Driving Behavioral Change for Information Management through Data-Driven Gree...
Extracting What We Think and How We Feel from What We Say in Social Media
1. Extracting What
We Think and How
We Feel from What
We Say in Social
Media---- Subjective Information Extraction
Subjective Information Extraction, Lu Chen 1
Lu Chen
Kno.e.sis Center
Wright State University
http://cdryan.com/blog/think-feel/
2. Directions
• From coarse-grained to fine-grained
– Document level -> sentence level -> expression level
– General sentiment -> domain-dependent sentiment -> target-dependent sentiment
– Sentiment Subjective information
• Sentiment (positive/negative/neutral) -> emotion (happy, sad, angry, surprise, etc.)
• Other types of subjective information: Intent, suggestion/recommendation,
wish/expectation, outlook, viewpoint, etc.
• From static to dynamic
– Our attitude can be changed during social communication.
• Modeling, detecting, and tracking the change of attitude
• What leads to the change of attitude? E.g., persuasion campaign
Subjective Information Extraction, Lu Chen 2
static
dynamic
coarse-grained
fine-grained
subjective information
3. Subjective Information Extraction, Lu Chen 3
Extracting a diverse and richer
set of sentiment-bearing
expressions, including formal
and slang words/phrases
Assessing the
target-dependent polarity
of each sentiment
expression
A novel formulation of assigning
polarity to a sentiment expression
as a constrained optimization
problem over the tweet corpus
Extracting Diverse Sentiment Expressions
With Target-dependent Polarity from Twitter
Lu Chen, Wenbo Wang, Meenakshi Nagarajan, Shaojun Wang, and Amit P. Sheth
5. Extracting Candidate Expressions
• Root word: a word that is considered sentiment-bearing in general
sense.
• Collecting root words from
– General-purpose sentiment lexicons: MPQA, General Inquirer, and
SentiWordNet
– Slang dictionary: Urban Dictionary
• For each tweet, selecting the “on-target” root words, and extracting
all the n-grams that contain at least one selected root word as
candidates
Subjective Information Extraction, Lu Chen 5
6. Identifying Inter-Expression Relations
• Connecting the candidate expressions via two types of inter-
expression relations – consistency relation and inconsistency
relation
• Basic ideas:
– A sentiment expression is inconsistent with its negation; two sentiment
expressions linked by contrasting conjunctions are likely to be
inconsistent.
– Two adjacent expressions are consistent if they do not overlap, and
there is no extra negation applied to them or no contrasting conjunction
connecting them.
Subjective Information Extraction, Lu Chen 6
7. An Example
1. I saw The Avengers yesterday evening. It was long but it was very good!
2. I do enjoy The Avengers, but it's both overrated and problematic.
3. Saw the avengers last night. Mad overrated. Cheesy lines and horrible
writing. Very predictable.
4. The avengers was good but the plot was just simple minded and predictable.
5. The Avengers was good. I was not disappointed.
Subjective Information Extraction, Lu Chen 7
8. Assessing Target-dependent Polarity
• For each candidate expression ,
– P-Probability – the probability that indicates positive
sentiment
– N-Probability – the probability that indicates negative
sentiment
• For each pair of candidate expressions and ,
– Consistency probability – the probability that and have the same
polarity:
– Inconsistency probability – the probability that and have
different polarities:
Subjective Information Extraction, Lu Chen 8
ic
)(Pr i
P
c
)(Pr i
N
c
ic
ic
1)(Pr)(Pr i
N
i
P
cc
ic jc
ic jc
)(Pr)(Pr)(Pr)(Pr),(Pr j
N
i
N
j
P
i
P
ji
cons
cccccc
ic jc
)(Pr)(Pr)(Pr)(Pr),(Pr j
P
i
N
j
N
i
P
ji
incons
cccccc
9. An Optimization Model
• We want the consistency and inconsistency probabilities derived
from the the P-Probabilities and N-Probabilities of the candidates
will be closest to their expectations suggested by the relation
networks.
• Objective Function:
Subjective Information Extraction, Lu Chen 9
1
1
22
),(Pr1),(Pr1minimize
n
i
n
ij
ji
inconsincons
ijji
conscons
ij ccwccw
where and are the weights of the edges (the frequency of the
relations) between and in the consistency and inconsistency
relation networks, and n is the total number of candidate expressions.
ic jc
cons
ijw incons
ijw
11. Evaluation
• Datasets:
– 168,005 tweets about movies
– 258,655 tweets about persons
• Gold standard:
– 1,500 tweets labeled with sentiment expressions and overall polarities for
the movie targets
– 1,500 tweets labeled with sentiment expressions and overall polarities for
the person targets
• Baseline methods:
– MPQA, GI, SWN: For each extracted root word regarding the target, simply
look up its polarity in MPQA, General Inquirer and SentiWordNet,
respectively.
