Dragomir Radev

Wrocław, Poland

July 29, 2009

Computational Linguistics

Example (from a famous movie)

Dave Bowman: Open the pod bay doors, HAL.HAL: I’m sorry Dave. I’m afraid I can’t do that.

Instructor

• Dragomir Radev, Professor, Computer Science and Information, Linguistics, University of Michigan

• radev@umich.edu

Natural Language Understanding

• … about teaching computers to make sense of naturally occurring text.

• … involves programming, linguistics, artificial intelligence, etc.

• …includes machine translation, question answering, dialogue systems, database access, information extraction, game playing, etc.

Example

• How many different interpretations does the above sentence have? How many of them are reasonable/grammatical?

I saw her fall

Silly sentences

• Children make delicious snacks• Stolen painting found by tree• I saw the Grand Canyon flying to New York• Court to try shooting defendant• Ban on nude dancing on Governor’s desk• Red tape holds up new bridges• Iraqi head seeks arms• Blair wins on budget, more lies ahead• Local high school dropouts cut in half• Hospitals are sued by seven foot doctors• In America a woman has a baby every 15 minutes. How

does she do that?

Types of ambiguity• Morphological: Joe is quite impossible. Joe is quite important.• Phonetic: Joe’s finger got number.• Part of speech: Joe won the first round.• Syntactic: Call Joe a taxi.• Pp attachment: Joe ate pizza with a fork. Joe ate pizza with meatballs. Joe

ate pizza with Mike. Joe ate pizza with pleasure.• Sense: Joe took the bar exam.• Modality: Joe may win the lottery.• Subjectivity: Joe believes that stocks will rise.• Scoping: Joe likes ripe apples and pears.• Negation: Joe likes his pizza with no cheese and tomatoes.• Referential: Joe yelled at Mike. He had broken the bike.

Joe yelled at Mike. He was angry at him.• Reflexive: John bought him a present. John bought himself a present.• Ellipsis and parallelism: Joe gave Mike a beer and Jeremy a glass of wine.• Metonymy: Boston called and left a message for Joe.

NLP• Information extraction• Named entity recognition• Trend analysis• Subjectivity analysis• Text classification• Anaphora resolution, alias resolution• Cross-document crossreference• Parsing• Semantic analysis• Word sense disambiguation• Word clustering• Question answering• Summarization• Document retrieval (filtering, routing)• Structured text (relational tables)• Paraphrasing and paraphrasing/entailment ID• Text generation• Machine translation

Syntactic categories

• Substitution test:

Nathalie likes {

}

cats.

black Persian tabby small

easy to raise

• Open (lexical) and closed (functional) categories:

No-fly-zoneyadda yadda yadda

thein

Jabberwocky (Lewis Carroll)

Twas brillig, and the slithy tovesDid gyre and gimble in the wabe:All mimsy were the borogoves,And the mome raths outgrabe.

"Beware the Jabberwock, my son!The jaws that bite, the claws that catch!Beware the Jubjub bird, and shunThe frumious Bandersnatch!"

Phrase structure

S

NP VP

NPVBD

caught the butterfly

That man PP

IN NP

with a net

Sample phrase-structure grammar

S NP VPNP AT NNSNP AT NNNP NP PPVP VP PP VP VBD VP VBD NP P IN NP

AT theNNS children NNS students NNS mountains VBD slept VBD ate VBD saw IN in IN of NN cake

Phrase structure grammars

• Local dependencies• Non-local dependencies• Subject-verb agreement

The women who found the wallet were given a reward.

• wh-extraction

Should Peter buy a book?Which book should Peter buy?

• Empty nodes

Subcategorization Subject: The children eat candy.

Object: The children eat candy.Prepositional phrase: She put the book on the table.Predicative adjective: We made the man angry.Bare infinitive: She helped me walk.To-infinitive: She likes to walk.Participial phrase: She stopped singing that tune at the end.That-clause: She thinks that it will rain tomorrow.Question-form clauses: She asked me what book I was reading.

Phrase structure ambiguity

• Grammars are used for generating and parsing sentences

• Parses• Syntactic ambiguity• Attachment ambiguity: Our company is training

workers.• The children ate the cake with a spoon.• High vs. low attachment• Garden path sentences: The horse raced past

the barn fell. Is the book on the table red?

