SENTIMENT ANALYSIS

TUTORIAL
Prof. Ronen Feldman
Hebrew University, JERUSALEM
Digital Trowel, Empire State Building
[email protected]
The Text Mining Handbook
CACM Article
 4

INTRODUCTION TO
SENTIMENT ANALYSIS
Based on slides from Bing Liu
and some of our work
 5

Introduction
Sentiment analysis
Computational study of opinions, sentiments, evaluations, attitudes,
appraisal, affects, views, emotions, subjectivity, etc., expressed in
text.
7H[W 5HYLHZVEORJVGLVFXVVLRQVQHZVFRPPHQWVIHHGEDFN

Sometimes called opinion mining

6

Typical sentiment analysis usage

Extract from text how people feel about different products
Sentiment analysis can be tricky
Honda Accords and Toyota Camrys are nice sedans
Honda Accords and Toyota Camrys are nice sedans,
but hardly the best car on the road
8

Sentiment Analysis is Hot

Dozens of companies and over a thousand research
papers

LingPipe
 9

Opinions are widely stated

Organization internal data
Customer feedback from emails, call centers, etc.
News and reports
Opinions in news articles and commentaries
Word-of-mouth on the Web
Personal experiences and opinions about anything in
reviews, forums, blogs, Twitter, micro-blogs, etc
Comments about articles, issues, topics, reviews, etc.
Postings at social networking sites, e.g., Facebook.
 10

Sentiment analysis applications

Businesses and organizations
Benchmark products and services; market intelligence.
Businesses spend a huge amount of money to find consumer opinions
using consultants, surveys and focus groups, etc

Individuals
Make decisions to purchase products or to use services
Find public opinions about political candidates and issues

Ad placement: e.g. in social media

Place an ad if one praises a product.
Place an ad from a competitor if one criticizes a product.

Opinion retrieval: provide general search for opinions.

 11

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Opinion summarization
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 12

Problem Statement: Abstraction

It consists of two parts
(1) Opinion definition. What is an opinion?
(2) Opinion summarization
Opinions are subjective. An opinion from a single person
(unless a VIP) is often not sufficient for action.
We need opinions from many people, and thus opinion
summarization.
 13

Abstraction (1): what is an opinion?


Id: Abc123 on 5-1-2008 I bought an iPhone a few days
ago. It is such a nice phone. The touch screen is really cool.
The voice quality is clear too. It is much better than my old
Blackberry, which was a terrible phone and so difficult to
type with its tiny keys. However, my mother was mad with
me as I did not tell her before I bought the phone. She also
thought the phone was too expensive
One can look at this review/blog at the
document level, i.e., is this review + or -?
sentence level, i.e., is each sentence + or -?
entity and feature/aspect level
 14

Entity and aspect/feature level

Id: Abc123 on 5-1-2008 I bought an iPhone a few days
ago. It is such a nice phone. The touch screen is really cool.
The voice quality is clear too. It is much better than my old
Blackberry, which was a terrible phone and so difficult to
type with its tiny keys. However, my mother was mad with
me as I did not tell her before I bought the phone. She also
thought the phone was too expensive
What do we see?
Opinion targets: entities and their features/aspects
Sentiments: positive and negative
Opinion holders: persons who hold the opinions
Time: when opinions are expressed
 15

Two main types of opinions

(Jindal and Liu 2006; Liu, 2010)
Regular opinions: Sentiment/opinion expressions
on some target entities
Direct opinions:
7KH touch screen LVUHDOO\FRRO
Indirect opinions:
$IWHUWDNLQJWKHGUXJP\SDLQKDVJRQH

Comparative opinions: Comparisons of more than

one entity.
(JL3KRQHLVEHWWHUWKDQ%ODFNEHUU\
We focus on regular opinions first, and just call
them opinions.
 16

A (regular) opinion
An opinion has the following basic components
(gi, soijkl, hi, tl),
where
gj is a target
soijl is the sentiment value of the opinion from opinion holder hi on
target gj at time tl. soijl is positive, negative or neutral, or a
rating score
hi is an opinion holder.
tl is the time when the opinion is expressed.
 17

Opinion target
In some cases, opinion target is a single entity or
topic.
,ORYHiPhoneand ,VXSSRUWtax cut

But in many other cases, it is more complex.

I bought an iPhone a few days ago. It is such a nice
phone. The touch screen is really cool.
Opinion target of the 3rd sentence is not just touch screen, but the
touch screen of iPhone
I support tax cut for the middle class, but for WKHULFK
We decompose the opinion target
 18

Entity and aspect (Hu and Liu, 2004; Liu, 2006)


Definition (entity): An entity e is a product, person,
event, organization, or topic. e is represented as
a hierarchy of components, sub-components, and so on.
Each node represents a component and is associated
with a set of attributes of the component.

An opinion can be expressed on any node or attribute

of the node.
For simplicity, we use the term aspects (features) to
represent both components and attributes.
 19

Opinion definition (Liu, Ch. in NLP handbook, 2010)



An opinion is a quintuple
(ej, ajk, soijkl, hi, tl),
where
ej is a target entity.
ajk is an aspect/feature of the entity ej.
soijkl is the sentiment value of the opinion from the opinion holder hi
on aspect ajk of entity ej at time tl. soijkl is +ve, -ve, or neu, or a more
granular rating.
hi is an opinion holder.
tl is the time when the opinion is expressed.
 20

Some remarks about the definition

Although introduced using a product review, the
definition is generic
Applicable to other domains,
E.g., politics, social events, services, topics, etc.
(ej, ajk) is also called the opinion target
Opinion without knowing the target is of limited use.
The five components in (ej, ajk, soijkl, hi, tl) must
correspond to one another. Very hard to achieve
The five components are essential. Without any of
them, it is problematic in general.
 21

Our example blog in quintuples

Id: Abc123 on 5-1-2008 I bought an iPhone a few days
ago. It is such a nice phone. The touch screen is really cool.
The voice quality is clear too. It is much better than my old
Blackberry, which was a terrible phone and so difficult to
type with its tiny keys. However, my mother was mad with
me as I did not tell her before I bought the phone. She also
thought the phone was too expensive
In quintuples
(iPhone, GENERAL, +, Abc123, 5-1-2008)
(iPhone, touch_screen, +, Abc123, 5-1-2008)

We will discuss comparative opinions later.
 22

&RQIXVLQJWHUPLQRORJLHV
Entity is also called object.
Aspect is also called feature, attribute, facet, etc
Opinion holder is also called opinion source
 23

5HDGHUVVWDQGLQJSRLQW
See this sentence
,DPVRKDSS\WKDW*RRJOHSULFHVKRWXSWRGD\
Although the sentence gives an explicit sentiment,
different readers may feel very differently.
If a reader sold his Google shares yesterday, he will not
be that happy.
If a reader bought a lot of Google shares yesterday, he
will be very happy.
Current research mostly ignores the issue.
 24

Structure the unstructured

Goal: Given an opinionated document,
Discover all quintuples (ej, ajk, soijkl, hi, tl),
Or, solve some simpler form of the problem
E.g. classify the sentiment of the entire document

With the quintuples,

Unstructured Text o Structured Data
Traditional data and visualization tools can be used to
slice, dice and visualize the results.
Enables qualitative and quantitative analysis.
 25

Subjectivity
Sentence subjectivity: An objective sentence presents
some factual information, while a subjective sentence
expresses some personal opinions, beliefs, views,
feelings, or emotions.
Not the same as emotion
 26

Subjectivity
Subjective expressions come in many forms, e.g.,
opinions, allegations, desires, beliefs, suspicions,
and speculations (Wiebe, 2000; Riloff et al 2005).
A subjective sentence may contain a positive or negative
opinion
Most opinionated sentences are subjective, but
objective (factual) sentences can imply opinions too
(Liu, 2010)
7KHPDFKLQHVWRSSHGZRUNLQJLQWKHVHFRQGGD\
:HEURXJKWWKHPDWWUHVV\HVWHUGD\DQGDERG\
LPSUHVVLRQKDVIRUPHG
$IWHUWDNLQJWKHGUXJWKHUHLVQRPRUHSDLQ
 27

Rational and emotional evaluations

Rational evaluation: Many evaluation/opinion sentences
express no emotion
HJ7KHYRLFHRIWKLVSKRQHLVFOHDU
Emotional evaluation
HJ,ORYHWKLVSKRQH
7KHYRLFHRIWKLVSKRQHLVFU\VWDOFOHDU"
Some emotion sentences express no (positive or
negative) opinion/sentiment
HJ,DPVRVXUSULVHGWRVHH\RX
 28

Abstraction (2): opinion summary


With a lot of opinions, a summary is necessary.
A multi-document summarization task
For factual texts, summarization is to select the
most important facts and present them in a
sensible order while avoiding repetition
1 fact = any number of the same fact
But for opinion documents, it is different because
opinions have a quantitative side & have targets
1 opinion z a number of opinions
Aspect-based summary is more suitable
Quintuples form the basis for opinion summarization
 29

Feature-based opinion summary1

(Hu & Liu, 2004)
I bought an iPhone a few days Feature Based Summary of
ago. It is such a nice phone. The iPhone:
touch screen is really cool. The
voice quality is clear too. It is Feature1: Touch screen
much better than my old Positive: 212
Blackberry, which was a terrible The touch screen was really cool.
phone and so difficult to type with The touch screen was so easy to use
its tiny keys. However, my mother and can do amazing things.
was mad with me as I did not tell
her before I bought the phone. Negative: 6
The screen is easily scratched.
She also thought the phone was
I have a lot of difficulty in removing
too H[SHQVLYH finger marks from the touch screen.

1. Feature2: voice quality


Note: We omit opinion holders
 30

Opinion Observer (Liu et al. 2005)

+
Summary of
reviews of
Cell Phone 1

_
Voice Screen Battery Size Weight

Comparison of +
reviews of
Cell Phone 1
Cell Phone 2
_
 31

Aspect-based opinion summary

 32

Google Product Search (Blair-Goldensohn et al 2008

?)
 33

Aggregate opinion trend

 34

Sentiment mining requires solving coupled

IE problems
(ej, ajk, soijkl, hi, tl),
ej - a target entity: Named Entity Extraction (more)
ajk an aspect/feature of ej: Information Extraction
soijkl is sentiment: Sentiment Identification
hi is an opinion holder: Information/Data Extraction
tl is the time: Information/Data Extraction
5 pieces of information must match
Coreference and entity resolution
 35

Easier and harder problems

Tweets from Twitter are probably the easiest
short and thus usually straight to the point
Stocktwits are much harder! (more on that later)
Reviews are next
entities are given (almost) and there is little noise
Discussions, comments, and blogs are hard.
Multiple entities, comparisons, noisy, sarcasm, etc
Extracting entities and aspects, and determining
sentiments/opinions about them are hard.
Combining them is harder.
 36

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Opinion summarization
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 37

Sentiment classification
Classify a whole opinion document (e.g., a
review) based on the overall sentiment of the
opinion holder (Pang et al 2002; Turney 2002)
Classes: Positive, negative (possibly neutral)
Neutral or no opinion is hard. Most papers ignore it.
An example review:
,ERXJKWDQL3KRQHDIHZGD\VDJR,WLVVXFKDQLFH
phone, although a little large. The touch screen is cool.
7KHYRLFHTXDOLW\LVFOHDUWRR,VLPSO\ORYHLW
Classification: positive or negative?
Perhaps the most widely studied problem.
 38

Some Amazon reviews

 39

A text classification task

It is basically a text classification problem
But different from topic-based text classification.
In topic-based text classification (e.g., computer, sport,
science), topic words are important.
In sentiment classification, opinion/sentiment words are
more important, e.g., great, excellent, horrible, bad,
worst, etc.
 40

Assumption and goal

Assumption: The doc is written by a single person
and express opinion/sentiment on a single entity.
Goal: discover (_, _, so, _, _),
where e, a, h, and t are ignored
Reviews usually satisfy the assumption.
Almost all papers use reviews
Positive: 4 or 5 stars, negative: 1 or 2 stars
Many forum postings and blogs do not
They can mention and compare multiple entities
Many such postings express no sentiments
 41

Supervised learning example

Training and test data
Movie reviews with star ratings
4-5 stars as positive
1-2 stars as negative

Neutral is ignored.
SVM gives the best classification accuracy based on
balanced training data
Typical result: 80-90% accuracy
Features: unigrams (bag of individual words)
 42

Features for supervised learning

The problem has been studied by numerous
researchers
Including domain adaption and cross-lingual, etc.

