Feuding Families and Former Friends: Unsupervised Learning ...
Transcript of Feuding Families and Former Friends: Unsupervised Learning ...
![Page 1: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/1.jpg)
Feuding Families and Former Friends:Unsupervised Learning for
Dynamic Fictional RelationshipsMohit Iyyer, Anupam Guha, Snigdha Chaturvedi,
Jordan Boyd-Graber, and Hal Daumé III
University of Maryland, College Park University of Colorado, Boulder
1
![Page 2: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/2.jpg)
How can we describe a fictional relationship between two characters?
2
![Page 3: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/3.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?
3
![Page 4: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/4.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
4
![Page 5: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/5.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
5
![Page 6: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/6.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
Frodo and Sam (Lord of the Rings)
6
![Page 7: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/7.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
Frodo and Sam (Lord of the Rings)
7
![Page 8: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/8.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
Frodo and Sam (Lord of the Rings) Winston and Julia (1984)
8
![Page 9: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/9.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
Frodo and Sam (Lord of the Rings) Winston and Julia (1984) ???
9
![Page 10: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/10.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
Frodo and Sam (Lord of the Rings) Winston and Julia (1984) ???
Harry Potter and Sirius (Prisoner of Azkaban)
10
![Page 11: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/11.jpg)
How can we describe a fictional relationship between two characters?
• isn’t this easy? we can assign it a single label (or relationship descriptor) from a predetermined set
Friend or foe?Peter Pan and Captain Hook (Peter Pan)
Frodo and Sam (Lord of the Rings) Winston and Julia (1984) ???
Harry Potter and Sirius (Prisoner of Azkaban) ???
11
![Page 12: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/12.jpg)
How can we describe a fictional relationship between two characters?
• what if we treat relationships as sequences (or trajectories) of descriptors? (Chaturvedi et al., 2016)
Tom Sawyer and Becky Thatcher: friends -> foes -> friends
12
![Page 13: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/13.jpg)
How can we describe a fictional relationship between two characters?
• what if we treat relationships as sequences (or trajectories) of descriptors? (Chaturvedi et al., 2016)
Tom Sawyer and Becky Thatcher: friends -> foes -> friends
• limited by fixed descriptor set
• required expensive annotations
• limited to plot summaries
13
![Page 14: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/14.jpg)
passage of time
Arthur and Lucy (Dracula)
![Page 15: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/15.jpg)
passage of time
I love him more than ever. We are to be married on 28 September.
Arthur and Lucy (Dracula)
![Page 16: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/16.jpg)
passage of time
love
joy
marriage
I love him more than ever. We are to be married on 28 September.
Arthur and Lucy (Dracula)
![Page 17: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/17.jpg)
passage of time
love
joy
marriage
I love him more than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
Arthur and Lucy (Dracula)
![Page 18: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/18.jpg)
passage of time
love
joy
marriage
love
sadness
sicknessI love him more
than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
Arthur and Lucy (Dracula)
![Page 19: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/19.jpg)
passage of time
love
joy
marriage
love
sadness
sicknessI love him more
than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
poor girl, there is peace for her at last. It is the end!
Arthur and Lucy (Dracula)
![Page 20: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/20.jpg)
passage of time
love
joy
marriage
love
sadness
sickness
love
fantasy
sickness
sadness
deathI love him more than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
poor girl, there is peace for her at last. It is the end!
Arthur and Lucy (Dracula)
![Page 21: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/21.jpg)
passage of time
love
joy
marriage
love
sadness
sickness
love
fantasy
sickness
sadness
deathI love him more than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
poor girl, there is peace for her at last. It is the end!
Arthur placed the stake over her heart … he struck with all his might. The Thing in the coffin writhed …
Arthur and Lucy (Dracula)
![Page 22: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/22.jpg)
passage of time
love
joy
marriage
love
sadness
sickness
love
fantasy
sickness
sadness
death
love
fantasy
death
murder
I love him more than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
poor girl, there is peace for her at last. It is the end!
