Practical Sentiment Analysis

Practical Sentiment Analysis Tutorial

Jason Baldridge @jasonbaldridge

Sentiment Analysis Symposium 2014

Associate Professor Co-founder & Chief Scientist

Wednesday, March 5, 14

About the presenter

Associate Professor, Linguistics Department, The University of Texas at Austin (2005-present)

Ph.D., Informatics, The University of Edinburgh, 2002

MA (Linguistics), MSE (Computer Science), The University of Pennsylvania, 1998

Co-founder & Chief Scientist, People Pattern (2013-present)

Built Converseon’s Convey text analytics engine, with Philip Resnik and Converseon programmers.

Why NLP is hardSentiment analysis overview

Document classificationbreak

Aspect-based sentiment analysisVisualization

Semi-supervised learningbreak

Stylistics & author modelingBeyond text

Wrap up

Document classificationAspect-based sentiment analysis

VisualizationSemi-supervised learning

Wrap up

Text is pervasive = big opportunities

Texts as bags of words (with apologies) (http://www.wordle.net/)

http://www.wired.com/magazine/2010/12/ff_ai_essay_airevolution/

That is of course not the full story...

Texts are not just bags-of-words.

Order and syntax affect interpretation of utterances.

manthe

legonmanthe dogbit thethe

legonmanthe dogbit thethe mandog

Subject

Object

Modifier

Subject

Object

Location

What does this sentence mean?

I saw her duck with a telescope.

Slide by Lillian Lee

[http://casablancapa.blogspot.com/2010/05/fore.htm]l

[http://casablancapa.blogspot.com/2010/05/fore.htm]l [http://www.supercoloring.com/pages/duck-outline/]

verb noun

[http://www.clker.com/clipart-green-eyes-3.html]

verb noun

[http://www.clker.com/clipart-3163.html]

verb noun

[http://www.simonpalfrader.com/category/tournament-poker]

verb noun

Ambiguity is pervasive

the a are of I [Steve Abney]

Ambiguity is pervasive

the a are of I [Steve Abney]

an “are” (100 m2)

another “are”

And it goes further...

Rhetorical structure affects the interpretation of the text as a whole.

Max fell. John pushed him.

Max fell. John pushed him.(Because)

Explanation

Max fell. John pushed him.(Then)

Continuation

Explanation

Continuation

pushing precedes

falling

Explanation

Continuation

pushing precedes

falling

falling precedes pushing

What’s hard about this story? [Slide from Jason Eisner]

John stopped at the donut store on his way home from work. He thought a coffee was good every few hours.

But it turned out to be too expensive there.

To get a spare tire (donut) for his car?

store where donuts shop? or is run by donuts?

or looks like a big donut? or made of donut?

or has an emptiness at its core?

I stopped smoking freshman year, but

Describes where the store is? Or when he stopped?

Well, actually, he stopped there from hunger and exhaustion, not just from work.

At that moment, or habitually?

That’s how often he thought it?

But actually, a coffee only stays good for about 10

minutes before it gets cold.

Similarly: In America a woman has a baby every 15 minutes. Our job is to

find that woman and stop her.

the particular coffee that was good every few hours? the donut store? the situation?

too expensive for what? what are we supposed to conclude about

what John did?

how do we connect “it” to “expensive”?

NLP has come a long way

Sentiment analysis overview

Why NLP is hard

Wrap up

Sentiment analysis: background [slide from Lillian Lee]

People search for and are affected by online opinions.

TripAdvisor, Rotten Tomatoes, Yelp, Amazon, eBay, YouTube, blogs, Q&A and discussion sites

According to a Comscore ’07 report and an ’08 Pew survey:

60% of US residents have done online product research, and 15% do so on a typical day.

73%-87% of US readers of online reviews of services say the reviews were significant influences. (more on economics later)

But, 58% of US internet users report that online information was missing, impossible to find, confusing, and/or overwhelming.

Creating technologies that find and analyze reviews would answer a tremendous information need.

Broader implications: economics [slide from Lillian Lee]

Consumers report being willing to pay from 20% to 99% more for a 5-star-rated item than a 4-star-rated item. [comScore]

But, does the polarity and/or volume of reviews have measurable, significant influence on actual consumer purchasing?

Implications for bang-for-the-buck, manipulation, etc.

Social media analytics: acting on sentiment

Richard Lawrence, Prem Melville, Claudia Perlich, Vikas Sindhwani, Estepan Meliksetian et al.In ORMS Today, Volume 37, Number 1, February, 2010.

Polarity classification [slide from Lillian Lee]

Consider just classifying an avowedly subjective text unit as either positive or negative (“thumbs up or “thumbs down”).

One application: review summarization.

Elvis Mitchell, May 12, 2000: It may be a bit early to make such judgments, but Battlefield Earth may well turn out to be the worst movie of this century.

Can’t we just look for words like “great”, “terrible”, “worst”?

Yes, but ... learning a sufficient set of such words or phrases is an active challenge.

Using a lexicon [slide from Lillian Lee]

From a small scale human study:

Proposed word lists Accuracy

Subject 1

Positive: dazzling, brilliant, phenomenal, excellent, fantasticNegative: suck, terrible, awful, unwatchable, hideous 58%

Subject 2

Positive: gripping, mesmerizing, riveting, spectacular, cool, awesome, thrilling, badass, excellent, moving, exciting Negative: bad, cliched, sucks, boring, stupid, slow

Automatically determined (from data)

Positive: love, wonderful, best, great, superb, beautiful, still Negative: bad, worst, stupid, waste, boring, ?, !

Polarity words are not enough [slide from Lillian Lee]

Can’t we just look for words like “great” or “terrible”?

Yes, but ...

This laptop is a great deal.

A great deal of media attention surrounded the release of the new laptop.

This laptop is a great deal ... and I’ve got a nice bridge you might be interested in.

Polarity words are not enough

Polarity flippers: some words change positive expressions into negative ones and vice versa.

Negation: America still needs to be focused on job creation. Not among Obama's great accomplishments since coming to office !! [From a tweet in 2010]

Contrastive discourse connectives: I used to HATE it. But this stuff is yummmmmy :) [From a tweet in 2011 -- the tweeter had already bolded “HATE” and “But”!]

Multiword expressions: other words in context can make a negative word positive:

That movie was shit. [negative]

That movie was the shit. [positive] (American slang from the 1990’s)

More subtle sentiment (from Pang and Lee)

With many texts, no ostensibly negative words occur, yet they indicate strong negative polarity.

“If you are reading this because it is your darling fragrance, please wear it at home exclusively, and tape the windows shut.” (review by Luca Turin and Tania Sanchez of the Givenchy perfume Amarige, in Perfumes: The Guide, Viking 2008.)

“She runs the gamut of emotions from A to B.” (Dorothy Parker, speaking about Katharine Hepburn.)

“Jane Austen’s books madden me so that I can’t conceal my frenzy from the reader. Every time I read ‘Pride and Prejudice’ I want to dig her up and beat her over the skull with her own shin-bone.” (Mark Twain.)

Thwarted expectations (from Pang and Lee)

This film should be brilliant. It sounds like a great plot, the actors are first grade, and the supporting cast is good as well, and Stallone is attempting to deliver a good performance. However, it can’t hold up.

There are also highly negative texts that use lots of positive words, but ultimately are reversed by the final sentence. For example

This is referred to as a thwarted expectations narrative because in the final sentence the author sets up a deliberate contrast to the preceding discourse, giving it more impact.

Polarity classification: it’s more than positive and negative

Positive: “As a used vehicle, the Ford Focus represents a solid pick.”

Negative: “Still, the Focus' interior doesn't quite measure up to those offered by some of its competitors, both in terms of materials quality and design aesthetic.”

Neutral: “The Ford Focus has been Ford's entry-level car since the start of the new millennium.”

Mixed: “The current Focus has much to offer in the area of value, if not refinement.”

http://www.edmunds.com/ford/focus/review.html

Other dimensions of sentiment analysis

Subjectivity: is an opinion even being expressed? Many statements are simply factual.

Target: what exactly is an opinion being expressed about?

Important for aggregating interesting and meaningful statistics about sentiment.

Also, it affects how the language use indicates polarity: e.g, unpredictable is usually positive for movie reviews, but is very negative for a car’s steering

Ratings: rather than a binary decision, it is often of interest to provide or interpret predictions about sentiment on a scale, such as a 5-star system.

Other dimensions of sentiment analysis

Perspective: an opinion can be positive or negative depending on who is saying it

entry-level could be good or bad for different people

it also affects how an author describes a topic: e.g. pro-choice vs pro-life, affordable health care vs obamacare.

Authority: was the text written by someone whose opinion matters more than others?

it is more important to identify and address negative sentiment expressed by a popular blogger than a one-off commenter or supplier of a product reviewer on a sales site

follower graphs (where applicable) are very useful in this regard

Spam: is the text even valid or at least something of interest?

many tweets and blog post comments are just spammers trying to drive traffic to their sites

Document Classification

Aspect-based sentiment analysisVisualization

Semi-supervised learningStylistics & author modeling

Beyond textWrap up

Text analysis, in brief

f( , ,... ) = [ , ,... ]

Sentiment labelsParts-of-speechNamed EntitiesTopic assignmentsGeo-coordinatesSyntactic structuresTranslations

f( , ,... ) = [ , ,... ]

Sentiment labelsParts-of-speechNamed EntitiesTopic assignmentsGeo-coordinatesSyntactic structuresTranslations

RulesAnnotation & Learning - annotated examples - annotated knowledge - interactive annotation and learningScalable human annotation

Document classification: automatically label some text

Language identification: determine the language that a text is written in

Spam filtering: label emails, tweets, blog comments as spam (undesired) or ham (desired)

Routing: label emails to an organization based on which department should respond to them (e.g. complaints, tech support, order status)

Sentiment analysis: label some text as being positive or negative (polarity classification)

Georeferencing: identify the location (latitude and longitude) associated with a text

Desiderata for text analysis function f

task is well-defined

outputs are meaningful

precision, recall, etc. are measurable and sufficient for desired use

Performant

Desiderata for text analysis function f

task is well-defined

outputs are meaningful

precision, recall, etc. are measurable and sufficient for desired use

Performant

affordable access to annotated examples and/or knowledge sources

able to exploit indirect or noisy annotations

access to unlabeled examples and ability to exploit them

tools to learn f are available or can be built within budget

Reasonable cost (time & money)

Four sentiment datasets

Dataset Topic Year # Train # Dev #Test Reference

Debate08Obama vs McCain debate

2008 795 795 795Shamma, et al. (2009) "Tweet the Debates: Understanding Community Annotation of

Uncollected Sources."

