Recognizing Stances in Online Debates Debate: iPhone vs. Blackberry iPhone of course. Blackberry is...

Recognizing Stances in Online Debates

Debate: iPhone vs. Blackberry

iPhone of course. Blackberry is now for the senior businessmen market! The iPhone incarnate the 21st century whereas Blackberry symbolizes an outdated technology. The iPhone can reach a very diversified clientele …

ACL 2009 [email protected]





Side Classification: pro-iPhone stance





Arguing why their stance is correct





Justifying why the opposite side is not good


http://www.convinceme.net/


http://www.convinceme.net/

Side Classification: pro-iPhone stance Side Classification: pro-

Blackberry stance


Topics:1. iPhone2. Blackberry

Sides/ Stances:1. Pro-iPhone2. Pro-Blackberry

Dual-topic, Dual-sided debates regarding Named Entities

Goal

Debate stance recognition using opinion analysis

Learn debating preferences from the web

Exploited in an unsupervised approach Combines the individual pieces of information to

classify the overall stance


Challenges


The iPhone incarnate the 21st century whereas Blackberry symbolizes an outdated technology

Challenges



Positive and negative opinions are employed to argue for a side


Challenges




Opinions towards both topics within a post


Challenges






+ towards iPhone

- towards Blackberry

Challenges






+ towards iPhone

- towards Blackberry

We need to consider not only positive and negative opinionsbut also what they are about (targets)

Challenges

Pro-blackberry

The Pearl does music and video nicely …

First, you still can't beat the full QWERTY keyboard for quick, effortless typing.

Pro-iPhone

Well, Apple has always been a well known company.

Its MAC OS is also a unique thing.


Challenges

Pro-blackberry



Pro-iPhone



Debate topics are evoked in a variety of ways: Opinions explicitly toward the named topics are not enough

Type of Blackberry

Feature of Blackberry

Maker of iPhone

Feature of iPhone


Challenges

Pro-blackberry



Pro-iPhone



We need to consider not only opinions towards topics,But also opinions towards aspects


Challenges

Pro-blackberry



Pro-iPhone



We need to consider not only opinions towards topics,But also opinions towards aspects

Unique Aspects


Challenges

iPhone and Blackberry, both Have e-mail facilities Can be used to take photos Operate on batteries Etc.

Both sides share aspects


Challenges

… I love the ability to receive emails from my work account…


People expressing positive opinions regarding emails (generally) prefer Blackberry

Challenges



Certain shared aspects may be perceived to be better in one side• Email on Blackberry

Value for shared aspects depends on personal preferences• Emailing – pro-Blackberry people will argue via Email+

• That is, Email+ is often a strategy for arguing for the pro-Blackberry stance.

• Or, Browsing+ for iPhone

Challenges



We need to find what a preference/dislike for an individual target means towards the debate stance as a whole

Challenges

While the iPhone looks nice and does play a decent amount of music, it can't compare in functionality to the BB.

Concessionary opinions can be misleading


Side Classification: pro-Blackberry stance

Challenges

While the iPhone looks nice and does play a decent amount of music, it can't compare in functionality to the BB.


Side Classification: pro-Blackberry stance

We need to detect and handle concessionary opinions

Challenges: Summary

For debate stance recognition we need to:

Consider not only positive and negative opinions, but also what they are about (targets).

Consider not only opinions towards topics, but also opinions towards aspects

Find what a preference/dislike for an individual target means towards the debate stance as a whole

Detect and handle concessionary opinions


Challenges: Summary







Turney, 2002; Pang et al, 2002; Dave et al, 2003; Yu and Hatzivassiloglou, 2003, Pang and Lee 2005, Wilson et al 2005, Goldberg and Zhu, 2006, Kim and Hovy 2006 …

Challenges: Summary







Hu and Liu, 2004; Popescu and Etzioni., 2005; Bloom et al. 2007, Stoyanov and Cardie 2008; Xu et al., 2008 …