Arthur placed the stake over her heart … he struck with all his might. The Thing in the coffin writhed …
Arthur and Lucy (Dracula)
![Page 23: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/23.jpg)
Why is this a worthwhile problem?
• “Distant reading” (Moretti, 2005) can help humanities scholars collect examples of specific relationship types
15
“Do Jane Austen’s female and male protagonists have a pattern in their evolving relationship (e.g., mutual disdain followed by romantic love)?” (Butler, 1975; Stovel, 1987; Hinant, 2006)
“Do certain authors or novels portray relationships of desire more than others?” (Polhemus, 1990)
“Can we detect positive or negative subtext underlying meals between two characters?” (Foster, 2009; Cognard-Black et al., 2014)
![Page 24: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/24.jpg)
Outline• Dataset: character interactions
• RMN: relationship modeling network
• Experiments: coherent descriptors, interpretable trajectories
• Analysis: RMN’s strengths and weaknesses
16
![Page 25: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/25.jpg)
A Dataset of Character Interactions
17
• For each pair of characters in a particular book, we extract all spans of text that contain mentions to both characters
![Page 26: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/26.jpg)
A Dataset of Character Interactions
17
“If anyone was ever minding his business, it was I," Ignatius breathed. "Please. We must stop. I think I'm going to have a hemorrhage.” "Okay." Mrs. Reilly looked at her son's reddening face and realized that he would very happily collapse at her feet just to prove his point.”
t=0
• For each pair of characters in a particular book, we extract all spans of text that contain mentions to both characters
![Page 27: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/27.jpg)
A Dataset of Character Interactions
17
“If anyone was ever minding his business, it was I," Ignatius breathed. "Please. We must stop. I think I'm going to have a hemorrhage.” "Okay." Mrs. Reilly looked at her son's reddening face and realized that he would very happily collapse at her feet just to prove his point.”
t=0
“Ignatius belched the gas of a dozen brownies trapped by his valve. "Grant me a little peace.…” "You know I appreciate you, babe," Mrs. Reilly sniffed. "Come on and gimme a little goodbye kiss like a good boy.”
t=1
• For each pair of characters in a particular book, we extract all spans of text that contain mentions to both characters
![Page 28: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/28.jpg)
A Dataset of Character Interactions
17
“If anyone was ever minding his business, it was I," Ignatius breathed. "Please. We must stop. I think I'm going to have a hemorrhage.” "Okay." Mrs. Reilly looked at her son's reddening face and realized that he would very happily collapse at her feet just to prove his point.”
t=0
“Ignatius belched the gas of a dozen brownies trapped by his valve. "Grant me a little peace.…” "You know I appreciate you, babe," Mrs. Reilly sniffed. "Come on and gimme a little goodbye kiss like a good boy.”
t=1
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
t=2
• For each pair of characters in a particular book, we extract all spans of text that contain mentions to both characters
![Page 29: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/29.jpg)
A Dataset of Character Interactions
• 1,383 novels from Project Gutenberg and other Internet sources • Genres represented include romance, mystery, and fantasy • Preprocessed with David Bamman’s BookNLP pipeline • Each span is a 200-token window centered around a
character mention
• 20,013 unique character pairs and 380,408 spans
18
![Page 30: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/30.jpg)
• recurrent autoencoder with dictionary learning
Relationship Modeling Network (RMN)
19
![Page 31: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/31.jpg)
• recurrent autoencoder with dictionary learning
Relationship Modeling Network (RMN)
20
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
“Ignatius belched the gas of a dozen brownies trapped by his valve. "Grant me a little peace.…”"You know I appreciate you, babe," Mrs. Reilly sniffed. "Come on and gimme a little goodbye kiss like a good boy.”
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Everyday I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
![Page 32: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/32.jpg)
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
“Ignatius belched the gas of a dozen brownies trapped by his valve. "Grant me a little peace.…”"You know I appreciate you, babe," Mrs. Reilly sniffed. "Come on and gimme a little goodbye kiss like a good boy.”
