CS 544 Lecture 3 Syntax and Compositional Semantics

CS 544: Lecture 3: Syntax and Compositional Semantics of the Noun Phrase March 1, 2011 Jerry R. Hobbs USC/ISI Marina del Rey, CA

Noun Phrases S NP Last time we elaborated on VPs. VP V NP This time we’ll elaborate on NPs.

The Relevant Parts of Speech Noun (NN, NNS): singular and plural; “The _____ is. . . ” Verb (VBN, VBG): broken, breaking; “_____-ing” Adjective (JJ, JJR, JJS): good, better, best; “The ____ thing is. . . ” Adverb (RB): very; often “____-ly” Pronoun (PP): I, you, he, she, it, we, they, anyone, . . . Determiner and quantifies (DT): a, the, some, . . Preposition (IN): of, in, to, during, before, . . .

The Structure of the NP the three tall government buildings in Chicago that I saw left modifiers head noun right modifiers, noun complements noun group the three tall gray U. S. government buildings determiners, numbers, quantifiers adjectives prenominal nouns head noun In this order. NP --> DT* JJ* NN* {NN | NNS}. . You can use this structure to tell what part of speech a word is

Nominal Compounds One or more nouns can be placed before the head noun; This says only that some relation occurs between the adjacent nouns or noun sequences turpentine jar pension fund air strike city lights language origin grammar induction human languages ==> turpentine(x) & nn(x, y) & jar(y) Bracketing ambiguities: [Stanford [Research Institute]] vs. [[cancer research] institute] ==> cancer(x) & research(y) & nn(x, y) What about: & institute(z) & nn(y, z) State pension funds Tree. Bank does not tag these -Grammar induction techniques “too hard to explain to annotators” Wall Street reform effort California Public Employees Retirement System Noun-like adjectives can appear in this position too.

The Adjective Position A sequence of adjectives in their absolute, comparative, or superlative forms: big investors the new chief executive a ferocious fire fight the nation’s biggest public pension fund the Taliban’s ideological agenda a cursory look different sentences unannotated textual input Present and Past Participles can occur in the adjective position: supporting air strikes resulting brawl stepped-up reform effort numbed reflection Note incorporated particle Present Participles can occur in prenominal noun position: parking spaces its cratering record Adverbs can modify the adjectives or participles: past internally driven geologic activity JJ Systematic ambiguity between adjective and prenominal noun: criminal lawyer resulting brawl vs. parking spaces ==> criminal(x) & lawyer(x) vs. criminal(x) & nn(x, y) & lawyer(y) NN

Other Adjective Position Constructions Measure phrases: 10 -year fund veteran, 12. 2% rise, six-foot high fence Note use of singular Hyphenated passives with nouns: Los Angeles-based company cement-lined ramp Hyphenated present participle (pre-participle noun is object): money-losing investments Hyphenated or unhyphenated adjective-noun: medium-resolution images higher level units finite state languages

Determiners The order of determiners (quantifiers, numbers) is a very random phenomenon that defies general principles: not as few as ten too many other books All the forest was destroyed * All a forest was destroyed the three other books the other three books * other the three books * other three books any such book * such any book * a such book such a book all three books * both two books

Some Determiner Phrases Definite, Indefinite, Demonstrative: the, a, that, those Possessive Pronouns: its, his, our, their, your Quantifiers: each, all, such, An Ambiguity: some situations vs some computer program little (evidence) Determiner or Adjective? other, our own Numbers and Related Words: two, 67 billion, several, many, most, four and a half (glasses) Complex Determiners: all our, all that, only a few, the two, just one or two, more than 100, a net 1, 000, every last (one), the past 25 (millennia) Possessive NPs also function as determiners: the nation’s biggest public pension fund Saturn’s satellites Afghanistan’s Uruzgan Valley

Count Nouns vs. Mass Nouns Count nouns in the singular require a determiner: I showed him an example. * I showed him example. Mass nouns don’t require a determiner in the singular: I gave him wine. When mass nouns do have a determiner, it often means a kind rather than an instance. I gave him a good wine. (= a good kind of wine) But there are many idiomatic exceptions: For example, by surprise, in front, in mind, He is president

