Interpreting the Temporal Aspects of Language Naushad UzZaman Advisor: James Allen Computer Science Department University of Rochester, Rochester, NY PhD Defense 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Research Context PhD Defense 2 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Understanding In 1492, Genoese explorer Christopher Columbus, under contract to the Spanish crown, reached several Caribbean islands, making first contact with the indigenous people. On April 2, 1513, Spanish conquistador Juan Ponce de León landed on what he called "La Florida"—the first documented European arrival on what would become the U.S. mainland. Spanish settlements in the region were followed by ones in the present-day southwestern United States that drew thousands through Mexico. French fur traders established outposts of New France around the Great Lakes; France eventually claimed much of the North American interior, down to the Gulf of Mexico. The first successful English settlements were the Virginia Colony in Jamestown in 1607 and the Pilgrims' Plymouth Colony in 1620. The 1628 chartering of the Massachusetts Bay Colony resulted in a wave of migration; by 1634, New England had been settled by some 10,000 Puritans. Between the late 1610s and the American Revolution, about 50,000 convicts were shipped to Britain's American colonies. Beginning in 1614, the Dutch settled along the lower Hudson River, including New Amsterdam on Manhattan Island. PhD Defense 3 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Structure 1492 April 2, 1513 Christopher Columbus reached Caribbean islands Juan Ponce de León landed La Florida documented Spanish settlements southwestern United States thousands through Mexico French fur traders 1607 are related with each other first European arrival landed drew France • Understand how entities established claimed outposts of New France much of the North American interior first successful English settlements were Virginia Colony in Jamestown 1610s along the lower Hudson River, 1614 Dutch settled including New Amsterdam on Manhattan Island. 1620 first successful English settlements 1628 chartering of the Massachusetts Bay Colony 1634 10,000 Puritans settled were Pilgrims' Plymouth Colony resulted migration New England American Revolution 50,000 convicts PhD Defense shipped Britain's American colonies • We can: ‣ Summarize ‣ Answer questions ‣ Visualize • Can benefit many domains and applications 4 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Scenario - Medical Domain • All clinical notes are becoming electronic • Notes from different doctors visit • Doctors can easily get answers like: QA ‣ Did the patient get follow-up CT after PET/CT ‣ Did the patient had any abdominal pain earlier PhD Defense 5 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Scenario - News domain • Chronological summary of events in a news ‣ Helps busy people to skim through articles ‣ Helps people with problem in understanding Summary Iraq war Aug 2010 - Google launched voice actions 2001 - September 11 incident 2003 - US invaded Iraq Siri vs Oct 2011 - Apple released Siri with Majel iPhone 4s 2006 - Saddam Hossain hanged 2008 - Obama became president Dec 2011 - Google released Majel to compete against Siri 2011 - Iraq War ended PhD Defense 6 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Scenario - History domain • Visualization to educate visitors in museum ‣ In front of Columbus section in museum ‣ Can check online wikipedia ‣ Automatically create the timeline http://upload.wikimedia.org/wikipedia/commons/1/1f/Antillas_%28orthographic_projection%29.svg 4/23/10 9:39 PM reached 1492 Christopher Columbus Caribbean islands Page 1 of 1 landed April 2, 1513 Juan Ponce de León La Florida first European arrival documented Visualization Spanish settlements drew landed southwestern United States thousands through Mexico http://upload.wikimedia.org/wikipedia/commons/8/8f/New_France_%28orthographic_projection%29.svg 4/23/10 10:18 PM km French fur traders established outposts of New France France much of the North American interior Page 1 of 1 claimed Llorens et al. 2011 1607 PhD Defense 7 first successful English settlements were Virginia Colony in Jamestown 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Information Understanding • Applications: Summarization,Visualization, Question Answering, and others • Domains: news, medical, history, and others PhD Defense 8 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Outline • Temporal Information Understanding • Evaluation of Temporal Information • Summary and Future Work PhD Defense 9 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Understanding in NL The advisor came to the department at 11 am. The student came before his advisor’s arrival. The meeting began at 11:30am. temporal expression event relation The student was productive in the meeting. 