Key: feature engineering. A large set of features

have been tried by researchers. E.g.,
Terms frequency and different IR weighting schemes
Part of speech (POS) tags
Opinion words and phrases
Negations
Syntactic dependency
 43

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Opinion summarization
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 44

Subjectivity classification
Document-level sentiment classification is too coarse
for most applications.
So do sentence level analysis
Assumes a single sentiment per sentence
not always true, so one can classify clauses instead
 45

Sentence sentiment analysis

Usually consists of two steps
Subjectivity classification
To identify subjective sentences
Sentiment classification of subjective sentences
As positive or negative

But bear in mind

Many objective sentences can imply sentiments
Many subjective sentences do not express positive or negative
sentiments/opinions
(J, EHOLHYHKHZHQWKRPH\HVWHUGD\
 46

Subjectivity classification using patterns

(Rilloff and Wiebe, 2003)

A bootstrapping approach.
A high precision classifier is first used to automatically identify
some subjective and objective sentences.
Two high precision (but low recall) classifiers are used,
A high precision subjective classifier
A high precision objective classifier
Based on manually collected lexical items, single words and n-grams,
which are good subjective clues.
A set of patterns are then learned from these identified
subjective and objective sentences.
Syntactic templates are provided to restrict the kinds of patterns to
be discovered, e.g., <subj> passive-verb.
The learned patterns are then used to extract more subjective
and objective sentences (the process can be repeated).
 47

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Feature/Aspect-based sentiment analysis
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 48

We need to go further

Sentiment classification at both the document
and sentence (or clause) levels are useful, but
They do not find what people liked and disliked.
They do not identify the targets of opinions, i.e.,
Entities and their aspects
Without knowing targets, opinions are of limited use.
We need to go to the entity and aspect level.
 49

Recall an opinion is a quintuple

An opinion is a quintuple
(ej, ajk, soijkl, hi, tl),
where
ej is a target entity.
ajk is an aspect/feature of the entity ej.
soijkl is the sentiment value of the opinion of the opinion holder hi on
feature ajk of entity ej at time tl. soijkl is +ve, -ve, or neu, or a more
granular rating.
hi is an opinion holder.
tl is the time when the opinion is expressed.
 50

Aspect-based sentiment analysis

Much of the research is based on online reviews
For reviews, aspect-based sentiment analysis is
easier because the entity (i.e., product name) is
usually known
Reviewers simply express positive and negative
opinions on different aspects of the entity.
For blogs, forum discussions, etc., it is harder:
both entity and aspects of entity are unknown,
there may also be many comparisons, and
there is also a lot of irrelevant information.
 51

Find entities (entity set expansion)

Although similar, it is somewhat different from the
traditional named entity recognition (NER).
E.g., one wants to study opinions on phones
given Motorola and Nokia, find all phone brands and models in a
corpus, e.g., Samsung, Moto,
 52

Feature/Aspect extraction
Extraction may use:
frequent nouns and noun phrases
Sometimes limited to a set known to be related to the entity of interest or
using part discriminators
e.g., for a scanner entiityRI VFDQQHUVFDQQHUKDV
opinion and target relations
Proximity or syntactic dependency
Standard IE methods
Rule-based or supervised learning
Often HMMs or CRFs (like standard IE)
 53

Extract aspects using DP (Qiu et al. 2009; 2011)



Use double propagation (DP)

Like co-training
an opinion should have a target, entity or aspect.
DP extracts both aspects and opinion words.
Knowing one helps find the other.
(JThe rooms are spacious
 54

The DP method
DP is a bootstrapping method
Input: a set of seed opinion words,
no aspect seeds needed
Based on dependency grammar (Tesniere 1959).
7KLVphone has good screen
 55

Roadmap
Opinion Mining Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 56

Sentiment (or opinion) lexicon

Sentiment lexicon: lists of words and expressions
XVHGWRH[SUHVVSHRSOHVVXEMHFWLYHIHHOLQJVDQG
sentiments/opinions.
Not just individual words, but also phrases and idioms,
HJFRVWDQDUPDQGDOHJ
Many sentiment lexica can be found on the web
They often have thousands of terms, and are quite
useful
 57

Sentiment lexicon
Sentiment words or phrases (also called polar words,
opinion bearing words, etc). E.g.,
Positive: beautiful, wonderful, good, amazing,
Negative: bad, poor, terrible, cost an arm and a leg.

Many of them are context dependent, not just

application domain dependent.
Three main ways to compile such lists:
Manual approach: not a bad idea for a one-time effort
Corpus-based approach
Dictionary-based approach
 58

Corpus- vs. Dictionary-based method

Corpus-based approaches
Often use a double propagation between opinion words and the
items they modify
require a large corpus to get good coverage
Dictionary-based methods
Typically use :RUG1HWV synsets and hierarchies to acquire opinion
words
usually do not give domain or context dependent meanings
 59

Corpus-based approaches
Rely on syntactic patterns in large corpora.
(Hazivassiloglou and McKeown, 1997; Turney, 2002; Yu
and Hazivassiloglou, 2003; Kanayama and Nasukawa,
2006; Ding, Liu and Yu, 2008)
Can find domain dependent orientations (positive, negative,
or neutral).
(Turney, 2002) and (Yu and Hazivassiloglou, 2003)
are similar.
Assign opinion orientations (polarities) to words/phrases.
(Yu and Hazivassiloglou, 2003) is slightly different from
(Turney, 2002)
use more seed words (rather than two) and use log-likelihood ratio
(rather than PMI).
 60

Corpus-based approaches (contd)

Sentiment consistency: Use conventions on
connectives to identify opinion words (Hazivassiloglou and
McKeown, 1997). E.g.,
Conjunction: conjoined adjectives usually have the same
orientation.
(J7KLVFDULVbeautiful and spaciousFRQMXQFWLRQ
AND, OR, BUT, EITHER-OR, and NEITHER-NOR have
similar constraints.
Learning
determine if two conjoined adjectives are of the same or different
orientations.
Clustering: produce two sets of words: positive and negative
 61

Find domain opinion words

A similar approach was also taken in (Kanayama and
Nasukawa, 2006) but for Japanese words:
Instead of only based on intra-sentence sentiment consistency,
the new method also looks at the previous and next sentence,
i.e., inter-sentence sentiment consistency.
Have an initial seed lexicon of positive and negative words.
 62

Context dependent opinion

Find domain opinion words is insufficient. A word
may indicate different opinions in same domain.
The battery life is longDQGIt takes a long time to focus-).

Ding, Liu and Yu (2008) and Ganapathibhotla and

Liu (2008) exploited sentiment consistency (both
inter and intra sentence) based on contexts
It finds context dependent opinions.
Context: (adjective, aspect), e.g., (long, battery_life)
It assigns an opinion orientation to the pair.
Other related work (Wu and Wen, 2010; Lu et al., 2011)
 63

The Double Propagation method

(Qiu et al 2009, 2011)

The DP method can also use dependency of

opinions & aspects to extract new opinion words.
Based on dependency relations
Knowing an aspect can find the opinion word that
modifies it
(J7KHrooms are VSDFLRXV

Knowing some opinion words can find more opinion

words
(J7KH rooms are spacious and beautiful

Jijkoun, Rijke and Weerkamp (2010) did similarly.

 64

Opinions implied by objective terms (Zhang

and Liu, 2011)

Most opinion words are adjectives and adverbs,

e.g., good, bad, etc
There are also many subjective and opinion verbs and
nouns, e.g., hate (VB), love (VB), crap (NN).
But objective nouns can imply opinions too.
(J$IWHUVOHHSLQJRQWKHPDWWUHVVIRURQHPRQWKD
valley/body impression KDVIRUPHGLQWKHPLGGOH
How to discover such nouns in a domain or
context?
 65

The technique
Sentiment analysis to determine whether the context is +ve
or ve.
(J,VDZDvalley LQWZRGD\VZKLFKLVWHUULEOH
This is a negative context.
Statistical test to find +ve and ve candidates.

Pruning to move those unlikely ones though sentiment

homogeneity.
 66

Pruning
For an aspect with an implied opinion, it has a fixed
opinion, either +ve or ve, but not both.
We find two direct modification relations using a
dependency parser.
Type 1: 22-'HS$
HJThis TV has good picture quality
Type 2: 22-'HS+$-'HS$
HJThe springs of the mattress are bad

If an aspect has mixed opinions based on the two

dependency relations, prune it.
 67

Opinions implied by resource usage (Zhang

and Liu, 2011)

Resource usage descriptions may also imply

opinions
(J7KLVZDVKHUXVHVDORWRIZDWHU

Two key roles played by resources usage:

An important aspect of an entity, e.g., water usage.
Imply a positive or negative opinion
Resource usages that imply opinions can often be
described by a triple.
(verb, quantifier, noun_term),
Verb: uses, quantifierDORWRInoun_term: water
 68

Dictionary-based methods
Typically use :RUG1HWV synsets and hierarchies to
acquire opinion words
Start with a small seed set of opinion words.
Bootstrap the set to search for synonyms and antonyms in
WordNet iteratively (Hu and Liu, 2004; Kim and Hovy, 2004;
Valitutti, Strapparava and Stock, 2004; Mohammad, Dunne
and Dorr, 2009).
Kamps et al., (2004) proposed a WordNet distance method to
determine the sentiment orientation of a given adjective.
 69

Semi-supervised learning
(Esuti and Sebastiani, 2005)
Use supervised learning
Given two seed sets: positive set P, negative set N
The two seed sets are then expanded using synonym
and antonymy relations in an online dictionary to
JHQHUDWHWKHH[SDQGHGVHWV3DQG1
3DQG1IRUPWKHWUDLQLQJVHWV
Using all the glosses in a dictionary for each term
LQ3 1DQGFRQYHUWLQJWKHPWRDYHFWRU
Build a binary classifier
Tried various learners.
 70

Multiple runs of bootstrapping

(Andreevskaia and Bergler, 2006)

Basic bootstrapping with given seeds sets

(adjectives)
First pass: seed sets are expanded using synonym,
antonyms, and hyponyms relations in WordNet.
Second pass: it goes through all WordNet glosses and
identifies the entries that contain in their definitions the
sentiment-bearing words from the extended seed set and
adds these head words to the corresponding category
(+ve, -ve, neutral)
Third pass: clean up using a POS tagger to make sure
the words are adjectives and remove contradictions.
 71

Multiple runs of bootstrapping (contd)


Each word is then assigned a fuzzy score
reflecting the degree of certainty that the word is
opinionated (+ve/-ve).
The method performs multiple runs of
bootstrapping using non-overlapping seed sets.
A net overlapping score for each word is computed based on how
many times the word is discovered in the runs as +ve (or ve)
The score is normalized based on the fuzzy membership.
 72

Which approach to use?

