RESEARCHOpenAccessACRF-basedsystemforrecognizingchemicalentitymentions - PDF document

RESEARCH OpenAccess ACRF-basedsystemforrecognizingchemical entitymentions(CEMs)inbiomedicalliterature ShuoXu 1 ,XinAn 2 ,LijunZhu 1 ,YunliangZhang 1* ,HaodongZhang 3 Abstract Background: Inordertoimproveinformationaccessonchemicalcompoundsanddrugs(chemicalentities) describedintextrepositories,itisverycrucialtobeabletoidentifychemicalentitymentions(CEMs)automatically correspondingsystemsthatareabletodetectmentionsofchemicalcompoundsanddrugs,whichhastwo subtasks:CDI(ChemicalDocumentIndexing)andCEM. Results: Oursystemprocessingpipelineconsistsofthreemajorcomponents:pre-processing(sentencedetection, tokenization),recognition(CRF-basedapproach),andpost-processing(rule-basedapproachandformatconversion). Inourpost-challengesystem,thecostparameterinCRFmodelwasoptimizedby10-foldcrossvalidationwithgrid search,andwordrepresentationsfeatureinducedbyBrownclusteringmethodwasintroduced.FortheCEM subtask,ourofficialrunswererankedintoppositionbyobtainingmaximum88.79%precision,69.08%recalland 77.70%balancedF-measure,whichwereimprovedfurtherto88.43%precision,76.48%recalland82.02%balanced F-measureinourpost-challengesystem. Conclusions: Inoursystem,insteadofextractingaCEMasawhole,weregardeditasasequencelabeling problem.Thoughourcurrentsystemhasmuchroomforimprovement,oursystemisvaluableinshowingthatthe performanceintermofbalancedF-measurecanbeimprovedlargelybyutilizinglargeamountsofrelatively inexpensiveun-annotatedPubMedabstractsandoptimizingthecostparameterinCRFmodel.Fromourpractice andlessons,ifonedirectlyutilizessomeopen-sourcenaturallanguageprocessing(NLP)toolkits,suchasOpenNLP, StandfordCoreNLP,falsepositive(FP)ratemaybeveryhigh.Itisbettertodevelopsomeadditionalrulesto minimizetheFPrateifonedoesnotwanttore-traintherelatedmodels.OurCEMrecognitionsystemisavailable at:http://www.SciTeMiner.org/XuShuo/Demo/CEM. Background Thereisanincreasinginteresttoimproveinformation accessonchemicalcompoundsanddrugs(chemicalenti- ties)describedintextrepositories,includingscientific articles,patents,healthagencyreports,ortheWeb[1].In ordertoachievethisgoal,itisverycrucialtobeableto identifychemicalentitymentions(CEMs)automatically fchemicalentitiesisalso crucialforothersubsequenttextprocessingtasks,such asdetectionofdrug-protei ninteractions[2],adverse effectsofchemicalcompoundsandtheirassociationsto toxicologicalendpoints,ortheextractionofpathwayand metabolicreactionrelationsandsoon.Thoughmany methodsandstrategiestorecognizechemicalsintext havebeenproposed[3],onlyaverylimitednumberof publiclyaccessibleCEMreco gnitionsystemshavebeen released[4]. TheBioCreative(CriticalAssessmentofInformation ExtractionSystemsinBiology)challengeisacommunity- wideefforttobuildanevaluationframeworkforasses- singtextminingsystemsinbiologicaldomains[5].The chemicalcompoundanddrugnamedentityrecognition (CHEMDNER)challengeinBioCreativeIVwasspecially designedtopromotetheimplementationofsystemsthat areabletodetectmentionsofchemicalcompoundsand drugs,whichhastwosubtasks,CDI(ChemicalDocument Indexing)subtaskandCEM(ChemicalEntityMention) subtask.CDIsubtaskisthetasktoreturnarankedlistof 1 InformationTechnologySupportingCenter,InstituteofScientificand TechnicalInformationofChina,No.15FuxingRd.,HaidianDistrict,100038 Beijing,PRChina Fulllistofauthorinformationisavailableattheendofthearticle Xu etal . JournalofCheminformatics 2015, 7 (Suppl1):S11 http://www.jcheminf.com/content/7/S1/S11 ©2015Xuetal.;licenseeSpringer.ThisisanOpenAccessarticledistributedunderthetermsoftheCreativeCommonsAttribution License(http://creativecommons.org/licenses/by/4.0),whichpermitsunrestricteduse,distribution,andreproductioninanymedium, providedtheoriginalworkisproperlycited.TheCreativeCommonsPublicDomainDedicationwaiver(http://creativecommons.org/ publicdomain/zero/1.0/)appliestothedatamadeavailableinthisarticle,unlessotherwisestated. chemicalentitiesdescribedwithinagivendocuments. CEMsubtaskisthetasktoprovideforagivendocument