Multi-lingual Concept Extraction with Linked Data and Human-in-the-Loop Alfredo Alba, Anni Coden, Anna Lisa Gentile, Daniel Gruhl, Petar Ristoski, Steve Welch IBM Research
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Multi-lingualConceptExtractionwithLinkedDataandHuman-in-the-Loop
AlfredoAlba,Anni Coden,AnnaLisaGentile,DanielGruhl,Petar Ristoski,SteveWelch
IBMResearch
Motivation
Motivation
§ extractinformationfromanovel corpus
§ whataretherelevantconcepts inthedomain?
§ limiteddomain andlanguage knowledge
§ IDEA:combinestatisticaltechniqueswithuser-in-the-loop
DomainLearningAssistant
• Startwithasmallnumberofseeds(1)
• Getsuggestionsofnewsurfaceforms
• Theuseraccept/reject
Findingconcept candidatesThesafetyandefficacyoffilgrastim aresimilarinadultsand childrenreceivingcytotoxicchemotherapy
Laeficacia ylaseguridad delfilgrastim sonsimilares en los adultos y en los niños tratados conquimioterapia citotóxica
Lasicurezza el’efficacia delfilgrastim sono simili negli adulti e nei bambinisottoposti achemioterapia citotossica
DieWirksamkeit undUnbedenklichkeit vonFilgrastim ist bei Erwachsenen undbei Kindern ,dieeine zytotoxische Chemotherapie erhalten ,vergleichbar
Findingconcept candidates
Plasmaeliminationhalf-lifeoforalpravastatin is1.5to2hours.
L’emivita plasmatica dieliminazione delpravastatin orale é compresa tra un’ora emezzoedueore.
Findingconcept candidatesCandidates:{eggs,flour}
“mixeggs andflour”àmix <candidate>and <candidate>
mix <candidate>and <candidate>à “mixsugarandbutter”
Candidates:{eggs,flour,sugar,butter}
“meltthebutter”àmeltthe<candidate>
…
Findingconcept candidatesCandidates:{uova,farina}
“amalgamare uova efarina”à amalgamare <candidate>e<candidate>
amalgamare <candidate>e<candidate>à “amalgamare zucchero eburro”
Candidates:{uova,farina,zucchero,burro}
“sciogliere il burro”à sciogliere il <candidate>
…
Multi-lingualexperimentHYPOTHESIS:samebehavior,regardlessofthelanguage
§ westartwithveryfewseeds(onecouldbesufficient)foreachlanguage§ weextractcontextpatternsandusethemtogeneratenewcandidates
§ weasktouser toaccept/reject thecandidates
§ werepeatforafixednumberofiterationsinalllanguages
Multi-lingualexperiment:DrugDiscovery§ DATA:parallelcorpusfromtheEuropeanMedicinesAgency(EMEA)§documentsrelatedtomedicinalproducts§translationsinto22officiallanguagesoftheEuropeanUnion§1,500documentsformostofthelanguages§weused4languages(en,es,it,de)
§ TASK:buildalexiconofclinicaldrugs
§user-in-the-loop simulatedbyconstructingaGoldStandard(GS)ofdrugsnamesextractedfromLinkedOpenData(weusedDBpediahttp://dbpedia.org)
DrugDiscovery:Oneseed
§ initialseeds:singleseed§Onedrugnamewhichappearsineachcorpus(e.g.“irbesartan”)
§ 20iterations
§ learningcurvesforalllanguagesarecomparable
Discovery growth for glimpse for English (en), Italian (it), Spanish (es) and German (de). Average correlation amongst all languages r = 0.998.
DrugDiscovery:LinkedDataseeds§ initialseeds:20%ofavailableLinkedData(DBpedia)§ 5-foldvalidation(randomlyselected20%,samedrugsforalllanguages)§ choiceofinitialseedsdoesnotimpactstheresults
Discovery growth with 5-fold cross validation on the EMEA dataset using DBpedia as seeds. Each plot shows the discovery growth for each of the randomly generated 5 folds and reports the Pearson correlation (r) amongst them.
DrugDiscovery:benefitofLinkedData§glimpseà onemanuallyprovidedseed
§glimpseLDàLinkedDataseeds
§in10iterationsglimpseLD cancoverthesamelexiconthatwouldtakemorethan20iterationswithglimpse
Human-in-the-loopexperimentwithasubjectmatterexpert(physician)
Multi-lingualexperiment:Colors§ DATA:Twitterstream1st-14thofJanuary2016– lang:En,De,Es, It§containatleastonementionofacolor§ goldstandardlistsofcolorsfromWikidata andDbpedia
§ balancedatasetssizeindifferentlanguages§ 155,828tweetsperlanguage
§ TASK:expandthelexiconofcolors
§ user-in-the-loop: 4nativespeakers,10iterations
Multi-lingualexperiment:Colors§ newcoloritemsextractedfromTwitterdata:§German:5§ Italian:5§English:19§Spanish:22§azulgrana§ rojo vivo§ “limn"(inplaceofthecolorlímon)
ConclusionsWHAT§ knowledgeresourcesarenevercomplete/exhaustive
§ construct/improvedictionariesfromtextcorpora
HOW§ iterativeandpurelystatistical algorithm§ nofeatureextractionrequired§ comparablebehaviorfor differentlanguages
§ organicallyincorporateshumanfeedback
Multi-lingualConceptExtractionwithLinkedDataandHuman-in-the-Loop
IBMResearch
[email protected] @AnLiGentile
AlfredoAlba,Anni Coden,AnnaLisaGentile,DanielGruhl,Petar Ristoski,SteveWelch
Related Documents
