polling the blogosphere: a rule-based approach to belief classification, by jason kessler
TRANSCRIPT
Polling the Blogosphere: a Rule-Based Approach to Belief Classification
Jason Kessler
Indiana University, Bloomington
Belief Analysis of Blogs
Polling the blogosphere on a controversial proposition Literal search on a proposition (e.g., “Obama is electable”) Which blog entries contain assert it? Which deny it? Aggregate results
243 bloggers assert it 616 bloggers deny it
Motivating Example
Polling for “the Moon landings were staged” “The theory that the Moon landings were staged is
complete nonsense.” The writer denies “the Moon landings were staged.”
Motivating Example
If Obama is electable, the country is in good shape. Writer takes no stance toward “Obama is
electable”.
Problem
When a writer uses a declarative finite clause, does that writer assert, deny, or take no stance toward its truth value?
This is the problem of identifying a writer’s stance toward a proposition.
Veridicity or facticity of a proposition.
Example
Everybody is sad that the bar closed. The writer asserts “the bar closed.” Belief != Sentiment
Negative sentiment toward “the bar closed” Positive stance.
Outline
System Description Given a proposition, sentence Dependency Parse
Syntactic Representation Hand written patterns over semantic classes
Veridicality Elements Veridicality Transformations
Evaluation Proof of concept Promising results
The theory that the Moon landings were staged is complete nonsense.
Dependency Parse
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
The theory that the Moon landings were staged is complete nonsense.
Veridicality Elements (VEs)
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
The theory that the Moon landings were staged is complete nonsense.
Veridicality Transformations (VTs)
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
The theory that the Moon landings were staged is complete nonsense.
Veridicality Transformations (VTs)
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
System Structure: Veridicality Elements
Find expressions that have the potential of changing the truth-value of a proposition or referring to it
Different classes affect truth values differently Examples:
Assertion – Positive The assertion that the sky is blue
Nonsense – Negative The idea that the sky is orange is nonsense
If – Neutral Pretend – Counter-factive
Finding Veridicality Elements
Manually created seed sets Search web for patterns likely to contain VEs “I agree with the assertion that”
“I * with the assertion that” “I quibble with the assertion that” “I take issue with the assertion that”
Manually classify matches, form new queries “I take issue with the * that”
“I take issue with the argument that”
Similar to Brin (1998)
System Structure:Veridicality Transformations
Relate these expressions to propositions Some expressions won’t be related to propositions Why bag-of-Veridicality-Elements fails
Templates over dependency graphs Select for a VE class and a proposition
System Structure:Veridicality Transformations
Examples Expression is a main verb, proposition is its comp. clause
John pretended the monkey was harmless. Cleft construction, expression is an adjective
It is inconceivable that two plus two equals five.
Another Example
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
If Bob goes to school, he realizes the Earth is round.
Another Example
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
If Bob goes to school, he realizes the Earth is round.
Another Example
Pipeline Stages:
2. Dependency Parse
3. Tag Veridicality Elements
4. Apply Veridicality Transformations
If Bob goes to school, he realizes the Earth is round.
Evaluation
Primitive, proof-of-concept evaluation Can we poll the blogosphere? Google blog search for “abortion is murder”
Unseen data Run the system on the first 100 hits. See if it does better baseline.
Evaluation
Exclude a number of results: Spam blogs Long, unparsable sentences Trivial sentences (no VEs)
Abortion is murder! Questions
Evaluation
Corpus Statistics: 48 Sentences
27 positive 3 negative 18 neutral
39 classified correctly (81% accuracy) Majority class was positive, giving a baseline of 56%
accuracy
Related Work
Nairn et al. (2006) focused on main verbs Complex behavior under negation
Work on contextual polarity for sentiment analysis. Wilson et al. (2005)
Statistical approach Polanyi and Zaenen (2006)
Theoretical approach
Related Work
Somasundaran et al. (2007) Statistical techniques used to detect presence of
“arguing” in a sentence. Arguing = writer takes a non-neutral stance toward
some content
Future Work
Annotate corpus Further testing Statistical approaches
Augment VE/VTs Integrate Nairn et al. (2006) Take into account questions
Takeaways
Belief analysis is a young field Bag-of-words is not enough Shallow linguistic methods show promise
Questions?
Thank you. References: Brin, S. 1998. Extracting patterns and relations from the world wide web. In
WebDB Workshop at 6th International Conference on Extending Database Technology, EDBT’98.
Nairn, R.; Condoravdi, C.; and Karttunen, L. 2006. Computing relative polarity for textual inference. In ICoS-5.
Polanyi, L.; and Zaenen, A. 2005. Contextual valence shifters. In Shanahan, J. G.; Qu, Y.; and Wiebe J., eds,. Computing Attitude and Affect in Text.
Somasundaran, S.; Wilson, T.; Wiebe, J.; and Stoyanov, V. 2007. QA with attitude: Exploiting opinion type analysis for improving question answering in on-line discussions and the news. In ICWSM.
Wilson, T.; Wiebe, J.; and Hoffmann, P. 2005. Recognizing contextual polarity in phrase-level sentiment analysis. In HLT/EMNLP.