Sentence-level constructions

• Declarative vs. imperative sentences

• Imperative sentences: S VP

• Yes-no questions: S Aux NP VP

• Wh-type questions: S Wh-NP VP• Fronting (less frequent):

On Tuesday, I would like to fly to San Diego

Semantics and pragmatics

• Lexical semantics and compositional semantics• Hypernyms, hyponyms, antonyms, meronyms

and holonyms (part-whole relationship, tire is a meronym of car), synonyms, homonyms

• Senses of words, polysemous words• Homophony (bass).• Collocations: white hair, white wine• Idioms: to kick the bucket

Discourse analysis

• Anaphoric relations:

1. Mary helped Peter get out of the car. He thanked her.

2. Mary helped the other passenger out of the car. The man had asked her for help because of his foot injury.

• Information extraction problems (entity crossreferencing)

Hurricane Hugo destroyed 20,000 Florida homes.At an estimated cost of one billion dollars, the disasterhas been the most costly in the state’s history.

Pragmatics

• The study of how knowledge about the world and language conventions interact with literal meaning.

• Speech acts

• Research issues: resolution of anaphoric relations, modeling of speech acts in dialogues

Coordination

• Coordinate noun phrases:– NP NP and NP– S S and S– Similar for VP, etc.

Agreement

• Examples:– Do any flights stop in Chicago?– Do I get dinner on this flight?– Does Delta fly from Atlanta to Boston?– What flights leave in the morning?– * What flight leave in the morning?

• Rules:– S Aux NP VP– S 3sgAux 3sgNP VP– S Non3sgAux Non3sgNP VP– 3sgAux does | has | can …– non3sgAux do | have | can …

Agreement

• We now need similar rules for pronouns, also for number agreement, etc.– 3SgNP (Det) (Card) (Ord) (Quant) (AP)

SgNominal– Non3SgNP (Det) (Card) (Ord) (Quant) (AP)

PlNominal– SgNominal SgNoun | SgNoun SgNoun– etc.

Combinatorial explosion

• What other phenomena will cause the grammar to expand?

• Solution: parameterization with feature structures (see Chapter 11)

Parsing as search

S NP VP Det that | this |a

S Aux NP VP Noun book | flight | meal | money

S VP Verb book | include | prefer

NP Det Nominal Aux does

Nominal Noun Proper-Noun Houston | TWA

Nominal Noun Nominal Prep from | to | on

NP Proper-Noun

VP Verb

VP Verb NP

Nominal Nominal PP

Parsing as search

Book that flight. S

VP

NP

Nom

Verb Det Noun

Book that flight

Two types of constraints on the parses: a) some that come from the input string,b) others that come from the grammar

Top-down parsing

S

NP VP

S

Aux VP

S

VP

S

NP

S

NP VP

Det Nom

S

NP VP

PropN

S

NP VP

Det Nom

S

VP

V NP

Aux

S

NP VPAux

PropN

S

VP

V

Book that flight

Book that flight

Noun Det Noun

Book that flight

Verb Det Noun

Book that flight

Noun Det Noun

Book that flight

Verb Det Noun

Book that flight

Noun Det Noun

Book that flight

Verb Det Noun

Book that flight

Verb Det Noun

Book that flight

Verb Det Noun

Book that flight

Verb Det Noun

NOM NOM NOM

NOMNOM NOM NOM

NOM NOM

VP NP

NP NP

VP

Bottom-up parsing

NP

VP

Grammatical Relationsand Free Ordering of Subject and Object

OSV- Кого же Вася увидел?.- Машу Вася увидел.- (Actually,) who did Vasya see?- Vasya saw Masha

SVO- Кого увидел Вася?- Вася увидел Машу.- Who did Vasya see?- Vasya saw Masha.

OVS- Да кого увидел Вася?- Машу увидел Вася- Well, whom did Vasya see?- It was Masha whom Vasya saw.

SOV- Кого же Вася увидел?- Вася Машу увидел- Who did Vasya see?- Vasya saw Masha

VSO- Увидел Вася кого? - Увидел Вася Машу .- Who did Vasya see?- Vasya saw Masha

VOS- Увидел Машу кто?- Увидел Машу Вася.- Who saw Masha, at the

end?- It was Vasya who saw

Masha

Slide from Lori Levin, originally by Leonid Iomdin

Features and unification

• Grammatical categories have properties• Constraint-based formalisms• Example: this flights: agreement is difficult to

handle at the level of grammatical categories• Example: many water: count/mass nouns• Sample rule that takes into account features: S

NP VP (but only if the number of the NP is equal to the number of the VP)

Feature structures

CAT NPNUMBER SINGULARPERSON 3

CAT NP

AGREEMENT NUMBER SG PERSON 3

Feature paths: {x agreement number}

Unification

[NUMBER SG] [NUMBER SG] +

[NUMBER SG] [NUMBER PL] -

[NUMBER SG] [NUMBER []] = [NUMBER SG]

[NUMBER SG] [PERSON 3] = ?