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Everyday I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
• recurrent autoencoder with dictionary learning
Relationship Modeling Network (RMN)
21
reconstruct inputs
![Page 33: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/33.jpg)
• recurrent autoencoder with dictionary learning
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
“Ignatius belched the gas of a dozen brownies trapped by his valve. "Grant me a little peace.…”"You know I appreciate you, babe," Mrs. Reilly sniffed. "Come on and gimme a little goodbye kiss like a good boy.”
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Everyday I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Relationship Modeling Network (RMN)
22
descriptor matrix
![Page 34: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/34.jpg)
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
“Ignatius belched the gas of a dozen brownies trapped by his valve. "Grant me a little peace.…”"You know I appreciate you, babe," Mrs. Reilly sniffed. "Come on and gimme a little goodbye kiss like a good boy.”
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Everyday I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
• recurrent autoencoder with dictionary learning
Relationship Modeling Network (RMN)
23
reconstruct inputs
![Page 35: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/35.jpg)
24
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with character
embeddings and book embeddings
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
![Page 36: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/36.jpg)
25
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with embeddings for characters and
books
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
![Page 37: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/37.jpg)
26
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with embeddings for characters and
books
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
![Page 38: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/38.jpg)
27
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with embeddings for characters and
books
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
linear interpolation
![Page 39: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/39.jpg)
28
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with embeddings for characters and
books
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
softmax activation
![Page 40: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/40.jpg)
29
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with embeddings for characters and
books
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
![Page 41: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/41.jpg)
30
rt = RTdt
Mrs. Reilly looked at her son slyly and asked, "Ignatius, you sure you not a communiss?" "Oh, my God!" Ignatius bellowed. "Every day I am subjected to a McCarthyite witchhunt in this crumbling building. No!"
Mrs. Reilly Ignatius “A Confederacy of Dunces”
ht = f(Wh · [vst ;vc1 ;vc2 ;vb])
vst vc1 vc2 vb
dt�1
R
dt = ↵ · softmax(Wd · [ht;dt�1])+
(1� ↵) · dt�1
: previous state
: descriptor matrix
: reconstruction of input span
: distribution over descriptors
6. make the reconstructed
vector close to the input span vector
1. word embedding
average
2. mix with embeddings for characters and
books
3. a recurrent connection that
copies over some of the previous hidden state
4. use a softmax activation function to make the hidden state a probability distribution over
descriptors…
5. multiply the hidden state by the descriptor
matrix to obtain a reconstruction of the
span vector
![Page 42: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/42.jpg)
Labeling the Learned Descriptors
• We compute the nearest word embeddings to each row of the descriptor matrix R, which humans use to provide external labels.
31
violence: grenades, guns, bullets
sadness: regretful, rueful, pity
politics: political, leadership, rule
fantasy: cosmic, realms, universe
suffering: fear, nightmares, suffer
![Page 43: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/43.jpg)
Relationship to Topic Models
• RMN outputs ≈ topic model latent variables: • descriptor matrix R ≈ topic-word matrices ϕ • descriptor weights dt at each timestep ≈
document-topic assignments z
• Baselines: • temporally-oblivious: LDA (Blei et al., 2001), Nubbi (Chang et al.,
2008) • temporally-aware: HTMM (Gruber et al., 2007)
32
![Page 44: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/44.jpg)
Experiment 1: Descriptor Coherence
33
![Page 45: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/45.jpg)
Do the Descriptors Make Sense?
• Goal: compare the descriptors learned by the RMN to the topics learned by our topic model baselines
• Task: word intrusion (Chang et al., 2009)
• Workers identify an “intruder” word from a set of words that—other than the intruder—come from the same descriptor
34
contempt malice condescend praise distaste mock
worship pray devote yourselves gods gather
![Page 46: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/46.jpg)
Do the Descriptors Make Sense?
• Goal: compare the descriptors learned by the RMN to the topics learned by our topic model baselines
• Task: word intrusion (Chang et al., 2009)
• Workers identify an “intruder” word from a set of words that—other than the intruder—come from the same descriptor
35
contempt malice condescend praise distaste mock
worship pray devote yourselves gods gather
![Page 47: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/47.jpg)
Do the Descriptors Make Sense?
