Question Answering (and Textual...

Preview:

Citation preview

QuestionAnswering(andTextualEntailment)

Prof.SameerSinghCS295:STATISTICALNLP

WINTER2017

March14,2017

BasedonslidesfromDanJurafsky,YejinChoi,StephenClark,DanKlein,Niranjan Balasubramanian,andeveryoneelsetheycopiedfrom.

Upcoming…

• Homework4wasdueonlastnight• Lowestgradeofthehomeworks willbedroppedHomework

• Finalreportdueinaweek:March20,2017• Instructionscomingsoon:ACLstyle,5pages(+references)Project

• Papersummariesduetonight• Summary2gradedSummaries

CS295:STATISTICALNLP(WINTER2017) 2

TA/InstructorEvaluationsareavailable!

Outline

QuestionAnswering

TextualEntailment

CS295:STATISTICALNLP(WINTER2017) 3

IR-BasedQASystem

OtherExtensions

Outline

QuestionAnswering

TextualEntailment

CS295:STATISTICALNLP(WINTER2017) 4

IR-BasedQASystem

OtherExtensions

QuestionsinModernSystemsFactoidquestions◦ Whowrote“TheUniversalDeclarationofHumanRights”?◦ Howmanycaloriesarethereintwoslicesofapplepie?◦ Whatistheaverageageoftheonsetofautism?◦ WhereisAppleComputerbased?

Complex(narrative)questions:◦ Inchildrenwithanacutefebrileillness,whatistheefficacyofacetaminopheninreducingfever?

◦ WhatdoscholarsthinkaboutJefferson’spositionondealingwithpirates?

CS295:STATISTICALNLP(WINTER2017) 5

Commercialsystems:mainlyfactoidquestions

WhereistheLouvreMuseumlocated? InParis,France

What’stheabbreviation forlimitedpartnership? L.P.

What arethenamesofOdin’sravens? Huginn andMuninn

What currencyisusedinChina? Theyuan

Whatkindofnutsareusedinmarzipan? almonds

WhatinstrumentdoesMaxRoachplay? drums

WhatisthetelephonenumberforStanfordUniversity? 650-723-2300

CS295:STATISTICALNLP(WINTER2017) 6

ParadigmsforQA

IR-basedapproaches◦ TREC;IBMWatson;Google

Knowledge-basedandHybridapproaches◦ IBMWatson;AppleSiri;WolframAlpha;TrueKnowledgeEvi

CS295:STATISTICALNLP(WINTER2017) 7

Manyquestionscanalreadybeansweredbywebsearch

CS295:STATISTICALNLP(WINTER2017) 8

IR-basedQuestionAnswering

CS295:STATISTICALNLP(WINTER2017) 9

IR-basedFactoidQA

CS295:STATISTICALNLP(WINTER2017) 10

DocumentDocumentDocument

DocumentDocume

ntDocumentDocume

ntDocument

Question Processing

PassageRetrieval

Query Formulation

Answer Type Detection

Question

Passage Retrieval

Document Retrieval

Answer Processing

Answer

passages

Indexing

RelevantDocs

DocumentDocumentDocument

Knowledge-basedQA(Siri)Buildasemanticrepresentationofthequery◦ Times,dates,locations,entities,numericquantities

Mapfromthissemanticstoquerystructureddataorresources◦ Geospatialdatabases◦ Ontologies(Wikipediainfoboxes,dbPedia,WordNet,Yago)◦ Restaurantreviewsourcesandreservationservices◦ Scientificdatabases

CS295:STATISTICALNLP(WINTER2017) 11

Hybridapproaches(Watson)Buildashallowsemanticrepresentationofthequery

GenerateanswercandidatesusingIRmethods◦ Augmentedwithontologiesandsemi-structureddata

Scoreeachcandidateusingricherknowledgesources◦ Geospatialdatabases◦ Temporalreasoning◦ Taxonomicalclassification

CS295:STATISTICALNLP(WINTER2017) 12

IBM’sWatson

WonJeopardy onFebruary16,2011!

