structures can occur at very specific levels too: deliver-V1: deliver a talk give-V12: give a talk Small Perspective Differences: Near Synonyms? hold: x holds e = cause(x, not(changeFrom(e))) hold that pose Specialize e to at: cause(x, not(changeFrom(at(y,z)))) hold the picture against the wall block: x blocks e = cause(x, not(changeTo(e))) The senator blocked the judge’s appointment Specialize e to at: cause(x, not(changeTo(at(y,z)))) He blocked my way These can describe the same situation: hold(x,e) = block(x,not(e)) “Capture” and “Seize” x holds y = cause(x, not(changeFrom(at(y,z)))) x captures y = cause(x, changeTo(hold(x,y))) “Seize”: Almost all senses of “seize” subsumed under seize-S1: x seizes y = cause’(e, x, changeTo(hold(x,y))) & forceful(e) near synonyms Text Entailment Example: Synonymy from Inference H: The captors let the hostage go free. let(x,e7) & go’(e17,y,e8) & free’(e8,y,c,e9) not(e5) & cause’(e5,x,e6) & not’(e6,e7) changeTo’(e7,e8) not’(e8,e10) & cause(e10,c,e11) & not’(e11,e9) & move’(e9,y,z,w) changeFrom’(e9,e12) & at’(e12,y,z) Rexist(e0) changeFrom’(e0,e1) & cause’(e1,x,e2) & not’(e2,e3) & changeFrom’(e3,e4) & at’(e4,y,z) release(x,y,z) T: A Filipino hostage in Iraq was released. “Blunder” and “Lapse” (Edmunds & Hirst) blunder(e,x) = error(e,x) & cause(e1,e) & stupid’(e1,x) lapse(e,x) = error(e,x) & cause(e1,e) & neglectful’(e1,x) Need to capture meaning of “error” in theory of composite entities as a mismatch between a composite entity and a pattern that serves as a norm. Need to capture meanings of “stupid” and “neglect” in theory of cognition; “stupid” in terms of ability to learn and reason “neglect” in terms of model of attention (Gordon & Hobbs) Near Synonyms? raise(x,y,z,w) = cause(x, change(at(y,z), at(y,w))) & above(w,z) rise(y,z,w) = change(at(y,z), at(y,w)) & above(w,z) Differ only by cause(x, ...) raise(x,y,z,w) = cause(x, change(at(y,z), at(y,w))) & above(w,z) lower(x,y,z,w) = cause(x, change(at(y,z), at(y,w))) & above(z,w) Differ only in relation of w and z To be near synonyms the words have to describe the same situations. Message The meanings of word senses can be expressed as axioms in terms of predicates from core theories of the relevant phenomena. The word senses of any word form a radial structure, where adjacent nodes differ by incremental changes in the axioms (specializations of predicates, constraints on arguments). Word senses of different words can have identical axiomatic decompositions; this is synonymy. Word senses of different words can have nearly identical axiomatic decompositions -- is this near synonymy? Is near synonymy a natural kind, about which we can make reliable judgments? More important than labeling word senses as near synonyms is capturing the distinctions formally.