Word Vector s to Take Figurative Language to New Heights Do or Do Not There Is No Try DiscourseLevel Style in Quotations Kyle Booten Andrea Gagliano Emily Paul Marti Hearst ID: 760364
Download Presentation The PPT/PDF document "Intersecting Cognitive Linguistics and" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Intersecting Cognitive Linguistics and
Word Vectors to Take Figurative Language to New HeightsDo or Do Not. There Is No Try:Discourse-Level Style in Quotations
Kyle Booten,
Andrea
Gagliano, Emily Paul,
Marti
Hearst
UC
Berkeley
Google Research, Oct 24, 2017
Slide2Motivation:
Support Creativity
Slide3eye
sun
Slide4eye
sun
orb
What would Shakespeare do?
Slide5Combine Two Ideas from Cognitive Linguistics
Family Resemblances
Every day Metaphor
Slide6Wittgensteinian family resemblances
Rosch and Mervis. 1975. Family resemblances: Studies in the internal structure of categories.
games
b
oard games
ball games
video games
golf
c
ard games
Common features, but no single unifying attribute
sun
solar
sphere
heat
light
orbit
orange
flame
Slide7In the middle of life’s road,I found myself in a dark wood.—Dante, Divine Comedy
Metaphor in poetry
And all our yesterdays have lighted foolsThe way to dusty death.—Shakespeare, Macbeth
Metaphor used: Life is a Journey
Lakoff
and Turner. 1989. More than cool reason: A field guide to poetic metaphor.
Slide8Metaphor “mak[es] use of structure imported from a completely different conceptual domain” —Lakoff and Turner
Lakoff
and Turner. 1989. More than cool reason: A field guide to poetic metaphor.
Slide9Blending of Semantic Spaces
In the middle of life’s road,I found myself in a dark wood.—Dante, Divine Comedy
life
road
Distinct semantic spaces
journey quest
way
. . .
Our goal: computationally suggest words to create a figurative relationship
Lakoff
and Turner. 1989. More than cool reason: A field guide to poetic metaphor.
Fauconnier and Turner. 2008. The way we think: Conceptual blending and the mind’s hidden complexities.
Dante poetically describes middle age as being lost in a wood while traveling down life’s path.
Slide10Some recent work on computational metaphor
“Surgeons are butchers”
Veale et al. (2000)Model based on graph isomorphism with slippage across analogical relations.
Veale
, O’Donoghue, and Keane. 2000. Computation and blending.
Slide11Recent work on computational metaphor
Search query:“as adj as a | an N” ->“As hot as an oven”“Solemn” yields related words: {monument, judge, owl, funeral, etc}
Veale and Hao (2007 & 2008)Develops a case base of similes via web search
Veale
and Hao. 2007. Comprehending and generating apt metaphors: a webdriven, case-based approach to figurative language.; Veale and Hao. 2008. A fluid knowledge representation for understanding and generating creative metaphors.
Slide12Recent work on computational metaphor
Harmon (2015)Builds a sentence based on meaning relations between two nounsRanks according to conventionalityPrefers words that are not conceptually similar
Harmon
. 2015. Figure8: A novel system for generating and evaluating figurative language.
Slide13Our Idea:
Using
word embeddings as
family resemblances
to blend
semantic
spaces
.
Slide14Mikolov, Yih, Corrado, and Dean. 2013c. Linguistic regularities in continuous space word representations.
Word2vec word embeddings
The king ruled over his land.
context
Conversion Step
Slide15Using word embeddings for family resemblances
Schütze. 1993. Word space.; Rosch and Mervis. 1975. Family resemblances: Studies in the internal structure of categories.
Slide16Semantic similarity from Word Embeddings
Mikolov, Chen, Corrado, Dean, and Jurafsky. 2013a. Efficient estimation of word representations in vector space.; Mikolov, Sutskever, Chen, Corrado, and Dean. 2013b. Distributed representations of words and phrases and their compositionality.; Mikolov, Yih, Corrado, and Dean. 2013c. Linguistic regularities in continuous space word representations.
Slide17Idea: use word embeddings for family resemblances
life
family
living
everyday
humanity
childhood
society
motherhood
No single unifying attribute
Slide18Mikolov, Chen, Corrado, Dean, and Jurafsky. 2013a. Efficient estimation of word representations in vector space.; Mikolov, Sutskever, Chen, Corrado, and Dean. 2013b. Distributed representations of words and phrases and their compositionality.; Mikolov, Yih, Corrado, and Dean. 2013c. Linguistic regularities in continuous space word representations.