11:00am camestudent 11:30am cameadvisor meeting1 arrival meeting2 •TimeML annotation scheme •TimeBank, AQUAINT corpus •TempEval competition productive PhD Defense 10 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Our Approach • Hybrid System • Semantic information from semantic parser • Machine learning classifier PhD Defense 11 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Suggestions from Semantic Parser Sentence He fought in the war (SPEECHACT V1 SA-TELL :CONTENT V2) Semantic parser (F V2 (:* FIGHTING FIGHT) :AGENT V3 :MODS (V4) :TMA ((TENSE PAST))) (PRO V3 (:* PERSON HE) :CONTEXT-REL HE) output (F V4 (:* SITUATED-IN IN) :OF V2 :VAL V5) (THE V5 (:* ACTION WAR)) ontology class ((THE ?x (? type SITUATION-ROOT)) 100+ -extract-noms> Extraction rules (EVENT ?x (? type SITUATION-ROOT) :pos NOUN :class OCCURRENCE )) Extracted with extraction rules <EVENT eid=V2 word=FIGHT pos=VERBAL ont-type=FIGHTING tense=PAST class=OCCURRENCE voice=ACTIVE aspect=NONE polarity=POSITIVE> <RLINK eventInstanceID=V2 ref-word=HE ref-ont-type=PERSON relType=AGENT> <SLINK signal=IN eventInstanceID=V2 subordinatedEventInstance=V5 relType=SITUATED-IN> <EVENT eid=V5 word=WAR pos=NOUN ont-type=ACTION class=OCCURRENCE voice=ACTIVE polarity=POSITIVE aspect=NONE tense=NONE> PhD Defense 12 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Tasks event-timex The advisor came to department at 11 am. main-sub The student came before his advisor’s arrival. They had a meeting from 11:30 am. main-main The student was productive in the meeting. 11:00am cameadvisor meeting1 cameadvisor meeting1 arrival PhD Defense meeting2 11:30am productive camestudent meeting2 13 arrival 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Events and its Attributes Event: something that happens or a dynamic property which holds the truth. Attributes: class, tense, aspect and pos. Class attribute: 1. 2. 3. 4. 5. 6. 7. PhD Defense Occurrence: die, crash, build State: on board, alive Reporting: say, report I-Action: attempt, try, promise I-State: believe, intend, want Aspectual: begin, stop, continue Perception: see, hear, watch, feel 14 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Event Extraction • Semantic parser suggestions and features • Train events and attributes on TimeML data • Markov Logic Network (MLN) classifier ‣ first order logic + markov network PhD Defense 15 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Event Extraction Features Word-level Semantic Features ontology type - WordNet class pos class tense voice polarity aspect modality penn tag verb pos sequence prev pos next pos pos x prev pos Lexical Features word base word constituent verb word sequence previous word of verb sequence prev word next word word x prev word index range Sentence-level Semantic Features Rlink relation type Rlink reference ont-type Slink core relation Slink reference relation Tlink reference Tlink signal PhD Defense 16 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Expression Extraction • Recognition with Conditional Random Field • Normalization with rule based system Temporal expression DCT (given): March 1, 1998; 14:11 hours Sunday last week mid afternoon nearly two years each month Type Value TIME DATE DATE TIME DURATION SET 1998-03-01T14:11:00 1998-03-01 1998-W08 1998-03-01TAF P2Y P1M cw -2 cw - 1 current word last next five ten months years init-list is number is time string Table 1: Examples of normalized values and types for temporal expressions according to TimeML PhD Defense 17 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Overview Text TRIPS Semantic parser Parser MLN event feature extractor CRF timex extractor MLN event event filtering extractor events timex normalizer event features timex timex features Event and Temporal Expression extraction from raw text: first step towards a temporally aware system. Naushad UzZaman and James F. Allen. International Journal of Semantic Computing, 2011. PhD Defense 18 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Identifying Temporal Relations event-timex The advisor came to department at 11 am. main-sub The student came before his advisor’s arrival. They had a meeting from 11:30 am. main-main The student was productive in the meeting. 