HCRHealth care

reform2010 839 838 839

Speriosu et al. (2011) "Twitter Polarity Classification with

Label Propagation over Lexical Links and the Follower Graph."

STS(Stanford)

Twitter Sentiment

2009 - 216 -Go et al. (2009) "Twitter

sentiment classification using distant supervision"

IMDBIMDB movie

reviews2011 25,000 25,000 -

Mas et al. (2011) "Learning Word Vectors for Sentiment

Analysis"

Rule-based classification

Identify words and patterns that are indicative of positive or negative sentiment:

polarity words: e.g. good, great, love; bad, terrible, hate

polarity ngrams: the shit (+), must buy (+), could care less (-)

casing: uppercase often indicates subjectivity

punctuation: lots of ! and ? indicates subjectivity (often negative)

emoticons: smiles like :) are generally positive, while frowns like :( are generally negative

Use each pattern as a rule; if present in the text, the rule indicates whether the text is positive or negative.

How to deal with conflicts? (E.g. multiple rules apply, but indicate both positive and negative?)

Simple: count number of matching rules and take the max.

Simplest polarity classifier ever?

def polarity(document) = if (document contains “good”) positive else if (document contains “bad”) negative else neutral

Debate08 HCR STS IMDB

20.5 21.6 19.4 27.4

No better than flipping a (three-way) coin?

Code and data here: https://github.com/utcompling/sastut

The confusion matrix

We need to look at the confusion matrix and breakdowns for each label. For example, here it is for Debate08:

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

+ is positive, - is negative, ~ is neutral

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

Machine predictions

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Total count of documents inthe corpus

Corpus labels

Machine predictions

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

Machine predictions

Correct predictions

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

Machine predictions

Correct predictions

(5+140+18)/795 = 0.205

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

Machine predictions

Correct predictions

Incorrect predictions

(5+140+18)/795 = 0.205

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

Machine predictions

Column showing outcomes of documents labeled negative by the machine

Correct predictions

(5+140+18)/795 = 0.205

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

Corpus labels

Machine predictions

Row showing outcomes of documents labeled negative in the corpus

Column showing outcomes of documents labeled negative by the machine

Correct predictions

(5+140+18)/795 = 0.205

Precision, Recall, and F-score: per category scores

Precision: the number of correct guesses (true positives) for the category divided by all guesses for it (true positives and false positives)

Recall: the number of correct guesses (true positives) for the category divided by all the true documents in that category (true positives plus false negatives)

F-score: derived measure combining precision and recall.

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

P R F-~+

83.3 1.1 2.2

18.3 99.3 30.1

72.0 9.0 16.0

57.9 36.5 16.4

P = TP/(TP+FP)

R = TP/(TP+FN)

F = 2PR/(P+R)

P~ = 140+442+182140 = .183

R- = 5+442+75 = .011

F+ = .72+.092 × .72 × .09 = .16

What does it tell us?

Overall accuracy is low, because the model overpredicts neutral.

Precision is pretty good for negative, and okay for positive. This means the simple rules “has the word ‘good’” and “has the word ‘bad’” are good predictors.

- ~ +-~+

5 442 7 454

1 140 0 141

0 182 18 200

6 764 25 795

P R F-~+

83.3 1.1 2.2

18.3 99.3 30.1

72.0 9.0 16.0

57.9 36.5 16.4

Where do the rules go wrong?

Confusion matrix for STS:

The one negative-labeled tweet that is actually positive, using the very positive expression “bad ass” (thus matching “bad”).

Booz Allen Hamilton has a bad ass homegrown social collaboration platform. Way cool! #ttiv

- ~ +-~+

0 73 2 75

0 31 2 33

1 96 11 108

1 200 15 216

A bigger lexicon (rule set) and a better rule

Good improvements for five minutes of effort!

Why such a large improvement for IMDB?37

pos_words = {"good","awesome","great","fantastic","wonderful"}neg_words = {"bad","terrible","worst","sucks","awful","dumb"}def polarity(document) =num_pos = count of words in document also in pos_wordsnum_neg = count of words in document also in neg_words

if (num_pos == 0 and num_neg == 0) neutral else if (num_pos > num_neg) positive else negative

Super simple

Small lexicon

20.5 21.6 19.4 27.4

21.5 22.1 25.5 51.4

IMDB: no neutrals!

Data is from 10 star movie ratings (>=7 are pos, <= 4 are neg)

Compare the confusion matrices!

“Good/Bad” rule Small lexicon with counting rule

2324 5476 4700 12500

0 0 0 0

651 7325 4524 12500

2975 12801 9224 25000

Accuracy: 27.4

5744 3316 3440 12500

0 0 0 0

1147 4247 7106 12500

6891 7563 10546 25000

Accuracy: 51.4

Sentiment lexicons: Bing Liu’s opinion lexicon

Bing Liu maintains and freely distributes a sentiment lexicon consisting of lists of strings.

Distribution page (direct link to rar archive)

Positive words: 2006

Negative words: 4783

Useful properties: includes mis-spellings, morphological variants, slang, and social-media mark-up

Note: may be used for academic and commercial purposes.

39Slide by Chris Potts

Sentiment lexicons: MPQA

The MPQA (Multi-Perspective Question Answering) Subjectivity Lexicon is maintained by Theresa Wilson, Janyce Wiebe, and Paul Hoffmann (Wiebe, Wilson, and Cardie 2005).

Note: distributed under a GNU Public License (not suitable for most commercial uses).

Other sentiment lexicons

SentiWordNet (Baccianella, Esuli, and Sebastiani 2010) attaches positive and negative real-valued sentiment scores to WordNet synsets (Fellbaum1998).

Note: recently changed license to permissive, commercial-friendly terms.

Harvard General Inquirer is a lexicon attaching syntactic, semantic, and pragmatic information to part-of-speech tagged words (Stone, Dunphry, Smith, and Ogilvie 1966).

Linguistic Inquiry and Word Counts (LIWC) is a proprietary database consisting of a lot of categorized regular expressions. Its classifications are highly correlated with those of the Harvard General Inquirer.

When you have a big lexicon, use it!

Super simple

Small lexicon

Opinion lexicon

20.5 21.6 19.4 27.4

21.5 22.1 25.5 51.4

47.8 42.3 62.0 73.6

Using Bing Liu’s Opinion Lexicon, scores across all datasets go up dramatically.

Well above (three-way) coin-flipping!

If you don’t have a big lexicon, bootstrap one

There is a reasonably large literature on creating sentiment lexicons, using various sources such as WordNet (knowledge source) and review data (domain-specific data source).

Advantage of review data: often able to obtain easily for many languages.

See Chris Potts’ 2011 SAS tutorial for more details:

http://sentiment.christopherpotts.net/lexicons.html

A simple, intuitive measure is the log-likelihood ratio, which I’ll show for IMDB data.

Log-likelihood ratio: basic recipe

Given: a corpus of positive texts, negative texts, and a held out corpus

For each word in the vocabulary, calculate its probability in each corpus. E.g. for the positive corpus:

Compute its log-liklihood ratio for positive vs negative documents:

Rank all words from highest LLR to lowest.

LLR examples computed from IMDB reviews

edie 16.069394855429000

antwone 15.85538381213240

din 15.747494864581600

goldsworthy 15.552434312463400

gunga 15.536930128612500

kornbluth -15.090106131301700

kareena -15.11542393212590

tashan -15.233206936755000

hobgoblins -15.233206936755000

slater -15.318364724832600

edie 16.069394855429000

antwone 15.85538381213240

din 15.747494864581600

goldsworthy 15.552434312463400

gunga 15.536930128612500

kornbluth -15.090106131301700

kareena -15.11542393212590

tashan -15.233206936755000

hobgoblins -15.233206936755000

slater -15.318364724832600

Filter:

edie 16.069394855429000

antwone 15.85538381213240

din 15.747494864581600

goldsworthy 15.552434312463400

gunga 15.536930128612500

kornbluth -15.090106131301700

kareena -15.11542393212590

tashan -15.233206936755000

hobgoblins -15.233206936755000

slater -15.318364724832600

perfection 2.204227744897310

captures 2.0551924704260400

wonderfully 2.020824971323010

powell 1.9933170865620900

refreshing 1.867299924519800

pointless -2.477406360270270

blah -2.57814744950696

waste -2.673668672544840

unfunny -2.7084876042405500

seagal -3.6618321047833000

Filter:

Top 25 filtered positive and negative words using LLR on IMDB

perfection captures wonderfully powell refreshing flynn delightful gripping beautifully underrated superb delight welles unforgettable touching favorites extraordinary stewart brilliantly friendship wonderful magnificent finest marie jackie

horrible unconvincing uninteresting insult uninspired sucks miserably boredom cannibal godzilla lame wasting remotely awful poorly laughable worst lousy redeeming atrocious pointless blah waste unfunny seagal

Some obvious film domain dependence, but also lots of generally good valence determinations.

Using the learned lexicon

There are various ways to use the LLR ranks:

Take the top N of positive and negative and use them as the positive and negative sets.

Combine the top N with another lexicon (e.g. the super small one or the Opinion Lexicon).

Take the top N and manually prune words that are not generally applicable.

Use the LLR values as the input to a more complex (and presumably more capable) algorithm.