Our Approach







Our Approach


Adopting from previous work,Opinion-target pairing using

Opinion Lexicons and Syntactic rules






Our Approach


Unsupervised system Learn Associations from web and incorporate

these towards stance recognition








Our Approach


Unsupervised system Learn Associations from web and incorporate

these towards stance recognition



Rule-based Concession Handler using PDTB connectives

Methodology

Learn associations from web data (weblogs) Process the web data to

Find opinion-target pairs Associate opinion-target pairs with each debate side

Utilize the associations to classify debate posts Process the debate posts to

Find opinion-target pairs in the post Handle concessionary opinions Optimize over all opinion-targets for a post-level stance

classification


Methodology




Find opinion-target pairs in the post Handle concessionary opinions Optimize over all opinion targets for a post-level stance

classification


Methodology: Learning associations


Debate title

Topic1 = iPhoneTopic2 = BB



Web search engine

Debate title


Yahoo search engine API



Web search engine

Debate title


Weblogs containing both topics

Yahoo search engine API



Web search engine

Debate title


Weblogs containing both topics Par

ser

Stanford parser



Web search engine

Debate title



ser

Parsed web documents

Stanford parser



Web search engine

Debate title



ser


Opinion-target pairing



Web search engine

Debate title



ser



Lexicon

like: +hate: -

Wilson et al., 2005



Web search engine

Debate title



ser



Lexicon

Syntactic Rules

like = +hate = -



Web search engine

Debate title



ser



Lexicon

Syntactic Rules

I like email = email+

like = +hate = -



Web search engine

Debate title



ser



Lexicon

Syntactic Rules


Associations with topic-

polarity

like = +hate = -

Topic1+

Topic1-Topic2-

Topic2+

targetj+

what does a positive opinion towards a target mean with respect to positive or negative opinions regarding either of the topics

Associations with topic-polarity



For each opinion-topic pair (topic1+, topic1-, topic2+, and topic2-) found in the web document

Find other opinion target pairs (targetjp) in its vicinity

For each opinion-target (targetjp) calculate its

association with each of the opinion-topics P(topic1+|targetj+)

P(topic1-|targetj+)

P(topic2+|targetj+)

P(topic2-|targetj+)


P(iPhone+ |email+)

P(iPhone- |email+)

P(BB+ |email+)

P(BB- |email+)



Web search engine

Debate title



ser



Lexicon

Syntactic Rules


Associations with topic-

polarity

P(iPhone- |email+)

P(BB- |email+)

P(iPhone+ |email+)

P(BB+ |email+)

like = +hate = -

Blackberry+

Blackberry-iPhone-

iPhone+

Storm- 0.062

0.8430.06

0.03



Methodology




Find opinion-target pairs in the post Handle concessionary opinions Optimize over all opinion targets for a post-level stance

classification


Methodology: Stance Classification


Debate PostDebate

PostDebate Post



Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate

PostParsed Debate Post



Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairing in the post

Lexicon

Syntactic Rules




Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairing in the post

Lexicon

Syntactic Rules


Association lookup, Side

Mapping



Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairing in the post

Lexicon

Syntactic Rules



Mapping

P(iPhone- |email+)P(BB- |email+)P(iPhone+ |email+)P(BB+ |email+)

Topic1+

Topic1-Topic2-

Topic2+

target+

Association of positive opinion towards a target to positive or negative opinions regarding either of the topics

Association Lookup

0.1

0.05

0.5

0.35


Side-1 Side-2

Topic1+

Topic1-Topic2-

Topic2+

target+

Side-1 = Topic1+ alternatively Topic2-Side-2 =Topic2+ alternatively Topic1-

Association Lookup, Side Mapping

0.1

0.05

0.5

0.35


target+Side-1 Side-2

0.150.85

Association of positive opinion towards a target to both of the stances

Association Lookup, Side Mapping




Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairingin the post

Lexicon

Syntactic Rules



Mapping




Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairingin the post

Lexicon

Syntactic Rules



Mapping

P(iPhone- |email+)P(BB- |email+)P(iPhone+ |email+)P(BB+ |email+) Assoc(Side-1,

email+)Assoc(Side-2, email+)



Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairing in the post

Lexicon

Syntactic Rules



Mapping


Concession Handling

Assoc(Side-1, email+)Assoc(Side-2, email+)

Concession Handling

Detecting concessionary opinions

Find Concession indicators Discourse connectives from Penn Discourse Treebank

(Prasad et al., 2007)

Use simple rules to find the conceded part of the sentence While the iPhone looks nice and does play a decent amount

of music, it can't compare in functionality to the BB.