thestartandendindicescorrespondingtoallthechemi- calentitiesmentionedinthedocument. Here,wepresentthemethod,theresultsandrecogni- tionsystemfromourparticipationintheCEMsubtask ofCHEMDNERchallenge[1,6]withsomepostchallenge systemsimprovement.Inourrecognitionsystem, insteadofextractingaCEMsuchas “ (+)-antiBP-7,8- diol-9,10-epoxide ” asawhole,weregarditasa sequencelabelingproblem.Ourmainfocusonthis improvedsystemwastoexploretheeffectivenessofcost parameteroptimization[7,8]andwordrepresentation-s [9-11]featureforourapproachtoCEMsubtask.The proposedmethodcombinesnaturallanguageprocessing (NLP)strategieswithmachinelearning(ML)techniques toutilizewordrepresentationsfeaturefromlarge amountsofrelativelyinexpensiveun-annotatedPubMed abstractsalongwithsmallamountsofannotatedones. AsshowninFigure1,oursystemfirstdetectssen- tenceboundariesonthePubMedabstracts,andthen tokenizeseachdetectedsentenceaspre-processing. Next,oursystemextractsCEMsfromtextwithacondi- tionalrandomfield(CRF)approach[12],followedby somepost-processingstep sincludingarule-based approachandaformatconversionstep.Wedescribe eachstepindetailinthefollo wingsections.Although currentapproachhasmuch roomforimprovement,it producedthetop-rankedperformanceamongallsub- mittedrunsintheCEMsubtaskofBioCreativeIV CHEMDNERchallenge. Theorganizationoftherestofthearticleisasfollows. Inthenextsection,wedescribetheresultsofoursub- missionandpost-challengerunsontheCEMsubtaskof BioCreativeIVCHEMDNERchallenge.Thisisfollowed bydiscussionandconclusionsdrawnfromourexperi- ence.Lastly,ourmethodsemployedareexplainedin detail. Resultsanddiscussion Weanalyzedthetraining,developmentandtestingdata setsandfoundthattherearemanynestedCEMsinthe developmentset,suchas “ polysorbate80 ” (offset:1138 to1152)and “ polysorbate ” (offset:1138to1149)inthe abstractofPMID:23064325.SeeTable1formore examplesofnestedCEMpairs.SincelinearCRFmodel, utilizedinthisarticle,cannotidentifythenestedCEMs, wejustomitthelessspannedCEMs.Inaddition,there maybesomeannotationerrorsinthedevelopmentset, suchasexamplesinTable2.Wealsomanuallycor- rectedtheseerrorsbeforetrainingourCRFmodel. Table3showsabriefoverviewofthecorrected CHEMDNERcorpus.Pleasesee[13]formoredetailsof CEMsannotating,classifyingandsplittingintotraining, developmentandtestdatasets. Toevaluatetheperformanceofsubmittedresults,the BioCreativeIVcompetitionreliedonthreeperformance measuresatentitylevel:recall,precisionandF-measure. Therecallistheproportionofcorrectpredictionof positiveCEMs.Theprecisionistheproportionofpre- dictedCEMsthatareactuallytrueCEMs.TheF-mea- sureprovidesamorebalancedevaluationbyaveraging precisionandrecall.Therecall,precisionandF-measure aredefinedformallyasfollows. r = TP TP + FN (1) p = TP TP + FP (2) F  =(1+  2 ) p × r  2 p + r (3) where TP (truepositive)isthenumberofthecorrect positivepredictions, FN (falsenegative)isthenumberof incorrectnegativepredictions(typeIIerrors),and FP is thenumberofincorrectpositivepredictions(typeI errors).ThebalancedF-measure( b =1),themaineva- luationmetricusedfortheCEMsubtaskoftheBioCrea- tiveIVCHEMDNERcompetition,canbesimplifiedto: F 1 =2 p × r p + r (4) Inordertomakethebestofannotatedcorpus,wepooled thetraininganddevelopmentdatasets.Theparticipating teamsareallowedtohave5daystogenerateuptofivedif- ferentannotations("runs ” )forthetestsetandtosubmitthe annotationstotheorganizers .Thus,participatingteamscan utilizedifferentsettings,modelsormethodswhengoldtest Figure1 Thesystemprocessingpipeline .Thesystemprocessingpipelinethatincludesthreemajorcomponents:pre-processing(sentence detection,tokenization),recognition(CRF-basedapproach)andpost-processing(rule-basedapproachandformatconversion). Xu etal . JournalofCheminformatics 2015, 7 (Suppl1):S11 http://www.jcheminf.com/content/7/S1/S11 Page2of9 