Standard word2vec usage
Slide19Standard word2vec usage
?
Slide20Standard word2vec usage
Slide21Standard word2vec usage
Slide22Standard word2vec usage
Slide23Standard word2vec usage
Slide24Standard word2vec usage
Slide25Standard word2vec usage
Slide26Example: surrendering & storm
Anchor words
Connector word
?
surrendering
storm
What words can
you
think of?
Slide27Using word2vec to approximate intersection of Lakoff & Turner metaphor frames
?
surrendering
losing
giving
yielding
conceding
relinquishing
storm
hurricane
tornado
squall
snowstorm
typhoon
Wittgenstein family resemblance
from semantically close vectors.
Lakoff and Turner importing from
different
conceptual domains
Slide28Baseline (Addition) Model vector representation
Slide29Baseline (Addition) Model vector representation
Slide30Baseline (Addition) Model vector representation
Slide31Sample output of Baseline (Addition) Model
hurricanetornadodelugefloodingdownpourrainrainstormtwistersquall...
Words most similar to:
Slide32Using word2vec to approximate intersection of Lakoff & Turner metaphor frames
?
surrendering
losing
giving
yielding
conceding
relinquishing
storm
hurricane
tornado
squall
snowstorm
typhoon
Wittgenstein family resemblance
from semantically close vectors.
Lakoff and Turner importing from
different
conceptual domains
Slide33eye
sun
Slide34eye
sun
orb
Slide35storm
hurricane
tornado
squall
snowstorm
typhoon
surrendering
losing
giving
yielding
conceding
relinquishing
NEW:
Intersection
Model
in word2vec
n = 1000
n = 1000
Slide36storm
hurricane
tornado
squall
snowstorm
typhoon
surrendering
losing
giving
yielding
conceding
relinquishing
NEW:
Intersection
Model
in word2vec
n = 1000
n = 1000
barrage
dissipating
onslaught
...
Slide37Intersection Model
vector representation
Slide38Intersection Model
vector representation
Slide39Intersection Model
vector representation
Slide40Sample output of Intersection Model
relinquishingyieldinggivingconcedinglosing
hurricane snowstormtornadosqualltyphoon
Words most similar to:
Slide41Sample output of Intersection Model
relinquishingyieldinggivingconcedinglosing… …barrage (n = 932)…
hurricane snowstormtornadosqualltyphoon...barrage (n = 288)………
Words most similar to:
Slide42Observing differences between models
Setup of example word pairsQuantitative observationsQualitative observations
Slide43Setup: Generate word pairs for observation
Poetic theme and concrete noun selected because Kao and Jurafsky (2015) found that professional poetry contains more concreteness
Anchor words
Connector word
Kao
and Jurafsky. 2015. A computational analysis of poetic style:
Imagism
and its influence on modern professional and amateur poetry.
Concrete noun
Poetic theme
?
Slide44Example: surrendering & storm
Anchor words
Connector word
?
surrendering
storm
Slide45Example: Connector words for surrendering & storm
Unique to Baseline Modelsqualltornadotyphoonsnowstormfloodingrainstormdelugehurricane...
Unique to Intersection Modelonslaughtstrandingblowingdissipatingbarrageregroupedbatteroutburst...
For any sets Baseline and Intersection of similar size, there are likely to be words unique to each set, as well as words shared between these sets; a proof is in the paper.
Slide46Observing differences between models
Step up of example word pairsQuantitative observationsQualitative observations
Slide47Unique to Intersection ModelSimilarity to stormSimilarity to surrenderingonslaught.30.20stranding.27.28blowing.24.29dissipating.23.22barrage.25.20regrouped.19.31batter.22.25outburst.21.20...
Average spread between similarity scores = 0.05
Unique to Baseline ModelSimilarity to stormSimilarity to surrenderingsquall.63-.03tornado.64-.02typhoon.62-.01snowstorm.64.01flooding.57.01rainstorm.57.07deluge.50.08hurricane.73.04...