11:00am cameadvisor meeting1 cameadvisor meeting1 arrival PhD Defense meeting2 11:30am productive camestudent meeting2 19 arrival 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Relations simultaneous identity during begins/ begun by Y before / after X iBefore / iAfter X overlaps/ overlapped by PhD Defense X ends/ ended by Y Y includes/ included by X X Y X Y X Y Y 20 20 July 2012 Motivation Feature category lexical features word-level semantics features sentence-level semantic features Source of Feature Semantic parser MLN classifier Rule based PhD Defense Temporal Understanding Features Event Class Event Tense Event Aspect Event Polarity Event Stem Event Word Event Constituent Event Ont-type Event LexAspect x Tense Event Pos Timex Word Timex Type Timex Value Timex DCT relation Event’s semantic role Event’s argument’s ont-type TLINK event-time signal SLINK event-event relation type Temporal Evaluation Task C event-dct Task D event-timex YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES YES Summary Task E main-main e1 x e2 e1 x e2 e1 x e2 e1 x e2 e1 x e2 YES e1 x e2 e1 x e2 e1 x e2 e1 x e2 Task F main-sub e1 x e2 e1 x e2 e1 x e2 e1 x e2 e1 x e2 YES e1 x e2 e1 x e2 e1 x e2 e1 x e2 YES Table 1: Features used for Tasks C, D, E and F. 20 July 2012 21 Motivation Temporal Understanding Temporal Evaluation Summary Performance in TempEval-2 Task Temporal Expression Event Event-Temporal Expression Event-DCT Consecutive Main Events Main-Subordinate Events Performance 0.85 (Fscore) 0.77 (Fscore) 0.65 (Precision) 0.79 (Precision) 0.56 (Precision) 0.6 (Precision) Comparison Second best Third best Best Second Best Best Second Best Table 1: Performance of TRIOS in TempEval-2 TRIPS and TRIOS System for TempEval-2: Extracting Temporal Information from Text. Naushad UzZaman and James F. Allen. International Workshop on Semantic Evaluations (SemEval-2010), Association for Computational Linguistics (ACL), Sweden, July 2010. PhD Defense 22 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Creating Temporal Structure 11:00am meeting2 11:30am cameadvisor meeting1 cameadvisor meeting1 productive meeting2 arrival 11:00am camestudent arrival camestudent 11:30am cameadvisor meeting1 arrival meeting2 productive PhD Defense 23 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Creating Temporal Structure • Temporal closure algorithm • Adopted from Timegraph ‣ ‣ ‣ ‣ Miller and Schubert (1990) Scalable and fast Convert time intervals to time points Ignore inconsistent relations PhD Defense 24 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Timegraph Example • A < B, B < C, D < B • a <a <b <b • b <b <c <c • d <d <b <b s s s e e e s s s e e e as ae bs be 10 20 30 40 as ae bs be cs ce 10 20 30 40 50 60 as ae bs be cs ce 10 ds 20 de 30 40 50 60 10 20 chain1 chain1 chain1 chain2 PhD Defense 25 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Structure 11:00am camestudent 11:30am cameadvisor meeting1 arrival meeting2 productive • Can infer how two entities are related • Can answer temporal questions from a document PhD Defense 26 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Summary of Temporal Understanding • Hybrid system • Automatically extract events, temporal expressions, identify temporal relations • Competitive performance in TempEval-2 • Temporal Structure based on Timegraph for temporal reasoning PhD Defense 27 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Outline • Temporal Information Understanding • Evaluation of Temporal Information • Summary and Future Work PhD Defense 28 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Evaluation of Temporal Information • Evaluation of Temporal Annotation • Evaluation of Temporal Understanding PhD Defense 29 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Problem with Current Metric • Precision/recall from IR community • Relation in different but equivalent ways system reference e1 e2 e1 PhD Defense e1 e3 e3 e3 30 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Related Work • Semantic matching of temporal relations ‣ for judging inter-annotator agreement • Setzer et al. (2003) ‣ use temporal closure; inaccurate • Tannier and Muller (2008) ‣ use only core relations; approximation PhD Defense 31 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Our Approach • Closure to verify equivalent relations • Single score to measure the temporal awareness of systems • Use temporal structure Temporal Evaluation. Naushad UzZaman and James F. Allen. Proc. of The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (Short Paper), USA, June 2011. PhD Defense 32 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Precision, Recall, Fscore Precision = (# of system temporal relations that can be verified from reference annotation temporal closure graph / # of temporal relations in system output) Recall = (# of reference temporal relations that can be verified from system temporal closure graph / # of temporal relations in reference annotation) a Sys+ = System Closure Graph Ref + = Reference Closure Graph a F score = PhD Defense a P recision = Recall = |Sys\Ref + | |Sys| |Ref \Sys+ | |Ref | 2⇤P recision⇤Recall P recision+Recall 33 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Example • All systems get 100% precision, including S2 with B<D edge • Similar to MUC-6 scoring for coreference resolution • They estimate the minimal number of missing links to complete chain to match human annotation • In S1, S3, we are missing 1 edge, hence 2/3 for recall is appropriate PhD Defense 34 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Example System Precision PhD Defense 35 Recall Fscore S1 2/2=1 2/3=0.66 0.8 S2 2/2=1 1/3=0.33 0.5 S3 2/2=1 2/3=0.66 0.8 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Properties of our Metric • Performance decreases with the decrease of information • If system fails to extract important entities the performance will decrease more • Implementation is scalable PhD Defense 36 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Summary • F score -- temporal awareness score • F score captures entity & relation performance • F score can help to find overall best system • Updated solution to reward implicit relations appropriately PhD Defense 37 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Evaluate Temporal Understanding • How annotation started? • TimeML for Temporal QA • Ideal case - evaluate Temporal QA? PhD Defense 38 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Why Evaluate with Temporal QA • Answering questions to judge understanding • Creating questions easier than annotating • Temporal IE performance might not reflect real task performance • Completeness of human annotations • Comparing human and system annotations PhD Defense 39 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Question Taxonomy • yes/no: Was Fein called after the killing? killing captured called • list: What happened after the crash? • list between: What happened between the crash and today? • factoid: When DT Inc. holders adopted a shareholder-rights plan? PhD Defense 40 during the crisis 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal QA System • Take TimeML annotation as input • Temporal reasoning with Timegraph (Miller and Schubert, 1990) - efficient • Maintains necessary relations in graph to infer relations between all entities • Can answer yes/no, list, factoid Evaluating Temporal Information Understanding with Temporal Question Answering. Naushad UzZaman, Hector Llorens and James F. Allen. Proceedings of IEEE International Conference on Semantic Computing, Italy, September 2012. PhD Defense 41 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal QA as Evaluation • 189 questions (79 yes/no, 63 list, 47 factoid) from 25 TimeML documents (1442 tlinks) • Evaluate how good are the system annotations • Get answer from gold, then compare with system • Compare against UzZaman and Allen (2011) metric Systems TIPSem TRIOS Yes/No 0.48 0.34 List 0.43 0.37 Factoid 0.53 0.22 IE-score 0.31 0.27 Table 1: Performance in Temporal QA and corpus based IE against annotation PhD Defense 42 