1. Laboratory for Knowledge Discovery in Databases Entity Extraction, Animal Disease-related Event Recognition and Classification from Web Presenter: Svitlana Volkova Adviser: William H. Hsu Committee: Dr. Doina Caragea, Dr. Gurdip Singh Supported by: K-State National Agricultural Biosecurity Center (NABC), US Department of Defense
2.
3. Sequence Labeling using Syntactic FeaturesDisease-related Event Recognition and Classification Summary & Future Work
4. Importance of the Problem influence on the travel and trade cause economic crises, political instability diseases, zoonotic in type can cause loss of life Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
5. Animal Disease Monitoring Systems - Automated Web Services Information retrieval system MedISys - http://medusa.jrc.it/medisys/homeedition/all/home.html Pattern-based Understanding and Learning System (PULS) - http://sysdb.cs.helsinki.fi/puls/jrc/all BioCaster - http://biocaster.nii.ac.jp/ HealthMap - http://healthmap.org/en EpiSpider- http://www.epispider.org/ Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
6. Limitations of the Existing Systems No timeline visualization (BioCaster) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
7. Problem Statement Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Introduce the following features to the framework for the epidemiological analysis: Classification of the disease-related documents collected from different domains Domain-specific entity extraction - animal disease names, viruses, disease serotypes Automated animal disease-related event recognition and classification from unstructured web data
8. Methodology Suppose we have a document collection D with documents collected from different domains C: news, web pages, scientific papers, medical literature, e-mails. We classify documents into two classes: disease-related documents DR; disease non-related document DNR. We extract a set of events E from every document di in DR for every domain cj. For every event ek in E we extract a set of domain-specific and domain-independent entities: disease, species, location, date, event status. We classify recognized events from E into: two classes – suspected or confirmed; three classes – susceptible, infected or recovered. Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
9. Related Work Approaches for text categorization: supervised, unsupervised and semi-supervised learning and different feature representations: “bag-of-words”, terms frequency, binary features, word bigrams, classification algorithms: lazy learners, decision trees, Naïve Bayes, Maximum Entropy. Entity extraction approaches: gazetteers, regular expressions, Hidden Markov Models and Conditional random Fields; ontology-based biomedical entity extraction. Relation extraction for automated ontology construction works. Animal disease-related event recognition methods. Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
10. Framework for Epidemiological Analytics Framework for Epidemiological Analysis Main Functional Components Data Collection (Document Relevance Classification) -> Data Sharing -> Search -> Data Analysis (Entity Extraction and Event Recognition) -> Visualization
11. Advantages of the Designed System Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
12. Phases of Data Processing Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
13. 1. Data Collection (1) Crawl the web using Heritrix crawler - http://crawler.archive.org/ set of seeds (ProMED-Mail, DEFRA etc.) set of terms (animal disease names from the ontology) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
14. 2. Data Sharing Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Document relevance classification Relevant Non-relevant
15. 3. Search Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Lucene-based* ranking Query-based keyword search Search by animal disease name and/or location *Lucene - http://lucene.apache.org
16. 4. Data Analysis Event example: “On 12 September 2007, a new foot-and-mouth disease outbreak was confirmed in Egham, Surrey” Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
17. 5. Visualization Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Map View GoogleMaps API - http://code.google.com/apis/maps/ TimeLine View SIMILE API - http://www.simile-widgets.org/timeline/
19. SupervisedLearningFramework New Documents DTest Feature Representation R1 … Feature Representation Rn Learned Model M1 … Learned Model Mk Crawled Documents DTrain Classifier Disease Related - DR (processed to the next phases) Disease Non-related – DNR (eliminated from the index) Feature Representations: R1 – “bag-of-words” binary, |R1|=28908 R2 – “bag-of-words” term frequency, |R2|=28908 R3 – “bag-of-words” bigrams, |R3|=99108 R4 – noun and verb keywords represented as binary counts, |R4|=2 R5 – noun and verb keywords normalized frequency, |R5|=2 Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
20. Experiment ADisease-related Document Classification ~1500 crawled documents Foot-and-mouth disease (FMD) Rift valley fever (RVF) Focused Crawl Terms [foot and mouth disease, FMD, rift valley fever, RVF] After labeling - 813 related and 752 non-related docs Testing with 10-fold cross validation + OR - Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
21. Classification Results: Precision, Recall,F-Measure, Area Under Curve Simplified Binary Counts as Features Simplified Noun and Verb Frequency as Features Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
22. Classification Results: Accuracy Comprehensive “Bag-of-words” Binary Features Comprehensive “Bag-of-words”, unigrams, bigrams and term frequency features Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
23. Summary (1) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 “Bag-of-words” representation gives higher accuracy; Generative approaches give the highest accuracy: Naïve Bayes together with comprehensive feature representation R3 using bigram as features – 0.97; Maximum Entropy classier using unigram “bag-of words” representation R2 – 0.96; Maximum Entropy classier using comprehensive binary counts as feature representation R1 – 0.94. Normalized term frequency is much better than just binary features.
