Lecture 13 Information Extraction

1. Lecture 13Information Extraction CSCE 771 Natural Language Processing • Topics • Name Entity Recognition • Relation detection • Temporal and Event Processing • Template Filling • Readings: Chapter 22 February 27, 2013

2. Overview • Last Time • Dialogues • Human conversations • Today • Slides from Lecture24 • Dialogue systems • Dialogue Manager Design • Finite State, Frame-based, Initiative: User, System, Mixed • VoiceXML • Information Extraction • Readings • Chapter 24, Chapter 22

3. Information extraction • Information extraction – turns unstructured information buried in texts into structured data • Extract proper nouns – “named entity recognition” • Reference resolution – \ • named entity mentions • Pronoun references • Relation Detection and classification • Event detection and classification • Temporal analysis • Template filling

4. Template Filling • Example template for “airfare raise”

5. Figure 22.1 List of Named Entity Types

6. Figure 22.2 Examples of Named Entity Types

7. Figure 22.3 Categorical Ambiguities

8. Figure 22.4 Categorical Ambiguity

9. Figure 22.5 Chunk Parser for Named Entities

10. Figure 22.6 Features used in Training NER • Gazetteers – lists of place names • www.geonames.com • www.census.gov

11. Figure 22.7 Selected Shape Features

12. Figure 22.8 Feature encoding for NER

13. Figure 22.9 NER as sequence labeling

14. Figure 22.10 Statistical Seq. Labeling

15. Evaluation of Named Entity Rec. Sys. • Recall terms from Information retreival • Recall = #correctly labeled / total # that should be labeled • Precision = # correctly labeled / total # labeled • F- measure where βweights preferences • β=1 balanced • β>1 favors recall • β<1 favors precision

16. NER Performance revisited • NER performance revisited • Recall, Precision, F • High performance systems • F ~ .92 for PERSONS and LOCATIONS and ~.84 for ORG • Practical NER • Make several passes on text • Start by using highest precision rules (maybe at expense of recall) make sure what you get is right • Search for substring matches or previously detected names using probabilistic searches string matching metrics(Chap 19) • Name lists focused on domain • Probabilistic sequence labeling techniques using previous tags

17. Relation Detection and classification • Consider Sample text: • Citing high fuel prices, [ORG United Airlines] said [TIME Friday] it has increased fares by [MONEY $6] per round trip on flights to some cities also served by lower-cost carriers. [ORG American Airlines], a unit of [ORG AMR Corp.], immediately matched the move, spokesman [PERSON Tim Wagner] said. [ORG United Airlines] an unit of [ORG UAL Corp.], said the increase took effect [TIME Thursday] and applies to most routes where it competes against discount carriers, such as [LOC Chicago] to [LOC Dallas] and [LOC Denver] to [LOC San Francisco]. • After identifying named entities what else can we extract? • Relations

18. Fig 22.11 Example semantic relations

19. Figure 22.12 Example Extraction

20. Figure 22.13 Supervised Learning Approaches to Relation Analysis • Algorithm two step process • Identify whether pair of named entities are related • Classifier is trained to label relations

21. Factors used in Classifying • Features of the named entities • Named entity types of the two arguments • Concatenation of the two entity types • Headwords of the arguments • Bag-of-words from each of the arguments • Words in text • Bag-of-words and Bag-of-digrams • Stemmed versions • Distance between named entities (words / named entities) • Syntactic structure • Parse related structures

22. Figure 22.14 a-part-of relation

23. Figure 22.15 Sample features Extracted

24. Bootstrapping Example “Has a hub at” • Consider the pattern • / * has a hub at * / • Google search • 22.4 Milwaukee-based Midwest has a hub at KCI • 22.5 Delta has a hub at LaGuardia • … • Two ways to fail • False positive: e.g. a star topology has a hub at its center • False negative? Just miss • 22.11 No frill rival easyJet, which has established a hub at Liverpool

25. Figure 22.16 Bootstrapping Relation Extraction

26. Using Features to restrict patterns • 22.13 Budget airline Ryanair, which uses Charleroi as a hub, scrapped all weekend flights • / [ORG] , which uses a hub at [LOC] /

27. Semantic Drift • Note it will be difficult (impossible) to get annotated materials for training • Accuracy of process is heavily dependant on initial sees • Semantic Drift – • Occurs when erroneous patterns(seeds) leads to the introduction of erroneous tuples

28. Fig 22.17 Temporal and Durational Expressions • Absolute temporal expressions • Relative temporal expressions

29. Fig 22.18 Temporal lexical triggers

30. Fig 22.19 MITRE’s tempEx tagger-perl

31. Fig 22.20 Features used to train IOB

32. Figure 22.21 TimeML temporal markup

33. Temporal Normalization • iSO 8601 - standard for encoding temporal values • YYYY-MM-DD

34. Figure 22.22 Sample ISO Patterns

35. Event Detection and Analysis • Event Detection and classification

36. Fig 22.23 Features for Event Detection • Features used in rule-based and statistical techniques

37. Fig 22.24 Allen’s 13 temporal Relations

38. Figure 22.24 continued

39. Figure 22.25 Example from Timebank Corpus

40. Template Filling

41. Figure 22.26 Templates produced by Faustus 1997

42. Figure 22.27 Levels of processing in Faustus

43. Figure 22.28 Faustus Stage 2

44. Figure 22.29 The 5 Partial Templates of Faustus

45. Figure 22.30 Articles in PubMed

46. Figure 22.31 biomedical classes of named entities