13

WILLIAMWILKINSON’S“ANACCOUNTOFTHEPRINCIPALITIESOF

WALLACHIAANDMOLDOVIA”INSPIREDTHISAUTHOR’SMOSTFAMOUSNOVEL

BramStoker

CS295:STATISTICALNLP(WINTER2017)

MotivationforWatson

CS295:STATISTICALNLP(WINTER2017) 14

SingleSourceisnotSufficient

CS295:STATISTICALNLP(WINTER2017) 15

WatsonArchitecture

CS295:STATISTICALNLP(WINTER2017) 16

. . .

Models

Answer & Confidence

Question

Evidence

Sources

Models

Models

Models

Models

ModelsPrimarySearch

CandidateAnswer

Generation

Answer Sources Evidence

RetrievalEvidence Scoring

Learned Modelshelp combine and

weigh the Evidence

Hypothesis and Evidence Scoring

Merging &

Ranking

Synthesis

WatsonPerformance

CS295:STATISTICALNLP(WINTER2017) 17

Outline

QuestionAnswering

TextualEntailment

CS295:STATISTICALNLP(WINTER2017) 18

IR-BasedQASystem

OtherExtensions

IR-basedFactoidQA

CS295:STATISTICALNLP(WINTER2017) 19

DocumentDocumentDocument

DocumentDocume

ntDocumentDocume

ntDocument

Question Processing

PassageRetrieval

Query Formulation

Answer Type Detection

Question

Passage Retrieval

Document Retrieval

Answer Processing

Answer

passages

Indexing

RelevantDocs

DocumentDocumentDocument

IR-basedFactoidQAQUESTIONPROCESSING◦ Detectquestiontype,answertype,focus,relations◦ Formulatequeriestosendtoasearchengine

PASSAGERETRIEVAL◦ Retrieverankeddocuments◦ Breakintosuitablepassagesandrerank

ANSWERPROCESSING◦ Extractcandidateanswers◦ Rankcandidates

◦ usingevidencefromthetextandexternalsources

CS295:STATISTICALNLP(WINTER2017) 20

FactoidQ/A

CS295:STATISTICALNLP(WINTER2017) 21

DocumentDocumentDocument

DocumentDocume

ntDocumentDocume

ntDocument

Question Processing

PassageRetrieval

Query Formulation

Answer Type Detection

Question

Passage Retrieval

Document Retrieval

Answer Processing

Answer

passages

Indexing

RelevantDocs

DocumentDocumentDocument

QuestionProcessingAnswerTypeDetection◦ Decidethenamedentitytype(person,place)oftheanswer

QueryFormulation◦ ChoosequerykeywordsfortheIRsystem

QuestionTypeclassification◦ Isthisadefinitionquestion,amathquestion,alistquestion?

FocusDetection◦ Findthequestionwordsthatarereplacedbytheanswer

RelationExtraction◦ Findrelationsbetweenentitiesinthequestion

CS295:STATISTICALNLP(WINTER2017) 22

QuestionProcessing

AnswerType:USstate

Query:twostates,border,Florida,north

Focus:thetwostates

Relations:borders(Florida,?x,north)

CS295:STATISTICALNLP(WINTER2017) 23

They’rethetwostatesyoucouldbereenteringifyou’recrossingFlorida’snorthernborder

AnswerTypes:NamedEntities

WhofoundedVirginAirlines?PERSON

WhatCanadiancityhasthelargestpopulation?CITY

CS295:STATISTICALNLP(WINTER2017) 24

PartofAnswerTypeTaxonomy

CS295:STATISTICALNLP(WINTER2017) 25

LOCATION

NUMERIC

ENTITY HUMAN

ABBREVIATIONDESCRIPTION

country city state

datepercent

money

sizedistance

individual

title

group

food

currency

animal

definition

reason expression

abbreviation

Li,Roth.LearningQuestionClassifiers.COLING(2002)

AnswerTypes

CS295:STATISTICALNLP(WINTER2017) 26

MoreAnswerTypes

CS295:STATISTICALNLP(WINTER2017) 27

AnswertypesinWatson

2500answertypesin20,000Jeopardyquestionsample◦ Themostfrequent200answertypescover<50%ofdata