Both corpus and dictionary based approaches
are needed.
Dictionary usually does not give domain or
context dependent meanings
Corpus is needed for that

Corpus-based approach is hard to find a very

large set of opinion words
Dictionary is good for that
In practice, corpus, dictionary and manual
approaches are all needed.
 73

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 74

Comparative Opinions
(Jindal and Liu, 2006)

Gradable
Non-Equal Gradable: Relations of the type greater or less than
([RSWLFVRIFDPHUD$LVEHWWHUWKDQWKDWRIFDPHUD
%
Equative: Relations of the type equal to
([camera A and camera B both come in 7MP
Superlative: Relations of the type greater or less than all others
([camera A is the cheapest in market
 75

An example
Consider the comparative sentence
&DQRQVRSWLFVLVEHWWHUWKDQWKRVHRI6RQ\DQG1LNRQ
Written by John in 2010.
The extracted comparative opinion/relation:
({Canon}, {Sony, Nikon}, {optics}, preferred:{Canon}, John, 2010)
 76

Common comparatives
In English, comparatives are usually formed by
adding -er and superlatives are formed by adding -
est to their base adjectives and adverbs
Adjectives and adverbs with two syllables or more
and not ending in y do not form comparatives or
superlatives by adding -er or -est.
Instead, more, most, less, and least are used before such
words, e.g., more beautiful.
Irregular comparatives and superlatives, i.e., more
most, less, least, better, best, worse, worst, etc
 77

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 78

Sentiment shifters (e.g., Polanyi and Zaenen 2004)



Sentiment/opinion shifters (also called valence shifters

are words and phrases that can shift or change opinion
orientations.
Negation words like not, never, cannot, etc., are the most
common type.
Many other words and phrases can also alter opinion
orientations. E.g., modal auxiliary verbs (e.g., would,
should, could, etc)
7KHEUDNHFRXOGEHLPSURYHG
 79

Sentiment shifters (contd)

Some presuppositional items also change
opinions, e.g., barely and hardly
,WKDUGO\ZRUNVFRPSDULQJWRLWZRUNV
It presupposes that better was expected.
Words like fail, omit, neglect behave similarly,
7KLVFDPHUDIDLOVWRLPSUHVVPH
Sarcasm changes orientation too
:KDWDJUHDWFDULWGLGQRWVWDUWWKHILUVWGD\
Jia, Yu and Meng (2009) designed some rules
based on parsing to find the scope of negation.
 80

Explicit and implicit aspects

(Hu and Liu 2004)

Explicit aspects: Aspects explicitly mentioned as

nouns or noun phrases in a sentence
The picture quality of this phone is great.
Implicit aspects: Aspects not explicitly mentioned
in a sentence but are implied
7KLVFDULVVRexpensive
7KLVSKRQHZLOOQRWHDVLO\fit in a pocket.
,QFOXGHG16MB LVVWLQJ\
Not much work has been done on mining or
mapping implicit aspects.
 81

Implicit aspect mapping

There are many types of implicit aspect
expressions. Adjectives and adverbs are perhaps
the most common type.
Most adjectives modify or describe specific attributes of
entities.
H[SHQVLYH DVSHFWSULFHEHDXWLIXO aspect
DSSHDUDQFHKHDY\ DVSHFWZHLJKW
Although manual mapping is possible, in different
contexts, the meaning can be different.
(J7KHFRPSXWDWLRQLVH[SHQVLYH
 82

Yet more complications

Spam sentiment
Common on user product reviews
All of NLP
Coreference resolution

 83

Roadmap
Sentiment Analysis Problem
Document sentiment classification
Sentence subjectivity & sentiment classification
Aspect-based sentiment analysis
Sentiment lexicon generation
Mining comparative opinions
Some complications
Generalized sentiment analysis
 84

Beyond positive/negative sentiments

Why are you buying/reading/tweeting this?
What leads to word-of-mouth?
What emotion does it evoke in you?
85

Estimating personality and well-being

from social media
Correlate words in Facebook posts, tweets and Google
search queries with personality, Age, Sex, IQ, happiness
Next: results from 100,000 individuals who took the
Big Five Personality test and their Facebook posts
Words that most highly correlate with traits

WWBP: Ungar, Schwartz, Johannes,

Kern, Ramones, Seligman
Female
Male
Extroverts
Introverts
90

Obama vs. Romney

 91

Sentiment Summary
Sentiment analysis has many sub-problems.
Despite the challenges, applications are flourishing!
Building a reasonably accurate domain specific system is
possible.
Building an accurate generic system is very hard.
But I am optimistic!
 92

Some interesting sentences

7U\LQJRXW*RRJOHFKURPHEHFDXVH)LUHIR[NHHSV
FUDVKLQJ
The opinion about Firefox is clearly negative, but for Google chrome,
there is no opinion.
We need to segment the sentence into clauses to decide that
FUDVKLQJRQO\DSSOLHVWR)LUHIR[
7U\LQJRXWDOVRLQGLFDWHVQRRSLQLRQ
How about this
,FKDQJHGWR$XGLEHFDXVH%0:LVVRH[SHQVLYH
 93

Some interesting sentences (contd)

Conditional sentences are hard to deal with (Narayanan et
al. 2009)
,I,FDQILQGDJRRGFDPHUD,ZLOOEX\LW
But conditional sentences can have opinions
,I\RXDUHORRNLQJIRUDJRRGSKRQHEX\1RNLD

Questions may or may not have opinions

No sentiment
$UHWKHUHDQ\JUHDWSHUNVIRUHPSOR\HHV"
With sentiment
$Q\LGHDKRZWRUHSDLUWKLVORXV\6RQ\FDPHUD"
 94

Some interesting sentences (contd)

Sarcastic sentences
:KDWDJUHDWFDULWVWRSSHGZRUNLQJLQWKHVHFRQGGD\
Sarcastic sentences are very common in political blogs,
comments and discussions.
They make political blogs difficult to handle
Many political aspects can also be quite complex and hard to
extract because they cannot be described using one or two words.
Some initial work by (Tsur, Davidov, Rappoport 2010)
 95

Some interesting sentences (contd)

See these two sentences in a medical domain:
,FRPHWRVHHP\GRFWRUEHFDXVHRIVHYHUHSDLQLQP\VWRPDFK
$IWHUtaking the drug,JRWVHYHUHSDLQLQP\VWRPDFK
If we are interested in opinions on a drug, the first sentence
has no opinion, but the second implies negative opinion on
the drug.
Some understanding seems to be needed?
 96

Some interesting sentences (contd)

The following two sentences are from reviews in the paint
domain.
)RUSDLQWB;RQHFRDWFDQFRYHUWKHZRRGFRORU
)RUSDLQWB<ZHQHHGWKUHHFRDWVWRFRYHUWKHZRRGFRORU
We know that paint_X is good and Paint_Y is not, but how
by a system.
Do we need commonsense knowledge and understanding of the
text?
 97

Some more interesting/hard sentences

0\JRDOLVWRKDYHDKLJKTXDOLW\tv with decent
VRXQG
7KHWRSRIWKHSLFWXUHZDVPXFKEULJKWHUWKDQ
WKHERWWRP
*RRJOHVWHDOVLGHDVIURP%LQJ%LQJVWHDOV
PDUNHWVKDUHIURP*RRJOH
:KHQ,ILUVWJRWWKHDLUEHGDFRXSOHRIZHHNVDJR
it was wonderful as all new things are, however
as the weeks progressed I liked it less and less
 98

UNSUPERVISED
INFORMATION
EXTRACTION
 99

Why do we need Text Mining

Information Overload
News Sites
Social Media
Scientific Articles
Text Mining and Information Extraction
Extract Key Events
Extract Relations
Detect Sentiment
 100

Traditional Text Mining is not cost effective

nor time efficient
h a team of experts led by the person

Why?
takes too long to develop
too expensive
not accurate enough
lacks complete coverage
The Evolution of Information Extraction Technology

Unsupervised
Care 2.0 + Corpus Based
IE
Learning

Care 2.0
Generic Grammar
Augmented IE

Hybrid Information Care 1.0

Extraction

HMM, CRF
Supervised Information Extraction

Rule Based Information Extraction DIAL

 102

Relation extraction
Relation Extraction (RE) is the task of recognizing
instances of specific relationships between two or more
entities in a natural language text.
In a traditional setting, the target relation types are known
to a RE system in advance, and it can be prepared for its
task either
1. by a knowledge engineer hand-crafting the extraction
rules
2. or by the system itself learning the rules from a set of
hand-labeled training examples.
Both ways require a large expenditure of manual labor.
 103

Relation extraction
In recent years, [Banko and Etzioni 2008]
introduced a new setting for the RE task, called
Open Information Extraction (Open IE).
In this setting, the RE system
1. does not know the target relations in advance, and
2. cannot have any relation-specific human input.
The task requires the system itself to identify the
target relations and to train itself for extracting them.
 104

To parse or not to parse?