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Observations • For yes/no: only 16.6% explicit, rest implicit • Manually answered yes/no questions Yes/No against human annotations Yes/No against human answers Gold 1.00 0.48 TIPSem 0.48 0.38 TRIOS 0.34 0.23 Table 1: Comparison of gold and system annotations against human answers • Human annotation not 100% (48%) - annotation coverage not complete • Difference between human and system lower PhD Defense 43 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Summary of Temporal QA QA can better evaluate temporal • Temporal information understanding ‣ Natural task; easy to create questions ‣ Can evaluate human temporal annotations • Developed a system for temporal QA ‣ Can reason with any TimeML annotations ‣ Released the toolkit PhD Defense 44 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Outline • Temporal Information Understanding • Evaluation of Temporal Information • Summary and Future Work PhD Defense 45 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Summary • System to extract temporal information ‣ Competitive performance at TempEval-2 • Metric to evaluate temporal annotation ‣ Adopted to evaluate participants at TempEval-3 • Temporal QA system with reasoning ‣ Can evaluate temporal understanding • TRIOS-TimeBank, Merging, TempEval-3 PhD Defense 46 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Future Work • Temporal Visualization or Timeline • Temporal QA+summarization+visualization • Future TempEval PhD Defense 47 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal Visualization or Timeline http://upload.wikimedia.org/wikipedia/commons/1/1f/Antillas_%28orthographic_projection%29.svg 1492 April 2, 1513 Christopher Columbus reached Caribbean islands Juan Ponce de León landed La Florida documented 4/23/10 9:39 PM reached 1492 Christopher Columbus Caribbean islands first European arrival Page 1 of 1 landed April 2, 1513 Spanish settlements southwestern United States landed Juan Ponce de León La Florida first European arrival documented thousands through Mexico drew French fur traders established outposts of New France Spanish settlements France claimed landed southwestern United States much of the North American interior drew thousands through Mexico http://upload.wikimedia.org/wikipedia/commons/8/8f/New_France_%28orthographic_projection%29.svg 1607 first successful English settlements were 4/23/10 10:18 PM Virginia Colony in Jamestown 1610s along the lower Hudson River, 1614 Dutch settled km French fur traders including New Amsterdam on Manhattan Island. established outposts of New France Page 1 of 1 1620 first successful English settlements 1628 chartering of the Massachusetts Bay Colony were Pilgrims' Plymouth Colony resulted France 10,000 Puritans settled much of the North American interior migration 1607 1634 claimed first successful English settlements were Virginia Colony in Jamestown New England American Revolution PhD Defense 50,000 convicts shipped Britain's American colonies 48 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Temporal QA+summarization+viz • Temporal Summary and Visualization with QA capability • Can benefit: doctors, students, people with reading disabilities PhD Defense 49 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Future TempEval • 1: given entities and attributes, classify relations • 2: different subtasks to extract entities and classify relations • 3: given raw text, annotate all information • Future: evaluate with temporal QA PhD Defense 50 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Publications at URCS Naushad UzZaman, Hector Llorens and James Allen. Evaluating Temporal Information Understanding with Temporal Question Answering. Proceedings of IEEE International Conference on Semantic Computing, Italy, September 2012. 1.Naushad UzZaman, Hector Llorens and James F. Allen. Merging Temporal Annotations. Proceedings of 19th International Symposium on Temporal Representation and Reasoning, Leicester, UK, September 2012. 2.Naushad UzZaman and James F. Allen. Temporal Evaluation. Proc. of The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (Short Paper), Portland, Oregon, USA, June 2011. 3.Naushad UzZaman and James F. Allen. Event and Temporal Expression extraction from raw text: first step towards a temporally aware system. International Journal of Semantic Computing, 2011. 