Enabling Exploration through Text Analytics

The 2009 Text Analytics Summit presentation discusses the importance of text analytics in enhancing exploratory search capabilities for various information-seeking needs, such as health and work-related queries. It highlights real-world examples and outlines key takeaways regarding text analytics' role in categorization, entity detection, and user engagement. It concludes by asserting that text analytics technology is currently available and essential for effective exploration, advocating for improvements in the domain.

Transcript

1. © 2009 Endeca Technologies, Inc. All rights reserved. Enabling Exploration through Text Analytics Daniel Tunkelang Chief Scientist, Endeca
2. © 2009 Endeca Technologies, Inc. All rights reserved. 2 overview information seeking tools need to support exploration text analytics can help you can do this here and now
3. © 2009 Endeca Technologies, Inc. All rights reserved. 3 real-world information seeking examples • looking for health information • looking for work-related information reminder search and text analytics are a means, not an end
4. © 2009 Endeca Technologies, Inc. All rights reserved. 4 example 1: looking for health information six months into my wife’s pregnancy, we discovered that she had gestational diabetes how to learn more?
5. © 2009 Endeca Technologies, Inc. All rights reserved. 5 google: the default option for most
6. © 2009 Endeca Technologies, Inc. All rights reserved. 6 in government we trust: fda.gov
7. © 2009 Endeca Technologies, Inc. All rights reserved. 7 maybe the private sector knows best: webmd powered by
8. © 2009 Endeca Technologies, Inc. All rights reserved. 8 success – and a sticky site powered by
9. © 2009 Endeca Technologies, Inc. All rights reserved. 9 example 2: looking for work-related information need to ramp up summer interns on text mining how to find a good book?
10. © 2009 Endeca Technologies, Inc. All rights reserved. 10 let’s try google again
11. © 2009 Endeca Technologies, Inc. All rights reserved. 11 google: the gateway to wikipedia?
12. © 2009 Endeca Technologies, Inc. All rights reserved. 12 the library of congress (loc.gov)
13. © 2009 Endeca Technologies, Inc. All rights reserved. 13 triangle research libraries: next-gen catalog powered by
14. © 2009 Endeca Technologies, Inc. All rights reserved. 14 faceted search enables query refinement powered by
15. © 2009 Endeca Technologies, Inc. All rights reserved. 15 take-away #1 exploratory search support: a must-have for many information needs
16. © 2009 Endeca Technologies, Inc. All rights reserved. 16 text analytics • categorization • named entity detection • term extraction • sentiment analysis vague term, lots of see-alsos text mining information extraction content enrichment
17. © 2009 Endeca Technologies, Inc. All rights reserved. 17 newssift: text analytics enabling exploration powered by categorization named entity detection term extraction sentiment analysis
18. © 2009 Endeca Technologies, Inc. All rights reserved. 18 exploring the news about facebook powered by
19. © 2009 Endeca Technologies, Inc. All rights reserved. 19 facebook: the good powered by Social Utility Iphone Application
20. © 2009 Endeca Technologies, Inc. All rights reserved. 20 facebook: the bad powered by Criminal Behavior Litigation And Settlement
21. © 2009 Endeca Technologies, Inc. All rights reserved. 21 take-away #2 text analytics enable exploratory search
22. © 2009 Endeca Technologies, Inc. All rights reserved. 22 text analytics is here and now ? ? ?
23. © 2009 Endeca Technologies, Inc. All rights reserved. 23 lots of off-the-shelf options and more!
24. © 2009 Endeca Technologies, Inc. All rights reserved. 24 caveats • rule-based techniques are domain-specific • statistical techniques rely on trained models • plan for errors, inconsistency • document vs. corpus analysis
