1. Connecting the Dots Between News Article
KDD‘10
Advisor: Jia Ling, Koh
Speaker: Yu Cheng, Hsieh

2. Outline Introduction Scoring a chain Formalize story coherence
Measuring influence
Finding a good chain
Evaluation
Interaction Model

3. Introduction
Users are constantly struggling to keep up with the large amounts of content that is being published every day.
With this much data, it is easy to miss the big picture.
Investigate methods for automatically connecting the dots.

4. Connecting the mortgage crisis to healthcare
This chain should be coherent
The user should gain better understanding of the progression of the story

5. Scoring a chain

7. Formalizing story coherence

8. Formalizing story coherence

9. Formalizing story coherence
Advantage:
- Positioning similar documents next to each other
- Rewards long stretches of words
Disadvantage:
- Overlook importance of a word
- Missing Words
- Overlook weak links

10. Formalizing story coherence

11. Formalizing story coherence

12. Formalizing story coherence

13. Formalizing story coherence

14. Formalizing story coherence
Jitteriness: topics that appear and disappear throughout the chain
- Only consider the longest continuous
stretch of each word.
- This way, going back-and-forth between
two topics provides no utility after the first
topic switch

15. Formalizing story coherence

16. Measuring influence

17. Measuring influence

18. Measuring influence

19. Measuring influence

20. Measuring influence

21. Finding a good chain

22. Finding a good chain Linear Programming - Chain Restriction
- Smoothness
- Activation Restriction
- Minmax Objective

23. Linear Programming

24. Linear Programming

25. Linear Programming

26. Linear Programming Minmax Objective
- Minedge is the minimum of all active edge scores

27. Evaluation More than half million real news articles were used.
Major news stories of recent years are considered.
For each story, selecting an initial subset of 500 – 10,000 candidate articles, based on keyword-search
Named entities and noun phrases were extracted from each article(remove infrequent name entities and non-informative noun phrase)

28. Evaluation Stories linking technique - Connecting-Dots - Shortest-path
- Google News Timeline(GNT)
- Event threading(TDT)

29. Evaluation Shortest path Google news timeline GNT
constructed a graph by connecting each document with its nearest neighbor based on Cosine similarity
Google news timeline GNT
- Using query string to get articles
- Construct query string for each story, based on
s and t
- Picked K equally-spaced documents between the dates
of the original query article

30. Evaluation

31. Evaluation 18 users with a pair of source an target articles
Gauged users familiarity with those articles
Ask whether they believe they knew a coherent story linking them together( on scale )
Ask user to indicate
- Relevance
- Coherence
- Non-Redundancy

32. Evaluation

33. Evaluation

34. Interaction Models Refinement:
- Users might be especially interested in a
specific part of the chain
- A refinement may consist of adding a
new article, or replacing an article

35. Interaction Models

36. Interaction Model

37. Evaluation Refinement
- Return two chains, obtained from the original chain by
(1) our local search
(2) adding an article chosen randomly from a subset of
candidate articles
- User preferred the local-search chains 72% of the
time

38. Evaluation User Interests - Two chains are showed to users
1 Obtained from the other by increasing the importance
of 2-3 words
2 Show them a list of ten words containing the words
(1) words whose importance we increased
(2) randomly chosen words
asked which words they would pick in order to obtain
the seconds chain from the first. The goal was to see if
users can identify at least some of the words
- User identified at least one word 63.3% of the time

39. Conclusion & Future Work
Describe problem of connecting the dots.
Explore different desired properties of a good story, formalized it as a linear program
Provided an efficient algorithm to connect two articles
Allowing more complex tasks