Average spread between similarity scores = 0.56
Observing quantitative differences of connector words
Slide48Balanced band of similarity
Anchor word pairs
Range of average similarity from words in
Baseline Model
to anchor words
Range of average similarity from words in
Intersection Model
to anchor words
flame & caring
0.13 – 0.58
0.22 – 0.30
color & earthly
0.17 – 0.55
0.28 – 0.32
hair & anguish
0.14 – 0.66
0.27 – 0.33
flame & killing
0.09 – 0.55
0.23 – 0.26
mouth & compassion
0.16 – 0.54
0.25 – 0.29
storm & surrendering
0.03 – 0.69
0.21 – 0.26
ring & mankind
0.11 – 0.57
0.21 – 0.34
...
Slide49Anchor word pairsRange of average similarity from words in Baseline Model to anchor wordsRange of average similarity from words in Intersection Model to anchor wordsflame & caring0.13 – 0.580.22 – 0.30color & earthly0.17 – 0.550.28 – 0.32hair & anguish0.14 – 0.660.27 – 0.33flame & killing0.09 – 0.550.23 – 0.26mouth & compassion0.16 – 0.540.25 – 0.29storm & surrendering0.03 – 0.690.21 – 0.26ring & mankind0.11 – 0.570.21 – 0.34...
Balanced band of similarity
Band of similarity
~
0.25 to
~
.30
Slide50Observing differences between models
Step up of example word pairsQuantitative observationsQualitative observations
Slide51Setup: Construct dataset of figurative relationships
List Bonslaughtstrandingblowingdissipatingbarrageregroupedbatteroutburst...
List Asqualltornadotyphoonsnowstormfloodingrainstormdelugehurricane...
OR
Mechanical Turkers were shown one of these lists and asked to select the word that “best connects the anchor words in a poetic sense (e.g. using a double meaning, creating a new image, creating an interesting relationship, etc.)”
?
surrendering
storm
Slide52Setup: Construct dataset of figurative relationships (continued)
Mechanical Turkers set A first picked the best word from each list.Set B then completed this sentence to draw the connection:“Barrage connects storm and surrendering because ___________”These were manually assessed to see how figurative their language is and if they blended semantic spaces.
surrendering
storm
Selected: barrage
Slide53Observing qualitative differences of connector words
“Hurricane connects storm and surrendering because it is a type of storm and those who surrender to it are spared, like grass and those who stand against it are devastated, like big trees.”“Barrage connects storm and surrendering because a storm is a barrage of bad weather like winds and rain people surrender when they feel a barrage of overwhelming things coming at them.”
.04
.20
.25
.73
Intersection connector word
Baseline connector word
surrendering
surrendering
hurricane
barrage
storm
storm
Slide54Observing qualitative differences of connector words
“Hues connects color and earthly because hues imply various colors, shades, or characteristics and hues can be earthly in tone, such as blues, greens and browns.”“Radiant connects color and earthly because radiant means a bright color that looks like it’s shining and at night, the earthly sky is radiant because it shines brightly with the stars.”
hues
.
09
.
22
.29
.
61
earthly
radiant
earthly
color
color
radiant
Intersection connector word
Baseline connector word
Slide55Conclusion
Slide56Conclusion
Balanced similarity scores leads to heightened effects.
Intersection model may achieve this more reliably:
=
Band of similarity
~0.25 to ~.30
Concrete noun
Poetic theme
?
Slide57Contributions
Word embeddings combined with classic cognitive linguistics to generate figurative language.
New way to compute using word2vec representation
Observations about balance of similarity scores
life
family
living
everyday
humanity
childhood
motherhood
Anchor word pairs
Range of average similarity from words in
Baseline Model
to anchor words
Range of average similarity from words in
Intersection Model
to anchor words
flame & caring
0.13 – 0.58
0.22 – 0.30
color & earthly
0.17 – 0.55
0.28 – 0.32
hair & anguish
0.14 – 0.66
0.27 – 0.33
Slide58Questions/Future Work
Hypothesis testing / evaluation around the band of similarity
Threshold testing on size of semantic space
n = ?
Input from poets vs. Mechanical Turkers
Stronger grounding in Lakoff & Turner metaphorical representation
Reassess anchor words of poetic theme and concrete noun
?
?
Anchor words
Band of similarity
~
0.25 to
~
.30
Slide59Intersecting Cognitive Linguistics and
Word Vectors to Take Figurative Language to New HeightsDo or Do Not. There Is No Try:Discourse-Level Style in Quotations
Kyle Booten, Andrea Gagliano, Emily Paul, Marti HearstUC BerkeleyGoogle Research, Oct 24, 2017
Best short paper runner-up, NAACL 2016
Slide60What is the role of literary discourse in the age of social media?