4.Naushad UzZaman, Jeffrey P. Bigham and James F. Allen. Multimodal Summarization of Complex Sentence. Proc. of International Conference on Intelligent User Interfaces (IUI), Palo Alto, California, 2011. 5.Naushad UzZaman, Jeffrey P. Bigham and James F. Allen. Multimodal Summarization for people with cognitive disabilities in reading, linguistic and verbal comprehension. 2010 Coleman Institute Conference, Denver, CO, 2010. 6.Naushad UzZaman and James F. Allen. Extracting Events and Temporal Expressions from Text. Fourth IEEE International Conference on Semantic Computing (IEEE ICSC2010), Pittsburgh, USA, September 2010. 7.Naushad UzZaman and James F. Allen. TRIPS and TRIOS System for TempEval-2: Extracting Temporal Information from Text. International Workshop on Semantic Evaluations (SemEval-2010), Association for Computational Linguistics (ACL), Sweden, July 2010. 8.Naushad UzZaman and James F. Allen. TRIOS-TimeBank Corpus: Extended TimeBank corpus with help of Deep Understanding of Text. Proceedings of The seventh international conference on Language Resources and Evaluation (LREC), Malta, May 2010. 9.Naushad UzZaman, Jeffrey P. Bigham and James F. Allen. Pictorial Temporal Structure of Documents to Help People who have Trouble Reading or Understanding. International Workshop on Design to Read, ACM Conference on Human Factors in Computing Systems (CHI), Atlanta, GA, April 2010. PhD Defense 51 20 July 2012 Acknowledgement Collaborators Advisor James Allen James Allen Jeffrey Bigham Committee Hector Llorens Jeffrey Bigham James Pustejovsky Greg Carlson Marc Verhagen Leon Derczynski Daniel Gildea Bosch RTC Lenhart Schubert Junling Hu Chair Microsoft Medical Lab Jeffrey Runner Michael Gillam Hank Rappaport & PD Singh And... Yahoo! Research Department staff Ben, George, Mary, Will Fall 2007 class Bangladeshi students at UR PhD Defense Peter Mika & My Family 52 Michael Matthews & Roi Blanco 20 July 2012 www.cs.rochester.edu/~naushad naushad@cs.rochester.edu PhD Defense 53 20 July 2012 Motivation Temporal Understanding Bigger Picture Temporal Evaluation Summary Natural Language Processing Natural Language Understanding Applications temporal information understanding, social media information extraction, question answering, multimodal summarization, game car prediction, dialog system Domains news medical PhD Defense 54 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Next ... • Mobility group of Nuance in Burlington, MA • Siri-like applications for ‣ smart phones ‣ smart cars ‣ smart tv and others 55 Motivation Temporal Understanding Temporal Evaluation Summary Multimodal Summarization In 1492, Genoese explorer Christopher Columbus, under contract to the Spanish crown, reached several Caribbean islands, making first contact with the indigenous people. Multimodal Summarization of Complex Sentence. Naushad UzZaman, Jeffrey P. Bigham and James F. Allen. Proc. of International Conference on Intelligent User Interfaces (IUI), Palo Alto, California, 2011. Multimodal Summarization for people with Cognitive Disabilities in reading, linguistic and verbal comprehension. Naushad UzZaman, Jeffrey P. Bigham and James F. Allen. 2010 Coleman Institute Conference, Denver, CO, 2010. PhD Defense 56 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary TwitterPaul: Retrieving, Ranking and Aggregating Twitter Predictions http://www.prefixmag.com/ rediff.com 57 Motivation Temporal Understanding Temporal Evaluation Summary Medical IE on Demand RECORD #106886 Discharge Date: 3/17/2006 ATTENDING: HEISSER III , MONROE JOSEPH MD SERVICE: Cardiac Surgery Service. PRINCIPAL DIAGNOSIS: Coronary artery disease. HISTORY OF PRESENT ILLNESS: Mr. Midget is a 56-year-old male ..... ..... ..... He had parkinson's disease since ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... .. ... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ..... ... RECORD #165885 Discharge Date: 10/20/1994 Attending: DENNIS O. STROHSCHEIN , M.D. CU6 SERVICE: Cardiac Surgery Service. PRINCIPAL DIAGNOSIS: Colon cancer status post sigmoid colectomy with a primary anastomosis HISTORY OF PRESENT ILLNESS: Patient is a 73 year old male who immigrated from Co Garlpaaran approximately fifteen years ago. He has had a several year history of tuberculosis ..... ..... ..... ..... RECORD #193097 Discharge Date: 9/4/1998 Attending: FRITZ E. BROMBERG , M.D. SF41 ED166/6793 PRINCIPAL DIAGNOSIS: 1. SEVERE PANCREATIC EXOCRINE AND ENDOCRINE INSUFFICIENCY. HISTORY OF PRESENT ILLNESS: Problem list: ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... ....... 