Title: CoCQA: CoTraining Over Questions and Answers with an Application to Predicting Question Subjectivity
1CoCQA Co-Training Over Questions and
Answerswith an Application to Predicting
Question Subjectivity Orientation
- Baoli Li, Yandong Liu, and Eugene Agichtein
- Emory University
2Community Question Answering
- An effective way of seeking information from
other users - Can be searched for resolved questions
3Community Question Answering (CQA)
- Yahoo! Answers
- Users
- Asker post questions
- Answerer post answers
- Voter vote for existing answers
- Questions
- Subject
- Detail
- Answers
- Answer text
- Votes
- Archive millions of questions and answers
4Lifecycle of a Question in CQA
Choose a category
Compose the question
Open question
Examine
Answer
Answer
Answer
Close question Choose best answers Give ratings
Find the answer?
Yes
Question is closed by system. Best answer is
chosen by voters
No
5Problem Statement
- How can we exploit structure of CQA to improve
question classification? - Case Study Question Subjectivity Prediction
- Subjective questions seek answers containing
private states such as personal opinion,
judgment, and experience - Objective questions are expected to be answered
with reliable or authoritative information
6Example Questions
- Subjective
- Has anyone got one of those home blood pressure
monitors? and if so what make is it and do you
think they are worth getting? - Objective
- What is the difference between chemotherapy and
radiation treatments?
7Motivation
- Guiding the CQA engine to process questions more
intelligently - Some Applications
- Ranking/filtering answers
- Improving question archive search
- Evaluating answers provided by users
- Inferring user intent
8Challenges
- Some challenges in online real question analysis
- Typically complex and subjective
- Can be ill-phrased and vague
- Not enough annotated data
9Key Observations
- Can we utilize the inherent structure of the CQA
interactions, and use the unlimited amounts of
unlabeled data to improve classification
performance?
10Natural Approach Co-Training
- Introduced by
- Combining labeled and unlabeled data with
co-training, Blum and Mitchell, 1998 - Two views of the data
- E.g. content and hyperlinks in web pages
- Provide complementary information for each other
- Iteratively construct additional labeled data
- Can often significantly improve accuracy
11Questions and Answers Two Views
- Example
- Q Has anyone got one of those home blood
pressure monitors? and if so what make is it and
do you think they are worth getting? - A My mom has one as she is diabetic so its
important for her to monitor it she finds it
useful. - Answers usually match/fit question
- My mom she finds
- Askers can usually identify matching answers by
selecting the best answer
12CoCQA A Co-Training Framework over Questions and
Answers
Unlabeled Data ?????????? ??????????
Unlabeled Data ?????????? ??????????
Labeled Data
Labeled Data
CQ
Q
Q
CA
A
A
Classify
---- ----
Validation (Holdout training data)
Stop
13Details of CoCQA implementation
- Base classifier
- LibSVM
- Term Frequency as Term Weight
- Also tried Binary, TFIDF
- Select top K examples with highest confidence
- Margin value in SVM
14Feature Set
- Character 3-grams
- has, any, nyo, yon, one
- Words
- Has, anyone, got, mom, she, finds
- Word with Character 3-grams
- Word n-grams (nlt3, i.e. Wi, WiWi1,
WiWi1Wi2) - Has anyone got, anyone got one, she finds it
- Word and POS n-gram (nlt3, i.e. Wi, WiWi1, Wi
POSi1, POSiWi1 , POSiPOSi1, etc.) - NP VBP, She PRP, VBP finds
15Overview of Experimental Setup
- Datasets
- From Yahoo! Answers
- Manually labeled data by Amazon Mechanical Turk
- Metrics
- Compare CQA to state-of-the semi-supervised method
16Dataset
- 1,000 Labeled Questions from Yahoo! Answers
- 5 categories (Arts, Education, Science, Health
Sports) - 200 questions from each category
- 10,000 Unlabeled Questions from Yahoo! Answers
- 2,000 questions from each category
- Data available at
- http//ir.mathcs.emory.edu/shared
17Manual Labeling
- Annotated using Amazons Mechanical Turk service
- Each question was judged by 5 Mechanical Turk
workers - 25 questions included in each HIT task
- Worker needs to pass the qualification test
- Majority vote to derive gold standard
- Discarded small fraction (22 out of 1000) of
nonsensical questions such as Upward Soccer
Shorts? and 11?fdgdgdfg by manual inspection
18Example HIT task
19Subjectivity Statistics by Category
Objective
Subjective
20Evaluation Metric
- Macro-Averaged F-1
- Prediction performance on both subjective
questions and objective questions is equally
important - F-1
- Averaged over subjective and objective classes
21Experimental Settings
- 5 fold cross validation
- Methods Compared
- Supervised LibSVM (Chang and Lin, 2001)
- Generalized Expectation (GE) (Mann and McCallum,
2007) - CoCQA our method
- Base classifier LibSVM
- View 1 question text View 2 answer text
22F1 for Supervised Learning
F1 with different sets of features
23Semi Supervised Learning Adding unlabeled data
Comparison between Supervised, GE and CoCQA
24CoCQA with varying K( new examples added in
each iteration)
25CoCQA for varying iterations
26CoCQA for varying amount of labeled data
27Conclusions and Future Work
- Problem Non-topical text classification in CQA
- CoCQA a co-training framework that can exploit
information from both question and answers - Case study subjectivity classification for real
questions in CQA - We plan to explore
- more sophisticated features
- related variants of semi-supervised learning
- other applications (Sentiment classification)
28Thank you!Baoli Li csblli_at_gmail.comYandong
Liu yandong.liu_at_emory.eduEugene
Agichtein eugene_at_mathcs.emory.edu
29Performance of Subjective vs. Objective classes
- Subjective class
- 80
- Objective class
- 60
30Related work
- Some related work
- Question Classification (Zhang and Lee, 2003)(
Tri et al., 2006) - Sentiment Analysis (Pang and Lee, 2004)
- (Yu and Hatzivassiloglou, 2003)
- (Somasundaran et al. 2007)
31Important words for Subjective, Objective classes
by Information Gain