The40mostfrequentJeopardyanswertypes

he,country,city,man,film,state,she,author,group,here,company,president,capital,star,novel,character,woman,river,island,king,song,part,series,sport,singer,actor,play,team,show,actress,animal,presidential,composer,musical,nation,book,title,leader,game

CS295:STATISTICALNLP(WINTER2017) 28Ferrucci etal.BuildingWatson:AnOverviewoftheDeepQA Project.AIMagazine.2010

AnswerTypeDetection

CS295:STATISTICALNLP(WINTER2017) 29

Hand-writtenRules

Regularexpression-basedrulescangetsomecases:◦ Who{is|was|are|were}PERSON

Otherrulesusethequestionheadword:◦ (theheadwordofthefirstnounphraseafterthewh-word)◦ WhichcityinChinahasmostforeignfinancialcompanies?◦ WhatisthestateflowerofCalifornia?

MachineLearning

QuestionwordsandphrasesPart-of-speechtagsParsefeatures(headwords)NamedEntitiesSemanticallyrelatedwords

FactoidQ/A

CS295:STATISTICALNLP(WINTER2017) 30

DocumentDocumentDocument

DocumentDocume

ntDocumentDocume

ntDocument

Question Processing

PassageRetrieval

Query Formulation

Answer Type Detection

Question

Passage Retrieval

Document Retrieval

Answer Processing

Answer

passages

Indexing

RelevantDocs

DocumentDocumentDocument

KeywordSelectionAlgorithm1.Selectallnon-stopwordsinquotations2.SelectallNNPwordsinrecognizednamedentities3.Selectallcomplexnominals withtheiradjectivalmodifiers4.Selectallothercomplexnominals5.Selectallnounswiththeiradjectivalmodifiers6.Selectallothernouns7.Selectallverbs8.Selectalladverbs9.SelecttheQFWword(skippedinallprevioussteps)10.Selectallotherwords

CS295:STATISTICALNLP(WINTER2017) 31Moldovan,Harabagiu,Pasca,Mihalcea,Goodrum,GirjuandRus.TREC(1999)

Choosingkeywords

CS295:STATISTICALNLP(WINTER2017) 32

Whocoinedtheterm“cyberspace”inhisnovel“Neuromancer”?

1 1

4 4

7

cyberspace/1Neuromancer/1term/4novel/4coined/7

FactoidQ/A

CS295:STATISTICALNLP(WINTER2017) 33

DocumentDocumentDocument

DocumentDocume

ntDocumentDocume

ntDocument

Question Processing

PassageRetrieval

Query Formulation

Answer Type Detection

Question

Passage Retrieval

Document Retrieval

Answer Processing

Answer

passages

Indexing

RelevantDocs

DocumentDocumentDocument

PassageRetrieval

RetrievedocumentsusingIR◦ querytermsaskeywords

CS295:STATISTICALNLP(WINTER2017) 34

Step1

Step2Segmentthedocumentsintoshorterunits◦ somethinglikeparagraphs

Step3Passageranking◦ Useanswertypetohelprerank passages

FeaturesforPassageRanking

NumberofNamedEntitiesoftherighttypeinpassage

Numberofquerywordsinpassage

NumberofquestionN-gramsalsoinpassage

Proximityofquerykeywordstoeachotherinpassage

Longestsequenceofquestionwords

Rankofthedocumentcontainingpassage

CS295:STATISTICALNLP(WINTER2017) 35

Eitherinrule-basedclassifiersorwithsupervisedmachinelearning

FactoidQ/A

CS295:STATISTICALNLP(WINTER2017) 36

DocumentDocumentDocument

DocumentDocume

ntDocumentDocume

ntDocument

Question Processing

PassageRetrieval

Query Formulation

Answer Type Detection

Question

Passage Retrieval

Document Retrieval

Answer Processing

Answer

passages

Indexing

RelevantDocs

DocumentDocumentDocument

AnswerExtractionRunananswer-typenamed-entitytaggeronthepassages◦ Eachanswertyperequiresanamed-entitytaggerthatdetectsit◦ IfanswertypeisCITY,taggerhastotagCITY◦ CanbefullNER,simpleregularexpressions,orhybrid

Returnthestringwiththerighttype:◦ WhoistheprimeministerofIndia(PERSON)◦ ManmohanSingh,PrimeMinisterofIndia,hadtoldleftleadersthatthedealwouldnotberenegotiated.