Slide61Question: what motivates people to choose the quotes they do to post on social media?
This question motivated us to investigate what is special about quotations vs prose.
Overall finding
: part of what unites quotes as a genre is their latent stylistic patterns.
Slide62Related work
Danescu-Niculescu-Mizil
et al (2012)
, “You had me at hello” used features to distinguish popular movie quotations from unmemorable lines from the same film.
The found that popular quotations tended to use less common words, but these were placed into more common syntactic contexts.
They hint at “common syntactic scaffolding”, which we try to uncover here.
They also find that quotations tend to contain linguistic features that make them more “generalizable”, such as tendency toward indefinite over definite articles.
Slide63Related work
Guerini
et al. (2015)
found that memorable quotes were more euphonic, with more instances of rhyme and alliteration than non-memorable counterparts.
Louis and
Nenkova
(2012)
, studying text coherence, found that certain types of sentences (in terms of syntactic structure) tend to follow certain other types of sentences.
They also demonstrated that sentences with similar communicate purposes are syntactically similar.
Slide64Slide65Slide66Slide67Tumblr Statistics
(In 2014)16th most popular site in US5th most popular social network>160M users
Chang et al. 2014. What is
tumblr
: A statistical overview and comparison. KDD 16(1)
Slide68The Ancient Arts of Rhetoric
Phonetic patterns: rhyme;alliteration
Rhetoric: ways of wielding language to make it persuasive or memorable.
A predecessor to modern linguistics was the description of rhetorical “tropes” or “figures”.
Syntactic patterns:
epistrophe
: successive clauses end with the same words;
p
ysma
: the speaker launches a series of sharp and vehement questions.
Slide69Idea: Focus on Two-Sentence Quotations
Compare first sentence to second sentence.
See if there is a difference between quotations and prose pairs.
Slide70Tumblr Quotes seem to have certain “styles”
“You cannot observe people through an ideology. Your ideology observes for you.”-- Philip Roth
Syntactic patterns:
A
ntimetabloe
:
:
words in the first clause appear reversed in the second.
Negative statement (cannot) followed by a positive one.
First sentence begins with generic “you”.
Slide71Tumblr Quotes seem to have certain “styles”
“You cannot observe people through an ideology. Your ideology observes for you.”-- Philip Roth
Syntactic patterns:Antimetabloe:: words in the first clause appear reversed in the second.Negative statement (cannot) preceded by a positive one.(Do. Or do not.)First sentence begins with generic “you”.
“Great things are done when men and mountains meet.
This is not done by jostling in the street.”
-- William Black
Slide72Other Features Observed and Detected
“Forgotten is forgiven.”
Syntactic features:
High-level syntax (from Feng et
al’s
2012 authorship detection work which they argue provides an interpretable representation of the sentence structure):
NP + VP + .
Slide73Other Features Observed and Detected
“Peace comes from within. Do not seek it without.”
Lexical features:
General/abstract words like “peace”.
Computed as “is most common sense of head word within 5 hops of
WordNet
synset
Abstraction.n.06?”
Similar for General.
Slide74Datasets
Quotations 1: Tumblr posts marked with the quote type and #quotations hashtag and precisely 2 sentences long. (N=4237) TrainingQuotations 2: Tumblr posts marked with the quote type and #quote hashtag and precisely 2 sentences long. (N=1846) Testing
Non-quotes 1:
Random 2-sentence long paragraphs from the Brown corpus (N=1846)
Non-quotes 2:
Randomly chosen 2-sentence
sequendes
from longer paragraphs in Brown (N=1846)
Which Features Preferentially Occur in Sentence 1 in Quotation Corpus?
Lexical features:“n’t” and “not”, signifying that negative sentences were more likely to occur in the first sentence of a quotation Similarly, “never”, as well as “do”
Do not worry about your difficulties in mathe- matics. I can assure you mine are still greater. (A. Einstein)
We are not human beings having a spiritual experience. We are spiritual beings having a human experience. (P. de
Chardin
)
Slide76Which Features Preferentially Occur in Sentence 1 in Quotation Corpus?
Syntactic features:QuestionsWhen“Sweeping Declarations”
Love is a trap. When it appears, we see only its light, not its shadows. (P. Coelho)
Slide77Which Features Preferentially Occur in Sentence 2 in Quotation Corpus?
Syntactic features:CC + NP + VP + .ButAndThis is not surprising given the role of coordinating conjunctions, but it was 6 times more likely for Quotes than for Non-Quotes 2 comparison corpus.IN + NP + VP + .