Here we’ll try three things:

IMDB100: the top 100 positive and 100 negative filtered words

IMDB1000: the top 1000 positive and 1000 negative filtered words

Opinion Lexicon + IMDB1000: take the union of positive terms in Opinion Lexicon and IMDB1000, and same for the negative terms.

Better lexicons can get pretty big improvements!

Super simple

Small lexicon

Opinion lexicon

IMDB100

IMDB1000

Opinion Lexicon + IMDB1000

20.5 21.6 19.4 27.4

21.5 22.1 25.5 51.4

47.8 42.3 62.0 73.6

24.1 22.6 35.7 77.9

58.0 45.6 50.5 66.0

62.4 49.1 56.0 66.1

Better lexicons can get pretty big improvements!

Super simple

Small lexicon

Opinion lexicon

IMDB100

IMDB1000

20.5 21.6 19.4 27.4

21.5 22.1 25.5 51.4

47.8 42.3 62.0 73.6

24.1 22.6 35.7 77.9

58.0 45.6 50.5 66.0

62.4 49.1 56.0 66.1

Nonetheless: for the reasons mentioned previously, this strategy eventually runs out of steam. It is a starting point.

Machine learning for classification

The rule-based approach requires defining a set of ad hoc rules and explicitly managing their interaction.

If we instead have lots of examples of texts of different categories, we can learn a function that maps new texts to one category or the other.

What were rules become features that are extracted from the input; their importance is extracted from statistics in a labeled training set.

These features are dimensions; their values for a given text plot it into space.

Idea: software learns from examples it has seen.

Find the boundary between different classes of things, such as spam versus not-spam emails.

HamSpam

Given a set of labeled points, there are many standard methods for learning linear classifiers. Some popular ones are:

Naive Bayes

Logistic Regression / Maximum Entropy

Perceptrons

Support Vector Machines (SVMs)

The properties of these classifier types are widely covered in tutorials, code, and homework problems.

There are various reasons to prefer one or the other of these, depending on amount of training material, tolerance for longer training times, and the complexity of features used.

Features for document classification

All of the linear classifiers require documents to be represented as points in some n-dimensional space.

Each dimension corresponds to a feature, or observation about a subpart of a document.

A feature’s value is typically the number of times it occurs.

Ex: Consider the document “That new 300 movie looks sooo friggin bad ass. Totally BAD ASS!” The feature “the lowercase form of the word ‘bad’” has a value of 2, and the feature “is_negative_word” would be 4 (“bad”,“ass”,“BAD”,“ASS”).

For many documentation classification tasks (e.g. spam classification), bag-of-words features are unreasonably effective.

However, for more subtle tasks, including polarity classification, we usually employ more interesting features.

Features for classification

That new 300 movie looks sooo friggin BAD ASS .

w=thatw=neww=300w=moview=looksw=sooow=frigginw=badw=ass

bi=<START>_that bi=that_newbi=new_300 bi=300_moviebi=movie_looksbi=looks_sooobi=sooo_frigginbi=friggin_badbi=bad_assbi=ass_.bi=._<END>

art adj noun noun verb adv adv adj noun punc

wt=that_artwt=new_adjwt=300_nounwt=movie_nounwt=looks_verbwt=sooo_advwt=friggin_advwt=bad_adjwt=ass_noun

subtree=S_NP_movie-S_VP_looks-S_VP_NP_bad_ass

subtree=NP_sooo_bad_ass

FEATUREENGINEERING...

(deep learning might help

ease the burden)

Complexity of features

Features can be defined on very deep aspects of the linguistic content, including syntactic and rhetorical structure.

The models for these can be quite complex, and often require significant training material to learn them, which means it is harder to employ them for languages without such resources.

I’ll show an example for part-of-speech tagging in a bit.

Also: the more fine-grained the feature, the more likely it is rare to see in one’s training corpus. This requires more training data, or effective semi-supervised learning methods.

Recall the four sentiment datasets

Dataset Topic Year # Train # Dev #Test Reference

Debate08Obama vs McCain debate

2008 795 795 795Shamma, et al. (2009) "Tweet the Debates: Understanding Community Annotation of

Uncollected Sources."

HCRHealth care

reform2010 839 838 839

Speriosu et al. (2011) "Twitter Polarity Classification with

Label Propagation over Lexical Links and the Follower Graph."

STS(Stanford)

Twitter Sentiment

2009 - 216 -Go et al. (2009) "Twitter

sentiment classification using distant supervision"

IMDBIMDB movie

reviews2011 25,000 25,000 -

Mas et al. (2011) "Learning Word Vectors for Sentiment

Analysis"

Logistic regression, in domain

Logistic Regressionw/ bag-of-words

Logistic Regressionw/ extended

features

62.4 49.1 56.0 66.1

60.9 56.0(no labeled training set) 86.7

70.2 60.5 -

When training on labeled documents from the same corpus.

Models trained with Liblinear (via ScalaNLP Nak)

Note: for IMDB, the logistic regression classifier only predicts positive or negative (because there are no neutral training examples), effectively making it easier than for the lexicon-based method.

Logistic regression (using extended features), cross-domain

Debate08 HCR STS

Debate08

Debate08+HCR

70.2 51.3 56.5

56.4 60.5 54.2

70.3 61.2 59.7Trai

Evaluation corpora

In domain training examples add 10-15% absolutely accuracy (56.4 -> 70.2 for Debate08, and 51.3 -> 60.5 for HRC).

More labeled examples almost always help, especially if you have no in-domain training data (e.g. 56.5/54.2 -> 59.7 for STS).

Accuracy isn’t enough, part 1

The class balance can shift considerably without affecting the accuracy!

58+24+47216

= 59.7

D08+HRC on STS- ~ +

58 12 5 75

7 24 2 33

34 27 47 108

99 63 54 216

8+15+106216

= 59.7

(Made up) Positive-heavy classifier

- ~ +-~+

8 12 55 75

7 24 11 33

1 1 106 108

16 28 172 216

Need to also consider the per-category precision, recall, and f-score.

- ~ +-~+

58 12 5 75

7 24 2 33

34 27 47 108

99 63 54 216

P R F-~+

58.6 77.3 66.7

38.1 72.7 50.0

87.0 43.5 58.0

61.2 64.5 58.2

Acc: 59.7 Big differences in precision for the three categories!

Errors on neutrals are typically less grievous than positive/negative errors, yet raw accuracy makes one pay the same penalty.

D08+HRC on STS

One solution: allow varying penalties such that no points are awarded for positive/negative errors, but some partial credit is given for positive/neutral and negative/neutral ones.

- ~ +-~+

58 12 5 75

7 24 2 33

34 27 47 108

99 63 54 216

Who says the gold standard is correct? There is often significant variation among human annotators, especially for positive vs neutral and negative vs neutral.

Solution one: work on your annotations (including creating conventions) until you get very high inter-annotator agreement.

This arguably reduces the linguistic variability/subtlety characterized in the annotations.

Also, humans often fail to get the intended sentiment, e.g. sarcasm.

Solution two: measure performance differently. For example, given a set of examples annotated by three or more human annotators and the machine, is the machine distinguishable from the humans in terms of the amount it disagrees with their annotations?

Often, what is of interest is an aggregate sentiment for some topic or target. E.g. given a corpus of tweets about cars, 80% of the mentions of the Ford Focus are positive while 70% of the mentions of the Chevy Malibu are positive.

Note: you can get the sentiment value wrong for some of the documents while still getting the overall, aggregate sentiment correct (as errors can cancel each other).

Note also: generally, this requires aspect-based analysis (more later).

Caveat emptor, part 1

In measuring accuracy, the methodology can vary dramatically from vendor to vendor, at times in unclear ways.

For example, some seem to measure accuracy by presenting a human judge with examples annotated by a machine. The human then marks which examples they believe were incorrect. Accuracy is then num_correct/num_examples.

Problem: people get lazy and often end up giving the machine the benefit of the doubt.

I have even heard that some vendors take their high-confidence examples and do the above exercise. This is basically cheating: high-confidence machine label assignments are on average more correct than low-confidence ones.

Performance on in-domain data is nearly always better than out-of-domain (see the previous experiments).

The nature of the world is that the language of today is a step away from the language of yesterday (when you developed your algorithm or trained your model).

Also, because there are so many things to talk about (and because people talk about everything), a given model is usually going to end up employed in domains it never saw in its training data.

With nice, controlled datasets like those given previously, the experimenter has total control over which documents her algorithm is applied too.

However, a deployed system will likely confront many irrelevant documents, e.g.

documents written in other languages

Sprint the company wants tweets by their customers, but also get many tweets of people talking about the activity of sprinting.

documents that match, but which are not about the target of interest

documents that should have matched, but were missed in retrieval

Thus, identification of relevant documents and even sub-documents with relevant targets, is an important component of end-to-end sentiment solutions.

Aspect-based sentiment analysis

Document classification

Wrap up

Is it coherent to ask what the sentiment of a document is?

Documents tend to discuss many entities and ideas, and they can express varying opinions, even toward the same entity.

This is true even in tweets, e.g.

positive towards the HCR bill

negative towards Mitch McConnell

Here's a #hcr proposal short enough for Mitch McConnell to read: pass the damn bill now

Fine-grained sentiment

Two products, iPhone and Blackberry

Overall positive to iPhone, negative to Blackberry

Postive aspect/features of iPhone: touch screen, voice quality. Negative (for the mother): expensive.

68Slide adapted from Bing Liu

Components of fine-grained analysis

Opinion targets: entities and their features/aspects

Sentiment orientations: positive, negative, or neutral

Opinion holders: persons holding the opinions

Time: when the opinions are expressed

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.

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

iPhone

screen battery

{cost,size,appearance,...}

{battery_life,size,...}{...} ...

Slide adapted from Bing Liu

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

An opinion is a quintuple (e,a,so,h,t) where:

e is a target entity.

a is an aspect/feature of the entity e.

so is the sentiment value of the opinion from the opinion holder h on feature a of entity e at time t. so is positive, negative or neutral (or more granular ratings).

h is an opinion holder.

t is the time when the opinion is expressed.

