Technical side of Watson, David Boloker

1. CTO, Emerging Technologies, IBM

2. About

3. Motivation

3.1. Offload more of the decision maker tasks to the engine

4. Strategy

4.1. read lots of texts

4.2. analyze subject-verb-object

4.3. build semantic network

4.4. find patterns

4.4.1. officials submit resignations (0.8)

5. Keywords match isn't enough

5.1. need

5.1.1. temporal reasoning

5.1.2. statistical paraphrasing

5.1.3. geospatial reasoning

6. Arch

6.1. decompose the question

6.2. run many many searches

6.2.1. our data

6.2.1.1. 2 FTS engines

6.2.1.1.1. Lucene

6.2.1.1.2. Indri

6.3. generate hypothesis

6.4. evidence sourcing

6.4.1. get confidence on hypothesis

6.5. deep evidence scoring

6.6. sythesis

6.7. apply machine learning model to get final confidence

6.7.1. learning from its mistakes

6.7.2. training data is the archive of all historic jeopardy games

6.8. output answer & confidence

7. Had iterations in architecture, each time improving accuracy

7.1. till they got 87%

8. HW

8.1. 2890 cores

8.1.1. very parallel architecture

8.2. 90 severs

8.3. 16TB RAM

8.4. 20TB disk

9. Challenges

9.1. Real language is full of slang, & metaphores

9.1.1. e.g., pun in question

10. History

10.1. PIQUANT

10.2. OpenEuphyra

11. uses UIMA for SOA

11.1. 100 annotators

11.1.1. in

11.1.1.1. Java

11.1.1.2. Prolog

11.2. reveal different kind of features

12. arch high-level

12.1. learn

12.1.1. ingest many sources

12.1.1.1. wikipedia

12.1.1.2. yago2

12.1.1.3. dbpedia

12.1.1.4. wordnet many more

12.1.2. store all in RDF store

12.2. Question analysis

12.3. Primary search

12.4. Shallow & deep scoring

12.5. Merging & ranking

13. Demo

13.1. Audience question

13.1.1. Country that has the largest solar dish?

13.1.1.1. Israel 24%

13.1.1.2. Untitled

13.1.1.3. Negev <10%

13.2. Demo machine uses only 7 servers

13.2.1. takes about 20 seconds