24. Summary (2) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
25. Entity Extraction in the Domain of Veterinary Medicine (1) Ontology-based Entity Extraction Automated Ontology Construction
26. Domain Meta-data Domain-independent knowledge Domain-specific knowledge Location hierarchy names of countries, states, cities; Time hierarchy canonical dates. Medical ontology diseases, serotypes, and viruses. Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
27.
28.
29. 100 manually labeled document for entity extractionThesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
30. Entity Extraction Results: ROC Curves Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
31. Entity Extraction Results: Learning Curves |OG|=754..1238 |OR|=772..1287 Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
32. Entity Extraction in the Domain of Veterinary Medicine (2) Sequence Labeling using Syntactic Features with Sliding Window
33. Syntactic Feature Extraction POS tag numeric word-level feature Capitalization binary word-level feature Capitalization inside binary word-level feature for identifying abbreviations Position in the sentence numeric document-level feature Position in the document numeric document-level feature Frequency numeric document-level feature Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
34. Sequence Labeling Approach Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
36. Experiment CSequence Labeling using Syntactic Features 100 manually labeled documents from Experiment B Number of the disease names is more that 5 per document Keep capitalization Remove stop words 202977 examples in the dataset 80% for training (approx. 160000 examples) 20% for testing (approx. 40000 examples) Results are averaged over 3 runs We do not report accuracy because the data set is unbalanced (approx. 8570 positive examples vs. approx. 194430 negative examples) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
37. Entity Extraction Results: Precision, Recall, AUC (1) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
38. Entity Extraction Results: Precision, Recall, AUC (2) Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
39. Summary BioCaster named entity recognition system 200 news articles F-score – 76.9 for all named entity classes SVM and feature window -2/+1 including surface word, orthography, biomedical prefixes/suffixes, lemma, head noun etc. DNA, RNA, cell type extraction SVM and orthographic features F-score – 79.9 during the identification phase and 66.5 during the classification phase; Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
41. Animal Disease-related Event Types Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
42. Event Recognition Methodology Step 1. Entity recognition from raw text. Step 2. Sentence classification from which entities are extracted as being related to an event or not; if they are related to an event we classify them as confirmed or suspected. Step 3. Combination of entities within an event sentence into the structured tuples and aggregation of tuples related to the same event into one comprehensive tuple. Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
43. Step 1.Entity Recognition Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Locate and classify atomic elements into predefined categories: Disease names:“foot and mouth disease”, “rift valley fever”; viruses: “picornavirus”; serotypes: “Asia-1”; Species: “sheep”, “pigs”, “cattle” and “livestock”; Locationsof events specified at different levels of geo-granularity: “United Kingdom", “eastern provinces of Shandong and Jiangsu, China”; Datesin different formats: “last Tuesday”, “two month ago”.