25. © 2009 Endeca Technologies, Inc. All rights reserved. 25 Person Location Organization ABDUL-KARIM KHALAF (1) ALTOONA, PA (1) ABC News Inc. (1) ABDULRAHMAN ABDULLAH (1) Afghanistan (7) Air Force (1) AL GORE (1) Africa (5) Amazon.com Inc. (1) ALEX TREBEK (1) Akihabara (1) American Airlines Inc. (1) ALI HASSAN AL (1) Alaska (3) Apple (1) AMANDA MARCOTTE (1) Allegheny (1) Arctic National Wildlife Refuge (1) AMY WINEHOUSE (1) Americas (17) Arianna Huffington (1) ANDERS ERICSSON (1) Appalachia (1) Australian Liberal Party (1) ANDREW LLOYD WEBBER (1) Argentina (1) Bad News Bears (1) ANTHONY MWANGI (1) Arizona (11) Bear Stearns (2) ANTONIN SCALIA (1) Arkansas (7) Big Apple Companies (1) ARYE BARAK (1) Arlington, Va. (2) BioDiversity Research Institute (1) Aaron Sorkin (1) Arrest (1) Bloomberg LP (3) Abbie Hoffman (1) Asia (1) Bob Dole (1) Abe Lincoln (1) Atlanta (2) Bocuse d’Or World Cuisine Contest (1) Abe Weiss (1) Austin (1) Boston Globe (1) Abraham Lincoln (1) Austin, Texas (1) Boston Tea Party (1) Adlai Stephenson (1) Australia (1) Budweiser (1) problems with entity extraction • moderate precision, but low recall • not just noisy, but inconsistent • corpus analysis can help! Arrest (1) Asia (1) ALTOONA, PA (1) Abe Lincoln (1) Bob Dole (1) Boston Tea Party (1) Abraham Lincoln (1)
26. © 2009 Endeca Technologies, Inc. All rights reserved. 26 look for ways to cheat! recall precision
27. © 2009 Endeca Technologies, Inc. All rights reserved. 27 division of labor people supply vocabulary machine annotates documents http://www.precolumbianwomen.com/images/inca-labor.10.gif
28. © 2009 Endeca Technologies, Inc. All rights reserved. 28 example: ACM digital library • opportunity – repository of (sometimes) author-tagged documents – high-precision tags: very few false positives • challenge – poor reuse of vocabulary: most tags unique – low-recall tags: 90% false negatives as is, tags were not useful for exploration
29. © 2009 Endeca Technologies, Inc. All rights reserved. 29 solution • bootstrap on author-supplied tags • prune 600K+ tags to 10K by – imposing frequency threshold – normalizing by case and singular/plural – eliminating infrequent subphrases • mine documents using resulting vocabulary • manually validate most frequently assigned tags
30. © 2009 Endeca Technologies, Inc. All rights reserved. 30 example: a search for boeing powered by
31. © 2009 Endeca Technologies, Inc. All rights reserved. 31 it’s a HITS!
32. © 2009 Endeca Technologies, Inc. All rights reserved. 32 if you prefer sports to computer science • no author-supplied tags • use search logs instead • supplement with authority files – team names – player names • mine documents using resulting vocabulary
33. © 2009 Endeca Technologies, Inc. All rights reserved. 33 roger clemens, then and now powered by
34. © 2009 Endeca Technologies, Inc. All rights reserved. 34 pivoting to a different view powered by
35. © 2009 Endeca Technologies, Inc. All rights reserved. 35 take-away #3 this is not vapor ware; text analytics to enable exploration is available here and now
36. © 2009 Endeca Technologies, Inc. All rights reserved. 36 looking forward • better tags are the beginning, not the end • improve with manual and automatic processing • give users control over precision / recall trade-off • help users and content creators help you
37. © 2009 Endeca Technologies, Inc. All rights reserved. 37 in closing exploratory search = must-have, not nice-to-have text analytics are a key enabler the technology is real, here, and now
38. © 2009 Endeca Technologies, Inc. All rights reserved. 38 thank you…and come to SIGIR! communication 1.0 email: dt@endeca.com communication 2.0 blog: http://thenoisychannel.com twitter: http://twitter.com/dtunkelang SIGIR: July 19-23 in Boston Industry Track on July 22nd!