Finding Question-Answer Pairs from Online Forums

1
Finding Question-Answer Pairs from Online Forums

• ACM , SIGIR 08

Gao Cong Aalborg University, Aalborg, Denmark
Long Wang Tianjin University, Tianjin, China
Chin-Yew Lin Microsoft Research Asia, Beijing,
China Young-In Song Korea University, Seoul,
South Korea Yueheng Sun Tianjin University,
Tianjin, China
2
Introduction

• Yahoo! Answers.
• Forums contain a huge amount of valuable user
  generated content on a variety of topics.
• Find Question-Answer pair in forums.

3
Algorithms

• Question Detection
• 5W1H
• Most of questions are not begin with 5W1H.
• Question Mark
• 30 questions do not end with question mark.
• I am wondering where I can buy cheap and good
  clothing in beijing.
• Labeled Sequential Pattern (LSP)

4
Graph based propagation method

• Building Graph
• Given a question q, and the set A_q of its
  candidate answers.
• For 2 candidate answers a1 a2 , compute KL(a1a2)
• If 1/(1KL(a1a2)) is lager than a threshold ?,
  then add an edge from a1 to a2.

5
Graph based propagation method

• Edge Weight
• Normalized
• ?0.01

6
Computing Propagated Scores

• Propagation without initial score
• Propagation with initial score

7
Answer Detection

• score(q,a)
• Cosine Similarity.
• Query likelihood language model.
• KL-divergence language model.

8
Experiment

• Data
• Select three forums of different scales to obtain
  source data.
• Two annotators
• The kappa statistic for identifying questions is
  0.96.
• The kappa statistic for linking answers and
  questions given a question is 0.69.

9
Experiment

• Q-Tinter intersection of two annotators.

10
Experiment

• 1,535 questions from 600 threads, 284 questions
  do not have answers.

11
Experiment

• Improved results on subsets
• Of 486 first questions, only 21 of them do not
  have answers for A-TUnion data and 45 for
  A-TInter data.

12
Experiment

• G_K Computing weight with KL-Divergence alone.
• G_1 Propagation without initial score.
• G_2 Propagation with initial score.

13
Experiment

• Data from three forum
• Tripadvisor, Lonely Planet, Bootsnall.

14

• Thank You.

15
Experiment