Examples from the previous passage:

(iPhone, GENERAL, +, Abc123, 5-1-2008) (iPhone, touch_screen, +, Abc123, 5-1-2008) (iPhone, cost, -, mother_of(Abc123), 5-1-2008)

Slide adapted from Bing Liu

The goal: turn unstructured text into structured opinions

Given an opinionated document (or set of documents)

discover all quintuples (e,a, so, h, t)

or solve a simpler form of it, such as the document level task considered earlier

Having extracted the quintuples, we can feed them into traditional visualization and analysis tools.

Several sub-problems

e is a target entity: Named Entity Extraction (more)

a is an aspect of e: Information Extraction

so is sentiment: Sentiment Identification

h is an opinion holder: Information/Data Extraction

t is the time: Information/Data Extraction

All of these tasks can make use of deep language processing methods, including parsing, coreference, word sense disambiguation, etc.

Named entity recognition

Given a document, identify all text spans that mention an entity (person, place, organization, or other named thing).

Requires having performed tokenization, and possibly part-of-speech tagging.

Though it is a bracketing task, it can be transformed into a sequence task using BIO labels (Begin, Inside, Outside)

Usually, discriminative sequence models like Maxent Markov Models and Conditional Random Fields are trained on such sequences, and used for prediction.

Mr. [John Smith]Person traveled to [New York City]Location to visit [ABC Corporation]Organization.

Mr. John Smith traveled to New York City to visit ABC CorporationO B-PER I-PER O O B-LOC I-LOC I-LOC O O B-ORG I-ORG

OpenNLP Pipeline demo

Sentence detection

Tokenization

Part-of-speech tagging

Chunking

NER: persons and organizations

Pierre Vinken, 61 years old, will join the board as a nonexecutive director Nov. 29. Mr. Vinken is chairman of Elsevier N.V., the Dutch publishing group. Rudolph Agnew, 55 years old and former chairman of Consolidated Gold Fields PLC, was named a director of this British industrial conglomerate.

Things are tricky in Twitterland - need domain adaptation

.@Peters4Michigan's camp tells me he'll appear w Obama in MI tomorrow. Not as scared as other Sen Dems of the prez:

.[@Peters4Michigan]PER 's camp tells me he'll appear w [Obama]PER in [MI]LOC tomorrow. Not as scared as other Sen [Dems]ORG of the prez:

Named entities referred to with @-mentions (makes things easier, but also harder for model solely trained on newswire text)

Tokenization: many new innovations, including “.@account” at begining of tweet (which blocks it being an @-reply to that account)

Abbreviations mess with features learned on standard text, e.g. “w” for with (as above), or even for George W. Bush:

And who changed that? Remember Dems, many on foreign soil, criticizing W vehemently? Speaking of rooting against a Prez ....

Identifying targets and aspects

We can specify targets, their sub-components, and their attributes:

But language is varied and evolving, so we are likely to miss many ways to refer to targets and their aspects.

E.g. A person declaring knowledge about phones might forget (or not even know) that “juice” is a way of referring to power consumption.

Also: there are many ways of referring to product lines (and their various releases, e.g. iPhone 4s) and their competitors, and we often want to identify these semi-automatically.

Much research has worked on bootstrapping these. See Bing Liu’s tutorial for an excellent overview:

http://www.cs.uic.edu/~liub/FBS/Sentiment-Analysis-tutorial-AAAI-2011.pdf

iPhone

screen battery

{cost,size,appearance,...}

{battery_life,size,...}{...} ...

Target-based feature engineering

Given a sentence like “We love how the Porsche Panamera drives, but its bulbous exterior is unfortunately ugly.”

NER to identify the “Porsche Panamera” as the target

Aspect identification to see that opinions are being expressed about the car’s driving and styling.

Sentiment analysis to identify positive sentiment toward the driving and negative toward the styling.

Targeted sentiment analysis require positional features

use string relationship to the target or aspect

or use features from a parse of the sentence (if you can get it)

In addition to the standard document-level features used previously, we build features particularized for each target.

These are just a subset of the many possible features.

Positional features

We love how the Porsche Panamera drives, but its bulbous exterior is unfortunately ugly.

Positional features

Challenges

Positional features greatly expands the space of possible features.

We need more training data to estimate parameters for such features.

Highly specific features increase the risk of overfitting to whatever training data you have.

Deep learning has a lot of potential to help with learning feature representations that are effective for the task by reducing the need for careful feature engineering.

But obviously: we need to be able to use this sort of evidence in order to do the job well via automated means.

Visualization

Semi-supervised learningStylistics & author modeling

Beyond textWrap up

Visualize, but be careful when doing so

It's often the case that a visualization can capture nuances in the data that numerical or linguistic summaries cannot easily capture.

Visualization is an art and a science in its own right. The following advice from Tufte (2001, 2006) is easy to keep in mind (if only so that your violations of it are conscious and motivated):

Draw attention to the data, not the visualization.

Use a minimum of ink.

Avoid creating graphical puzzles.

Use tables where possible.

Sentiment lexicons: SentiWordNet

Twitter Sentiment results for Netflix.

Twitrratr blends the data and summarization together

Relationships between modifiers in WordNet similar-to graph

Visualizing discussions: Wikipedia deletions [http://notabilia.net/]

Could be used as a visualization for evolving sentiment over time in a discussion among many individuals.

Semi-supervised Learning

Visualization

Wrap up

Scaling

Scaling for text analysis tasks typically requires more than big computation or big data. ➡ Most interesting tasks involve representations “below” the text itself.

Scaling

Being “big” helps when you know what you are computing and how you can compute it. ➡ GIGO, and “big” garbage is still garbage.

Scaling

Being “big” helps when you know what you are computing and how you can compute it. ➡ GIGO, and “big” garbage is still garbage.

Scaling often requires being creative about how to learn f from relatively little explicit information about the task. ➡ Semi-supervised methods and indirect supervision to the rescue.

Scaling annotations

Accurate toolExtremely low annotation

Annotation is relatively expensive

Accurate toolExtremely low annotation

Accurate toolExtremely low annotation ?

We lack sufficient resources for most languages, most domains and most problems.

Semi-supervised learning approaches become essential.➡ See Philip Resnik’s SAS 2011 keynote: http://vimeo.com/32506363

Example: Learning part-of-speech taggers

They often book flights .

The red book fell .

N Adv V N PUNC

D Adj N V PUNC

The red book fell .

N Adv V N PUNC

D Adj N V PUNC

The red book fell .

POS Taggers are usually trained on hundreds of thousands of annotated word tokens. What if we have almost nothing?

N Adv V N PUNC

D Adj N V PUNC

annotation HMM

The overall strategy: grow, shrink, learn

annotation HMMEM

annotation HMMTag Dict

Generalization EM

annotation HMMTag Dict

Generalization EM

cover the vocabulary94

annotation HMMModel

MinimizationTag Dict

Generalization EM

cover the vocabulary remove noise94

annotation HMMModel

MinimizationTag Dict

Generalization EM

cover the vocabulary remove noise train94

Extremely low annotation scenario [Garrette & Baldridge 2013]

Obtain word types or tokens annotated with their parts-of-speech by a linguist in under two hours

the Dbook N, Voften Advred Adj, NTypes: construct a tag dic.onary from scratch

(not simulated)

Tokens: standard word-‐by-‐word annota.on

They often book flights . N Adv V N PUNCThe red book fell . D Adj N V PUNC

Strategy: connect annotations to raw corpus and propagate them

Raw Corpus

Tokens

Raw Corpus

Tokens

Raw Corpus

Tokens

Raw Corpus

Tokens

Raw Corpus

Tokens

Label propagation for video recommendation, in brief

Basil Marceaux for Tennessee Governor

Jimmy Fallon: Whip My Hair

Radiohead: Paranoid Android

Pink Floyd: The Wall (Full Movie)

Pink Floyd: The Wall (Full Movie)Local updates, so scales easily!

TOK_the_1 TOK_dog_2TOK_the_4 TOK_thug_5

NEXT_walksPREV_<b> PREV_the

PRE1_tPRE2_th SUF1_g

TYPE_the TYPE_thug TYPE_dog

Tag dictionary generalization

Type Annotations________________________thedog

Token Annotations________________________

the dog walksDT NN VBZ

Token Annotations________________________

the dog walks

Result:

• a tag distribution on every token

Result:

• a tag distribution on every token

• an expanded tag dictionary (non-zero tags)

English Kinyarwanda Malagasy

EM only EM only+ Our approach + Our approach

Tokens TypesEM only EM only+ Our approach + Our approachEM only EM only+ Our approach + Our approachEM only EM only+ Our approach + Our approachEM only EM only+ Our approach + Our approach

TotalAccuracy

Results (two hours of annotation)

TotalAccuracy

With 4 hours + a bit more ➡ 90% [Garrette, Mielens, & Baldridge 2013]

Polarity classification for Twitter

Obama looks good. #tweetdebate #current+- McCain is not answering the questions #tweetdebate

Sen McCain would be a very popular President - $5000 tax refund per family! #tweetdebate+

- "it's like you can see Obama trying to remember all the "talking points" and get his slogans out there #tweetdebate"

Logistic regression... and... done!

What if instance labels aren’t there?

No explicitly labeled examples?

Positive/negative ratio using polarity lexicon.➡ Easy & works okay for many cases, but fails spectactularly elsewhere.

Emoticons as labels + logistic regression.➡ Easy, but emoticon to polarity mapping is actually vexed.

Label propagation using the above as seeds.➡ Noisy labels provide soft indicators, the graph smooths things out.

If you have annotations, you can use those too.➡ Including ordered labels like star ratings: see Talukdar & Crammer 2009

Using social interaction: Twitter sentiment

Obama is making the repubs look silly and petty

bird images from http://www.mytwitterlayout.com/

Papers: Speriosu et al. 2011; Tan et al. KDD 2011

“Obama”, “silly”, “petty”

is happy Obama is president

Obama’s doing great!