8. CORONARY ARTERY DISEASE , STATUS POST SILENT IMI IN 1988. ....... ....... ....... ....... ....... ....... ....... 1. In the HISTORY OF PRESENT ILLNESS section 2. UMLS semantic category Disease or Syndrome 58 Medical Media Lab Motivation Temporal Understanding Temporal Evaluation Summary In-Car Dialog System • Dialog assistant like iPhone’s Siri • Navigation and MP3 domain ‣ ‣ ‣ ‣ i want to go to dave's house i would like a chinese restaurant i want to go to gym on my way home play music by scorpion, raise volume • Developed syntactic and semantic grammar 59 Motivation Temporal Understanding Temporal Evaluation Summary Timegraph • Timegraph construction ‣ diamond 1 chain and pseudo time calculated when point first entered into the Timegraph ‣ relation between two points in the same chain is constant, by comparing pseudo-time ‣ relation between two points in different chain is found by searching in cross-chain links ‣ search in cross-chain links depends on # of chain + # of cross-chain links ‣ goal in Timegraph construction: minimize # of chain and # of cross-chains PhD Defense 60 (a) diamond 2 circle1 square1 circle 2 square 2 circle 3 circle 4 square 3 Chain 1 (circles) (b) Chain 2 (squares) Chain 3 (diamonds) Examples: (a) timegraph (each node represents a time point); (b) metagraph (each node represents a chain). ( ) in-chain links; (--->) cross-chain links. 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Problem with ACL 2010 < K A < B < C < D < < S1 A < B C D < S2 A < B C A < B S4 A < B C < < C D D < < S5 A < B < C D < < < S6 A < B C < D < < < S7 A < B < C Both S1 and S2 gets 100% precision, 1/3 recall • • • • S2 extracts extra implicit D < S3 • < No extra reward for implicit But S4 > S2 > S1 S3 = S4, but S2 > S1 D < < PhD Defense 61 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Metrics for Temporal Evaluation Evaluation Metric TempEval-2 Setzer et al. Tannier and Muller Our ACL’11 metric Our Updated metric 1 where, w = PhD Defense Recall Precision Ref \Sys Ref + Ref \Sys+ Ref + Ref \Sys Ref |Ref \Sys+ | |Ref | + |Ref \Sys |+w⇤|(Sys Ref )\Ref + | 1 |Ref | Sys\Ref Sys + Sys \Ref + Sys+ Sys \Ref Sys |Sys \Ref + | |Sys | |Sys \Ref + | |Sys | 0.99 (1+|#Ref + | |Ref \Sys+ |) 62 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Improvement |Sys \ Ref + | P recision = |Sys | Recall = |Ref * G = Reduced Graph of G, which includes only the core relations \ Sys+ | + w ⇤ |(Sys Ref ) \ (Ref + |Ref | Ref )| a * (Ref \ Sys+ ) captures system explicit relations that are found in gold explicit relations * (Sys Ref ) captures implicit system relations in terms of Ref * (Ref + Ref ) captures implicit gold relations * (Sys Ref )\(Ref + Ref ) captures implicit system relations that exists in gold implicit relations PhD Defense 63 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary Improvement |Sys \ Ref + | P recision = |Sys | Recall = |Ref Recall = |Ref * G = Reduced Graph of G, which includes only the core relations \ Sys+ | + w ⇤ |(Sys Ref ) \ (Ref + |Ref | \ Sys+ | + w ⇤ |(Sys |Ref | Ref )| Ref ) \ Ref + | a * (Sys Ref ) \ Ref + can be easily computed efficiently by checking, Sys relations that can not be found in Ref but can be verified in Timegraph of Ref a \ ) \ (Ref + \ Ref \) = (Sys Ref ) \ (Ref + Ref ) = (Sys \ Ref \ \ Ref + ) \ (Sys \ Ref \ \ Ref \ ) = (Sys \ Ref \ \ Ref + ) = (Sys \ Ref (Sys Ref ) \ Ref + = (Ref + Ref ) \ Sys PhD Defense 64 20 July 2012 Motivation Temporal Understanding Temporal Evaluation Summary How it solves the problem < K A < B < C < D < < S1 A < B C D < S2 A < B C D < S3 A < B S4 A < B C < < C D D < < S5 A < B < C D < < system S1 S2 S3 S4 S5 S6 S7 precision 1/1=1 2/2=1 2/2=1 2/2=1 3/3=1 4/4=1 3/3=1 recall (1+0.3*0)/3=0.333 (1+0.3*1)/3=0.433 (2+0.3*0)/3=0.666 (2+0.3*0)/3=0.666 (2+0.3*1)/3=0.766 (2+0.3*2)/3=0.866 (3+0.3*0)/3=1.000 fscore 0.5000 0.6043 0.8000 0.8000 0.8674 0.9281 1.0000 < S6 A < B C < D < D < Table 1: Updated precision, recall and fscore for systems < < S7 A < B < C < < PhD Defense 65 20 July 2012