IE systems that work on free text can perform
either a shallow or a deep parsing.
The advantages of shallow parsing are high
speed, simplicity of training and usage, and
consequent higher accuracy.
Most of the state-of-the-art IE systems do not use
deep parsing, because its obvious theoretical
advantages are offset by the practical limitations
of existing general-purpose parsers.
 105

To parse!!!
We do perform deep parsing, using a parser
which is built specifically for the task of
Information Extraction.
The SDUVHUVunderlying framework, called CARE-
II, is capable of parsing arbitrary weighted typed-
feature-structure-based context free grammars.
This gives the framework a unique power,
allowing it to use a high-level unification-based
grammar, such as HPSG, while still being able to
flexibly interface with feature-rich sequence
classifiers, such as CRF-based NER and PoS
 106

Parsing approach
The parser we built on top of the CARE-II framework
is generic it uses an HPSG-like grammar (derived
primarily from the example grammar in [Sag,
Wasow et al. 2003]).
The parser relies on the CRF-trained NER and PoS
sequence classifiers to provide weights for different
possible typed-feature structure assignments for
different words.
For any input sentence, the parser is able to
generate a single highest- weight parse the parse
which is the most consistent with the NER and POS
classifiers.
 107

An IE focused parser

The resulting parses contain many errors, and would
probably not compare favorably with the results of the
best existing standalone parsers.
However, the goal of the system is not to provide a
general-purpose parser, but to be easily adaptable for the
IE task.
Given a target relation type, the system only requires the
definitions for a small number of the relevant content
words in order to function as an accurate parser for the
relevant sentences the ones that actually contain
instances of the target relations.
 108

Sequence classifiers
A sequence classifier is a component that takes as input a
sequence of tokens (the input sentence) and selects for
each token a single label from a small predefined set of
labels.
For example:
1. a named entity recognizer (NER) would select the
ODEHOVIURPDVHWOLNH^3HUVRQ2UJDQL]DWLRQ
/RFDWLRQ1RQH`
2. a part-of-speech tagger (PoS) would select the labels
IURPDVHWOLNH^11113--9%9%*`
 109

Using the sequence classifiers

Instead of first running a classifier and then using
the produced single fixed sequence of labels, the
framework retrieves the actual weights that the
classifier assigns, and uses them directly as the
weights of the terminal symbols.
Instead RIWKHFODVVLILHUVVitterbi algorithm, it is
the more general WTFSCFG inference algorithm,
which finds the highest-scoring grammar parse,
which maximizes the sum of the weights of words,
rules, and classifier labels that participate in it.
 110

The actual parsing algorithm

An extension of the Agenda-based parser for
PCFG-s [Klein and Manning 2001]. The extension
allows arbitrary weights, as well as feature
structures (HPSG like) and their manipulation.
In the general case, inference with feature
structures is Turing-machine-powerful, so the
DOJRULWKPVZRUVW-case space and time complexity
is exponential. However, with careful writing of the
grammar rules and feature structures, it can be
made to stay efficiently polynomial.
 111

The generic grammar

The English grammar is based on HPSG, as described in
[Sag, Wasow et al. 2003] and adapted for the CARE-II
framework. It contains around thirty general rules and the
lexicon definitions for several hundred functional words:
determiners, pronouns, prepositions, auxiliary verbs, etc.
The content words are defined generically, with the help of
the PoS and NER sequence classifiers. The definitions
allow any word in the input sentence to be assigned to any
word class (in the HPSG sense), with the weights set as
specified by the classifiers
 112

example
Qualcomm has acquired Elata for 57 million in cash.
Acquisition(Acquirer = QualcommAcquired Elata
It is sufficient to define a single content word WKHYHUEWR
DFTXLUH7KHGHILQLWLRQcan look like this:
 113

Example - cont.
The additional positive weight encourages the parses
involving the YHUEDFTXLUHRYHUJHQHULFDOO\GHILQHGZRUGV
If this definition is added to the grammar, the sentence
above will be parsed into the following:
<1: RELN=Org> Qualcomm </1>
<2: RELN=perfect ARG=3> has </2>
<3: RELN=Acquire ACQUIRER=1 ACQUIRED=4
PAST=true> acquired </3>
<4: RELN=Org> Elata </4>
 114

Pattern Learning flow

1. Parse a large unlabeled corpus using the generic grammar.
This may result in many parsing mistakes, but genuine
meaningful patterns would still appear much more frequently
in the parsed text than random mistakes.
2. The extracted frequent patterns are converted into lexical
entries and the corpus is reparsed using the combined
grammar. This produces much more precise results and also
significantly increases recall.
3. The relations extracted by different lexical patterns are
filtered, compared to each other, and merged together into
single relation types if they have sufficient overlap.
4. Names are given to the relation types and their slots, and the
lexical entries are re-generated accordingly, producing the
final domain-specific lexicon.
 115

Pattern types
The patterns are pieces of parses, and the parses are
word dependency graphs. Thus, all patterns are
connected dependency sub-graphs, and each one
includes at least two entity placeholders. There are three
sets of pattern types:
1. verb-based
2. noun-based
3. BE-based.
 116

Verb based patterns

In these patterns the head word is a verb, and the entities
are either subjects, or objects, or connected to the verb or
to its object via a sequence of preposition phrases.
For example:
;2UJs DFTXLUHGF Y/Org
;2UJs merged m with c Y/Org
;2UJs completed c acquisition m of c
Y/Org
The link types:
s - subject
c - complement
m - modifier
 117

Noun based patterns

The noun-based patterns are headed by a noun, and the
entities are connected via preposition phrases,
possessives, or compounds.
For example:
(acquisition m of c X/Org) m by c 
Y/Org
X/Org poss Vm acquisition m by c
Y/Org
Merger m of c (X/Org conj and conj 
Y/Org).
 118

BE patterns
BE-patterns DUHKHDGHGE\WKHYHUEEHLQits predicative
(non-auxiliary) sense:
;3HUVRQLV mayor m of c Y/Loc
;3HUVRQ mayor m of c Y/Loc
;2UJposs V m headquarters (are) in c
Y/Loc
 119

Experimental evaluation ACMM

corpus
 120

Comparison to text runner

The TextRunner TXHU\IRULVPD\RUproduces such
results as:
1. /'$IURPWKHVHQWHQFHThe LDA is the 0D\RUV
agency for business and jobs.
2. \HDUIURPLater this year he is elected Mayor for the
first time.
3. EULJKWVSRWVIURPLook around, you'll see bright spots
in city Morganton News Herald - The following is
Morganton Mayor Mel Cohen's state of the city speech.
Even ZKHQWKHTXHU\LVFRQVWUDLQHGWR3HUVRQis mayor
RI&LW\ZKLFKUHGXFHVWKHUHFDOOIURPVHYHUDOthousands
to just 33 instances, the last sentence from the three
above still remains.
 'LJLWDO7URZHOV9LVXDO&$5(
Visual CARE
The silver bullet for the content industry
unsupervised machine learning
deep linguistic analysis
automatic pattern discovery
automatic rule definition
Accuracy above 95% vs. the industry
standard of 80%

121
'LJLWDO7URZHOV9LVXDO&$5(

Visual CARE the rules are written

for you!
in an intuitive language
review and identify the rules
relevant to you
keep and refine those you care
about and discard the rest
merge similar rules to a single
concept

122
Visual CARE game changing benefits
development time reduced by 99%
GLVFRYHUWKHUXOHV\RXGLGQWNQRZWRORRNIRU
unprecedented accuracy of over 95%
domain agnostic
A disruptive change in the way industries
are able to utilize content

123
 124

Visual Care Main window

 125
Visual Care Architecture
 126

Extractions: Medical forums

I was on adderall which was great, but would give me a
stomach ache for a short time after each dose, and bad night
sweats
SideEffect:
Drug = "adderall"
Symptom = ["stomach ache",
"bad night sweats"]
the dr wants me to go off the avandamet and just take straight
metformin for a week to see if it still causes me nausea
DrugReplacement:
Drug = "avandamet"
Replacement = "metformin"
Reason="to see if it still causes me nausea"
 127

Hpsg parser
CARE-II-HPSG is an English grammar written for the
CARE-II framework, based on the principles of HPSG
grammar theory.
The grammar's lexicon is largely underspecified. Only
the most frequent and functional words have full
definitions, while the open classes of words are defined
using generic underspecified lexical entries and tightly-
integrated feature-rich sequence classification models
for part-of-speech tagging (PoS) and named entity
recognition (NER)
For any input sentence, the parser generates a single
highest-weight parse the parse which is the most
consistent with both the grammar rules and the NER and
PoS classifiers.
 128

sentences IProblematic
I had severe knee swelling and pain from Levemir
insulin and the dr doesn't think Levemir had
anything to do with the severe pain because knee
swelling wasn't listed as a side effect even though
hand and foot swelling was.
The Charniak's parser missed the important
domain-specific relation between "knee swelling"
and "Levemir insulin" by forming a noun phrase
"pain from Levemir insulin and the dr" and
interpreting it as the subject of "doesn't think".
 129

&KDUQLDNV Parse
 130

9&V3DUVH
 131

sentences IIProblematic
Financial Systems Innovation LLC has entered
into a settlement agreement covering a patent
that applies to credit card fraud protection
technology with Lone Star Steakhouse, Inc.
The wide-coverage parser makes a PP-
attachment mistake, attaching "with Lone Star
Steakhouse, Inc" to the immediately preceding
NP instead of "settlement agreement".
 132

&KDUQLDNV Parse
 133

9&VSDUVH
 134

How Care handles these sentences

The focused domain-specific lexical entries, learned
automatically from simpler sentences with more
straightforward parses, have higher weight than the
generic lexical entries, and increase the chances of
parses that contain them to win over parses containing
generic words.
In the first sentence, the important domain-specific word
is the particular form of the verb "have", as in the pattern
"PersonX has SideEffectY".
In the second sentence, it is three words " enter",
"agreement", and the preposition "with", as in the pattern
"CompanyX enters into agreement with CompanyY".
 135

Generalization from simple sentences

The entries in the domain-specific lexicon, after
they were learned from simple patterns are able
to perform extraction in much more general
contexts.
The words participate in all the general linguistic
rules defined by the HPSG grammar, such as
agreement, passive voice, negation, rearranging
of preposition phrases order, conjunctions, etc.
 136

Development cycle for a domain-specific

Relation Extraction system
A newly created VC project contains only the generic
 grammar and the standard set of named entities (Person,
Organization, Location, and Date, available from a NER
sequence classifier, CRF-trained on the data from CoNLL-
2003 (Tjong, Sang et al. 2003) shared task).
If additional domain-specific entity types are needed, their
definitions must be supplied.
A corpus of domain-related sentences must be added to the
project. This starts the pattern extraction and lexicon
acquisition process.
The extracted patterns must be clustered, filtered, and
optionally renamed, which completes the relation identification
and lexicon acquisition processes.
 137

Entity extraction
There 4 different methods to define additional types of
entity types:
1. Using a separately-trained NER model
2. CARE-II rules
3. CARE-II-HPSG lexicon definitions
4. Lists of allowed values (for entity types for which the
sets of entities are closed).
Arbitrary mixing of these methods is also possible and
effective.
 138

 drug namesNER in the medical

domain
For the purposes of extracting DrugReplacement-s and
DrugSideEffect-s the following phrases are all equivalent:
"Byetta", "Byetta pill", "a dose of about 100 mg of Byetta a
day", etc.
However, in terms of the generic CARE-II-HPSG
grammar, they are not equivalent, because the heads of
the noun phrases are different: "Byetta" (DRUG), "pill"
(generic common noun), and "dose" (different generic
common noun), respectively.
In order to make the longer phrases equivalent to a
simple DRUG entity, it is sufficient to add the lexical
entries for the possible head words: "pill", "dose", "mg",
and several others.
 139

NER in the medical domain - symptoms

Extracting SYMPTOM-s is a more complex process. As with
the drugs, we started by building a list of known symptoms,
downloading it from www.drugs.com. Each drug there has a
description page with list of possible symptoms for it.
We combined the lists from all of the listed drugs. The resulting
dictionary was used to create a set of rules in the CARE-II
framework. These rules break down the symptoms to their
components:
1. the problem nouns (e.g., "acid", "bleeding", "ache")
2. problem adjectives ("abnormal", "allergic")
3. body_part ("abdomen", "ankle")
4. behavior ("appetite", "balance", "mood", "sleep").
These components were then added as domain-specific lexical
entries, with the semantics that would allow them to form full
symptom names by combining with each other in any
syntactically-licensed manner.
 140

Handling negation
Nestle S.A. (NESN.VX), the world's largest food and
beverages producer, Tuesday said it won't bid for U.K.-
based confectionery company Cadbury Plc (CBY).
The syntax and semantics of such forms of negation are
handled in the generic grammar in a way compatible with
the HPSG grammar theory.
In practice, if either the main verb or one of the slots of a
relation is modified by a negating modifier ("not" and its
various forms, and other negating words), then the
extracted relation is marked as negated.
 141