◦ HowtallisMt.Everest?(LENGTH)◦ TheofficialheightofMountEverestis29035feet

CS295:STATISTICALNLP(WINTER2017) 37

RankingCandidateAnswersButwhatiftherearemultiplecandidateanswers!

Q:WhowasQueenVictoria’ssecondson?

AnswerType:Person

CS295:STATISTICALNLP(WINTER2017) 38

• Passage:TheMariebiscuitisnamedafterMarieAlexandrovna,thedaughterofCzarAlexanderIIofRussiaandwifeofAlfred,thesecondsonofQueenVictoriaandPrinceAlbert

RankingCandidateAnswersButwhatiftherearemultiplecandidateanswers!

Q:WhowasQueenVictoria’ssecondson?

AnswerType:Person

CS295:STATISTICALNLP(WINTER2017) 39

• Passage:TheMariebiscuitisnamedafterMarieAlexandrovna,thedaughterofCzarAlexanderIIofRussiaandwifeofAlfred,thesecondsonofQueenVictoriaandPrinceAlbert

FeaturesforMLAnswertypematch:Candidatecontainsaphrasewiththecorrectanswertype.

Patternmatch:Regularexpressionpatternmatchesthecandidate.

Questionkeywords:#ofquestionkeywordsinthecandidate.

Keyworddistance:Distanceinwordsbetweenthecandidateandquerykeywords

Noveltyfactor:Awordinthecandidateisnotinthequery.

Appositionfeatures:Thecandidateisanappositivetoquestionterms

Punctuationlocation:Thecandidateisimmediatelyfollowedbyacomma,period,quotationmarks,semicolon,orexclamationmark.

Sequencesofquestionterms:Thelengthofthelongestsequenceofquestiontermsthatoccursinthecandidateanswer.

CS295:STATISTICALNLP(WINTER2017) 40

ScoringCandidatesinWatsonEachcandidateanswergetsscoresfrom>50components◦ (fromunstructuredtext,semi-structuredtext,triplestores)◦ logicalform(parse)matchbetweenquestionandcandidate◦ passagesourcereliability◦ geospatiallocation◦ Californiais”southwestofMontana”

◦ temporalrelationships◦ taxonomicclassification

CS295:STATISTICALNLP(WINTER2017) 41

Outline

QuestionAnswering

TextualEntailment

CS295:STATISTICALNLP(WINTER2017) 42

IR-BasedQASystem

OtherExtensions

AskMSR

CS295:STATISTICALNLP(WINTER2017) 43

1 2

3

45

QuestionProcessing Search

AnswerExtractionAnswerScoring

Dumais, Banko, Brill, Lin, Ng,SIGIR(2002)

Step1:RewriteQueries

CS295:STATISTICALNLP(WINTER2017) 44

Intuition: Questions are often syntactically quite closeto sentences with the answer• Where istheLouvreMuseumlocated?

• TheLouvreMuseumislocated in Paris• Who createdthecharacterofScrooge?

• Charles DickenscreatedthecharacterofScrooge

FeedbackLoops:FALCON

CS295:STATISTICALNLP(WINTER2017) 45

AllenAIScienceChallenge

CS295:STATISTICALNLP(WINTER2017) 46

Whichobjectisthebestconductorofelectricity?(A) awaxcrayon(B)aplasticspoon(C)arubbereraser(D)anironnail

AllenAIScienceChallenge

CS295:STATISTICALNLP(WINTER2017) 47

Whichobjectisthebestconductorofelectricity?(A) awaxcrayon(B)aplasticspoon(C)arubbereraser(D)anironnail

AllenAIScienceChallenge

CS295:STATISTICALNLP(WINTER2017) 48

Fourthgradersareplanningaroller-skaterace.Whichsurfacewouldbethebestforthisrace?(A)gravel(B)sand(C)blacktop(D)grass

AllenAIScienceChallenge

CS295:STATISTICALNLP(WINTER2017) 49

Fourthgradersareplanningaroller-skaterace.Whichsurfacewouldbethebestforthisrace?(A)gravel(B)sand(C)blacktop (D)grass