Structure of NPs: Left Modifiers all the many tall educated Washington presidential foreign policy advisors Head noun: advisor’(e 0, x) & Plural’(e 1, x, s) Prenominal nouns: Washington’(e 2, w) & nn’(e 3, w, x) Also noun-like Adjs: president’(e 3, y) & nn’(e 4, y, x) Adjectives: tall’(e 5, x) Also vbn, vbg: educate’(e 6, z, x) Quantifiers, Numbers: many’(e 7, s) (property of the set in plurals) Determiner: the’(e 8, x, . . &e 1) (relation between entity x and description given by rest of NP) Possessive NPs are determiner phrases: Amy’s => poss’(e 10, a, x) Predeterminer: all’(e 9, s, . . . &e 1) IN THIS ORDER (relation between set and description)

Pronouns and Proper Nouns It should be easy: No internal structure: NP --> Pronoun (PP) NP --> Proper. Noun (NNP, NNPS) I, you, he, she, anyone, . . . Uruzgun Valley, Enceladus But proper nouns sometimes function as nouns inside NPs, rather than full NPs: The Voyager 2 spacecraft The Cassini mission A resurgent Taliban So maybe it’s better to treat proper nouns as simply nouns: Noun --> Proper. Noun where in most uses it has no modifiers.

Headless NPs With some determiners, you don’t need a head noun: I like those. Give me some. You have many. I don’t have any. the most * I like the. * Give me a. * I want my. the fifth-highest They can also occur with right modifiers: Those with siblings should stand up for them. Many in America don’t question authority.

Recognizing Noun Groups with Finite-State Automata Noun groups can be recognized fairly reliably with a finite-state automaton: NG --> { (Detp) (Adjs) Ns | Detp[complete] | Pro | N[time] | that N[sing] } Detp[complete] --> { ( { Adv[pre-num] | another | { det | Pro[poss] } ( only (other) ) } ) Number | Q-er | (the) Q-est | another } | Det[complete, ~that] | Pro[poss, complete] } Detp[incomplete] --> { { { Det | Pro[poss] } only | a | Det[incomplete] | Pro[poss, incomplete] } (other) | Detp[complete] other } Adjs --> Adj ( ({ “, ” | and | “, ” and }) Adj )* Adj --> JJ | VBN | VBG | Number “-” N[measure] (“-” JJ) | N “-” VBN } Ns --> { ( JJ “-” ) N[sing, ~time, ~proper] }* N

Logical Form of Noun Groups Nouns convey three kinds of information: Singular nouns: “man”: the entity referred to: man(x) the property conveyed about that entity: man’(e, x) Plural nouns: “men”: the set s referred to the typical entity in that set: x A representation for “men”: man’(e, x) & plural(x, s) s is the set of men x is the typical man e is the property of x’s being a man “tall presidents” ==> tall(x) & president’(e, x) & plural(x, s) “numerous presidents” ==> numerous(s) & president’(e, x) & plural(x, s) “former presidents” ==> former(e) & president’(e, x) & plural(x, s) “the men ran” ==> run(x) “the men gathered” ==> gather(s)

Logical Form of Noun Groups Determiners and quantifiers are complicated in their logical form. Standard approach: all, every ==> (forall (x). . . ) some, a ==> (exists (x). . . ) the, these, such, . . ==> ? I would just say which of x, s, and e the word’s meaning depends on. many men ==> many(s) the man ==> the(x, e) [x can be identified by the description e] such men ==> such(x, e) Possessives: Pat’s book ==> Pat(x) & Poss(x, y) & book(y) Headless NPs: several arrived ==> several(s) & plural(x, s) & arrive(x) Verbal forms in adjective position: Note object position running man ==> run(x) & man(x) broken window ==> break(y, x) & window(x)