Obama’s doing great!=

(hopefully)

Twitter polarity graph with knowledge and noisy seeds

I love #NY! :)

Ahhh #Obamacare

We can’t pass this :(#killthebill

I hate #Obamacare! #killthebill

EveWe need health care!

Let’s get it passed! :)

I love #NY! :)

Ahhh #Obamacare

I love #NY! :)

Ahhh #Obamacare

I love #NY! :)

Ahhh #Obamacare

Let’s get it passed! :) Hashtags

killthebillobamacareny

Word n-gram

swe can’t

love ny

i love

I love #NY! :)

Ahhh #Obamacare

OpinionFindercare

Word n-gram

swe can’t

love ny

i love

I love #NY! :)

Ahhh #Obamacare

OpinionFindercare

Word n-gram

swe can’t

love ny

i love

Emoticons;-):(:)

I love #NY! :)

Ahhh #Obamacare

OpinionFindercare

Word n-gram

swe can’t

love ny

i love

Emoticons;-):(:)

I love #NY! :)

Ahhh #Obamacare

OpinionFindercare

Word n-gram

swe can’t

love ny

i love

Emoticons;-):(:)

I love #NY! :)

Ahhh #Obamacare

OpinionFindercare

Word n-gram

swe can’t

love ny

i love

Emoticons;-):(:)

I love #NY! :)

Ahhh #Obamacare

Results: polarity assignment (positive/negative, no neutral)

Stanford TwitterSentiment

Obama-McCainDebate

Health CareReform

Random

Lexicon Ratio

Emoticon-trained(Logistic regression)

Label propagation

50.0 50.0 50.0

72.1 59.1 58.1

83.1 61.3 62.9

84.7 66.7 71.2

Take-home message: label propagation can make effective use of labeled features (from external knowledge sources) and noisy annotations.

Let’s not forget scalable human annotationMechanical Turk: can work well, but also problematic (e.g. lack of workers for many languages).

Also real-time polling and reactions, e.g., ReactLabs.

Stylistics & author modeling

Beyond textWrap up

Content-based analysis vs stylistics

For general categorization tasks, the content words are the most helpful.

E.g. to know whether a document is about sports or finance, it helps to know it contains “baseball”, “umpire” and “game” versus “money”, “stocks”, and “bonds”.

Often we filter so-called stop-words when creating features for such tasks.

Stylistics: tasks of interest include authorship attribution, status, depression, deceit, demographics, and more.

Stylistics is different from content categorizaiton: the subtle differences in use of function words is very important.

E.g. in authorship attribution studies, content words are often filtered out. The “stop words” instead become the “keep words”!

Author comparison: match texts on left to those by same author on the right

His manner was not effusive. It seldom was; but he was glad, I think, to see me. With hardly a word spoken, but with a kindly eye, he waved me to an armchair, threw across his case of cigars, and indicated a spirit case and a gasogene in the corner. Then he stood before the fire and looked me over in his singular introspective fashion.

For all the preposterous hat and the vacuous face, there was something noble in the simple faith of our visitor which compelled our respect. She laid her little bundle of papers upon the table and went her way, with a promise to come again whenever she might be summoned.

He was invited to Kellynch Hall; he was talked of and expected all the rest of the year; but he never came. The following spring he was seen again in town, found equally agreeable, again encouraged, invited, and expected, and again he did not come; and the next tidings were that he was married.

There are many theories about what happened, but two general narratives seem to be gaining prominence, which we will call the greed narrative and the stupidity narrative. The two overlap, but they lead to different ways of thinking about where we go from here.

He was not an ill-disposed young man, unless to be rather cold hearted and rather selfish is to be ill-disposed: but he was, in general, well respected; for he conducted himself with propriety in the discharge of his ordinary duties. Had he married a more amiable woman, he might have been made still more respectable than he was.

Our moral and economic system is based on individual responsibility. It’s based on the idea that people have to live with the consequences of their decisions. This makes them more careful deciders. This means that society tends toward justice — people get what they deserve as much as possible.

Brooks, New York Times, Apr 2, 2009

Brooks, New York Times, Feb 19, 2009

Doyle, Sherlock Holmes, A Scandal in Bohemia, 1891

Brooks, New York Times, Feb 19, 2009

Doyle, Sherlock Holmes, A Case of Identity, 1891

Doyle, Sherlock Holmes, A Scandal in Bohemia, 1891

Austen, Persuasion, 1818 Brooks, New York Times, Feb 19, 2009

Austen, Sense and Sensibility, 1811

Doyle, Sherlock Holmes, A Case of Identity, 1891

Quantitative features that help discriminate the authors

Grammatical person: 1st (we/us/our, I/me/my)

Grammatical tense: present, past

Word frequencies: frequent use of “he”

Punctuation: use of colons and semi-colons

Average word and sentence length

Syntax: prepositional adverbial phrases (“With hardly...”, “For all the...”)

These must be counted in all texts. The texts of unknown authorship should then have values most similar to those of the texts of one of the known authors.

Forensic linguistics

Forensic linguistics is a branch of applied linguistics that applies linguistic theory, research and principles to real life language in the legal context.

Even more generally, it can be viewed as analyzing examples of language to discover properties that reveal more than just what is said.

authorship (same as other examples?, plagiarism)

psychological attributes of the author (deception, depression)

similarity to other examples (e.g., trademark disputes)

Linguistic fingerprinting and identification

There has been much interest in finding linguistic “fingerprints”, but there are problems with the concept:

language acquisition: language is learned and continually changing

linguistic homogeneity: education, mass media

register: the same person speaks differently in different contexts, with different people

No accepted definition of general linguistic fingerprint has so far been proposed, nor are we likely to see one for these reasons.

Nonetheless, people do exhibit regularities in their speech and writing that could distinguish them from others.

This allows us to compare a limited set of authors/speakers in certain restricted conditions, just as we did with the first page of these slides.

Identifying style

Every speaker uses language differently, leading to a unique style.

Style is both:

a collection of markers which can be observed and measured

a set of unconscious habits which can be observed and measured

Quantifying style:

word usage: presence/absence of words, relative word frequencies

type/token ratios

average word and sentence length

the number of unique words (hapax legomena)

Authorship attribution with machine learningMachine learning provides a class of algorithms that perform unsupervised clustering.

They don’t have labels for any of the data points (e.g., documents).

Based on properties measured from the data points, coherent clusters of documents with similar properties can be identified.

A cluster can correspond to many different things, including collections of documents by the same author.

Mixture models: a popular class of probabilistic algorithms for clustering.

collections of probability distributions over the data

“soft” cluster membership: points are proportionally part of multiple clusters

a mixture of Gaussian distributions (a.k.a. the normal distribution) are one of the most commonly used type of mixture model

The K-means algorithm is a related “hard” clustering algorithm that is simple and easy to understand.

Hard clustering into k groups

Assume you can measure various attributes for each data point, e.g.:

the weight and top speed of various vehicles

the average sentence length and average word length of various authors.

Next, you want to identify k groups of similar items based on these attributes.

How many groups? How to find them using an algorithm?

top speed

k=3?k=2? wei

top speed

k=3?k=2? wei

top speed

An authorship clustering problem

Texts from three authors (five documents each)

Arthur Conan Doyle (obtained from Project Gutenberg)

Jane Austen (obtained from Project Gutenberg)

Paul Krugman (obtained from New York Times website)

Measure the relative frequency of the words “I” and “the” in each document.

Document I theEmma 1.8 3.2

Mansfield 1.5 3.9Persuasion 1.3 4.0

Pride 1.7 3.6Sense 1.6 3.4

Document I theCity 2.1 4.7

Gerard 3.6 6.1Holmes 2.8 5.3Hound 2.5 5.6

Polestar 2.3 6.2

Document I the12-01-2008 0.0 6.712-07-2008 0.0 5.612-15-2008 0.3 6.712-19-2008 0.1 6.412-22-2008 0.4 6.3

Austen Doyle Krugman

Now, “forgetting” who the authors are, let’s see if they fall into distinct clusters based on these attributes.

Plot the attributes against each other

0 1 2 3 4

AustenDoyleKrugman

Relative frequency of “I”

of “

the”

K-means: intuition

We see the clusters quite clearly, but a computer doesn’t and we need to specify an algorithm that allows it to identify them.

The K-means algorithm is a simple algorithm for such tasks.

The basic idea:

the values for the attributes in each dimension will be similar for each document of the same author

each author is represented as the averages for the attributes of all the documents he or she wrote

but: we don’t know those averages since we “forgot” the authors!

so, we take a guess at the average for each author, and then see which documents each of our hypothesized authors were likely to have produced

these guesses will probably be wrong, but we can fix that by iteratively re-estimating them

K-means: algorithm

We are given N documents: D = d1, d2, ..., dN

We need to output K centroids: C = c1, c2, ..., cK

These centroids partition the documents into clusters based on which centroid is closest to each document.

K-means(D,K)C ←SelectRandomCenters(D,K)while C does change for k ← 1 to K gk ← {} for n ← 1 to N j ← argmini distance(ci,dn) gj ← gj ∪ {dn} for k ← 1 to K

ck ←

return C

d∑d in gk

[Based on Manning, Raghavan, and Schutze 2008]

K-means: algorithm

ck ←

return C

d∑d in gk

Pick K random points (could be some of the data points in D)

K-means: algorithm

ck ←

return C

d∑d in gk

Stopping criteria (when does the algorithm stop?)

K-means: algorithm

ck ←

return C

d∑d in gk

(Re)initialize the document clusters.

K-means: algorithm

ck ←

return C

d∑d in gk

Find the closest centroid for each document; put the document in that group.

K-means: algorithm

ck ←

return C

d∑d in gk

Find the closest centroid for each document; put the document in that group.

Recompute centroids based on the new document clusters (the gk’s).

Calculating distance

The documents are data points in some (possibly high-dimensional) space. We’ll work with 2D here.