44. Entity Recognition Tools Animal Disease Extractor* relies on a medical ontology, automatically-enriched with synonyms and causative viruses. Species Extractor* pattern matching on a stemmed dictionary of animal names from Wikipedia. Location Extractor Stanford NER Tool** (uses conditional random fields); NGA GEOnet Names Database (GNS)*** for location disambiguation and retrieving latitude/longitude. Date/Time Extractor set of regular expressions. *KDD KSU DSEx - http://fingolfin.user.cis.ksu.edu:8080/diseaseextractor/ **Stanford NER - http://nlp.stanford.edu/ner/index.shtml ***GNS - http://earth-info.nga.mil/gns/html/ Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
45. Step 2. Event Sentence Classification Constraint: True events should include a disease name together with a status verb from Google Sets* and WordNet** (eliminate event non-related sentences). “Foot and mouth disease is[V] a highly pathogenic animal disease”. Confirmed status verbs “happened” and verb phrases “strike out” “On 9 Jun 2009, the farm's owner reported[V] symptoms of FMD in more than 30 hogs”. Suspected status verbs “catch” and verb phrases “be taken in” “RVF is suspected[V] in Saudi Arabia in September 2000”. *GoogleSets - http://labs.google.com/sets **WordNet - http://wordnet.princeton.edu/ Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
46. Step 3. Event Tuple Generation Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Event attributes: disease date location species confirmation status Event tuple: Eventi = < disease; date; location; species; status > = <FMD, 9 Jun 2009, Taoyuan, hog, confirmed> Event tuple with missing attributes: Eventj = <FMD, ?, ?, ?, confirmed>
47. Event Recognition Workflow Step 1: Entity Recognition Foot-and-mouth disease[DIS]on hog[SP] farm in Taoyuan[LOC]. Taiwan's TVBS television station reports that agricultural authorities confirmed foot-and-mouth disease[DIS] on a hog[SP] farm in Taoyuan[LOC]. On 9 Jun 2009[DT], the farm's owner reported symptoms of FMD[DIS] in more than 30 hogs[SP]. Subsequent testing confirmed FMD[DIS]. Agricultural authorities asked the farmer to strengthen immunization. The outbreak has not affected other farms. Authorities stipulated that the affected hog[SP] farm may not sell pork for 2 weeks. Step 2: Sentence Classification YES 1. Foot-and-mouth disease[DIS]on hog[SP] farm in Taoyuan[LOC]. YES 2.Taiwan's TVBS television station reports that agricultural authorities confirmedfoot-and-mouth disease[DIS]on a hog[SP] farm in Taoyuan[LOC]. YES 3. On 9 Jun 2009[DT], the farm's owner reported symptoms of FMD[DIS] in more than 30 hogs[SP]. YES 4. Subsequent testing confirmedFMD[DIS]. NO 5. Agricultural authorities asked the farmer to strengthen immunization. NO 6. The outbreak has not affected other farms. NO 7. Authorities stipulated that the affected hog[SP] farm may not sell pork for 2 weeks. Step 3a: Tuple Generation E1 = <Foot-and-mouth disease, ?, Taoyuan, hog, ?> E3 = <FMD, 9 Jun 2009,?, hog, reported> E2 = <Foot-and-mouth disease, ?, Taoyuan, hog, confirmed > E4 = <FMD, ?, ?, ?, confirmed> Step 3b: Tuple Aggregation E = <disease, date, location, species, status> = <Foot-and-mouth disease, 9 Jun 2009, Taoyuan, hog, confirmed > Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
48. Experiment DEvent Recognition and Classification The First International Workshop on Web Science and Information Exchange in the Medical Web (MedEx 2010) ~100 event-related documents Foot-and-mouth disease (FMD) Rift valley fever (RVF) Manually created 2 sets of summaries for 100 docs DUCView Pyramid Scoring Tool* – Score [0..1] relies on multiple summaries to assign the significance weights to summarization content units (i.e., entities) to compare automatically generated event tuples with entities from human summaries. Scorei = < wddisease; wtdate; wllocation; wsspecies; wcstatus… >, subject to disease + status = 2
49. Event Score Distribution by Range We interpret the Pyramid score values as an event extraction accuracy: # of unique contributing entities (TP); # of entities not in the summary (FP); # of extra contributing entities from summary (FN). multiple summaries – majority voting for annotation. The First International Workshop on Web Science and Information Exchange in the Medical Web (MedEx 2010)