I like my music, and phone, but I don't want to carry a brick around in my pocket when I only need my phone.


Concession Handling

I like my music, and phone, but I don't want to carry a brick around in my pocket when I only need my phone.

Conceded opinions music+ phone+


Side-2Pro-Iphone

Side-1Pro-Blackberry

music+

phone+

1.0

0.5090.45

Original associations learnt from the web

Concession Handling


Side-2Pro-Iphone

Side-1Pro-Blackberry

music+

phone+

1.0

0.509 0.45

Concession Handling

Associations after concession handling

Conceded opinions are counted for the opposite side




Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairing in the post

Lexicon

Syntactic Rules



Mapping


Concession Handling

Post-level associationaggregation


Side-2Pro-Iphone

Side-1Pro-

Blackberry

Aggregation

target1+

target2+

target3+

target4+

Each opinion-target pair in the post has a bias toward one or the side

0.9

0.7

0.4

0.5

0.1

0.3

0.6

0.5


Side-2Pro-Iphone

Side-1Pro-

Blackberry

Aggregation

target1+

target2+

target3+

target4+

Each opinion-target pair in the post has a bias toward one or the other side

Optimize the post classification such that The side assigned to the post maximizes the association value of the post




Debate PostDebate

PostDebate Post

Par

ser

Debate PostDebate


Opinion-target

pairing in the post

Lexicon

Syntactic Rules



Mapping


Concession Handling

Post-level associationaggregation

Side= pro-Topic1


Blackberry+

Blackberry-iPhone-

iPhone+

Storm- 0.062

0.8430.06

0.03

Associations learnt from web data


Blackberry+

Blackberry-iPhone-

iPhone+

Storm+ 0.227

0.068

0.022

0.613


Both OpPMI, and Op-Pref agree with each other;Both learnt the IS-A relationship


Blackberry+

Blackberry-iPhone-

iPhone+

Keyboard+

0.718

0.0

0.12

0.09



Blackberry+

Blackberry-iPhone-

iPhone+

Keyboard- 0.25

0.25

0.125

0.375


Negative opinions towards keyboards are not useful for side discrimination

0.5 0.5


Summing Up

Looked at several tasks ranging from purely lexical to discourse classification Identify subjective words Classify their senses as subjective or objective Recognize, in a text or conversation, whether a word

is used with a subjective or objective sense Sense-aware contextual subjectivity and sentiment

analysis Contextual polarity recognition Discourse-Level Opinion Interpretation

Many ambiguities are involved in interpreting subjective language!

Summing Up

Many other ambiguities than these! Sarcasm and Irony

Yeah, he’s just wonderful. You’re no different from the mob! Oh, there’s a big

difference, Mrs. De Marco. The mob is run by murdering, thieving, lying, cheating psychopaths. We work for the President of the United States. [Married to the Mob]

Literal versus non-literal language He is a pain in the neck

Pointers

Please see http://www.cs.pitt.edu/~wiebe Publications OpinionFinder Subjectivity lexicon MPQA manually annotated corpus Tutorials Bibliography

Acknowledgements

Subjectivity Research Group, Pittsburgh Cem Akkaya, Yaw Gyamfi, Paul Hoffman, Josef Ruppenhofer, Swapna

Somasundaran, Theresa Wilson

Cornell: Claire Cardie, Eric Breck, Yejin Choi, Ves Stoyanov

Utah: Ellen Riloff, Sidd Patwardhan, Bill Phillips

UNT: Rada Mihalcea, Carmen Banea

NLP@Pitt: Wendy Chapman, Rebecca Hwa, Pam Jordan, Diane Litman, …

Thank you


Recognizing Stances in Online Debates Debate: iPhone vs. Blackberry iPhone of course. Blackberry is...

Documents

Transcript of Recognizing Stances in Online Debates Debate: iPhone vs. Blackberry iPhone of course. Blackberry is...