§ Informationretrievalmethodsfail§ Wordco-occurrencemethodsstruggle

Graders are commonly used in the construction and maintenance of dirt roads and gravel roads

Also strong correlations between:grass « racegravel « surface

AllenAIScienceChallenge

CS295:STATISTICALNLP(WINTER2017) 50

Astudentputstwoidenticalplantsinthesametypeandamountofsoil.Shegivesthemthesameamountofwater.Sheputsoneoftheseplantsnearasunnywindowandtheotherinadarkroom.Thisexperimenttestshowtheplantsrespondto

(A)light(B)air(C)water(D)soil

AllenAIScienceChallenge

CS295:STATISTICALNLP(WINTER2017) 51

Astudentputstwoidenticalplantsinthesametypeandamountofsoil.Shegivesthemthesameamountofwater.Sheputsoneoftheseplantsnearasunnywindowandtheotherinadarkroom.Thisexperimenttestshowtheplantsrespondto

(A)light (B)air(C)water(D)soil

Knowledge needed(forexample):§ nearsunnywindow® receivelight§ inadarkroom® nolight§ testX’sresponsetoY

® compareX+YwithX+not(Y)

Outline

QuestionAnswering

TextualEntailment

CS295:STATISTICALNLP(WINTER2017) 52

IR-BasedQASystem

OtherExtensions

NaturalLanguage&Meaning

CS295:STATISTICALNLP(WINTER2017) 53

Meaning

Language

AmbiguityVariability

interpretatio

n expression

InferencevsEntailment

CS295:STATISTICALNLP(WINTER2017) 54

MeaningRepresentation

Natural Language

Inference

TextualEntailment

TextualEntailment

CS295:STATISTICALNLP(WINTER2017) 55

• Adirectionalrelationbetweentwotextfragments:Text(t) andHypothesis(h):

t entails h (t Þ h) if humans reading t will infer that h is most likely true

• Assuming“commonbackgroundknowledge”–whichisindeedexpectedfromapplications

Example

CS295:STATISTICALNLP(WINTER2017) 56https://cogcomp.cs.illinois.edu/page/resource_view/9

MoreSentencePairs

CS295:STATISTICALNLP(WINTER2017) 57

1. Some students came to school by car.Some students came to school.

2. No students came to school by car. Some students came to school.

3. John drove legally. John drove.

4. John drove predictably.John drove.

5. Legally, John could drive.John drove.

EntailmentwithKnowledge

CS295:STATISTICALNLP(WINTER2017) 58

Fortextualentailmenttoholdwerequire:◦ text ANDknowledgeÞ h,but◦ knowledgeshouldnotentailh alone

Systemsare notsupposedtovalidateh’struthregardlessoft(e.g.bysearchinghontheweb)

t entails h (t Þ h) if humans reading t will infer that h is most likely true

Example

CS295:STATISTICALNLP(WINTER2017) 59

TEXT: …WhilenooneaccusesMadonnaofdoinganythingillegalinadoptingthe4-year-oldgirl,reportedlynamedMercy,therearequestionsnonethelessabouthowMadonnaisabletonavigateMalawi's18-to-24monthvettingperiodinjustamatterofdaysorweeks…

HYPOTHESIS:

Madonnais50yearsold.

ThirdLabel:Contradictions

CS295:STATISTICALNLP(WINTER2017) 60

t contradicts h (t ^ h) if humans reading t will find the relations/events described by h to be highly unlikely given t.

Example

CS295:STATISTICALNLP(WINTER2017) 61

MoreExamples

CS295:STATISTICALNLP(WINTER2017) 62

Applications

CS295:STATISTICALNLP(WINTER2017) 63

QuestionAnswering

InformationExtraction

MachineTranslation

InformationRetrieval

Question Expected answer formWho bought Overture? >> X bought Overture

Overture’s acquisitionby Yahoo

Yahoo bought Overture

text hypothesized answer

entails

Heabhorredthemen’sunctuousways.

Hedislikedthemen’sflatteringways.

Similaritybasedonwhetherdocumententailsthequery.

Recommended