Agreement

• S NP VP{NP AGREEMENT} = {VP AGREEMENT}

• Does this flight serve breakfast?• Do these flights serve breakfast?

• S Aux NP VP{Aux AGREEMENT} = {NP AGREEMENT}

Agreement

• These flights• This flight

• NP Det Nominal{Det AGREEMENT} = {Nominal AGREEMENT}

• Verb serve{Verb AGREEMENT NUMBER} = PL

• Verb serves{Verb AGREEMENT NUMBER} = SG

Subcategorization

• VP Verb{VP HEAD} = {Verb HEAD}{VP HEAD SUBCAT} = INTRANS

• VP Verb NP{VP HEAD} = {Verb HEAD}{VP HEAD SUBCAT} = TRANS

• VP Verb NP NP{VP HEAD} = {Verb HEAD}{VP HEAD SUBCAT} = DITRANS

Eliza [Weizenbaum, 1966]

User: Men are all alike

ELIZA: IN WHAT WAY

User: They’re always bugging us about something or other

ELIZA: CAN YOU THINK OF A SPECIFIC EXAMPLE?

User: Well, my boyfriend made me come here

ELIZA: YOUR BOYFRIEND MADE YOU COME HERE

User: He says I’m depressed much of the time

ELIZA: I AM SORRY TO HEAR THAT YOU ARE DEPRESSED

Eliza-style regular expressions

s/.* YOU ARE (depressed|sad) .*/I AM SORRY TO HEAR YOU ARE \1/

s/.* YOU ARE (depressed|sad) .*/WHY DO YOU THINK YOU ARE \1/

s/.* all .*/IN WHAT WAY/

s/.* always .*/CAN YOU THINK OF A SPECIFIC EXAMPLE/

Step 1: replace first person references with second person referencesStep 2: use additional regular expressions to generate repliesStep 3: use scores to rank possible transformations

Finite-state automata

• Finite-state automata (FSA)

• Regular languages

• Regular expressions

Finite-state automata (machines)

baa!baaa!baaaa!baaaaa!...

q0 q1 q2 q3 q4

b a a !

a

baa+!

state transition finalstate

Input tape

a b a ! b

q0

Finite-state automata

• Q: a finite set of N states q0, q1, … qN

: a finite input alphabet of symbols

• q0: the start state

• F: the set of final states(q,i): transition function

State-transition tables

Input

State

b a !

0 1 0 0

1 0 2 0

2 0 3 0

3 0 3 4

4 0 0 0

Morphemes

• Stems, affixes• Affixes: prefixes, suffixes, infixes: hingi

(borrow) – humingi (agent) in Tagalog, circumfixes: sagen – gesagt in German

• Concatenative morphology • Templatic morphology (Semitic languages): lmd (learn), lamad (he studied), limed (he

taught), lumad (he was taught)

Morphological analysis

• rewrites

• unbelievably

Inflectional morphology

• Tense, number, person, mood, aspect

• Five verb forms in English

• 40+ forms in French

• Six cases in Russian, seven in Polish

• Up to 40,000 forms in Turkish (you will cause X to cause Y to … do Z)

Derivational morphology

• Nominalization: computerization, appointee, killer, fuzziness

• Formation of adjectives: computational, embraceable, clueless

Finite-state morphological parsing

• Cats: cat +N +PL• Cat: cat +N +SG• Cities: city +N +PL• Geese: goose +N +PL• Ducks: (duck +N +PL) or (duck +V +3SG)• Merging: +V +PRES-PART• Caught: (catch +V +PAST-PART) or (catch +V

+PAST)

Phonetic symbols

• IPA• Arpabet• Examples

Using WFST for language modeling

• Phonetic representation• Part-of-speech tagging

Dependency grammars

• Lexical dependencies between head words

• Top-level predicate of a sentence is the root

• Useful for free word order languages

• Also simpler to parse

Dependencies

John likes tabby cats

NNP VBS JJ NNS

NP

VP

NP

S

Discourse, dialogue, anaphora

• Example: John went to Bill’s car dealership to check out an Acura Integra. He looked at it for about half an hour.

• Example: I’d like to get from Boston to San Francisco, on either December 5th or December 6th. It’s okay if it stops in another city along the way.

Information extraction and discourse analysis

• Example: First Union Corp. is continuing to wrestle with severe problems unleashed by a botched merger and a troubled business strategy. According to industry insiders at Paine Webber, their president, John R. Georgius, is planning to retire by the end of the year.