Domain specific lexicon acquisition

We add to the project an unlabeled corpus a set of
domain-relevant sentences and run the pattern learning
and lexicon acquisition process. VC uses the unlabeled
corpus for discovering linguistic patterns that can be
directly translated into CARE-II-HPSG lexicon definitions.
The patterns whose corpus frequency exceeds the
threshold (which is set to 2 in this study) get "promoted"
into lexicon definitions, added to the project, and the
affected parts of the corpus are automatically reparsed.
 142

example
The Rel_ORG_enter_into_agreement_with_ORG pattern adds
four entries: the prepositions "into" and "with", the noun
"agreement", and the verb "enter".
1. "Into" DQGZLWKDUHGHILQHGas argument prepositions
2. "agreement" is defined as a noun with a special
SYN.HEAD.FORM and without any non-generic semantics
3. "enter" is defined as a verb with three complements and with
the output relation semantics.
The definitions of non-pattern-specific words, such as "into",
"with" and "agreement" can be reused by many patterns.
The pattern names show only the complements the required
pieces of the pattern. The actual instances of the relation may
also contain optional pieces, specified by modifiers.
 143

Relation clustering
The automatic pattern clustering uses a variant of the
HAC (Hierarchical Agglomerative Clustering) algorithm
with single linkage, which was shown in (Rosenfeld and
Feldman 2007) to be superior for the relation identification
task.
The direct similarity estimation between patterns takes
into account:
1. the structural similarity between patterns, including
standard syntactic transformations;
2. identity of slots' entity types;
3. identity or synonymy or other association between
specific words in the patterns, discovered using
WordNet (Fellbaum 1998).
 144

Clustering example
 145

Extending relations by utilizing

modifiers
 The lexicon acquisition component is currently able to identify
and learn two common modifier patterns: preposition phrase
modifiers and possessive construction modifiers.
Preposition phrase (PP) modifiers have the form of a
preposition complemented by a noun phrase (<prep> NP), and
can modify verb and noun phrases. The generic CARE-II-
HPSG grammar contains default lexical entries for all
prepositions, which allow any PP to modify any NP or VP.
Possessive construction modifiers always modify noun
phrases. They have three syntactically different but
semantically identical forms: possessive determiner form (X's
<noun>), compound noun form (X <noun>), and of-preposition
phrase form (of X).
 146

examples
In order for a non-default domain-specific modifier to be useful,
and therefore learnable, its NP part must contain a relation
argument extractable entity either directly, as in:
In January 1997, Hays bought German distributor Daufenbach
for 30 million GBP, ...
or via a PP-NP chain, as in:
Ms. Bennett succeeds William J. Viveen, Jr., as a member of
the Interleukin Board of Directors.
assuming the verb relation patterns Rel_ORG_buy_ORG and
Rel_PERSON_succeed_PERSON are already learned, the two
sentences above would generate the PP modifier patterns
Mod_verb_in_DATE and Mod_verb_as_member_of_ORG.
 147

Modifiers are cross-pattern

The same PP modifier can attach to phrases extracting
different relations, and the same possessive modifier can
attach to nouns of the same type within different patterns.
It is possible to precisely fine-tune the scope of a modifier
to specify the relations and clusters to which the modifier
is applicable, and, for each relation and cluster, the slot
name of the argument that the modifier extracts.
Thus, the scope of Mod_verb_as_member_of_ORG can
be limited to the ManagementChange cluster, and the
extracted slot can be specified as EMPLOYER.
 148

Co-reference resolution
we implemented the co-reference resolution system of (Lee,
Peirsman et al. 2011), which was the best-performing in
CoNLL-2011 shared task, and which is perfectly adaptable to
the CARE-II-HPSG/VC environment.
The method for resolution of co-references is based on locating
all noun phrases, identifying their properties, and then
clustering them in several deterministic iterations (called
sieves), starting with the highest-confidence rules and moving
to lower-confidence higher-recall ones. In each iteration, the
order of candidate checks is deterministic, and any matching
noun phrases with matching properties are immediately
clustered together.
Despite its simplicity, the method showed state-of-the-art
accuracy, outperforming the other systems in the CoNLL
shared task.
 149

Co-reference resolution
The method is especially suitable for the our framework,
because all information the method requires is already
extracted: the noun phrases are located from the parses,
together with their properties, which are identified from
HPSG feature structures.
Co-reference resolution module tries to resolve all noun
phrases, although non-entity ones are discarded. This is
necessary for improving the accuracy on the relevant
entity mentions, by removing irrelevant candidates.
 150

Examples
The entity to resolve to is the first found one with matching features. The order of checking
is strictly by distance (closest first), except for pronominal/generic-nominal references, for
which
the entities in subject positions are checked before the others.

Microsoft announced it plans to acquire Visio. The company said it will finalize its plans
within a week.

Only pronomial and generic nominal references are here. The first "LWVWUDLJKWIRUZDUGO\
matches "Microsoft". "The company" also matches "Microsoft" first, because it's a
pronominal reference, and "Microsoft" is in the subject position. The second "it" matches
"the company", and so is also resolved to "Microsoft".

Mark said that he used Symlin and it caused him to get a rash. He said that it bothered
him.

First "he" resolves to "Mark". "It" does not match "Mark", and so is resolved to
"Symlin". Other "Him" and "he" are also resolved to "Mark". The second "it" is resolved to
"a rash", because it is closer and neither of the matching candidates - "Symlin" and "a
 151

Post processing
The post-processing phase tried to combines a few simple
relation to one complex relation.
The slots of any adjacent relation-pieces are merged, if the
pieces are compatible, and if none of the slots contradict each
other.
For the DrugReplacement relation, StopUsingDrug can merge
with StartUsingDrug and with UseDrug. For the SideEffect
relation, HasSymptom can merge with any of the three:
StartUsingDrug, StopUsingDrug, or UseDrug.
After merging, all StartUsingDrug, UseDrug, and HasSymptom
relations that were not identified as parts of a larger relation
can be removed, since we are interested only in the full
DrugReplacement and SideEffect relations in this project.
 152

evaluation
The input corpus for the medical domain consists of about
650,000 messages downloaded from the archives of
several public medical forums.
The contents of the messages were separated into
sentences, filtered and deduplicated, leaving about
300,000 sentences totaling in about 40MB of ASCII text.
 153

Evaluation results
 Without Coreference With Coreference

Relation Precision Count Recall est. Prec Count Recall est.

DrugReplacement 0.89 2732 0.65 0.88 3039 0.74

 2850  3163  0.98

Slot: Drug 0.96 0.98 0.95

 638  825  0.80

Slot: Replacement 0.93 0.70 0.93
Slot: Reason 0.93  294 0.76  
SideEffect 0.85 2278 0.66   

Acquisition 0.88 191 0.48 0.85 341 0.81

 154

Evaluation methodology
Precision
We evaluated on a random set of 200 extractions.
Recall
We checked only the sentences that contained at least two
DRUG entities for the DrugReplacement and at least one
DRUG and at least one SYMPTOM entity for the SideEffect,
together with an indicative word, such as "replace", "switch",
"change", or one of their derivatives for the DrugReplacement.
 155

Financial domain evaluation

In the financial domain, we used two corpora: "Corpus A" of 200,000 sentences selected randomly
from Dow Jones newswire articles published between 1990 and 2010, and "Corpus B" of 40,000
sentences from Dow Jones in 2011.
First, we used VC to learn as many relations as it could from the Corpus A. The system found 919
different patterns, of which 481, clustered into 172 relation clusters, were judged interesting (by manual
inspection).
Here are the ten top (according to the number of instances found in the corpus) identified clusters:
1. Acquisition (Organization, Organization)
2. Replacement (Person, Person)
3. Employment (Person, Organization)
4. NewProduct (Product, Organization)
5. Agreement (Organization, Organization)
6. ORG_own_interest_in_ORG (Org, Org)
7. HeadPosition (Person, Organization)
8. Resignation (Person, Organization)
9. Sale (Organization, Organization)
10. HeadquartersIn(Organization, Location
 156

Coverage experiment
we used VC to learn relations from Corpus B
independently from Corpus A, and compared the sets of
relations extracted from the Corpus B by domain-specific
lexicons learned from itself and from Corpus A.
We have found that out of 1742 relation instances found
in Corpus B, we were able to extract 1367 of them
(78.4%) using the patterns learned from Corpus A.
Sentiment Analysis of Stocks from News
Sites
 

So, How Can We Utilize NLP for

Making Money?
Goal: sentiment analysis of financial texts as an aid for
stock investment
1. Tagging positive and 2. Article scoring
negative sentiment in articles

3. Score aggregation:
daily and cumulative
score
 

The Need for Event Based SA

Toyota announces voluntary recall of their highly successful top
selling 2010 model-year cars

Phrase-level SA:
highly successful top selling positive
Or at best neutral
Taking into account voluntary recall negative
1HHGWRUHFRJQL]HWKHZKROHVHQWHQFHDVDSURGXFWUHFDOOHYHQW
 

CaRE extraction
Engine
 Template Based Approach to Content
Filtering
Hybrid Sentiment Analysis

Events
(Predicate)

Patterns Dictionaries
(Phrasal) (Lexical)

All levels are part of the same rulebook, and are therefore
considered simultaneously by CaRE


Dictionary-based sentiment
Started with available sentiment lexicons
Domain-specific and general
Improved by our content experts
Examples
Modifiers: attractive, superior, inefficient, risky
Verbs: invents, advancing, failed, lost
Nouns: opportunity, success, weakness, crisis
Expressions: exceeding expectations, chapter 11
Emphasis and reversal
successful, extremely successful,
far from successful


Event-Based Sentiment
Product release/approval/recall, litigations, acquisitions, workforce
change, analyst recommendations and many more
Semantic role matters:
Google is being VXHGLVVXLQJ
Need to address historical/speculative events
Google acquired YouTube in 2006
What if Google buys Yahoo and the software giant Microsoft
remains a single company fighting for the power of the Internet?