Recall the Pythagorean theorem: c2 = a2 + b2

Here, the “a” is the distance on the x-axis and the “b” is the distance on the y-axis between points di and dj.

distance(di,dj) = (xi - xj)2 + (yi - yj)2

Consider two data points d1 = (5,4) and d2 = (1,2).

distance(d1, d2) = (x1 - x2)2 + (y1 - y2)2 = (5-1)2 + (4-2)2 = 42 + 22 = 20

Note: we could take the square root, but it doesn’t matter since we are just comparing a bunch of squared distances.

Simplified problem: just two authors and four documents

Let’s apply the K-means algorithm to four documents

Keep in mind that we are acting like we don’t know who is the author of each document.

Document I the

Mansfield 1.5 3.9

Persuasion 1.3 4.0

Document I the

Gerard 3.6 6.1

Holmes 2.8 5.3

Austen Doyle

D = {d1, d2, d3, d4} = { (1.5,3.9), (1.3,4.0), (3.6,6.1), (2.8,5.3) }

Choose K = 2 (i.e., 2 authors)

Choose C = {c1, c2} = { (3.6,6.1), (2.8,5.3) } as initial seed centroids.

Here’s what it looks like

1 1.75 2.5 3.25 4

Computing the groups

Calculate the nearest centroid for each document and put it in the group for that centroid.

distance(c1,d1) = (3.6-1.5)2 + (6.1-3.9)2 = 9.25

distance(c2,d1) = (2.8-1.5)2 + (5.3-3.9)2 = 3.65

c2 is closer, so d1 is in g2

Doing this for d2, d3, and d4, we find that:

g1 = {d3} and g2 = {d1,d2,d4}

1 1.75 2.5 3.25 4

New centroids

Next, we need to compute the new centroids based on these groups.

g2 has multiple elements:

sum of the x-values: 1.5+1.3+2.8 = 5.6

sum of the y-values: 3.9+4.0+5.3 = 13.2

size of g2 is 3, so c2 = (5.6/3, 13.2/3) = (1.9, 4.4)

g1 stays the same:

size of g1 is 1, so we have c1 = (3.6/1, 6.1/1) = (3.6,6.1)

1 1.75 2.5 3.25 4

Re-assign groups based on new centroids

We then keep iterating until the centroids stay the same

calculate nearest centroid for each document, put in in the group

recalculate centroids for new groups

Notice that d4 is now closer to c1

distance(c1,d4) = (3.6-2.8)2 + (6.1-5.3)2 = 1.28

distance(c2,d4) = (1.9-2.8)2 + (4.4-5.3)2 = 1.62

1 1.75 2.5 3.25 4

The next round would be...

1 1.75 2.5 3.25 4

With the right groups

1 1.75 2.5 3.25 4

Running k-means on all the documents

0 1 2 3 4Relative frequency of “I”

of “

the”

of “

the”

of “

the”

of “

the”

Wrong cluster!

Back to authorship identification

Features were extracted from the documents and used as values for plotting each document in a multi-dimensional space.

Documents were then clustered according to K-means (other algorithms could be used).

K-means gave us a set of centroids, so we can plot other documents into the same multi-dimensional space and compute which one is closest.

We are given new documents of known authorship

of “

the”

of “

the”

of “

the”

of “

the”

(Austen)

Attribution

The known documents provide evidence that the clusters we found are sets of documents produced by the author(s) of the known documents.

We can estimate the confidence in our authorship assignments based on how close the known documents are to each center.

A case study: The Federalist Papers

85 essays written in 1787 and 1788 arguing for the ratification of the new US constitution by the individual states.

Three authors, Alexander Hamilton, John Jay and James Madison, all writing under the pseudonym “Publius”.

Later, Hamilton and Madison both claimed to have written a number of the same articles. Scholarship in the 20th century revealed most of them to be Madison’s.

Alexander Hamilton1st US Secretary of the Treasury.51 articles (nos. 1, 6–9, 11–13, 15–17, 21–36, 59–61, and 65–85); co-authored 18, 19 & 20 w/ Madison.

James Madison4th US President, “Father of the Constitution”29 articles (nos. 10, 14, 37–58 and 62–63); co-authored 18, 19 & 20 w/ Hamilton.

John Jay1st Chief Justice of the US.5 articles (nos. 2-5 and 64)

Authorship of the Federalist Papers

Historian Douglas Adair in 1944 argued that Madison was the author of many of the disputed papers.

This was confirmed by Mosteller and Wallace in 1964 using a Bayesian classification model.

Adair’s authorship determinations are still generally accepted, though twelve essays are still disputed over by some scholars.

Experiment: cluster the documents based on all words that occur 5 or more times and k-means. (With some principal components analysis in between).

Extracting “features”: frequent words and their counts

FEDERALIST No. 1

General Introduction

For the Independent Journal. Saturday, October 27, 1787

HAMILTON

To the People of the State of New York:

AFTER an unequivocal experience of the inefficacy of the subsistingfederal government, you are called upon to deliberate on a newConstitution for the United States of America. The subject speaks itsown importance; comprehending in its consequences nothing less than theexistence of the UNION, the safety and welfare of the parts of which itis composed, the fate of an empire in many respects the most interestingin the world. It has been frequently remarked that it seems to have beenreserved to the people of this country, by their conduct and example,to decide the important question, whether societies of men are reallycapable or not of establishing good government from reflection andchoice, or whether they are forever destined to depend for theirpolitical constitutions on accident and force. If there be any truthin the remark, the crisis at which we are arrived may with propriety beregarded as the era in which that decision is to be made; and a wrongelection of the part we shall act may, in this view, deserve to beconsidered as the general misfortune of mankind.

.... <the rest of the essays> ....

ID Author NumWords . people for jury macedon one power with an as at to more states its be by this upon government them they has the not that than a ; but state courts , is it in if may have executive would been no constitution and any on of or there all his are from their which will other1 HAMILTON 1771 49 5 12 0 0 4 2 6 11 10 8 71 7 2 10 34 14 14 6 9 2 6 6 130 14 28 11 25 7 2 5 0 105 13 20 27 4 11 10 0 2 3 3 8 40 6 9 104 6 2 9 0 12 11 14 18 25 32 JAY 1841 40 21 13 0 0 10 1 13 1 15 10 52 5 2 5 15 10 14 1 9 4 22 6 105 10 44 5 29 11 8 0 0 122 16 38 34 3 4 17 0 5 8 1 0 83 1 8 81 10 0 4 0 6 4 21 11 2 43 JAY 1604 37 7 11 0 0 8 3 10 3 24 1 55 13 11 1 31 18 6 0 16 8 5 5 91 13 20 8 13 7 7 7 3 117 7 21 25 6 6 7 1 2 2 2 0 60 5 6 60 32 1 4 2 8 15 11 11 24 74 JAY 1806 32 7 12 0 0 13 2 12 3 20 2 50 13 1 9 26 14 1 0 16 12 17 1 84 14 17 9 16 10 10 5 0 125 10 28 24 12 10 9 0 17 2 1 0 90 5 11 70 24 3 4 2 11 8 19 10 15 115 JAY 1475 35 2 7 0 0 10 1 11 3 3 4 44 11 1 4 31 10 6 0 2 11 11 0 64 8 23 9 9 8 4 1 0 86 7 20 28 3 2 1 0 37 0 2 0 72 3 5 51 10 0 4 0 3 11 11 10 6 46 HAMILTON 2528 77 3 13 0 0 3 4 15 12 19 6 60 10 7 1 18 12 11 4 2 3 13 10 187 17 24 6 52 10 5 5 0 174 8 11 60 5 3 27 0 6 13 0 1 81 0 5 166 19 8 6 3 18 16 15 24 6 107 HAMILTON 2562 81 0 14 0 0 4 2 12 15 19 11 81 6 31 2 47 28 22 11 2 12 7 4 205 14 27 5 48 16 3 7 0 158 14 17 43 12 3 23 0 51 8 3 1 51 7 13 156 10 9 12 0 7 11 18 24 1 98 HAMILTON 2306 72 8 9 0 0 6 5 13 13 16 11 79 8 10 10 35 11 19 3 3 11 13 8 156 12 18 3 46 17 10 6 0 164 21 21 42 7 8 18 2 27 16 4 4 54 2 11 133 17 2 9 0 14 9 16 26 11 69 HAMILTON 2213 68 2 10 0 0 12 3 14 12 18 10 70 6 12 5 26 13 14 4 17 5 20 12 168 13 24 1 46 14 6 6 1 133 16 23 37 6 5 24 0 8 15 2 3 45 4 9 153 11 3 8 4 14 7 14 25 7 510 MADISON 3337 79 4 18 0 0 8 1 11 14 20 8 99 12 3 9 61 39 11 0 13 8 11 4 259 14 31 13 78 32 11 2 0 221 41 47 63 6 16 16 0 6 9 5 2 121 4 18 155 22 6 4 8 26 12 21 39 30 2211 HAMILTON 2766 78 0 18 0 0 5 3 21 15 21 11 82 9 13 6 44 20 24 6 5 6 10 8 186 20 24 6 69 8 8 4 0 173 9 21 66 6 9 15 0 50 3 5 0 70 4 5 178 12 8 13 1 9 19 11 25 17 1012 HAMILTON 2410 72 3 8 0 0 5 1 18 11 17 13 81 5 10 8 40 15 17 7 8 6 8 13 174 6 27 4 47 15 8 11 0 161 22 26 54 7 3 13 0 22 8 2 0 62 7 12 139 7 9 5 2 15 19 14 23 11 1413 HAMILTON 1045 28 2 1 0 0 8 5 10 3 8 1 42 6 9 2 25 5 5 2 7 1 3 2 72 5 18 10 26 6 2 3 0 53 9 6 14 6 8 4 0 14 2 4 0 17 3 3 56 3 9 3 0 4 5 1 11 9 114 MADISON 2357 53 5 20 0 0 2 3 7 9 24 7 71 8 12 8 45 18 13 0 9 4 17 8 200 13 33 5 37 19 4 1 0 144 25 31 40 6 16 17 0 5 9 10 2 60 3 17 122 11 0 6 0 6 11 19 40 26 715 HAMILTON 3395 83 4 18 0 0 1 9 19 18 15 24 116 6 11 8 44 32 24 10 14 4 15 14 251 16 38 10 69 25 8 5 1 186 48 30 73 7 7 31 0 13 14 6 4 74 3 10 194 38 18 21 1 22 22 7 55 19 6< ... 70 more lines for the other essays ...>

Dimensionality reduction: principal components analysis

Author Color

Hamilton green

Madison purple

Jay cyan

Hamilton & Madison red

K-means clustering

Three clusters (k=3): the model thinks the collaborations were most similar to Madison, plus three of Hamilton’s.