Logical Form of Noun Groups Verbal forms in adjective position: Note object position running man ==> run(x) & man(x) broken window ==> break(y, x) & window(x) Possible positions for present participle: Adjective: supporting strike ==> support(x, y) & strike(x) Prenominal noun: parking space ==> park’(e, y, z) & nn(e, x) & space(x) Head noun: a brutal shooting ==> brutal(e) & shoot’(e, x, y) Three types of adjectives (wrt logical form): Standard: red car ==> red(x) & car(x) Opaque: accidental guerilla ==> accidental(e) & guerilla’(e, x) (The person is not accidental, only the property of being a guerilla) Noun-like: Saturnian system ==> Saturn(x) & nn(x, y) & system(y)

Attributives Some attributive adjectives have an implicit comparison set or scale: A small elephant is bigger than a big mosquito. That mosquito is big. mosquito(x) & big(x, s) The implicit comparison set or scale, which must be determined from context

Proper Names Proper names: Could treat them as constants: Springfield is the capital of Illinois. ==> capital(Springfield, Illinois) But there are many Springfields; we could treat it as a predicate true of any town named Springfield: capital(x, y) & Springfield(x) & Illinois(y) Or we could treat the name as a string, related to the entity by the predicate name: capital(x, y) & name(“Springfield”, x) & name(“Illinois”, y)

Indexicals An indexical or deictic is a word or phrase that requires knowledge of the situation of utterance for its interpretation. “I”, “you”, “we”, “here”, “now”, some uses of “this”, “that”, . . . The property of being “I” is being the speaker of the current utterance Indexicals require an argument for the utterance or the speech situation. I(x, u): x is the speaker of utterance u you(x, u): x is the intended hearer of utterance u we(s, u): s is a set of people containing the speaker of utterance u here(x, u): x is the place of utterance u now(t, u): t is the time of utterance u ago(t, d, u): t is a duration of d before the time of the utterance u from the quotation marks Chris said, “I see you now. ” ==> say(Chris, u) & content(e, u) & see’(e, x, y) & I(x, u) & you(y, u) & at. Time(e, t) & now(t, u)

Quantifier Ambiguities Every man loves a woman. ==> ( m)( w) love(m, w) ==> ( w)( m) love(m, w) i. e. , his wife i. e. , Britney Spears Most politicians in most countries can fool most of the people on most issues most of the time. This has 120 possible readings, all distinct. e. g. , different issues for each country, or same issues? different people for each issue, or same people? Do we need to generate each separate reading? Quasi-logical form (Alshawi), typical elements (me) Specify the predicate-argument relations, remain silent about quantifier scoping every(m, e 1) & man’(e 1, m) & love(m, w) & a(w, e 2) & woman’(e 2, w) May later learn: Functionally. Dependent(w, m) i. e. , ( m)( w)

Right Modifiers or Noun Complements Anything that can be a predicate complement can be a right modifier: PP: the man in the black coat Adjective phrase: the people responsible, investors nationwide Past participle: the ingredients required for the recipe Present participle: those people seeking reelection NP (Appositive): Barack Obama, president of the U. S. Think of these as “reduced relatives”: the man (who is) in the black coat Also relative clauses: the birds that I am able to identify () And a few other, less common constructions: Infinitival complements: an attempt to escape Sentential complements: the belief that America deserves better

Prepositional Phrases Some examples: a vital part of Cal. PERS’ efforts the Cassini mission to Saturn a ferocious fire fight with the Taliban a cautionary tale about fights breaking out the regularity of chunks over different sentences PP attachment ambiguities are the biggest source of ambiguity in English: I [saw [the whales in the bay] with binoculars. I [saw [the whales] [in [the bay with binoculars]]]. etc. Usually attachments don’t cross: Specify the length, in bytes, of the word. Lots of ambiguities are benign: I never spent enough time in one place. build a life for myself away from the job. bear little resemblance to the usual phrase structure.

Adjective and Participle Phrases Some examples: big investors nationwide the interest generated by Voyager’s visit A classic ambiguity in computational linguistics: I saw the Grand Canyon, flying to New York.