annotationsetisunknown.WesubmittedfiverunsfortheCEMsubtask,eachusingthesamepipeline,butwithdiffer-entvaluesforthecostparameterintheCRFmodel[12,14].Duetotimeconstraints,wejustsetthecostparametertoeachelementin{2,2,2}.Table4presentstheofficialperformancescoresofoursubmittedruns.Run5performedthebestintermsofrecallandbalancedF-mea-sure.Run1performedthebestintermofprecision. Table1NestedCEMpairsinthedevelopmentsetoftheCHEMDNERcorpusA567A567Foreachrow,theCEMwithoffsetincolumn6-7isnestedintheCEMwithoffsetincolumn4-5.TheCEMswithrespectiveoffsetsincolumn6-7areomitteddirectlywhentrainingourCRFmodels.etalJournalofCheminformatics(Suppl1):S11http://www.jcheminf.com/content/7/S1/S11Page3of9 Infact,thecostparametertradesthebalancebetweenover-fittingandunder-fitting[12,14].Withlargercostparametervalue,CRFtendstoover-fittothegiventrainingcorpus.FromTable4,onecaneasilyseethatthepredictedresultsweresignificantlyinfluencedbythisparameter.Inourpost-challengeimprovedsystems,10-foldcrossvalidationatdocumentlevelisutilizedtooptimizethecostparameterwithgridsearch[7,8].Spe-cifically,thepooledtraininganddevelopmentdatasetsarerandomlydividedinto10sub-corpusofnearlyequalsize.Foreachcost,2,2},aCRFmodelisinduced10times,eachtimeleavingoutoneofthesub-corpusesthatisthenusedtocalculatethebalancedF-measure.Anoptimalvalueofcostsisselectedfromthisgridsearch.Inourpost-challengeimprovedsystem,wereobtainedfiverunsfortheCEMsubtask,eachusingthesamepipelineasofficialsubmissions,butwithdifferentfea-turessets(Table5).FromTable3,CHEMDNERcorpusincludeslargeamountsofrelativelyinexpensiveun-annotatedPubMedabstracts.Inordertoreducedatasparsityandimprovefurthertheperformanceofoursystem,wordrepresentationsfeatureisusedinourpost-challengesystem,sinceitisasimpleandgeneralmethodforsemi-supervisedlearning[11].Previousstu-dies[11,15,16]showthatwordrepresentationsfeatureisaveryimportantfeaturetoimprovethebalancedF-measureofpre-definedcategoriesofpropernamesandbio-entityrecognition.Here,thetraining,development,testandbackgrounddatasetsarepooledtoinducewordrepresentationsofeachtokenbyBrownclusteringmethod[10,17]with500,1000,1500and2000clusters,respectively.Figure2showsthebalancedF-measureforpostchallengerunswith10-foldcrossvalidationbygridsearch[7,8].Table6reportstheperformanceresultswiththeoptimalvalueforthecostparameter.FromFigure2andbycomparingTable4andTable6,itisnotdifficulttoseethatthewordrepresentationsfeatureimprovedlargelytheperfor-manceofoursystemintermsofbalancedF-measureandrecall,butwithalittleperformancedegradationintermofprecision.Run1,Run4andRun3performedthebestintermofprecision,recall,balancedF-measure,ThoughtheannotatedCEMsareclassifiedintoeight={SYSTEMATIC,IDENTIFIER,FORMULA,TRIVIAL,ABBREVIATION,FAMILY,MULTIPLE,NOCLASS},theannotationsoftheindividualCEMclassesaredisregardedinourpost-challengesystem.InordertohighlighttheexistinggapsintheCEMrecognitionsystem,performanceresultsforeachcategoryinCarealsogiveninTable4andTable6intermofprecision.AsforofficialperformancescoresinTable4,oursystemworkedbestonrecognizingtheFORMULACEMsforRun1,Run2andRun3,andSYSTEMATICCEMsforRun4andRun5.FromTable6,onecanseethatour Table2NestedCEMpairsinthedevelopmentsetoftheCHEMDNERcorpusIDPMIDT/AStartEndStartEnd123412114A977984977985223572392T42554256323414800T69896889423411224A278288277288523401298A438502438501Theoffsetsincolumn4-5arecorrectedtotheonesincolumn6-7. Table3TheoverviewofthecorrectedCHEMDNERcorpusintermsofthenumberofPubMedabstracts(#Articles),thenumberofCEMs(#CEMs),andthenumberofCEMsforeachoftheCEMclassesinC={SYSTEMATIC,IDENTIFIER,FORMULA,TRIVIAL,ABBREVIATION,FAMILY,MULTIPLE,NOCLASS}×meanstheresultingfigureisTrainingDevelopmentTestBackground#Articles3,5003,5003,00017,000#CEMs29,47829,48525,351×ABBREVIATION4,5384,5174,059×FAMILY4,0904,2123,622×FORMULA4,4484,1173,443×IDENTIFIER672639513×MULTIPLE202187199×SYSTEMATIC6,6566,8145,666×TRIVIAL8,8328,9677,808×NOCLASS403241× Table4OfficialscoresfortheCEMsubtaskintheBioCreativeIVCHEMDNERcompetitionRun1Run2Run3Run4Run5cost222-TP15,82116,53116,99117,32817,5121,8342,0072,0092,1292,2119,5308,8208,3608,0237,839Precision(%)89.6189.1789.4389.0688.79Recall(%)62.4165.2167.0268.3569.08score(%)73.5875.3376.6277.3477.70ABBREVIATION53.9055.7556.7457.6358.24FAMILY59.5062.2064.8066.2367.28FORMULA72.7674.3074.8875.4675.84IDENTIFIER68.6270.1869.9869.7969.59MULTIPLE27.6432.1628.6431.6631.66SYSTEMATIC69.5772.6174.3275.6176.35TRIVIAL58.9562.7365.4867.4168.26NOCLASS51.2251.2256.1058.5458.54etalJournalofCheminformatics(Suppl1):S11http://www.jcheminf.com/content/7/S1/S11Page4of9 