AuthorCluster COLLABMH HAMILTON JAY MADISON 1 3 3 0 26 2 0 48 0 0 3 0 0 5 0

K-means clustering

With no changing of the features or parameters, amazingly close to the authorship attributions given by Adair.

Three clusters (k=3): the model thinks the collaborations were most similar to Madison, plus three of Hamilton’s.

AuthorCluster COLLABMH HAMILTON JAY MADISON 1 3 3 0 26 2 0 48 0 0 3 0 0 5 0

AuthorCluster COLLABMH HAMILTON JAY MADISON 1 3 0 0 1 2 0 0 5 0 3 0 0 0 25 4 0 51 0 0

Four clusters (k=4): Cleanly clustered. One article (#47) attributed to Madison is astray, but it sensibly goes with the collaborative articles.

Web-scale authorship attribution

Narayanan et al. (2012) “On the Feasibility of Internet-Scale Author Identification”

The paper shows, for the first time, that large-scale authorship attribution is feasible (in their case, testing with 100,000 authors) and accurate enough to be useful in authorship attribution use cases.

For example, in their experiments, in 35% of their test cases, the correct author is in the top 20 authors predicted by their model.

Given that this is out of 100,000 authors, that is quite significant.

Supervised learning and clustering

Simple note: clusters are often used as features for learning supervised classifiers.

E.g. cluster words according to their syntactic contexts and use the cluster ids as the features. This can help with some word-sense disambiguation without using a predefined set of word senses.

He paid a lot of interest after he took out that bank loan to buy his house.

He paid a lot of interest to the lecture he saw on YouTube.

C2 C2 C5C8

C6 C6 C42 C12

Clusters can also themselves be labeled and used for training!

Relative status: who has the lead?Communicative behaviors are “patterned and coordinated, like a dance” [Niederhoffer and Pennebaker, ‘02]

adah ja ad to the adajkj the

adah ja ad at a adajkj the

adah ja ad of adajkj the

adah to ja ad an adajkj gh

adah ja ad the adajkj forhgh

adah ja ad to the adajkj the

adah ja ad at a adajkj the

adah ja ad of adajkj the

adah to ja ad an adajkj gh

adah ja ad the adajkj forhgh

Those with less power tend to immediately match the function-word choices of those with more power. [C. Danescu-Niculescu-Mizil et al. WWW 2012]

Content words still matter: detecting dodging by debaters

149Slide by Philip Resnik

Demographics

Predict categories such as age, race, and gender based on what a person writes.

Mixes both content and stylistics features

E.g. men and women tend to talk about different topics and they also use function words at different rates. (See James Pennebaker’s work)

Usually also relies on other inputs, such as a person’s name, when they write (e.g. tweet times), and their social network.

http://www.tweetolife.com/gender/ [no longer working]

Example: People Pattern audience demographics (3.8k accounts)

Beyond text

Stylistics & author modeling

Wrap up

What does “barbecue” mean?

What does “barbecue” mean? Barbecue’

What I thought semantics was before 2005

From: John Enrico and Jason Baldridge. 2011. Possessor Raising, Demonstrative Raising, Quantifier Float and Number Float in Haida. International Journal of American Linguistics. 77(2):185-218

Updated perspective a la Ray Mooney (UT Austin CS)

http://www.cs.utexas.edu/users/ml/slides/chen-icml08.ppt

http://www.lib.utexas.edu/books/travel/index.htmlTravel at the Turn of the 20th Century

Motivation: Google Lit Trips [http://www.googlelittrips.com/]

Grapes of Wrath in Google Earth

http://www.googlelittrips.com/GoogleLit/9-12/Entries/2006/11/1_The_Grapes_of_Wrath_by_John_Steinbeck.html

Motivation: Google Lit Trips [http://www.googlelittrips.com/]

Grapes of Wrath in Google Earth

http://www.googlelittrips.com/GoogleLit/9-12/Entries/2006/11/1_The_Grapes_of_Wrath_by_John_Steinbeck.html

Crisis response: Haiti earthquake

Look, Mom, no hands! (Err, um... no metadata.)

Topics with a clear, circumscribed geographic focus emerge!

But, of course, certain kinds of metadata are now plentiful.

Geotagged Wikipedia

30° 17′ N 97° 44′ W

01:55:55 RT @USER_dc5e5498: Drop and give me 50....

05:09:29 I said u got a swisher from redmond!? He said nah kirkland! Lol..ooooooooOkay!

05:57:35 Lmao!:) havin a good ol time after work! Unexpected! #goodtimes

06:00:09 RT @USER_d5d93fec: #letsbereal .. No seriously, #letsbereal>>lol. Don't start.

06:00:37 On my way to get @USER_60939380 yeee! She want some of this strawberry! Sexy!

47°31’41’’ N 122°11’52’’ W162

Geotagged Twitter

01:55:55 RT @USER_dc5e5498: Drop and give me 50....

05:09:29 I said u got a swisher from redmond!? He said nah kirkland! Lol..ooooooooOkay!

05:57:35 Lmao!:) havin a good ol time after work! Unexpected! #goodtimes

06:00:09 RT @USER_d5d93fec: #letsbereal .. No seriously, #letsbereal>>lol. Don't start.

06:00:37 On my way to get @USER_60939380 yeee! She want some of this strawberry! Sexy!

47°31’41’’ N 122°11’52’’ W162

Geotagged Twitter

Document geolocation: where is this person?

Document geolocation

Language-model-based Information Retrieval (LMIR) [Ponte and Croft 1998, Zhai and Lafferty 2001]

Given texts annotated with locations, construct language models for points or regions on the earth’s surface.

(There are many choices for doing this.)

Given a new (unlabeled) text, rank all locations w.r.t. how good a match they are.

Give the ranking of locations, choose a single coordinate as the location for the unlabeled text.

[Serdyukov, Murdock, & van Zwol 2009; Cheng, Caverlee, & Lee 2010; Wing & Baldridge 2011]

Amsterdam, Zaandam, Amstelveen, Diemen, Landsmeer ...

Frankfurt, Frechen, Hürth, Brühl, Wesseling, ...

Construct pseudo-documents from a geodesic grid

Grid: equal degree/area cells; rectangular or otherwise?

Pseudoc location is cell center or centroid of all documents in cell.

Construct pseudo-documents from a geodesic grid

Grid: equal degree/area cells; rectangular or otherwise?

Pseudoc location is cell center or centroid of all documents in cell.

Generate a language model for each pseudo-document

Interpolate or smooth against the LM for the entire collection, neighboring pseudo-documents, or both.

Generate a language model for a query document

May optionally be smoothed, depending on the ranking method.

longitude

latitu

− log( rank )

−10−8−6−4

Rank all pseudo-documents w.r.t. “query” text

Kullback-Leibler divergence

Query Likelihood

longitude

latitu

− log( rank )

−10−8−6−4

Choose the best matching pseudo-document

Highest ranking document

Mean shift over top-K

Reranking (Learning to Rank)

Locations of Twitter users are not uniformly distributed!

(Small) GeoUT (Twitter) plotted on Google Earth, one pin per user.

Density of (all) documents in GeoUT

over the USA (390 million tweets)

k-d tree for geotagged Wikipedia, looking at N. America

Pre-grid clustering [Erik Skiles, MA thesis, UT Austin, Ling]

Four clusters on GeoUT (390 million tweets)

West coast East coast Midwest & South Spanish language

All tweets

Performance (kd-tree with clustering)

Wikipedia (entire world)Half of documents geotagged within 12 km of truthPercent of documents within 166km (100 miles): 91%

Twitter (USA)Half of users geotagged within 330 km of truthPercent of documents within 166km (100 miles): 40%

For better or worse, it soon might not matter whether you have location turned on or not... what

you say is where you are / are from. (Also, other factors, e.g. who you are linked to, of course.)

Geographic meaning of words (based on Wikipedia)

mountainbeach

wine barbecue

Image geo-location: http://graphics.cs.cmu.edu/projects/im2gps/

Entity linking

It is one thing to identify that a span of text is a named entity, and another thing entirely to pick the unique entity in the world that it refers to.

John Smith went to London to visit Apple last year.

Which John Smith?

London in UK, Canada, or elsewhere?

Apple Inc. or Apple Records?

2013? 2012?1999? 1850?

Toponym (place name) resolution: geographic entity linking

They visit Portland every year.

Toponym (place name) resolution: geographic entity linking

They visit Portland every year.

Which Portland? (Also: Canada, Australia, Ireland...)

Toponym resolution in context

Although Elisha Newman made the first land entry in the township of Portland (June, 1833), he did not become a settler until three years later, by which time a few settlers had located in the town. From Mr. Newman's story, it appears that early in 1833, he was visiting friends in Ann Arbor, and during an evening conversation discussed with others the subject of unlocated lands lying west of Ann Arbor. One of the company (Joseph Wood) remarked that he had been out with the party sent to survey Ionia and other counties, and that the surveyors were struck by the valuable water-power at the mouth of the Looking Glass River, saying there would surely be a village there some day.Mr. Newman was at once taken with the idea of locating lands at the mouth of the Looking Glass. Following up his impulse, he made ready to start at once, and, accompanied by James Newman and Joseph Wood, went out to the Looking Glass on a tour of inspection. Being satisfied with the location, he returned Eastward with his companions, and at White Pigeon made his land entry.Newman did not return for a permanent settlement until the spring of 1836, and meanwhile, in November, 1833, Philo Bogue bought a piece of land on section 28, in the bend of the Grand River, where he proposed to set up a trading post. Unaided he rolled up a log cabin near where the Detroit, Lansing, and Northern depot was located, and when he brought the house into decent shape went over to Hunt's at Lyons for his family, whom he had left there against such time as he should have affairs prepared for their comfort.