Where a goat can go, a man can go. And where a man can go, he can drag a gun. (William Phillips)
One of
the
most adventurous things left us is
to go to bed. For no one can lay a hand on our dreams. (E.V. Lucas)
Slide78Which Features Preferentially Occur in Sentence 2 in Quotation Corpus?
Lexical features:Simply: a specific rhetorical pattern that emphasizes the second sentence’s proposition with respect to the first’s.
I used to dream about escaping my ordinary life, but my life was never ordinary. I had
simply
failed to notice how extraordinary it was. (R. Riggs)
Slide79Quote Sentence Ordering Task
Training: see quotes with sentences either in order or reversed (S1 S2 or S2 S1).
Testing
: try to predict the order.
Hypothesis
: can better predict the order of quotations than prose.
Slide80Datasets
Quotations 1: Tumblr posts marked with the quote type and #quotations hashtag and precisely 2 sentences long. (N=4237) TrainingQuotations 2: Tumblr posts marked with the quote type and #quote hashtag and precisely 2 sentences long. (N=1846) Testing
Non-quotes 1:
Random 2-sentence long paragraphs from the Brown corpus (N=1846)
Non-quotes 2:
Randomly chosen 2-sentence
sequendes
from longer paragraphs in Brown (N=1846) T
Quote Sentence Ordering Task
Training: see quotes with sentences either in order or reversed (S1 S2 or S2 S1).
Testing: try to predict the order.Hypothesis: can better predict the order of quotations than prose.
Significant, two-tailed t-test, p<.01
This is evidence that quotations as a genre are more “formulaic” than other textual sequences, their order more easily
predicted.
Confirmed!
Slide82Conclusions
This stylistic patterning may be especially strong in quotations.
Analyzed linguistic style not merely as the presence of features, but also their order across sentences.
In quotations, certain words as well as categories of words and syntactic patterns are more likely to appear in the first or second of a pair of two-sentence texts.
Slide83Another Project:NewsLens
Philippe Laban, Marti Hearst,
ACL 2017 Workshop on Events and Stories in the News
Slide84Slide85Slide86http://
newslens.berkeley.edu
/lanes
/
Slide87A Major ACL Initiative:
arXiv
/ Preprint Policy
Slide88Highlights of the NewACL Policies for Submission, Review, & Citation
https://www.aclweb.org/adminwiki/index.php?title=ACL_Policies_for_Submission,_Review_and_Citation
Submission
Review / Citation
*ACL and TACL submissions must be anonymousA submission is not considered anonymous if posted to a preprint server within an anonymity periodNon-anonymized submissions before the anonymity period are allowed, but discouraged
When reviewing, read the paper first and form an opinion before any searching.
For citation, refereed citations take priority over preprints
Papers (refereed or not) that appear within 3 months of submission should be considered contemporaneous
Slide89Appendices
Slide90Slide91Proof
Starting with the set A, from the Addition Model:
Alpha is the minimum similarity threshold resulting from selecting top n.
Slide92Proof (continued)
The set I, from the Intersection Model:
Beta and Gamma are the minimum similarity threshold resulting from selecting top n.
Slide93Proof (continued)
If we were finding the single word vector that maximized (1) and (2), the two equations would be equivalent (Levy and Goldberg 2014), such that (3) would need to be satisfied:
Note, (3) assumes the vectors are length normalized.
Slide94Proof (continued)
We then expand (3) as follows:
Slide95Proof (continued)
We can then solve (4) as follows:
But, (5) is not necessarily always true. Thus, the initial assumption that the intersection and addition models contain the same word vectors is contradicted. This confirms that the set A does not equal the set I.
Slide96Poetic themes and concrete nouns
Poetic themesloss melancholy angeranimals calmness compassionconfusion death envy faith fear forgiveness freedom friendship godgrace gratitude griefhate hope immortalityjealousy joy lifemothers nature peacepeople religion remembrance...
Concrete nouns
bed ear finger
horse sand hair
bell grass rock
book rose breast
ship blood window
wing girl snow
wood ring body
room wine ground
mouth garden stone
storm brain flame
...
Slide97Anchor word pairs selected
flame & caring storm & surrendering
color & earthly ring & mankind
hair & anguish hair & envied
flame & killing book & liberties
mouth & compassion town & grieving
Slide98eye
sun
Slide99eye
sun