• Problems with summarization and generation

Reference resolution

• The process of reference (associating “John” with “he”).

• Referring expressions and referents.

• Needed: discourse models

• Problem: many types of reference!

Example (from Webber 91)

• According to John, Bob bought Sue an Integra, and Sue bough Fred a legend.

• But that turned out to be a lie. - referent is a speech act.

• But that was false. - proposition• That struck me as a funny way to describe the

situation. - manner of description• That caused Sue to become rather poor. - event• That caused them both to become rather poor. -

combination of several events.

Reference phenomena

• Indefinite noun phrases: I saw an Acura Integra today.

• Definite noun phrases: The Integra was white.• Pronouns: It was white.• Demonstratives: this Acura.• Inferrables: I almost bought an Acura Integra

today, but a door had a dent and the engine seemed noisy.

• Mix the flour, butter, and water. Kneed the dough until smooth and shiny.

Constraints on coreference

• Number agreement: John has an Acura. It is red.• Person and case agreement: (*) John and Mary have

Acuras. We love them (where We=John and Mary)• Gender agreement: John has an Acura. He/it/she is

attractive.• Syntactic constraints:

– John bought himself a new Acura.– John bought him a new Acura.– John told Bill to buy him a new Acura.– John told Bill to buy himself a new Acura– He told Bill to buy John a new Acura.

Preferences in pronoun interpretation

• Recency: John has an Integra. Bill has a Legend. Mary likes to drive it.

• Grammatical role: John went to the Acura dealership with Bill. He bought an Integra.

• (?) John and Bill went to the Acura dealership. He bought an Integra.

• Repeated mention: John needed a car to go to his new job. He decided that he wanted something sporty. Bill went to the Acura dealership with him. He bought an Integra.

Preferences in pronoun interpretation

• Parallelism: Mary went with Sue to the Acura dealership. Sally went with her to the Mazda dealership.

• ??? Mary went with Sue to the Acura dealership. Sally told her not to buy anything.

• Verb semantics: John telephoned Bill. He lost his pamphlet on Acuras. John criticized Bill. He lost his pamphlet on Acuras.

Salience weights in Lappin and Leass

Sentence recency 100

Subject emphasis 80

Existential emphasis 70

Accusative emphasis 50

Indirect object and oblique complement emphasis

40

Non-adverbial emphasis 50

Head noun emphasis 80

Lappin and Leass (cont’d)

• Recency: weights are cut in half after each sentence is processed.

• Examples:– An Acura Integra is parked in the lot. (subject)– There is an Acura Integra parked in the lot.

(existential predicate nominal)– John parked an Acura Integra in the lot. (object)– John gave Susan an Acura Integra. (indirect object)– In his Acura Integra, John showed Susan his new CD

player. (demarcated adverbial PP)

Algorithm

1. Collect the potential referents (up to four sentences back).

2. Remove potential referents that do not agree in number or gender with the pronoun.

3. Remove potential referents that do not pass intrasentential syntactic coreference constraints.

4. Compute the total salience value of the referent by adding any applicable values for role parallelism (+35) or cataphora (-175).

5. Select the referent with the highest salience value. In case of a tie, select the closest referent in terms of string position.

Example

• John saw a beautiful Acura Integra at the dealership last week. He showed it to Bill. He bought it.

Rec Subj Exist ObjIndObj

NonAdv

HeadN

Total

John 100 80 50 80 310

Integra 100 50 50 80 280

dealership 100 50 80 230

Example (cont’d)

Referent Phrases Value

John {John} 155

Integra {a beautiful Acura Integra} 140

dealership {the dealership} 115

Example (cont’d)

Referent Phrases Value

John {John, he1} 465

Integra {a beautiful Acura Integra} 140

dealership {the dealership} 115

Example (cont’d)

Referent Phrases Value

John {John, he1} 465

Integra {a beautiful Acura Integra, it} 420

dealership {the dealership} 115

Example (cont’d)

Referent Phrases Value

John {John, he1} 465

Integra {a beautiful Acura Integra, it} 420

Bill {Bill} 270

dealership {the dealership} 115

Example (cont’d)

Referent Phrases Value

John {John, he1} 232.5

Integra{a beautiful Acura Integra, it1} 210

Bill {Bill} 135

dealership {the dealership} 57.5

Observations

• Lappin & Leass - tested on computer manuals - 86% accuracy on unseen data.

• Centering (Grosz, Josh, Weinstein): additional concept of a “center” – at any time in discourse, an entity is centered.