postchallengeimprovedsystemidentifiedSYSTEMATIC CEMsatthebest.What ’ smore,itseemsbeverydiffi- culttorecognizeMULTIPLECEMsinbothsystems. Mainreasonmaybethatthenumberofannotated CEMsisnotsufficefortheMULTIPLEcategory(202, 187,199fortraining,developmentandtestdatasets, respectivelyinTable3). Conclusions Inthearticle,wepresentourpost-challengesystemand itsperformancefortheCEMsubtaskofBioCreativeIV CHEMDNERchallenge.Oursystemprocessingpipeline consistsofthreemajorcomponents:preprocessing(sen- tencedetection,tokenization),recognition(CRF-based approach),andpost-processing(rulebasedapproachand formatconversion).Ourmainfocusonthisimproved systemwastoexploretheeffectivenessofthecostpara- meteroptimizationandwordrepresentationsfeaturefor theCEMsubtask. Inourpost-challengeimprovedsystem,insteadof extractingaCEMasawhole,weregardeditasa sequencelabelingproblem.ThefamousCRFmodelis utilizedtosolvethesequencelabelingproblem,whose costparameterisoptimizedby10-foldcrossvalidation withgridsearch.Differentfeaturetypes,includinggen- erallinguistic,character,casepattern,contextual,and wordrepresentationsfeatures,wereexploitedforour runs.Inordertoreducedatasparsityintheannotated traininganddevelopmentdatasets,wordrepresentations wereinducedfrompooledtraining,development,test andbackgrounddatasetsbyBrownclusteringmethod. Finkel&Manning[18]proposedamodelspecifically forrecognizingnestednamedentitiesbyusingadiscri- minativeconstituencyparser.Themodelexplicitly representsthenestedstruc ture,allowingentitiestobe influencednotjustbythelabelsofthetokenssurround- ingthem,asinaCRF,butalsobytheentitiescontained inthem,andinwhichtheyarecontained.Inongoing work,themodelwillbeintroducedforrecognizing nestedCEMs. Thoughourcurrentsystemhasmuchroomfor improvement,oursystemisvaluableinshowingthat Table5Featurecombinationsusedforpost-challengerunsontheCEMsubtask. WordRepresentation GeneralLinguisticCharacterCasePatternContextual500100015002000 Run1  Run2  Run3  Run4  Run5  Figure2 ThebalancedF-measureforpost-challengerunswith10-foldcrossvalidationbygridsearch . Xu etal . JournalofCheminformatics 2015, 7 (Suppl1):S11 http://www.jcheminf.com/content/7/S1/S11 Page5of9 theperformanceintermofbalancedF-measurecanbeimprovedlargelybyutilizinglargeamountsofrelativelyinexpensiveun-annotatedPubMedabstracts.Fromourpracticeandlesson,ifwedirectlyusesomeopen-sourceNLPtoolkits,suchasOpenNLP,StanfordCoreNLP,falsepositiveratemaybeveryhigh.Itisbettertodevelopsomeadditionalrulestominimizethefalsepositiverateifonedontwanttore-traintherelatedMethodsPre-processing:sentencedetection&tokenizationAsentencedetectorcanidentifyifapunctuationcharac-termarkstheendofasentenceornot.Here,thesen-tencedetectorinOpenNLP[19]isutilized.However,sentenceboundaryidentificationischallengingbecausepunctuationmarksareoftenambiguous[20].Inordertoimprovefurthertheperformanceofthesentencedetec-tion,wecollectedmanyabbreviations,suchassyn.,etc.fromthetraininganddevelopmentsets.Thenwegeneratedseveralrules,suchasifcurrentsen-tenceendswiththeseabbreviationsorcomma,ornextsentencestartswithlower-caseletter.Inthiscase,thecurrentandnextsentencesaremergedintoanewone.Atokenizerdivideseachobtainedsentenceaboveintotokens,whichusuallycorrespondtowords,punctuation,numbers,etc.However,tocaptureindividualcompo-nentswithinaCEM,similartoWeietal.