51. ENTITY EXTRACTION Document 3, sentence s31 Almost 2000 cattle[SP] are waiting to be slaughtered on 02/28/2001[DATE]since the resurgence of FMD[DIS] in Northumberland[LOC]. Document 2, sentence s21 The UK Ministry of Agriculture confirmed on 2/20/01[DATE] that 27 pigs[SP] found with vesicles in an abattoir near Brentwood, Essex[LOC] have FMD[DIS]. Document 1, sentence s11, s12 The signs suggested the 27 pigs[SP] could be suffering from foot and mouth disease[DIS] in Anglesey, Wales[LOC].It was reported on 02/18/01[DATE]. … … EVENT TUPLE GENERATION e11 = [27 pigs, FMD, ?, Anglesey, Wales, “suggest”] e12 = [?,?, 02/18/01, ?, “report”] e21 = [27 pigs, FMD, 2/20/01, Brentwood, Essex, “confirm”] e31 = [2000 cattle, FMD, 2/28/01, Northumberland, “slaughter”] EVENT TUPLE CLASSIFICATION Susceptible Recovered Infected EVENT TUPLE AGGREGATION E2 = [27 pigs, FMD, 2/20/01, Brentwood, Essex, Infected] E3 = [2000 cattle, FMD, 2/28/01, Northumberland, Recovered] E1= [27 pigs, FMD, 02/18/01, Anglesey, Wales, Susceptible] Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
52. The spread of foot-and-mouth disease outbreak in UK, 2001 118 ProMed-Mail reports yellow - susceptible red - infected green - recovered Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
53. Summary The accuracy of the event recognition depends on the separate entity extraction accuracy The event aggregation and deduplication requires much comprehensive heuristics and additional knowledge, for example co-reference resolution BioCaster 950 disease-location pairs per month reported results - 887/950 correct disease-location pairs and 0.934 precision MedISys/PULS 100 English-language documents with 156 events Reported results – 0.88 precision Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
54. Conclusions, Contributions and Future Work Summary: 1. Disease-related Document Classification 2. Ontology-based Entity Extraction 3. Entity Extraction using Sequence Labeling 4. Event Recognition and Classification
55. Conclusions Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Disease-related Document Classification supervised framework feature representations and classification algorithms Ontology-based Domain-specific Entity Extraction semantic relationship extraction approach sequence labeling using syntactic patterns Event Recognition and Classification novel sentence-based approach
56. Contributions Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Paper “Computational Knowledge and Information Management in Veterinary Epidemiology” IEEE Intelligence and Security Informatics Conference (ISI'10), 23-26 May 2010, Vancouver, BC, Canada Paper “Animal Disease Event Recognition and Classification” First International Workshop on Web Science and Information Exchange in the Medical Web (MedEx'10), WWW Conference, 26-30 April 2010, Raleigh, NC, USA Paper “Boosting Biomedical Entity Extraction by Using Syntactic Patterns for Semantic Relation Discovery” (to appear) 2010 IEEE/WIC/ACM International Conference on Web Intelligence (WI'10), August 31 - September 3, York University, Toronto, Canada Poster “Named Entity Recognition and Tagging in the Domain of Epizootics” Women in Machine Learning Workshop (WiML'09) Workshop, 6-7 Dec 2009, Vancouver, Canada ACM Poster Presentation Competition “Automated Event Extraction and Named Entity Recognition in the Domain of Veterinary Medicine” 2010 Grace Hopper Celebration of Women in Computing (GHC'10),September 28 - October 1, Atlanta, Georgia, USA
57. Future Work Domain-specific Entity Extraction multilingual ontology construction using Wikipedia. Automated Ontology Construction generalize for other named entities Event Recognition and Classification deeper syntactic analysis co-reference resolution Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010
58. Acknowledgments Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010 Faculty: Dr. William H. Hsu Dr. Doina Caragea Dr. Gurdip Singh KDD Lab alumni: Tim Weninger (crawler deployment) and Jing Xia (rule-based event extraction) KDD Lab assistants: Information Extraction Team: John Drouhard, Landon Fowles, Swathi Bujuru Spatial Data Mining Team: Wesam Elshamy, AndrewBerggren Topic Detection & Tracking Team: Danny Jones, Srinivas Reddy Fulbright Program supported by the US Department of State's Bureau of Education and Cultural Affairs
59. Thank you! Svitlana Volkova, svitlana.volkova@gmail.com http://people.cis.ksu.edu/~svitlana Thesis "Entity Extraction, Animal Disease-related Event Recognition and Classification from Web", July 30 2010