• Backwards looking center; forward looking centers (a set).

• Centering has not been automatically tested on actual data.

Part of speech tagging

• Problems: transport, object, discount, address• More problems: content• French: est, président, fils• “Book that flight” – what is the part of speech

associated with “book”?• POS tagging: assigning parts of speech to words

in a text.• Three main techniques: rule-based tagging,

stochastic tagging, transformation-based tagging

Rule-based POS tagging

• Use dictionary or FST to find all possible parts of speech

• Use disambiguation rules (e.g., ART+V)

• Typically hundreds of constraints can be designed manually

Example in French

<S> ^ beginning of sentence

La rf b nms u article

teneur nfs nms noun feminine singular

Moyenne jfs nfs v1s v2s v3s adjective feminine singular

en p a b preposition

uranium nms noun masculine singular

des p r preposition

rivi`eres nfp noun feminine plural

, x punctuation

bien_que cs subordinating conjunction

délicate jfs adjective feminine singular

À p preposition

calculer v verb

Sample rules

BS3 BI1: A BS3 (3rd person subject personal pronoun) cannot be followed by a BI1 (1st person indirect personal pronoun). In the example: ``il nous faut'' ({\it we need}) - ``il'' has the tag BS3MS and ``nous'' has the tags [BD1P BI1P BJ1P BR1P BS1P]. The negative constraint ``BS3 BI1'' rules out ``BI1P'', and thus leaves only 4 alternatives for the word ``nous''.

N K: The tag N (noun) cannot be followed by a tag K (interrogative pronoun); an example in the test corpus would be: ``... fleuve qui ...'' (...river, that...). Since ``qui'' can be tagged both as an ``E'' (relative pronoun) and a ``K'' (interrogative pronoun), the ``E'' will be chosen by the tagger since an interrogative pronoun cannot follow a noun (``N'').

R V:A word tagged with R (article) cannot be followed by a word tagged with V (verb): for example ``l' appelle'' (calls him/her). The word ``appelle'' can only be a verb, but ``l''' can be either an article or a personal pronoun. Thus, the rule will eliminate the article tag, giving preference to the pronoun.

Confusion matrix

IN JJ NN NNP RB VBD VBN

IN - .2 .7

JJ .2 - 3.3 2.1 1.7 .2 2.7

NN 8.7 - .2

NNP .2 3.3 4.1 - .2

RB 2.2 2.0 .5 -

VBD .3 .5 - 4.4

VBN 2.8 2.6 -

Most confusing: NN vs. NNP vs. JJ, VBD vs. VBN vs. JJ

HMM Tagging

• T = argmax P(T|W), where T=t1,t2,…,tn

• By Bayes’s theorem: P(T|W) = P(T)P(W|T)/P(W)• Thus we are attempting to choose the sequence

of tags that maximizes the rhs of the equation• P(W) can be ignored• P(T)P(W|T) = ?• P(T) is called the prior, P(W|T) is called the

likelihood.

HMM tagging (cont’d)

• P(T)P(W|T) = P(wi|w1t1…wi-1ti-1ti)P(ti|t1…ti-2ti-1)

• Simplification 1: P(W|T) = P(wi|ti)

• Simplification 2: P(T)= P(ti|ti-1)

• T = argmax P(T|W) = argmax P(wi|ti) P(ti|ti-

1)

Estimates

• P(NN|DT) = C(DT,NN)/C(DT)=56509/116454 = .49

• P(is|VBZ = C(VBZ,is)/C(VBZ)=10073/21627=.47

Example

• Secretariat/NNP is/VBZ expected/VBN to/TO race/VB tomorrow/NR

• People/NNS continue/VBP to/TO inquire/VB the/AT reason/NN for/IN the/AT race/NN for/IN outer/JJ space/NN

• TO: to+VB (to sleep), to+NN (to school)

Example

NNP VBZ VBN TO VB NR

Secretariat is expected race tomorrowto

NNP VBZ VBN TO NN NR

Secretariat is expected race tomorrowto

Example (cont’d)

• P(NN|TO) = .00047• P(VB|TO) = .83• P(race|NN) = .00057• P(race|VB) = .00012• P(NR|VB) = .0027• P(NR|NN) = .0012• P(VB|TO)P(NR|VB)P(race|VB) = .00000027• P(NN|TO)P(NR|NN)P(race|NN) = .00000000032

Decoding

• Finding what sequence of states is the source of a sequence of observations

• Viterbi decoding (dynamic programming) – finding the optimal sequence of tags

• Input: HMM and sequence of words, output: sequence of states