Extract sentiment and events of public companies from online
news articles to enable better trading decisions

The Stock Sonar

 166

TEVA
 167

MOS
SNDK
GM
MU
CLF
0DF\V
JC Penny
POT
Ford
Monsanto
Mining Medical User Forums
 178

The Text Mining Process

html-pages are downloaded from a given forum site
Downloading

html-like tags and non-textual information like images,

Cleaning FRPPHUFLDOVHWFDUHFOHDQHGIURPWKHGRZQORDGHGWH[W

The textual parts are divided into informative units like

Chunking threads, messages, and sentences

Products and product attributes are extracted from the

Information
Extraction messages

Comparisons are made either by using co-occurrence

Comparisons analysis or by utilizing learned comparison patterns
 179

The Text Mining Process


Downloading

We downloaded messages from 5 different consumer

forums

diabetesforums.com
Cleaning

healthboards.com

Chunking
forum.lowcarber.org

diabetes.blog.com**
Information
Extraction

diabetesdaily.com
Comparison
s
** Messages in Diabets.blog.com were focused mainly on Byetta
 Drug Analysis
Drug Co-Occurrence - Spring Graph Perceptual Map
180

Several
Pockets of
drugs that were
mentioned
frequently
together in a
message were
identified
Byetta was
mentioned
frequently with:
Glucotrol
Januvia
Amaryl
Actos
Avandia
Prandin
Symlin
Lifts larger than 3
Width of edge reflects how frequently the two drugs appeared together over and beyond what
one would have expected by chance
 Drug Usage Analysis
Drug Co-taking 'UXJVPHQWLRQHGDV7DNHQ7RJHWKHU
181
There are two main clusters
of drugs that are mentioned
DVWDNHQWRJHWKHU

Byetta was mentioned as

WDNHQWRJHWKHUZLWK
Januvia
Symlin
Metformin
Amaryl
Starlix

Pairs of drugs that are

taken frequently together
include:
Glucotrol--Glucophage
Glucophage--Stralix
Byetta--Januvia
Avandia--Actos
Glucophage--Avandia
Lifts larger than 1
Width of edge reflects how frequently the
two drugs appeared together over and
beyond what one would have expected by
chance
 Drug Usage Analysis
Drug Switching 'UXJVPHQWLRQHGDV6ZLWFKHGWRDQGIURP
182

There are two main clusters of

diabetes drugs within which
consumers mentioned frequently
WKDWWKH\VZLWFKHGIURPRQHGUXJ
to another
%\HWWDZDVPHQWLRQHGDVVZLWFKHGWR
and from:
Symlin
Januvia
Metformin

Lifts larger than 1

Width of edge reflects how frequently the two drugs
appeared together over and beyond what one would
have expected by chance
 Drug Terms Analysis
183
Byetta - Side Effects Analysis

Byetta appeared much

more than chance with
the following side
effects:
1RVHUXQQLQJRU
UXQQ\QRVH
1RDSSHWLWH
:HLJKWJDLQ
$FLGVWRPDFK
9RPLW
1DXVHD
+LYHV
 Drug Terms Analysis
Drug Comparisons on Side Effects
184
Lifts larger than 1
Width of edge reflects how frequently
the two drugs appeared together over
and beyond what one would have
expected by chance

The main side effects

discussed with Januvia:
Thyroid
Respiratory infections
Sore throat

The main side effects

discussed with Levemir :
Note that only Byetta No appetite
is mentioned Hives
frequently with terms
OLNHYRPLWDFLG The main side effects
VWRPDFKDQG discussed with Lantus:
GLDUUKHD Weight gain
Byetta shares with Byetta shares with Byetta shares with Nose running
Januvia the side effects: Levemir the side effects: Lantus the side effects:
Pain
Runny nose No appetite Nose running
Nausea Hives Weight gain
Stomach ache
Hives
 Drug Terms Analysis
185
Byetta Positive Sentiments

Byetta appeared much

more than chance (lift>2)
with the following positive
sentiments:
+HOSVZLWKKXQJHU
1RQDXVHD
(DV\WRXVH
:RUNV
+HOSVORVLQJZHLJKW
1RVLGHHIIHFWV
 Drug Terms Analysis
Drug Comparisons on Positive Sentiments
186
Lifts larger than 0.5
Width of edge reflects how frequently
the two drugs appeared together over
and beyond what one would have
expected by chance
The main positive sentiments
discussed with Januvia:
1RQDXVHD
%HWWHUEORRGVXJDU
:RUNV
Note that only Byetta 1RVLGHHIIHFWV
is mentioned
IUHTXHQWO\ZLWKKHOSV The main positive sentiments
ZLWKKXQJHUSRLQW discussed with Levemir :
of difference) (DV\WRXVH
)DVWDFWLQJ

The main positive sentiments

Byetta shares with Byetta shares with discussed with Lantus:
Byetta shares with
Januvia: Levemir: )DVWDFWLQJ
Lantus:
%HWWHUEORRGVXJDU (DV\WRXVH :RUNV
(DV\WRXVH
1RQDXVHD +HOSVORVHZHLJKW 1RVLGHHIIHFWV
+HOSVORVHZHLJKW 1RVLGHHIIHFWV :RUNV
1RVLGHHIIHFWV :RUNV
:RUNV
 
Example of Unsupervised IE Results
Actos; weight gain (40 (P: 38, N: 2))
Rel_take_DRU_has_SYM(DRUG, SYMPTOM)
Negative (1)
I've been taking 15 mg of Actos for just over a year now and so far (knock on wood) I haven't had the
weight gain that
some others have reported as a side effect.
Positive (8)
I also have read here about some of you who have been on the Actos and the weight gain you had
experienced.
We saw an endo because of all of the weight gain and side effects from taking actos.
He was on Actos but went off of it because of weight gain and stomach bloating.
I really don't want to go back on Actos because of weight gain/fluid retention.
My doctor wanted me to start Actos for awhile, until the Byetta kicks in, but I stopped Actos in the first
place because
of weight gain and I said no to restarting that.
I started taking Actos first on May 2, 2007 and I started Metformin 3 weeks later I can not take the
Spironolactone till
Aug but I have noticed that I have gained weight with these 2 drugs instead of losing
and I got a treadmill and do 30 min every morning when I get up and lately I have been doing 30 min at
night too
because of the weight gain.
I have experienced weight gain as well and i am on Actos and insulin and glucophage.
I guess that everything comes with a price, but I'm wondering if most folks who have tried Actos have
experienced
weight gain and the other side effects (edema, headaches, nausea, fatigue, etc.).
Rel_SYMofDRU(SYMPTOM, DRUG)
Positive (5)
I do notice that it increases my hunger, so it is possible that Actos weight gain issues may be from hunger
being
stimulated.
I don't think that a lot of us had made the Actos induced weight gain connection.
One reported side effect of Actos is weight gain.
 
Example of Unsupervised IE Results - FRQW

Rel_cause_DRUvSYM(DRUG, SYMPTOM)
Negative (1)
Actos hasn't caused any weight gain, I am still losing some.
Positive (25)
I also am on Synthroid, Atenolol, Diovan, Lotrel, Lexapro, Vitorin and Prilosec OTC. I didn't realize that Actos can cause
a weight gain
as I had never read it as a side effect; however, after reading all of the
comments on this site, I now know why my weight has increased over the past few months since taking on it.
I don't take any oral meds, but from what I have read here, Actos causes weight gain because of water retention.
why does the endo think you're a type 1? oral meds are usually given only to type 2's,as type 2's have insulin resistance.
oral meds
treat the insulin resistance. type 1's require insulin.....i take actoplus met- which is actos and metformin.actos is like
avandia and
i've had no heart issues.....tho-avandia and actos can cause weight gain....take care,trish
Actos causes edema and weight gain also.
Actos can cause weight gain (so can Avandia, it's cousin)
Now I have started to see a lot of reports of Actos causing weight gain, among other things.
for the record, Actos can, and does, cause weight gain/water retention.
I'm on both - what did you hate about Metformin? (Actos causes weight gain, metformin weight loss)
Also I hear that the Actos causes weight gain, so now I am afraid the new pill will cause me to gain weight.
I'm type 1 so only on insulin, but I have heard that Actos can cause weight gain.
Avandia & Actos, especially in combination with insulin, causes fluid retention and/or fat weight gain.
My endocrinologist warned me that Actos can cause significant weight gain.
Actos caused weight gain and fluid retention in my chest.
Metformin causes weight loss, Avandia and Actos causes the birth of new fat cells and weight gain.

Side Effects


Side Effects and Remedies
Red lines side effects/symptoms
Blue lines - Remedies

See what causes

symptoms and
what relieves them

See what positive

and negative
effects a drug has

See which
symptoms are most
complained about
Drugs Taken in Combination
 192

XML output of VC
USING MULTI-VIEW LEARNING
TO IMPROVE DETECTION
OF INVESTOR SENTIMENTS ON
TWITTER
Motivation
Stocks-related messages on social media, Twitter in
particular, have several interesting properties with regards
to the sentiment analysis task.
The analysis is particularly challenging:
frequent typos
bad grammar
idiosyncratic expressions specific to the domain and the media.
Stocks-related messages almost always refer to the state
of specific entities companies and their stocks at
specific times (times of sending). This state is an objective
property and even has a measurable numeric
characteristic, namely the stock price.
Multi View Learning
Given a large dataset of twitter messages ("twits"), it is
possible to create two separate "views" on the dataset, by
analyzing the text of the twits and their external properties
separately.
Using the two views, we can expand the coverage from
sentiments detectable using generic sentiment analysis
tools to many other twits in the dataset.
We can use the results to learn new sentiment
expressions.
Agenda
1. Complexity and challenges associated with analyzing
finance-related message.
2. Related work.
3. Methodological approach for multi-view learning.
4. The SA system.
5. Experimental design and results.
6. Concludes and Future Research.
 Complexity and challenges associated
with analyzing finance-related message
Language Challenges
Sarcasm - ,W
VUXPRUHGWKDW)DFHERRNZLOODQQRXQFHD
new kind of free stock dividend tonight, namely
CrashVille. $FB $ZNGA #SoCalledEarnings
Price Dependency - 6,1$ZDWFK57.50 which now is a
resistance, $59.21 that will open the door to $62.48
before earning, if not we will break $50.
If the price at the time of the analysis is 59.21, the sentiment is very
positive, as the author believes it can reach 62.48,
otherwise, it is negative as the author believes it can drop to $50.
General SA Issues
Sentiment PRGLILHUVHJKLJKO\
Emphasis PRGLILHUVHJPRVWO\
Opposite PRGLILHUVHJIDUIURP
Sentiment VKLIWHUVHJQRW
Anaphora resolution is also a challenge when conducting
sentiment analysis, although, the short messages on
social media often bear little anaphora.
Trading MU (Micron)
 Related work Text mining and
Finance
Engelberg (2008)
Found that qualitative earnings information has additional predictability for
asset prices beyond the predictability in quantitative information.
Loughran and McDonald (2011)
Uses textual analysis to examine the sentiment of corporate 10-K reports.
Feldman at al. (2011)
The Stock Sonar uses state-of-the-art information extraction technologies to
analyze news relating to stocks and aggregate the sentiments into a score
and graphically present the information.
Boudoukh et al. (2012)
Found evidence of a strong relationship between stock price changes
and information arrival by using relevant news articles as a proxy.
Information arriving to investors sooner can provide a significant competitive
advantage to those investors. Vis--vis, it is reasonable to assume that
social media response to news is much quicker than professional media.
Related work Social networks
Sprenger & Welpe (2010)
Found that tweet sentiment associated with abnormal stock returns and
message volume to be a good predictor for next-day trading volume.
Bar-haim et al. (2011)
Proposed a general framework for identifying expert investors, and used it
as a basis for several models that predict stock rise from stock microblogging
messages.
Bollen, Mao, and Zeng (2011)
Investigated whether measurements of collective mood states derived from
large-scale Twitter feeds are correlated to the value of the DJIA over time.
Their results suggested that the accuracy of DJIA predictions can be
significantly improved by including specific public mood dimensions.
Xu (2012)
Used NLP to get the public sentiment on individual stocks from social media.
)RXQGWKDWXVHUVactivity overnight significantly correlates positively to the
stock trading volume the next business day.
Found Granger Causality (in 9 out of 15 stocks studied) between collective
sentiments for afterhours and the change of stock price for the next day
Analysis
One particular point in time that should be noted in the
figure is that on April 1st 2011, there were both positive
and negative messages about the company.
However, the positive messages discussed the stock past
performance (which was, indeed, positive for the
referenced period),
while the negative messages reflect the investors current
and future expectations from the stock (which appears to
FRUUHODWHZLWKWKHDFWXDOVWRFNSULFHVWUHQGVLQWKH
presented example).
Basic Learning Model
Assume first, that we have a large corpus T = {t1, t2, ...} of text
messages. Each message t has a true polarity
Pol(t) POLS={POS, NEG, NEUTRAL}, which can always be
identified (by people) by reading and understanding the text.
We make further assumption that the polarity of a message is
determined by occurrence of various sentiment expressions
within the message's text.
The nature of the expressions is not essential at this point.
They can be individual words, multi-word sequences,
sequences with gaps, syntactic patterns, etc. The important
point is that given a message ti, it must be easy to list all
expressions occurring within it. For the purposes of sentiment
expression learning, we represent message texts as bags-of-
expressions: ti = {wi1, wi2, ...}.
Observations
If the true polarity Pol(t) were known for each message tT,
there would be a natural way to search for sentimental
expressions by simply counting their occurrences. However, in
a large unlabeled corpus, the true polarities of messages are
unknown.
On the other hand, the polarity of a message is causally-
independently influenced by external (relative to the message's
text) factors:
if a given stock is doing good or bad on a given day, then the polarity
of the messages about that stock would tend to be correspondingly
positive or negative;
messages from the same author about the same stock would tend to
have the same polarities; etc.
Thus, we have two parallel views on a set of twits, which allows
multi-view learning.
Parallel Views Textual View
Given a large corpus T we first process it with some text-
based SA (sentiment analysis) system, which performs as
a function SA : T o POLS, producing a classification
T = TSA-POS TSA-NEG TSA-NEUTRAL.
The SA system is assumed to have a relatively high
precision for polarized twits, but insufficient recall.
TSA-POS and TSA-NEG generally contain mostly positive and
negative messages, respectively
TSA-NEUTRAL cannot be assumed to contain only neutral
twits. It is also much bigger than the two polarized sets.
Parallel Views : The Second View
We process the corpus T with a feature extractor, which
generates a real-valued high-dimensional vector for each
message, using any possible property of it that is conditionally
independent from its text content given its polarity.
Using this representation, we train a binary SVM classifier with
TSA-POS and TSA-NEG as the training data. This classifier then
produces a score f(t) for each message t in TSA-NEUTRAL.
The significant properties of f(t) are:
(1) because it is grounded on generic SA and external
properties, its sign and magnitude correlates with the true
polarity of t, but
(2) it is independent from the text patterns within t (conditional
on t's true polarity).
Estimating P(Pol(w) = A) - I
Let there be a previously unknown text pattern w, appearing in
TSA-NEUTRAL. We are interested in probabilistically estimating the
polarity of w, that is, in the value of P(Pol(w) = A), where
A{POS, NEG}.
Let Tw = { tTSA-NEUTRAL : wt } be the set of all messages
containing w. Then the probability P(Pol(w) = A) can be
estimated from f(t) scores of messages in Tw:



The constant prior P(Pol(w) = A) can be ignored, assuming the
set Tw is sufficiently large. In the main factor, we can safely
assume that the different twits in Tw are independent from each
other, so:




Estimating P(Pol(w) = A) - II
The marginal P(f(t)) can be estimated directly from the
SVM classifier's scores on T. In the other part of the
formula above, we are only dealing with twits that contain
w whose polarity is non-neutral. According to our
simplifying assumptions, we proceed as if there could be
no conflicts, and the polarities of all twits in Tw are equal to
the polarity of w. And then, the likelihood P(f(t) | Pol(w)=A)
can be reduced to:

 

The constant marginal P(Pol(t) = A) can be directly
estimated from a manually labeled development test set.
And the rest can be estimated using the SA-polarized
sets.
Estimating P(Pol(w) = A) - III
Due to the conditional independence of SA(t) from f(t),
given the true polarity of t, we have:

 

  



P(f(t) | SA(t)=B) and P(SA(t)=B) are estimated directly
from the SA results and SVM results.
P(SA(t) = B | Pol(t) = A) is estimated on a development
test set.
architecture of the learning system

Unlabeled
Corpus

Seed SA SA-Neutrals

SA-POS SA-NEG
Trained Classifier based on
External Features

External Features Extractor

Stock prices Messages together with their
information from classification scores
Google Finance
SVM Training

Extracted Sentiment Expressions

sorted by probability
Features for the "External" View

The external view can use any message properties that
are independent from the message text.
1. Stock-price-related features. These are related to the
price of the stock referenced in the messages, within
some time frame of the message post time. The
numerical values of the stock prices cannot be used
directly, because they vary widely from stock to stock.
However, we can identify and use 'price change events'
the points in time where the price of a stock
significantly changed from one day to the next.
2. Seed SA-related features. These features more
directly utilize connections between messages
established by identity of their authors and/or subject.
Stock-price-related features
There are many different possible adjustable parameters for
identifying the useful price changes, and there is no a priori reason to
select any particular numbers. Therefore, we choose several different
reasonable values, and let an SVM classifier training algorithm
choose the best. We use all possible combinations of the following:
1. SMALL is when the price is changed by 2-5%, LARGE is when the
change is at least 5%, NOCHANGE if the change is less than 1%.
2. YESTERDAY is when the change occurs between yesterday's
closing price (relative to the message post time) and today's
opening price. TODAY is for changes between today's opening and
closing prices, and TOMORROW is for changes between today's
closing and tomorrow's opening prices.
3. PLUS when the price is increased, and MINUS when it decreased.
We use all combinations of these properties as binary features, and
also all possible intersections of pairs of them.
Seed SA-related features
Let POS, NEG, NEUTRAL stand for the binary property of
some message's (not the target's!) having overall positive,
overall negative, and overall neutral sentiment labeling,
respectively, according the seed SA. Also, let HASPOS,
HASNEG, and HASANY stand for there being some
positive, some negative, and some polarized sentiment
expression within. (Thus, for example, POS and HASNEG
may both be true for a message if a negative sentiment
expression occurs within, but the overall sentiment is
positive).
Given some set of messages, let EXIST_X be a binary
property, of at least one of the messages satisfying X from
(i). Also, let SUM_X, and AVERAGE_X be the real-valued
sum and average of X over the messages in the set.
Seed SA-related features
Given a target message, let DAYBEFORE, DAYAFTER,
WITHINDAY, WITHINHOUR, WITHIN10MIN be the sets of
messages posted the day before the message, the day
after, within the same day, within the same hour, and
within 10 minutes, respectively.
Given a target message, let SAMEAUTHOR and
ANYAUTHOR be the sets of messages posted by the
same author and by any author, respectively.
Again, we use all possible combinations of these
properties as features.
Pattern-based Filtering
This is a pattern-centric learning method, different from the
message-based learning described above. The method uses a
different learning model. It is based on the observation that
different sentiment expressions occurring within the same
message generally tend to be of the same polarity. Thus, given
a candidate expression, we can ascertain its polarity by
observing all messages that contain the expression which were
labeled by the seed SA.
The method cannot be directly incorporated into the above-
described learning model, because it is directly using the
message text, and does not satisfy the independence
condition. However, the method can be used to perform an
additional filtering step for the learned expressions, significantly
improving the precision of the overall learning process.
Sentiment Analysis Systems
The basic architecture of the SA system determines the
kinds of sentiment expressions that can be used and
learned. It is also the seed SA used for starting off the
learning process and for calculating the SA-related
features.
Since the goal is seeking new polarity expressions, the
SA system must be able to use them directly. This
immediately eliminates from consideration some SA
architectures, such as bag-of-words classification-based.
(They should also be eliminated on independent grounds)
Mainly, we experiment with a SA architecture, which is
based on CARE-II-HPSG parser and relation extraction
system.
Sentiment Analysis Systems
We experiment with two versions of the system:
GenericSA, a general-purpose SA system, which contains
only language-wide sentiment expressions, and
FinancialSA, which is an extension of GenericSA, created
by manually adding many financial domain-specific
sentiment expressions.
For baseline, we also use DictionarySA, a simple system
that classifies a message into positive, negative, or
neutral categories according to the number of sentimental
expressions that occur within the message. The
expressions are the words and multi-word sequences
taken from the GenericSA, without the additional syntactic
information.
Care-II
CARE-II is a domain-independent framework for building
Information Extraction systems. The framework includes a
grammar description language and the supporting tools.
The core of the framework is a parser, which is capable
of parsing arbitrary weighted typed-feature-structure
context-free grammars (WTFSCFG-s).
These are weighted CFG-s, in which every matched
symbol, terminal or non-terminal, carries a typed feature
structure; the grammar rules have access to the feature
structures of their component symbols, building from them
the feature structures for their heads, by applying the
operations of unification, slot extraction, and slot removal.
Care-II HPSG
CARE-II-HPSG is an English grammar written for the
CARE-II framework, based on the principles of HPSG
grammar theory.
The grammar's lexicon is largely underspecified. Only the
most frequent and functional words have full definitions,
while the open classes of words are defined using generic
underspecified lexical entries and tightly-integrated
feature-rich sequence classification models for parts-of-
speech (POS) and named entities recognition (NER).
The models provide weights for different possible typed-
feature-structure assignments. Then, for any input
sentence, the parser generates a single highest-weight
parse the parse which is the most consistent with both
the grammar rules and the NER and POS classifiers.
 Using CARE-II-HPSG for Sentiment
Analysis
 When used for sentiment analysis, either general purpose
or domain specific, the lexicon is extended to include
sentiment words and expressions, which include
sentiment labels in the semantic parts of their HPSG
feature structures. The labels may indicate polarity of
expressions, their intensity, and their combining
properties, such as behavior under negation.
After a parse of a sentence is generated, it is post-
processed by the SA post-processor, which merges
sentiments from related expressions, performs
coreference resolution, and attaches the sentiments to
their targets, where appropriate. The post-processor is
rule-based and deterministic.
To Parse or not to Parse?
Benefits in using full parsing for the SA task:
1. Precise identification of sentiment target in cases where several
entities are available as possible targets.
2. Principled and uniform combining interdependent sentiment
expressions and processing of negation.
3. Disambiguation for the cases where polarity of an expression
depends on its syntactic role and/or part-of-speech (as in, for
example, "fine" as a positive adjective vs. "fine" as a negative
noun).
4. Principled and uniform way of defining multi-word sentiment
expressions - using syntactic and semantic links instead of simple
words proximity.
The disadvantages of using full parsing are: slower processing speed,
possible problems with bad grammar and typos, and generally low
quality of parses, which may introduce SA errors instead of solving
them. However, with a robust parser, these disadvantages should be
minimized.
 GenericSA, FinancialSA, and
DictionarySA