[21],weper-formedtokenizationonafinerlevel.Specifically,specialcharactersinTable7,numbers,andGreeksymbolsaredividedasseparatetokens.AnexampleisshowninTable8.Pluralupper-caseabbreviationsarealsosepa-ratedintotwotokens,suchasintoAsamatteroffact,beforeanypre-processing,wealsomergedsomespecialcharacterswiththesamemeaning,suchas,etc.Recognition:CRF-basedapproachAsmentionedinBackground,weseetheCEMrecogni-tionproblemasasequencelabelingone(seeTable8).Asatypeofdiscriminativeundirectedprobabilisticmodel,CRFs[12,14]areoftenusedforlabelingorpar-singofsequentialdata,suchasnaturallanguagetextorbiologicalsequences.CRFs[22-24]hasbeenappliedsuc-cessfullytoidentifyvariousbio-entities,suchasgene,proteinandsoon,andshownagoodperformance.Giventokensequence  x=(x1,x2,···,xN) ,CRFdefinestheconditionalprobabilitydistribution Pr(  y|  x) oflabelsequence  y=(y1,y2,···,yN) asfollows. Pr(exp( Here, w=(w1,w2,···,wM)T isaglobalfeatureweightvector, )=( isalocalfeaturevectorfunction,andMisthenumberoffeaturefunctions.Theweightvectorwcanbeobtainedfromthetraininganddevelopmentsetsbyalimited-memoryBroyden-F(L-BFGS)[25]method.ThetraditionalBIEOlabelsetisusedinourpost-chal-lengeimprovedsystem.Thatistosay,eachtokenislabeledasbeingthebeginningof(B),theinsideof(I),theendof(E)orentirelyoutside(O)ofaspanofinterest.Here,CRF++[26]isadoptedfortheactualimplementa-tion.InCRF++,thereare4majorparameters(and)tocontrolthetrainingcondition.Inoursub-mittedpredictionsandpost-challengeones,thepara-metersandwereconsistentlysettoCRF-L2,2and4,respectively.Theoptionisoptimizedwith10-foldcrossvalidation,asintroducedabove.FeaturesforourCRFmodelOursystemexploitsfourdifferenttypesoffeatures:GenerallinguisticfeaturesOursystemincludestheoriginaluni-tokensandbi-tokens,aswellasstemmeduni-tokens,bi-tokensandtri-tokens,asfeaturesusingthePortersstemmer[27]fromStanfordCoreNLP[28].CharacterfeaturesSincemanyCEMscontainnumbers,Greekletters,Romannumbers,aminoacids,chemicalelements,andspecialcharacters,oursystemcalculatesseveralstatisticsasfeaturesforeachtoken,includingitsnumberofdigi-tals,numberofupper-andlower-caseletters,numberofallcharactersandpresenceorabsenceofspecificcharac-tersorGreekletters,Romannumbers,aminoacids,orchemicalelements. Table6Performanceresultsinourpost-challengeimprovedsystemfortheCEMsubtaskintheBioCreativeIVCHEMDNERcompetitionRun1Run2Run3Run4Run5cost222218,02519,25919,38919,49519,3552,3122,6712,5372,6942,5057,3266,0925,9625,8565,996Precision(%)88.6387.8288.4387.8688.5471.1075.9776.4876.9076.35score(%)78.9181.4782.0282.0281.99ABBREVIATION59.7763.3764.2865.8565.9872.3673.5572.9273.9472.97FORMULA73.0273.4574.1274.0174.30IDENTIFIER64.7263.3566.6764.3366.86MULTIPLE23.6232.1635.1833.1730.65SYSTEMATIC79.1982.8683.2583.2282.5671.4181.7682.3882.7081.56NOCLASS53.6663.4163.4168.2963.41etalJournalofCheminformatics(Suppl1):S11http://www.jcheminf.com/content/7/S1/S11Page6of9 Casepatternfeatures Similarto[21],anyuppercasealphabeticcharacteris replacedby ‘ A ’ ,anylowercaseoneisreplacedby ‘ a ’ , andanynumber(0-9)isreplacedby ‘ 0 ’ .Moreover,our systemalsomergeconsecutivelettersandnumbersand generatedadditionalsingleletter ‘ a ’ andnumber ‘ 0 ’ features. Contextualfeatures Foreachtoken,oursystemincludesacombinationof thecurrentoutputtokena ndpreviousoutputtoken (bigram). Wordrepresentationfeatures Onecommonapproachtoinducingunsupervisedword representationistouseclustering,perhapshierarchical, suchasBrownclusteringmethod[17],Collobertand Westonembeddings[29],hierarchicallog-bilinear model(HLBL)embeddings[30]andsoon.Here,the Brownclusteringmethodisused.Theimplementation ofBrownclusteringmethodbyLiang[31]isadoptedin ourpost-challengesystem. TheresultofrunningtheBrownclusteringmethodis abinarytree,whereeachtokenoccupiesasingleleaf node,andwhereeachleafnodecontainsasingletoken. Therootnodedefinesaclustercontainingtheentire tokenset.Interiornodesrepresentintermediatesize clusterscontainingallofthetokensthattheydominate. Thus,nodeslowerinthebinarytreecorrespondto smallertokenclusters,whilehighernodescorrespondto largertokenclusters.AccordingtoHuffmancoding [32],aparticulartokencanbeassignedabinarystring byfollowingthetraversalpathfromtheroottoitsleaf, assigninga0foreachleftbranch,anda1foreachright branch. Intuitively,theBrownclusteringmethodwillmerge thetokenswithsimilarcontextsintothesamecluster. Thus,themoresimilartheprefixofthetoken ’ sHuff- mancoding,themoresimilarthetokens.Table9shows sometokenexamplesandtheirbinarystringrepresenta- tionswith500clusters.Let ’ stakeTable9asanexam- ple.AccordingtomainideaoftheBrownclustering method,thetoken “ interpeak ” (01100110110)ismore similarthanthetoken “ aquaporine ” (01101110011)with thetoken “ florbetapir ” (0110011010). Post-processing:rule-basedapproach&format conversion Oncloserexamination,wefindthattheresultsofCRF approachincludesomefalsepositiveCEMs,suchas “ 25 (3),186-193 ” , “ 1-D,2-D ” andsoon.So,wedeveloped severaladditionalregularexpressestoremovethem.In addition,ourpost-processingstepalsohelpsadjusttext spansofCEMs,suchasaddingamissingclosingpar- enthesis,suchas “ [4Fe-4S](2+ ” into “ [4Fe-4S](2+) ” .All oftheadjustmentrulesarelistedinTable10.Here,#(·, str)meansthenumberofoccurrencesofthestringstr intheinterestedCEM,right(·, n )andleft(·, n )denote thesubstringwiththelengthof n rightorlefttothe interestedCEM,andoffset(·,start)andoffset(·,left)indi- catethestartorendoffsetoftheinterestedCEM.Let ’ s takethefirstrowinTable10asanexample.Itmeans thatifthenumberoftheoccurrencesof “ ( ” ishigher thanthatof “ ) ” intheinterestedCEM,andifthesub- stringwiththelengthof1righttotheinterestedCEM is “ ) ” ,thenstartoffsetoftheinterestedCEMismoved onecharacterfurthertotheright. Finally,weconvertedtherecognizedCEMsintothe officialformatwiththeresultingconfidencescores.In oursystem,theconfidencescoreissimplysettoaver- agedconditionalprobablyofeachtokenscomposedof theinterestedCEM,formallydefinedasfollows. score(CEM)= 1 | CEM |  t  CEM CondProb(t) (6) Table7Specialcharactersincludedinourtokenizer ()[]{�}  ,./\ ‘ ™ @·© ® “ := 
 +-?_ |  ¬®    ±~*% ‰  #&;!£ € ¥$ ×÷ ‡† ...    \b
\t\n\f
\b\t Table8AnexampleofCEMcomponentlabelsinan excerpt “ ...[C(8)mim][PF(6)]... “ inPMID:23265515 token ...[C( label OBI I conditionalprob....0.9944560.9972410.999912 token 8)mim] label I I I I conditionalprob.0.9999140.9998530.9972440.996372 token [PF(6 label I I I I conditionalprob.0.9961100.9959400.9967330.996693 token ) ]... label I EO conditionalprob.0.8257820.731261... Table9Sampletokensandtheirresultingbinarystring representationswith500clusters. ID Token BinaryString 1 gracile 010011 2 quintile 010010 3 florbetapir 0110011010 4 interpeak 01100110110 5 aquaporine 01101110011 Xu etal . JournalofCheminformatics 2015, 7 (Suppl1):S11 http://www.jcheminf.com/content/7/S1/S11 Page7of9 where|CEM|meansthenumberoftokencomponents ofaCEM.Take “ [C(8)mim][PF(6)] ” inTable8asan example.Itsconfidencescoreiscalculatedasfollows. score([C(8)mim][PF(6)]) = 1 13  t  [C(8)mim][PF(6)] CondProb(t) =0.963655 (7) Competinginterests Theauthorsdeclarethattheyhavenocompetinginterests. Authors ’ contributions SXandYZdevelopedtheCEMrecognitionsystem.SXandHZconducted extensiveexperimentsanddraftedthemanuscript.LZandXAconceivedof thesupportingprojects,andparticipatedintheresultingdesignand coordinationandhelpeddraftthemanuscript.Allauthorsreadand approvedthefinalmanuscript. Declarations ThisworkwassupportedpartiallybyFundamentalResearchFundsforthe CentralUniversities:ResearchonForestPropertyCirculationMechanismin CollectiveForestArea(JGTD2014-04),BeijingForestryUniversityYoung ScientistFund:ResearchonEconometricMethodsofAuctionwiththeir ApplicationsintheCirculationofCollectiveForestRight(BLX2011028),the NationalScienceFoundationofChina:ResearchonTechnologyOpportunity DetectionbasedonPaperandPatentInformationResources(71403255),Key TechnologiesR&DProgramofChinese12thFive-YearPlan(2011-2015): STKOSCollaborativeConstructionSystemandAuxiliaryToolDevelopment (2011BAH10B02),andKeyWorkProjectofInstituteofScientificand TechnicalInformationofChina(ISTIC):IntelligentAnalysisServicePlatform andApplicationDemonstrationforMulti-SourceScienceandTechnology LiteratureintheEraofBigData(ZD2014-7-1). Thisarticlehasbeenpublishedaspartof JournalofCheminformatics Volume 7Supplement1,2015:TextminingforchemistryandtheCHEMDNERtrack. Thefullcontentsofthesupplementareavailableonlineathttp://www. jcheminf.com/supplements/7/S1. Authors ’ details 1 InformationTechnologySupportingCenter,InstituteofScientificand TechnicalInformationofChina,No.15FuxingRd.,HaidianDistrict,100038 Beijing,PRChina. 