– PROP: a propagation approach proposed by Qiu et al. (2009)
– COM-const: Assign 0.5 to all the candidates as their initial P-Probabilities.
– COM-gelex: Initialize the candidates’ polarities according to the root word
set.
Subjective Information Extraction, Lu Chen 11
Reference: Qiu, G.; Liu, B.; Bu, J.; and Chen, C. 2009. Expanding domain sentiment lexicon through double propagation. In Proc. of IJCAI.
15. Subjective Information Extraction, Lu Chen 15
Relevance of User Groups Based on Demographics and
Participation to Social Media Based Prediction
-- -- A Case Study of 2012 U.S. Republican Presidential Primaries
Lu Chen, Wenbo Wang, and Amit P. Sheth
• Existing studies on predicting election result are under the
assumption that all the users should be treated equally.
• How could different groups of users be different in predicting
election results?
1. Providing a detailed analysis of the social media users on different
dimensions
2. Estimating the “vote” of each user by analyzing his/her tweets, and
predicted the results based on “vote-counting”
3. Examining the predictive power of different user groups in predicting
the results of Super Tuesday races in 10 states
17. Electoral Prediction with Different User Groups
Subjective Information Extraction, Lu Chen 17
Revealing the challenge of identifying
the vote intent of “silent majority”
Retweets may not necessarily reflect
users' attitude.
18. Electoral Prediction with Different User Groups
Subjective Information Extraction, Lu Chen 18
Prediction of user’s vote based on
more opinion tweets is not
necessarily more accurate than the
prediction using more information
tweets
The right-leaning user group provides
the most accurate prediction result. It
correctly predict the winners in 8 out
of 10 states.
To some extent, it demonstrates the
importance of identifying likely voters
in electoral prediction.
19. Emotion
• Discovering Fine-grained Sentiment in Suicide Notes: Classify each
sentence from suicide notes into 15 emotional categories, e.g., love,
pride, guilt, blame, hopelessness, etc.
• Emotion Identification from Twitter Data: 7 emotion categories,
including joy, sadness, anger, lover, fear, thankfulness, and surprise
– Can we automatically create a large emotion dataset with high quality
labels from Twitter? How?
– What features can effectively improve the performance of supervised
machine learning algorithms?
– How much performance will be gained by increasing the size of the
training data?
– Can the system developed on Twitter data be directly applied to identify
emotions from other datasets?
Subjective Information Extraction, Lu Chen 19
20. What’s next?
Subjective Information Extraction, Lu Chen 20
static
dynamic
coarse-grained fine-grained
subjective information
Detecting the
change of
attitude during
persuasive
communication
Discriminating
other types of
subjective information
from sentiment,
e.g., wish,
intent
Research goalGenerally speaking, the information of what we think and how we feel is subjective information, so such information extraction task can be called subjective information extractionA large part of information we post in social media is about what we think and how we feel. For example, #nervous, movie The Avengers, it’s important for sentiment analysis to provide more actionable insight.
coordinate frameImagine the original point is the “traditional sentiment analysis”, the basic task of which is to classify the polarity of a given text.Basically, I am working towards two directions, one aims to get more fine-grained subjective information, the other one aims to capture the dynamics of subjective information.“Fine-grained” information can be pursued from different angle. For example, in sentiment analysis, identifying the polarity of sentiment expressions in a document is more fine-grained than classifying the overall polarity of a document. It is well-known that sentiment is sensitive to the domain or even the target. Identifying the sentiment associated with a specific target (e.g., a specific person or product feature) is more fine-grained than assessing the general sentiment. Another angle to pursue the fine-grained subjectivity is to consider sentiment as one type of subjective information. For example, emotion, as another type of subjective information, can be more fine-grained. In addition, we can also discriminate other types of subjective information from sentiment. In social media, people often post something like “On the way to watch a movie, hope it’s good”, or “A friend recommended me a new restaurant, I want to try it.” Currently, they are likely to be classified as positive sentiment. But the first one is actually expectation and the second one is intent.Another direction is from static to dynamic. Have your attitude about something or some person even been changed during social communication? I guess we all have. It is appealing to have tools to detect and track such changes, and discover what leads to such changes. One specific research topic along this line is to study the persuasion campaign in social media.