Leveraging Knowledge Bases for Contextual Entity Exploration

Joonseok Lee, Google Ariel Fuxman, Google Bo Zhao, LinkedIn Yuanhua Lv, Microsoft

1 Contextual Insight Problem

2 Contextual Insight Problem

Within application

Silas Deane Silas Deane (Dec 24, 1737 – Sep 23, 1789) was an American merchant, politician and diplomat. Entity linking Revolutionary War Silas Deane took an active part in the movements in Connecticut preceding the American Revolutionary war. Contextual Benjamin Franklin Exploration Subsequently Deane became, with Benjamin Franklin and Arthur Lee, one of the regularly accredited commissioners to France.

3 Query Rewriting Approaches

User Context Query Filtering Selection Extraction Rewriting

Search Ticonderoga Engine Connecticut Benedict Arnold Silas Deane Silas Deane (Dec 24, 1737 – Sep 23, 1789) was an American merchant, politician and diplomat. “Silas Deane” rankonly:Connecticut rankonly:Ticonderoga Silas Deane Silas Deane House Silas Deane house is in Wethersfield. James Easton Connecticut was the home of Silas Deane.

Samuel Herrick

Massachusetts Committee of Safety Green Mountain Boys

4 Proposed Approach: Leveraging Semantics in KB

Ticonderoga Connecticut Benedict Arnold

Continental Congress Silas Deane American Revolutionary War James Easton

Samuel Herrick B. Franklin

Massachusetts Committee of Safety Green Mountain Boys

5 Problem Definition

Context User Selection

Contextual Insight System Knowledge Base (Wikipedia graph) Recommendation List

Previous work leveraging Wikipedia graph: [Yeh2009] WikiWalk: random walks on Wikipedia for semantic relatedness. [Agrawal2014] Similarity Search using Concept Graphs. [Milne2008] An Effective, Low-Cost Measure of Semantic Relatedness Obtained from Wikipedia Links. 6 Focused Sub-graph Construction

Green Mt. Boys Fort Ticonderoga Connecticut

American Revolutionary War Silas Deane House

Silas Deane

Arthur Lee John Adams User Selection Benjamin Franklin Context Nodes

General Nodes 7 Personalized Random Walk

0.04% 0.06% 0.05% Green Mt. Boys Fort Ticonderoga Connecticut

0.36% 0.17% American Revolutionary War Silas Deane House

28.71% Silas Deane 0.48% 0.46% Arthur Lee John Adams 0.57% Benjamin Franklin

8 Incorporating Contextual Signals

Green Mt. Boys Fort Ticonderoga Connecticut

American Revolutionary War Silas Deane House

Silas Deane

Arthur Lee John Adams

Benjamin Franklin

9 Context-Selection Betweenness

• Retrieve nodes that play an important role connecting the user selection and context nodes. • Idea from Betweenness Centrality: the number of shortest paths from all vertices to all others that pass through that node. [Freeman1977] • Credit each node when a shortest path from the user selection to each context node passes through it.

10 Context-Selection Betweenness Example

0.33 0.33 0.33 Green Mt. Boys Fort Ticonderoga Connecticut

0.670.33 American Revolutionary War Silas Deane House

Silas Deane

Arthur Lee John Adams

Benjamin Franklin

11 Score Aggregation

Random Walk score: Context-Selection Betweenness score: How many times the Expected number of shortest paths user is likely to visit the from user selection to any context page page v. through v.

12 Evaluation: Crowd-sourced Data

• Corpora • 2,600 textbooks from Wikibooks

• Foci of attention • Workers selected phrases for which they would like to learn more. • 100 workers per paragraph (500 paragraphs per corpus).

• Relevance labeling • Each (focus, result) pair evaluated by 10 crowd workers. • Large scale: over 145,000 labels.

13 Evaluation

14 Comparing Proposed Components

15 Score decomposition with Silas Deane Example

Rank Page Title CSB RW Total 1 Silas Deane 402.26 6442.53 6844.78 2 American Revolutionary War 49.58 82.42 131.99 3 Benjamin Franklin 2.46 128.17 130.63 4 Arthur Lee (diplomat) 0.00 107.99 107.99 5 Thomas Jefferson 2.46 98.82 101.28 6 Continental Congress 19.58 74.02 93.60 7 Continental Army 30.25 61.45 91.70 8 Capture of Fort Ticonderoga 49.58 39.56 89.14

16 Examples

Entity linking Two independent expeditions to capture Ticonderoga – one Silas Deane out of Massachusetts and the other from Connecticut – were organized. At Cambridge, Massachusetts, Benedict Silas Deane House Current Arnold told the Massachusetts Committee of Safety about Connecticut Route 99 Leibniz the cannon and other military stores at the lightly defended fort. On May 3, 1775, the Committee gave Arnold a Benjamin Franklin colonel's commission and authorized him to command a Arthur Lee (diplomat) Random Walk secret mission to capture the fort. Ethan Allen demanding (Semantic) the surrender of Fort Ticonderoga Meanwhile, in Hartford, John Adams Connecticut, Silas Deane and others had organized an Thomas Jefferson expedition of their own. Ethan Allen assembled over 100 of his Green Mountain Boys, about 50 men were raised by American Revolutionary War James Easton at Pittsfield, Massachusetts, and an additional Continental Congress Context-Selection 20 men from Connecticut volunteered. This force of about Betweenness Continental Army 170 gathered on May 7 at Castleton, Vermont. (Context) Capture of Fort Ticonderoga

17 Examples Entity linking In 1895, when Joyce was in his third year at Belvedere Oxford English Dictionary College, he chose Ulysses as his subject for an essay entitled "My Favourite Hero". In English Ulysses is sometimes Oxford spelling Current stressed on the second syllable, and this is the Dictionary Leibniz pronunciation required in most verse translations of the Homeric epics. Joyce, however, always referred to his novel New Oxford American Dictionary as YOOL-i-seez, with the stress on the first syllable. This Concise Oxford English Dictionary pronunciation is sanctioned by the Oxford English Dictionary and is used almost universally in Ireland when one is English language Random Walk referring to the book. In his design for the cover of the 1949 Greek language Random House edition of Ulysses, the American artist (Semantic) Edward McKnight Kauffer emphasized the initial UL, "giving Old English graphic form to the phonetic structure of the title with its Standard English Context-Selection accent on the first syllable." Received Pronunciation Betweenness British English (Context)

18 Take-home Messages

• Contextual entity exploration problem

• Knowledge-graph-based solution • Semantics incorporated • Context-relevant retrieval • Nice performance in IR measures

• Crowd-sourced evaluation

19 Thank you for your attention.

20