2 SchoolofEconomicsandManagement,BeijingForestry University,No.35QinghuaEastRd.,HaidianDistrict,100083Beijing,PR China. 3 NetworkCenter,ScienceandTechnologyDaily,No.15FuxingRd., HaidianDistrict,100038Beijing,PRChina. Published:19January2015 References 1.KrallingerM,LeitnerF,RabalO,VazquezM,MiguelJ,ValenciaA: CHEMDNER:Thedrugsandchemicalnamesextractionchallenge. JCheminform 2015, 7(Suppl1) :S1. 2.LiJ,ZhuX,ChenJY: Buildingdisease-specificdrug-proteinconnectivity mapsfrommolecularinteractionnetworksandpubmedabstracts. PLoS ComputationalBiology 2009, 5(7) :1000450,doi:10.1371/journal.pcbi.1000450. 3.EltyebS,SalimN: Chemicalnamedentitiesrecognition:Areviewon approachesandapplications. JournalofCheminformatics 2014, 6(17) :1-12, doi:10.1186/1758-2946-6-17. 4.VazquezM,KrallingerM,LeitnerF,ValenciaA: Textminingfordrugsand chemicalcompound:Methods,toolsandapplications. Molecular Informatics 2011, 30(6-7) :506-519,doi:10.1002/minf.201100005. 5.KrallingerM,MorganA,SmithL,LeitnerF,TanabeL,WilburJ,HirschmanL, ValenciaA: Evaluationoftext-miningsystemsforbiology:Overviewof thesecondBioCreativecommunitychallenge. GenomeBiology 2008, 9(Suppl2) :1,doi:10.1186/gb-2008-9-S2-S1. 6.XuS,AnX,ZhuL,ZhangY,ZhangH: ACRF-basedsystemforrecognizing chemicalentitiesinbiomedicalliterature. In Proceedingsofthe4th BioCreativeChallengeEvaluationWorkshop KrallingerM,LeitnerF,RabalO, VazquezM,OyarzabalJ,ValenciaA2013, 2 :152-157. 7.XuS,MaF,TaoL: Learnfromtheinformationcontainedinthefalse splicesitesaswellasinthetruesplicesitesusingSVM. where|CEM| meansthenumberoftokencomponentsofaCEM.Take “ [C(8)mim][PF(6)] ” in Table8asanProceedingsoftheInternationalConferenceonIntelligent SystemsandKnowledgeEngineering AtlantisPress,Amsterdam,Netherlands; 2007,1360-1366,doi:10.2991/iske.2007.13. 8.XuS: Selenoproteingenespredictioninsilicobasedonmachine learningapproaches. PhDthesis ChinaAgriculturalUniversity;2008. 9.MikolovT,ChenK,CorradoG,DeanJ: Efficientestimationofword representationsinvectorspace. ProceedingsoftheInternationalConference onLearningRepresentations 2013. Table10TheAdjustmentRulesoftheTextSpansintheBioCreativeIVCHEMDNERcompetition. ID IFCondition Action 1 #(·, “ ( ” )==#(·, “ ) ” )+1  right(·,1)== “ ) ” offset(·,end)=offset(·,end)+1 2 #(·, “ ( ” )==#(·, “ ) ” )-1  left(·,1)== “ ( ” offset(·,start)=offset(·,start)-1 3 #(·, “ [ ” )==#(·, “ ] ” )+1  right(·,1)== “ ] ” offset(·,end)=offset(·,end)+1 4 #(·, “ [ ” )==#(·, “ ] ” )-1  left(·,1)== “ [ ” offset(·,start)=offset(·,start)-1 5 #(·, “ { ” )==#(·, “ } ” )+1  right(·,1)== “ } ” offset(·,end)=offset(·,end)+1 6 #(·, “ { ” )==#(·, “ } ” )-1  left(·,1)== “ { ” offset(·,start)=offset(·,start)-1 7 #(·, “ &#xsc00; “ )==#(·, “ &#x/sc0; “ )+1  right(·,5)== “ &#x/sc0; “ offset(·, end)=offset(·,end)+5 8 #(·, “ &#xsc00; “ )==#(·, “ &#x/sc0; “ )-1  left(·,4)== “ &#xsc00; “ offset(·,start)=offset(·,start)-4 9 #(·, “ &#xi000; “ )==#(·, “ &#x/i00; “ )+1  right(·,4)== “ &#x/i00; “ offset(·,end)=offset(·,end)+4 10 #(·, “ &#xi000; “ )==#(·, “ &#x/i00; “ )-1  left(·,3)== “ &#xi000; “ offset(·,start)=offset(·,start)-3 11 #(·, “ &#xsup0; “ )==#(·, “ &#x/sup; “ )+1  right(·,6)== “ &#x/sup; “ offset(·,end)=offset(·,end)+6 12 #(·, “ &#xsup0; “ )==#(·, “ &#x/sup; “ )-1  left(·,5)== “ &#xsup0; “ offset(·,start)=offset(·,start)-5 13 #(·, “ &#xsub0; “ )==#(·, “ &#x/sub; “ )+1  right(·,6)== “ &#x/sub; “ offset(·,end)=offset(·,end)+6 14 #(·, “ &#xsub0; “ )==#(·, “ &#x/sub; “ )- 1  left(·,5)== “ &#xsub0; “ offset(·,start)=offset(·,start)-5 #(·, str )meansthenumberofoccurrencesofthestring str intheinterestedCEM,right(·, n )andleft(·, n )denotethesubstringwiththelengthof n rightorleftto theinterestedCEM,andoffset(·,start)andoffset(·,left)indicatethestartorendoffsetoftheinterestedCEM. Xu etal . JournalofCheminformatics 2015, 7 (Suppl1):S11 http://www.jcheminf.com/content/7/S1/S11 Page8of9 