online debate summarization using topic directed sentiment analysis

34
+ Online Debate Summarization Using Topic Directed Sentiment Analysis 2014/10/31 (Fri.) Chang Wei-Yuan @ MakeLab Group Meeting Sarvesh Ranade, Jayant Gupta, Vasudeva Varma, Radhika Mamidi WISDOM ‘13

Upload: chang-wei-yuan

Post on 01-Jul-2015

283 views

Category:

Data & Analytics


2 download

DESCRIPTION

Online debate summarization using topic directed sentiment analysis @ WISDOM'13

TRANSCRIPT

Page 1: Online Debate Summarization using Topic Directed Sentiment Analysis

+

Online Debate Summarization Using Topic Directed Sentiment Analysis

2014/10/31 (Fri.)�Chang Wei-Yuan @ MakeLab Group Meeting

Sarvesh Ranade, Jayant Gupta, Vasudeva Varma, Radhika Mamidi �WISDOM ‘13

Page 2: Online Debate Summarization using Topic Directed Sentiment Analysis

+Outline

n Introduction �

n Method �

n Experiment �

n Conclusion �

n Thought

2

Page 3: Online Debate Summarization using Topic Directed Sentiment Analysis

+Outline

n Introduction �

n Method �

n Experiment �

n Conclusion �

n Thought

3

Page 4: Online Debate Summarization using Topic Directed Sentiment Analysis

+ Introduction 4

Page 5: Online Debate Summarization using Topic Directed Sentiment Analysis

+ Introduction

n Online debate forum�n take a stance and argue debate topics �n dynamic and increase rapidly�

n This paper aims to summarize online debates. �n extracting highly topic relevant �n sentiment rich sentences �

n Effective opinion summarization without going through the entire debate.

5

Page 6: Online Debate Summarization using Topic Directed Sentiment Analysis

+Outline

n Introduction �

n Method �

n Experiment �

n Conclusion �

n Thought

6

Page 7: Online Debate Summarization using Topic Directed Sentiment Analysis

+Method

n Extractive summaries are generated by ranking the Dialogue Acts (DAs) from the original documents. �n DA is a smallest unit of debate �

7

Page 8: Online Debate Summarization using Topic Directed Sentiment Analysis

+ 8

Score

Score

Score

Score

Score

Page 9: Online Debate Summarization using Topic Directed Sentiment Analysis

+

n Most highly ranked DAs are chosen until summary length constraint is satisfied. �

n Scores of DAs n 

9

Equation.(Scores of DAs)�

where λ is weighted, s is a DA of the Document, D is the Document.

Page 10: Online Debate Summarization using Topic Directed Sentiment Analysis

+

Feature Category Feature Names

Topic Relevance Topic Directed Sentiment Score Topic Co-occurrence

Document Relevance tf-idf Sentiment Score

Sentiment Relevance

Number of Sentiment Words Sentiment Strength

Context Relevance

Sentence position Sentence length

10

n Argument Structure Examples

Page 11: Online Debate Summarization using Topic Directed Sentiment Analysis

+Word Sentiment Score �

n Word Sentiment Score �n Parsing dependency1 parse of the DAs. �n Calculating word sentiment score2. �n Updating word sentiment score. �

n  good student, great warrior���1Stanford dependency parse http://nlp.stanford.edu:8080/parser/ �2Sentiment lexicon SentiWordNet http://sentiwordnet.isti.cnr.it/�

11

Page 12: Online Debate Summarization using Topic Directed Sentiment Analysis

+Word Sentiment Score �

n Parsing dependency parse of the DAs. �n “ A large company needs a sustainable

business model. ”

12

Page 13: Online Debate Summarization using Topic Directed Sentiment Analysis

+Word Sentiment Score �

n Calculating word sentiment score. �

13

+0

+0

+0

+0

+0 +1 +0 +1

Page 14: Online Debate Summarization using Topic Directed Sentiment Analysis

+Word Sentiment Score �

n Updating word sentiment score. �

14

+0

+0

+0 +1 +0 +1

+1 +1

Page 15: Online Debate Summarization using Topic Directed Sentiment Analysis

+Extended Targets

n Extended Targets (ET)�n Extended targets are the entities closely

related to debate topics. �n To extract the extended targets, we capture

named entities (NE) from Wikipedia page of the debate topic.