FinancialSA is the same as GenericSA, but extended by
manually adding many domain-specific lexical entries (for the
financial domain, investor sentiments in particular).
DictionarySA is provided as a baseline. It is a very simple SA
system that contains a dictionary of sentiment words and word
sequences, and classifies a text by counting the number of
occurrences of various polarity expressions within it. Whichever
polarity occurs most frequently wins. If the number of
occurrences is equal, the message is considered neutral. The
initial dictionary is the same as used in GenericSA, without the
additional syntactic information, such as parts-of-speech,
valence, etc.
The type of SA determines what kind of sentiment expressions
is available for learning. The simple DictionarySA can only
learn words and multi-word sequences. The parser-based SA
systems are able to learn more complex patterns.
Pattern Types
1. Word patterns: individual and compound non-proper
nouns, verbs, and adjectives.
2. Valence patterns: head word together with its Valentes,
which can be noun phrases or preposition phrases
complemented by noun phrases. The noun phrases are
identified by their head noun, which can be either
unrestricted, or restricted to a specific common noun, or
restricted to the sentiment target entity type (Company
name or stock symbol for the financial domain).
3. Modifier patterns: a head word modified by an
adjectival phrase or a prepositional phrase
complemented by a noun phrase. Same restrictions
apply as for valence patterns.
Experimental Evaluation
We use a corpus of several million stocks-related twitter
messages collected between May and October 2011. We
only use the messages related to stocks for which we
were able to collect the price information from Google
Finance.
For the test set we use a manually-labeled set of
randomly chosen 1500 twits. Another set of 500 twits was
used as a development test set for estimating the
marginal probabilities and for tuning the final threshold
parameter.
Baseline
In the baseline experiment we compare the results
produced by the three seed SA systems that we use. For
reference, we also show the results produced by Bing's
system on the same test set. The system is one of the
state-of-the-art general-purpose SA systems.
Ding, X. et al. 2008. A Holistic Lexicon-Based Approach to
Opinion Mining. the Conference on Web Search and Web
Data Mining (WSDM-2008) (2008), 231239.
Initial Results on StockTwits (Base Line)

 TP FP FN Prec Recall F1

 Bing's 365 295  418 0.553 0.466 0.506

 Dictionary 308 223 474 0.580 0.393 0.468

 GenericSA 159 94 506 0.628 0.239 0.346

FinancialSA 284 116 381 0.710 0.427 0.533

Analysis
As can be seen from the table, generic SA systems have
relatively low accuracy, due to specifics of the domain.
Also notable, that the simple dictionary-based SA is not
worse in precision from a much more sophisticated Bing's
system in this domain (although much worse in recall).
Note, that in the Table 1 we only consider positive and
negative sentiments when counting the "true positives"
(TP). Neutral sentiments are not included in the
evaluation, except when they contribute to "false
positives".
Results with neutral messages included

 TP FP FN Prec Recall F1

 Bing's1239 295  418 0.808 0.748 0.777

 Dictionary 1183 223 474 0.841 0.713 0.771

 GenericSA 964 94 506 0.911 0.655 0.762

FinancialSA1066 116 381 0.901 0.736 0.810

Learning Experiment
In this experiment we compare the results produced by learning new
sentimental expressions from the twitter messages corpus.
In addition to comparing the learning capabilities of the three SA
systems, we also compare three sets of external features:
(1) PriceOnly set, which contains only features based on stock price,
(2) SeedSA-Only set, which contains only features based on using
Seed-SA-produced classification of messages related to the same
entities, and
(3) Full set, containing both Price-related and SeedSA-related
features, as well as their intersections.
The experiments are performed in the following way: given a SeedSA
and an external feature set, we process the corpus with the SA,
producing three separated sets of messages: SA-POS, SA-NEG, and
SA-NEUTRAL. Then we train an SVM classifier using the
representations in the external feature set of SA-POS and SA-NEG
as training data. This classifier is applied on the representations of the
messages in SA-NEUTRAL, producing a score for each message
within.
Learning Experiment
Then, for every candidate expression that appears at
least ten times for different stocks (so as to eliminate local
biases), we calculate its two final probability scores (of
being positive and of being negative), using the formulas
in Section 4. Given a probability threshold, we append all
discovered expressions that pass the threshold to the
corresponding seed system, and perform the test.
We found that the final score number, while good for
ordering the expressions, is not very indicative
numerically. Consequently, we select the best thresholds
using a development test set.
Finally, we apply the extended SA system on the test set.
The results are shown in Table 3.
Learning Experiment
As can be seen from the table, improvements in accuracy
are achieved for all of the seed SA-s using any of the
feature sets. Somewhat surprisingly, FinancialSA shows
the most improvement, even though it is the best of the
systems from the beginning. This is probably due to the
fact that it produces significantly better initial training sets.
For GenericSA and FinancialSA, as expected, the Full
feature set produces significantly better results than using
either Price or SA-based feature sets separately.
Unexpectedly, for DictionarySA, all three feature sets
produce very similar results. The reason for the difference
is unclear, and is under investigation.
 Results of the Learning Experiment
Dictionary Generic SA Financial SA

Prec Recall F1 Prec Recall F1 Prec Recall F1

0.580 0.393 0.468 0.628 0.239 0.346 0.710 0.427 0.533
Baseline
0.599 0.427 0.498 0.635 0.252 0.361 0.697 0.443 0.541
Price

0.588 0.423 0.492 0.643 0.254 0.364 0.699 0.443 0.542

SA-based
0.589
0.428 0.495 0.635 0.262 0.370 0.700 0.518 0.595
(+0.9
(+3.5%) (+2.7%) (+0.7%) (+2.3%) (+2.4%) (-1.0%) (+9.1%) (+6.2%)
Full %)
Conclusions and Future Work
We experimented with several SA systems and different
external features to identify domain-specific expressions in
financial messages on social media. We proposed a novel
unsupervised multi-view-based approach that uses a seed SA
system together with domain-specific external information in
order to mine a large corpus of messages for domain-specific
sentiment expressions, which, in turn, can extend the SA to
improve its accuracy.
We contribute to the body of knowledge on sentiment analysis,
in general, and on sentiment analysis of financial social media
messages, in particular. The proposed unsupervised
methodological approach to sentimental analysis, which uses
multiple views, may be adapted to other studies from other
domains requiring the solving of sophisticated sentiment
analysis problems.
Conclusions and Future Work
Our experimental results indicate that our method is
successful in integrating diverse sources of external
information for the learning purposes. The sources we
compare are stock prices on the one hand, and SA results
on messages related by subject or by author, on the other.
Our results show that when combined in our approach,
the sources produce much better accuracy than
individually, at least for the best-performing SA systems.
)XWXUHVWXGLHVPD\LQFRUSRUDWHIXUWKHUH[WHUQDOYLHZV
features, such as events known to have a positive or
negative effect on companies, or may address both the
time reference mentioned in the messages and the price
movement in a corresponding period in order to generate
the domain-specific lexicon.
 236

Content Profiling

Social Media Mining

 237

Concept Mining
Understand the negative and positive concepts
that consumers associate with top shows in their
tweets, Facebook and Google+ updates
Visualize and track the trending concepts
associated with each show over time
 238

Concept Sentiment
A&E S how Related Messages

Concept
Identification

Sentiment
Processing
Negative Concept Positive Concept
Associations Associations
 239

Positive Expression Categories A&E Shows

Great Show Emotional Connection

Beyond Scared Straight

Hoarders

The First 48

Longmire

Intervention

Dog the Bounty Hunter

Storage Wars

Criminal Minds

Bates Motel

Duck Dynasty

0 500 1000 1500 2000 2500 3000 3500 4000

 240

Positive Concept Categories

Opinions/Feedback
161
Emotional
What I love about Connection
it
1,990 2,556

Great Show
13,889
 22.8%

19.0%

17.0%

12.9%

10.9%

4.9%

4.1%

2.9%

1.2%

0.9%

0.6%

0.5%
DYNASTY

0.5%

0.5%

0.4%

0.3%
WHAT I LOVE ABOUT DUCK

0.2%

0.2%

0.2%
241

0.2%

0.1%
 242

Emotional Connection Great Show

Devotion Great Show
1.3% 4.9%

Makes
Wednesdays
Better Funniest Show
9.8% 10.4%

Makes Me Happy Favorite Show

Makes Things Better 41.2% 10.6%
18.5% I Love It
54.2%

Best Show on TV
Opinions/Feedback 19.9%
The Perfect Date Beards are back
Love Reruns
29.3% 2.5%
3.1%
2 new episodes
4.3%

Pay talent
more
9.3%
Great video
43.5%
Scripted
12.4%

Darius Rucker
24.8%
 243

RESEARCH INSIGHTS
 244

How We Can Impact Research

Provide real time information regarding your
target audience
Identify issues before longer term research
can be fielded and reported
Opportunity to utilize social media chatter to
establish the drivers of popularitynot just
WKDWWKHUHDUHFRQYHUVDWLRQVEXWZKDWVEHLQJ
talked about
Provide overall chatter to inspire what comes
next topics
 245

Research Insights
Social Media Popularity Tracking
Track daily by show the total mentions, positive and
negative across Facebook, Google+ and Twitter
8QGHUVWDQGHDFKVKRZVGDLO\Pos/Neg ratio the true measure of a
SURJUDPVUHVRQDQFHZLWK\RXUDXGLHQFH

Crossover Profiling
Understand the off-network shows that consumers
mention most frequently with your shows across social
media
Understand both positive and negative frequencies
 246

Social Media Mining

Popularity Tracking
 247

The Social Media Popularity Ratio

The Popularity Ratio (Positive mentions/Negative mentions) is a much better
LQGLFDWRURIDVKRZVSRSXODULW\WKDQEX]]DORQHWRWDOPHQWLRQV
Popularit
Sorted Mentions Positives Negatives Neutral Pos% Neg% y
Shipping Wars 102 20 2 80 19.6% 2.0% 10.0
Southie Rules 69 32 6 31 46.4% 8.7% 5.3
Duck Dynasty 7972 1827 567 5578 22.9% 7.1% 3.2
Bates Motel 6578 1089 489 5000 16.6% 7.4% 2.2
Criminal Minds 4688 1063 492 3133 22.7% 10.5% 2.2
Storage Wars 1677 335 219 1123 20.0% 13.1% 1.5
Hoarders 1402 142 96 1164 10.1% 6.8% 1.5
Cold Case Files 38 8 6 24 21.1% 15.8% 1.3
Intervention 1775 285 214 1276 16.1% 12.1% 1.3
Barter Kings 72 15 12 45 20.8% 16.7% 1.3
Be the Boss 158 23 19 116 14.6% 12.0% 1.2
Beyond Scared Straight 470 57 56 357 12.1% 11.9% 1.0
American Hoggers 11 2 2 7 18.2% 18.2% 1.0
The First 48 651 80 96 475 12.3% 14.7% 0.8
Dog the Bounty Hunter 362 25 43 294 6.9% 11.9% 0.6
Longmire 156 26 62 68 16.7% 39.7% 0.4
Storage Wars: Texas 249 31 123 95 12.4% 49.4% 0.3
The Glades 45 1 11 33 2.2% 24.4% 0.1
POPULARITY TRACKING 248

%LJJHVW)DQVDQG%LJJHVW
Identify the authors who make the most comments about
your shows: both positive and negative.
Total PositiveNegative
Most Mentions (by single author) 98 24 94

Author Stats Show

Most Active Authors BBallbags Storage Wars: Texas
einberg_kristi Criminal Minds
nellynichole Criminal Minds
Most Negative Authors BBallbags Storage Wars: Texas
BBallbags Storage Wars
fl0k_r0ck Storage Wars: New York
Most Positive Authors supremestream Southie Rules
Bobby6740 Storage Wars
supremestream Bates Motel