10.LiangP:Semi-supervisedlearningfornaturallanguage.sthesisMassachusettsInstituteofTechnology;2005.11.TurianJ,RatinovL,BengioY:Wordrepresentations:Asimpleandgeneralmethodforsemi-supervisedlearning.Proceedingsofthe48thAnnualMeetingoftheAssociationforComputationalLinguistics.AssociationforComputationalLinguistics,Stroudsburg,PA,USA2010,384-394.12.LaffertyJ,McCallumA,PereiraF:Conditionalrandomfields:Probabilisticmodelsforsegmentingandlabelingsequencedata.Proceedingsofthe18thInternationalConferenceonMachineLearningMorganKaufmannPublishersInc.,SanFrancisco,CA,USA;2001,282-289.13.KrallingerM,RabalO,LeitnerF,VazquezM,SalgadoD,LuZ,LeamanR,LuY,JiD,LoweDM,SayleRA,Batista-NavarroRT,RakR,HuberT,RocktaschelT,MatosS,CamposD,TangB,XuH,MunkhdalaiT,RyuKH,RamananSV,NathanS,ZitnikS,BajecM,WeberL,IrmerM,AkhondiSA,KorsJA,XuS,AnX,SikdarUK,EkbalA,YoshiokaM,DiebTM,ChoiM,VerspoorK,KhabsaM,GilesCL,LiuH,RavikumarKE,LamuriasA,CoutoFM,DaiH,TsaiRT,AtaC,CanT,UsieA,AlvesR,Segura-BedmarI,MartinezP,OryzabalJ,ValenciaA:TheCHEMDNERcorpusofchemicalsanddrugsanditsannotationprinciples.JCheminform7(Suppl1)14.ShaF,PereiraF:Shallowparsingwithconidtionalrandomfields.ProceedingsoftheHumanLanguageTechnologyConferenceoftheNorthAmericanChapteroftheAssociationforComputationalLinguistics.AssociationforComputationalLingustics,Stroudsburg,PA,USA2003,213-220,15.MillerS,GuinnessJ,ZamanianA:Nametaggingwithwordclustersanddiscriminativetraining.ProceedingsofConferenceonHumanLanguageTechnology/NorthAmericanChapteroftheAssociationforComputationalLinguiusticsAnnualMeetingAssociationforComputationalLinguistics,Boston,Massachusetts;2004,337-342.16.GanchevK,CrammerK,PereiraF,MannG,BellareK,McCallumA,CarrollS,JinY,WhiteP:Penn/Umass/CHOPBioCreativeIIsystems.Proceedingsofthe2ndBioCreativeChallengeEvaluationWorkshop17.BrownPF,deSouzaPV,MercerRL,PietraVJD,LaiJC:Class-basedn-grammodelsofnaturallanguage.ComputationalLinguistics18.FinkelJR,ManningCD:Nestednamedentityrecognition.Proceedingsofthe2009ConferenceonEmpiricalMethodsinNaturalLanguageProcessing.AssociationforComputationalLingustics,Stroudsburg,PA,USA2009,141-150.TheApacheOpenNLPLibrary.Library.20.ReadJ,DridanR,OepenS,SolbergLJ:Sentenceboundarydetection:Alongsolvedproblem?Proceedingsofthe24ndInternationalConferenceonComputationalLinguistics.IndianInstituteofTechnologyBombay,Mumbai,Maharashtra,India;KayM,BoitetC2012:985-994.21.WeiC-H,HarrisBR,KaoH-Y,LuZ:tmVar:Atextminingapproachforextractingsequencevariantsinbiomedicalliterature.22.McDonaldR,PereiraF:Identifyinggeneandproteinmmentionintextusingconditionalrandomfields.BMCBioinformatics6(Suppl1)23.HuangH-S,LinY-S,LinK-T,KuoC-J,ChangY-M,YangB-H,ChungI-F,HsuC-N:High-recallgenementionrecognitionbyunificationofmultiplebackgroundparsingmodels.Proceedingsofthe2ndBioCreativeChallengeEvaluationWorkshop24.KlingerR,FriedrichCM,FluckJ,Hofmann-ApitiusM:Namedentityrecognitionwithcombinationsofconditionalrandomfields.Proceedingsofthe2ndBioCreativeChallengeEvaluationWorkshopHirschmannL,KrallingerM,ValenciaA2007,89-92.25.LiuDC,NocedalJ:OnthelimitedmemoryBFGSmethodforlargescaleMathematicalProgramming:503-528,doi:10.1007/26.KudoT:CRF++:YetAnotherCRFToolkit.Toolkit.trunk/doc/index.html].27.PorterMF:Analgorithmforsuffixstripping.28.ManningC,BauerJ:StanfordCoreNLP-ASuiteofNLPTools.Tools.stanford.edu/software/corenlp.shtml].29.CollobertR,WestonJ:Aunifiedarchitecturefornaturallanguageprocessing:Deepneuralnetworkswithmultitasklearning.Proceedingsofthe25thInternationalConferenceonMachineLearning30.MnihA,AndriyG:Ascalablehierarchicaldistributedlanguagemodel.AdvancesinNeuralInformationProcessingSystems21.MITPress,Cambridge,MA;KollerD,SchuurmansD,BengioY,BottouL2009:1081-1088.31.LiangP:C++ImplementationoftheBrownWordClusteringAlgorithm.Algorithm.32.HuffmanDA:Amethodfortheconstructionofminimum-redundancyProceedingsoftheI.R.E:1098-1101,doi:10.1109/doi:10.1186/1758-2946-7-S1-S11Citethisarticleas:etalACRF-basedsystemforrecognizingchemicalentitymentions(CEMs)inbiomedicalliterature.JournalofCheminformatics(Suppl1):S11. W. Jeffery Hurst,The Hershey Company. http://www.chemistrycentral.com/manuscript/ etalJournalofCheminformatics(Suppl1):S11http://www.jcheminf.com/content/7/S1/S11Page9of9