15

Page 16: Online Debate Summarization using Topic Directed Sentiment Analysis

+Topic Relevance

n Topic Directed Sentiment Score

16

Equation.(Topic Directed Sentiment Score)�

where w is a word in DA, ET() is Extended Targets.

Page 17: Online Debate Summarization using Topic Directed Sentiment Analysis

+Topic Relevance

n Topic Co-occurrence �

17

Equation.(Topic Co-occurrence Score)�

Page 18: Online Debate Summarization using Topic Directed Sentiment Analysis

+Document Relevance

n tf-idf Sentiment Score �

18

Equation.(Topic Co-occurrence Score)�

Page 19: Online Debate Summarization using Topic Directed Sentiment Analysis

+Sentiment Relevance �

n Number of Sentiment Words �n  �

n Sentiment Strength

19

Equation.(Sentiment Strength Score)�

Page 20: Online Debate Summarization using Topic Directed Sentiment Analysis

+Context Relevance �

n Sentence position �n In debates, initial and ending DAs of the

debate posts are more important than the middle ones. �

n Sentence length �n As the longer sentences tend to contain more

information, we have used sentence length as document context feature.

20

Page 21: Online Debate Summarization using Topic Directed Sentiment Analysis

+Method

n Scores of DAs

21

Equation.(Scores of DAs)�

where λ is weighted, s is a DA of the Document, D is the Document.

Page 22: Online Debate Summarization using Topic Directed Sentiment Analysis

+Outline

n Introduction �

n Method �

n Experiment �

n Conclusion �

n Thought

22

Page 23: Online Debate Summarization using Topic Directed Sentiment Analysis

+Experiment

n This paper extracted 10 online debate discussions from www.convinceme.net. �n Number of users:1168 �n Number of posts :1945 �n Number of DA:23681

23

Page 24: Online Debate Summarization using Topic Directed Sentiment Analysis

+Experiment

n Following values gave the best results as indicated by ROUGE results by grid search. �n λtopicRel = 0.3 �n λdocRel = 0.1 �n λsentiRel = 0.5 �n λconRel = 0.1

24

Page 25: Online Debate Summarization using Topic Directed Sentiment Analysis

+Experiment

n We compared our system to the following systems: �n Max-length �n Lead �n pHAL �n tf-Idf�n OpinionSumm�

n  document similarity, topic relevance, sentiment and length

25

Page 26: Online Debate Summarization using Topic Directed Sentiment Analysis

+Result 26

Page 27: Online Debate Summarization using Topic Directed Sentiment Analysis

+Result 27

Page 28: Online Debate Summarization using Topic Directed Sentiment Analysis

+Outline

n Introduction �

n Method �

n Experiment �

n Conclusion �

n Thought

28

Page 29: Online Debate Summarization using Topic Directed Sentiment Analysis

+Conclusion

n This paper focuses on summarizing the on-line debates. �n topic directed sentiment �n topic related information �

n The results show that our system beats all these systems comprehensively.

29

Page 30: Online Debate Summarization using Topic Directed Sentiment Analysis

+Future Work

n Sentiment scores �n word sense disambiguation �n domain specific sentiment analysis �

n Creating users' profile by capturing their intention. �

30

Page 31: Online Debate Summarization using Topic Directed Sentiment Analysis

+Outline

n Introduction �

n Method �

n Experiment �

n Conclusion �

n Thought

31

Page 32: Online Debate Summarization using Topic Directed Sentiment Analysis

+

Thought Debate Comparison

32

Page 33: Online Debate Summarization using Topic Directed Sentiment Analysis

+ Introduction 33

Page 34: Online Debate Summarization using Topic Directed Sentiment Analysis

+Thanks for listening. 2014 / 10 / 31 (Fri.) @ MakeLab Group Meeting �[email protected]