Slide 1IntroductionDuration is Non-trivialSystem DesignFeature AnalysisFeature AnalysisResultsResultsFeature SelectionUnsupervised ClusteringConclusionA Novel Approach to Event Duration PredictionPranav KhaitanDivye Raj KhilnaniYe JinIntroductionPredicting event duration has been a challenging problem and can solve some major challenges being faced in question answering systems. Examples:Liverpool will be playing inter-Milan this Friday.The United States has been fighting a cold war with the Soviet Union. duration of the match is in hoursduration of the war was in decadesMore Features Duration is Non-trivialSame event can have different bounds in different contexts. James watched a movie.James watched the birds fly.HourMinuteSubjectObjectGrammaticalPart of SpeechTenseModalityContextClassHypernymAspectSystem DesignFeature ExtractionFeature Selection•X2 score•MI score•Emperical observationLearning and ClassificationSupervised learning:•Naïve Bayes•Logistic Regression•Maximum EntropyUnsupervised Learning•Agglomerative Clustering•Multinomial clustering•Parse Tree•Web Count•Hypernym•Named Entity RecognitionEvaluation•Precision•Recall •F1•Kappa•Approximate AgreementFeature Analysis•Subject-object•Jonathan is watching a movie vs Jonathan is watching an advertisement•Base verb lemmatization•eating, ate, has eaten, will be eating•Tense•Jonathan will play football in the evening vs Jonathan has been playing football for the past ten years•Sentential Dependencies•He read the report quickly vs He read the report slowly•Part of speech tagging•The government’s move was anticipated•Named Entity Recognition•The body will define the role of the United NationsFeature Analysis•Hypernyms•Contextual Features•Web Counts•Generic Features: Modality, Aspect, Class•Contextual Features•Report FeatureResultsResultsFeature Selection000000.010.01MI Score ChartTotal extracted features: 10,000+. Need to scale down. MI score for features drops quicklyEffectiveness of feature selection500 2500 5000 10000 none0.710.720.730.740.750.760.770.78Feature Selection Chartmaxentnaivelogisticnumber of featuresscoreUnsupervised Clustering1 2 3 40102030405060Clusters using Agglomerative ClusteringSecondsMinutesHoursDaysWeeksMonthsYearsClusterPe rc e ntage o f in stan c e in c l uste r1 2 30102030405060708090Clusters using mixture of multinomialsSeconds MinutesHours DaysWeeks MonthsYearsClusterPe rce nta ge o f in sta n c e in clu ste rCluster 1Less than a dayGreater than a dayCluster 2Cluster 3Conclusion-Significant gain in event duration prediction accuracy using supervised learning-Unsupervised learning results look promising and gives opportunity to do duration prediction across domains with little annotated data-Important to automatically select features and reduce human involvementClassification TaskOur Results Feng Pan et alHuman AgreementCoarse Grain 75.16% 70.3% 87.7%Fine Grain 63.69% 65.8%