Appositive NPs Some examples: Proper names can be the head, or the appositive. Anne Stausboll, a 10 -year fund veteran, the Australian counterinsurgency expert David Kilcullen a term, “accidental guerilla” chunking – recognizing higher level units of structure

Logical Forms of Right Modifiers PP: the cat in the hat ==> cat(x) & in(x, y) & hat(y) Sometimes arguments in relation nouns: mother of Pat ==> mother(x, y) & Pat(y) Adjective phrase: the man responsible ==> man(x) & responsible(x, y) Past participle: the window broken by the boy ==> window(x) & break(y, x) & boy(y) Present participle: a student studying logic ==> student(x) & study(x, y) & logic(y) Appositives: NP NP Barack Obama, president of the U. S. Use the predicate “be”: Obama(x) & be(x, y) & president(y, z) & US(z) or identity of variables: Obama(x) & president(x, z) & US(z) or predicate “Appos”: Obama(x) & Appos(x, y) & president(y, z) & US(z) “Appos” usually means “be”, but can mean other things

Some Other Right Modifiers Time NPs: the summit last March Not an appositive because summit =/= last March More like a PP ==> at. Time(s, t) Infinitival complements: the desire to fly ==> desire(x, e) & fly’(e, x) “That” clause: the belief that the world is round ==> belief(x, e) & round’(e, y) & world(y) Like an appositive because belief = that the world is round Not a relative clause because nothing’s missing

Relative Clauses One of the arguments or adjuncts is missing from the relative clause; the relative clause is a right adjunct on some head noun; that head noun “supplies” the missing argument or adjunct. Simplest and most common case: the subject is missing: fund which ( ) has lost more than a quarter of its value the most momentous thing that ( ) had happened in their valley storms that ( ) set the sky close to the ground images that ( ) confounded planetary scientists the filler the gap

Relative Clauses Relative clauses in which the direct object or some other complement is missing: the man I saw ( ) some form of chunked representation of sentences, which we might interpret ( ) as a phrase structure tree

Relative Clauses Relative clauses in which an adjunct is missing: beliefs for which our forbears fought ( ) ways to turn these villagers into allies ( ) one way to capture the regularity of chunks ( )

Why “Long-Distance Dependencies”? NP NP SBAR IN S NP VP S V VP NP the book that Chris thinks Pat said Kim V NP read ()

Other Long-Distance Dependencies Relative clauses are examples of long-distance dependencies, because the gap and its filler can be arbitrarily far apart in the parse tree. the man that Mary believes John asked Susan to tell George to meet with ( ) To get the right logical form, we have to pass information about the gap up to the filler (or pass information about the filler down to the gap). Other long-distance dependency constructions: Questions: Who does Mary believe John asked George to meet with ( )? Wh-nominalizations: Whoever Mary believes John asked George to meet with ( ) is here what was most revealing about the battle was the fact that. . . asked later why they’d done so the question of how to learn structure from unannotated textual input

Introducing Gaps in Clauses The elegant way: NP[gap=x] --> “ ” The efficient way: VP[gap=x] --> V[compl 1=np] S[gap=x] --> VP x is some identifier of the node that is gapped Gives us a gap in the object Gives us a gap in the subject Passing gaps up to higher nodes: S[gap=x] --> NP VP[gap=x] --> V NP (that) S[gap=x] These work for gaps in questions and wh-nominalizations too.

Filling the Gaps The “wh-phrase” introducing the relative clause can be quite complex: the man who I met ( ) the man the brother of whose father I met ( ) Rule: Replace the wh-word with the head noun. Fill the gap with the whole wh-phrase Similar rules for questions and wh-nominalizations

Logical Forms for Long-Distance Dependencies the man who arrived ==> man(x) & arrive(x) the man whom Pat met ==> man(x) & Pat(y) & meet(y, x) the man for whom Pat works ==> man(x) & Pat(y) & work’(e, y) & for(e, x) the man whose cousin arrived ==> man(x) & cousin(y, x) & arrive(y) whoever studies will pass ==> study(x) & pass’(e, x) & Future(e)