Spatial minimality

Although Elisha Newman made the first land entry in the township of Portland (June, 1833), he did not become a settler until three years later, by which time a few settlers had located in the town. From Mr. Newman's story, it appears that early in 1833, he was visiting friends in Ann Arbor, and during an evening conversation discussed with others the subject of unlocated lands lying west of Ann Arbor. One of the company (Joseph Wood) remarked that he had been out with the party sent to survey Ionia and other counties, and that the surveyors were struck by the valuable water-power at the mouth of the Looking Glass River, saying there would surely be a village there some day.

Mr. Newman was at once taken with the idea of locating lands at the mouth of the Looking Glass. Following up his impulse, he made ready to start at once, and, accompanied by James Newman and Joseph Wood, went out to the Looking Glass on a tour of inspection. Being satisfied with the location, he returned Eastward with his companions, and at White Pigeon made his land entry.

Newman did not return for a permanent settlement until the spring of 1836, and meanwhile, in November, 1833, Philo Bogue bought a piece of land on section 28, in the bend of the Grand River, where he proposed to set up a trading post. Unaided he rolled up a log cabin near where the Detroit, Lansing, and Northern depot was located, and when he brought the house into decent shape went over to Hunt's at Lyons for his family, whom he had left there against such time as he should have affairs prepared for their comfort.

4048392 Portland Mills Portland Mills 39.7781 -87.00918 P PPL US IN 133 0 223 218 America/Indiana/Indianapolis 2010-02-154084605 Portland Portland 32.15459 -87.1686 P PPL US AL 047 0 30 41 America/Chicago 2006-01-154127143 Portland Portland Portlend,Портленд 33.2379 -91.51151 P PPL US AR 003 430 38 39 America/Chicago 2011-05-144169227 Portland Portland 30.51242 -86.19578 P PPL US FL 131 0 8 14 America/Chicago 2006-01-154217115 Portland Portland 34.05732 -85.03634 P PPL US GA 233 0 229 228 America/New_York 2010-09-054277586 Portland Portland 37.0778 -97.31227 P PPL US KS 191 0 362 364 America/Chicago 2006-01-154305000 Portland Portland 37.12062 -85.44608 P PPL US KY 001 0 220 223 America/Chicago 2006-01-154305001 Portland Portland 38.26924 -85.8108 P PPL US KY 111 0 135 138 America/Kentucky/Louisville 2006-01-154305002 Portland Portland 38.74812 -84.44772 P PPL US KY 191 0 265 266 America/New_York 2006-01-15404289 Portland Portland Portlend,Портленд 38.71088 -91.71767 P PPL US MO 027 0 170 172 America/Chicago 2010-01-294521811 Portland Portland Portlend,Портленд 39.00341 -81.77124 P PPL US OH 105 0 187 188 America/New_York 2010-01-294650946 Portland Portland Portlend,Портленд 36.58171 -86.51638 P PPL US TN 165 11480 244 245 America/Chicago 2011-05-144720131 Portland Portland Portlend,Портленд 27.87725 -97.32388 P PPL US TX 409 15099 13 11 America/Chicago 2011-05-144841001 Portland Portland Portlend,Портленд 41.57288 -72.64065 P PPL US CT 007 5862 24 27 America/New_York 2011-05-144871855 Portland Portland 43.12858 -93.12354 P PPL US IA 033 35 327 330 America/Chicago 2011-05-144906524 Portland Portland 41.66253 -89.98012 P PPL US IL 195 0 190 190 America/Chicago 2006-01-155006314 Portland Portland Portlend,Портленд 42.8692 -84.90305 P PPL US MI 067 3883 221 223 America/Detroit 2011-05-145746545 Portland Portland 45.52345 -122.67621 P PPLA2 US OR 051 583776 12 15 America/Los_Angeles 2011-05-14

Spatial minimality

Ann ArborDetroit

IoniaLyons

PortlandWhite Pigeon

>15>17

# LocationsToponym

Spatial minimality

PortlandLyonsIonia

White Pigeon

Ann ArborDetroit

IoniaLyons

PortlandWhite Pigeon

>15>17

# LocationsToponym

Spatial minimality often fails

I moved from Encinitas, CA, a nice beach town in North San Diego County to Asheville, NC. By far, Ashville is more hip, especially West Asheville. Asheville has a lot in common with Portland. Austin, I've never been to so I cannot comment. But what makes a place cool and hip, in my opinion are that give a area "punch". There are a lot of ingredients. One is geography. Add a college or university (and all that they bring- and draw), good restaurants, a good music scene, a progressive attitude and tolerance. Hmmm. I'm sure there are many more to ponder. But that's my start. Oh, lots of bars!

From: http://www.city-data.com/forum/austin/1694181-what-makes-city-like-austin-portland-3.html

City-data.com incorrectly marks “West” and “Portland” as the cities in Texas -- presumably because of their textual and spatial proximity to “Austin”.

Spatial minimality often fails

From: http://www.city-data.com/forum/austin/1694181-what-makes-city-like-austin-portland-3.html

City-data.com incorrectly marks “West” and “Portland” as the cities in Texas -- presumably because of their textual and spatial proximity to “Austin”.

But: it is clear from the text that Portland, Oregon and Austin, Texas are the referents, though their states are never mentioned and are far from the other locations!

Toponym classifiers

Strategy: build a textual classifier per toponym by obtaining indirectly labeled examples from Wikipedia.

Toponym classifiers

P(Portland-OR|music) > P(Portland-ME|music)P(Portland-OR|wharf ) < P(Portland-ME|wharf )

Results: disambiguating toponyms

Average error distance

Accuracy Average error distance

Accuracy

Population

SPIDER(spatial minimality)

WISTR(Wiki supervised)

SPIDER+WISTR

216 81.0 1749 59.7

2180 30.9 266 57.5

279 82.3 855 69.1

430 81.8 201 85.9

TR-CoNLLReuters News Texts

August 1996

Perseus Civil War CorpusBooks

Late 19th Century

Take-home message: text classifiers are very effective & can be boosted by spatial minimality algorithms.

Back to grounding

Grounding often involves connecting text to knowledge sources and other modalities (image, video) & bootstrapping.

Back to grounding

Semi-supervised learning methods such as label propagation help bridge the annotation gap.

Back to grounding

Semi-supervised learning methods such as label propagation help bridge the annotation gap.

Also, they can help us create models for deeper aspects of language, such as syntactic structure and logical form.

Temporality of words, by hour http://www.tweetolife.com/hour/ (no longer working)

Temporality of expressions, by day: http://www.google.com/trends

Temporality of expressions, by year: http://ngrams.googlelabs.com/

slavetrenches aircraft

Temporal resolution [Kumar, Lease, and Baldridge 2011]

Lexical brain decoding [Yarkoni, Poldrack, Nichols, Van Essen & Wager (2011)]

More modalities: videos [Motwani & Mooney, 2012]

Beyond word co-occurences for vector-space models

bear boat car cow hadoop snow water wrench

3 234 42 4 1 2 325 0beach

3 234 42 4 1 2 325 0

Combining distributional models with logics

Erk (2013): “Towards a semantics for distributional representations.”

Garrette et al (2012): “A formal approach to linking logical form and vector-space lexical semantics”Beltagy et al (2013): “Montague Meets Markov: Deep Semantics with Probabilistic Logical Form”

Multi-component structured vector-space models

beachchildren

the children visit the beach

Agent Patient

Language learning in context [Kim & Mooney, 2013]

Wrap up

All your meaning are belong to us

ALPAC yer bags

It’s cold out there, after the hype

The Hype Cycle (Gartner)

The Hype Cycle (Gartner, 2013)

http://www.gartner.com/newsroom/id/2575515

The Hype Cycle (Gartner, 2013)

http://www.gartner.com/newsroom/id/2575515

http://davidrothman.net/2009/09/02/all-your-healthbase-are-belong-to-us-want-em-back/

Grounding matters

Being right can be consequential too

Narayanan et al 2012: “On the Feasibility of Internet-Scale Author Identification”

Tremendous challenges and tremendous opportunities

• It is possible to create automated methods for opinion mining tasks that make reasonably accurate predictions that are useful for summarizing attitudes and opinions at a glance.

• The deeper you go, the better you can discriminate detailed opinions, but also the harder it is to have the necessary data to build accurate models. (The problem is “NLP-complete”.)

• Semi-supervised methods provide a lot of promise for these endeavors.

• Methods that reduce the burden of feature engineering, such as deep learning, are also important.

• A big component of this is grounding texts in the real world.

Practical Sentiment Analysis

philo bogue

equal degreearea

adah toja

gj gj dn

evening conversation

singular introspective

adah ja ad

porsche panamera

Data & Analytics

SentiRuEval: Testing Object-Oriented Sentiment Analysis...

Tree Communication Models for Sentiment Analysis ·...

Sentiment Analysis & Computational...

Redefining Sentiment Analysis

Sentiment Analysis & Opinion Mining€¦ · Sentiment...

Sentiment Analysis - Stanford University · PDF file ·...

Sentiment Analysis Introduction to -...

Practical Sentiment Analysis Tutorial -...

Sentiment Analysis and Opinion...

Introduction to Sentiment Analysis - ETH Z · sentiment !.....

Part 4: Practical Sentiment Analysis in GATE · Practical.....

Geo-spatial Multimedia Sentiment Analysis in Disasters ·.....

Practical Sentiment Analysis

Sentiment Analysis in Organizational...

SENTIMENT ANALYSIS · Sentiment Analysis •Sentiment...

Sentiment Analysis and the Consumer...