• Goal: compare the descriptors learned by the RMN to the topics learned by our topic model baselines
• Task: word intrusion (Chang et al., 2009)
• Workers identify an “intruder” word from a set of words that—other than the intruder—come from the same descriptor
36
contempt malice condescend praise distaste mock
worship pray devote yourselves gods gather
![Page 48: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/48.jpg)
Do the Descriptors Make Sense?
37
![Page 49: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/49.jpg)
K=10 K=30 K=50
0.4
0.5
0.6
0.7
0.8
Mod
el P
reci
sion
LDA Nubbi HTMM GRMN RMN
Do the Descriptors Make Sense?
38
![Page 50: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/50.jpg)
Coherent Descriptors
39
HTMMcrime: blood knife pain legs stealfood: kitchen mouth glass food bread violence: sword shot blood shouted swung boats: ship boat captain deck crew outdoors: stone rock path darkness desert
RMNoutdoors: outdoors trail trails hillside grassy slopes sadness: regretful rueful pity pained despondent education: teaching graduate year teacher attended love: love delightful happiness enjoyed enjoyable murder: autopsy arrested homicide murdered
![Page 51: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/51.jpg)
Coherent Descriptors
40
HTMMcrime: blood knife pain legs stealfood: kitchen mouth glass food bread violence: sword shot blood shouted swung boats: ship boat captain deck crew outdoors: stone rock path darkness desert
RMNoutdoors: outdoors trail trails hillside grassy slopes sadness: regretful rueful pity pained despondent education: teaching graduate year teacher attended love: love delightful happiness enjoyed enjoyable murder: autopsy arrested homicide murdered
![Page 52: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/52.jpg)
Experiment 2: Trajectory Quality
41
![Page 53: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/53.jpg)
Visualizing Trajectories• for all time steps t,
compute argmax of dt and stack vertically
42
![Page 54: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/54.jpg)
• for all time steps t, compute argmax of dt and stack vertically
Visualizing Trajectories
43
time love death money crime0 0.95 0.01 0.03 0.011 0.8 0.01 0.18 0.012 0.4 0.01 0.5 0.093 0.3 0.01 0.2 0.54 0.2 0.7 0.05 0.05
![Page 55: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/55.jpg)
Visualizing Trajectories
44love
time
time love death money crime0 0.95 0.01 0.03 0.011 0.8 0.01 0.18 0.012 0.4 0.01 0.5 0.093 0.3 0.01 0.2 0.54 0.2 0.7 0.05 0.05
• for all time steps t, compute argmax of dt and stack vertically
![Page 56: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/56.jpg)
Visualizing Trajectories
45
love
time
time love death money crime0 0.95 0.01 0.03 0.011 0.8 0.01 0.18 0.012 0.4 0.01 0.5 0.093 0.3 0.01 0.2 0.54 0.2 0.7 0.05 0.05
• for all time steps t, compute argmax of dt and stack vertically
![Page 57: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/57.jpg)
Visualizing Trajectories
46
moneylove
time
time love death money crime0 0.95 0.01 0.03 0.011 0.8 0.01 0.18 0.012 0.4 0.01 0.5 0.093 0.3 0.01 0.2 0.54 0.2 0.7 0.05 0.05
• for all time steps t, compute argmax of dt and stack vertically
![Page 58: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/58.jpg)
Visualizing Trajectories
47
moneylove
time
time love death money crime0 0.95 0.01 0.03 0.011 0.8 0.01 0.18 0.012 0.4 0.01 0.5 0.093 0.3 0.01 0.2 0.54 0.2 0.7 0.05 0.05
crime
• for all time steps t, compute argmax of dt and stack vertically
![Page 59: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/59.jpg)
Visualizing Trajectories
48
death
moneylove
time
time love death money crime0 0.95 0.01 0.03 0.011 0.8 0.01 0.18 0.012 0.4 0.01 0.5 0.093 0.3 0.01 0.2 0.54 0.2 0.7 0.05 0.05
crime
• for all time steps t, compute argmax of dt and stack vertically
![Page 60: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/60.jpg)
Do the Trajectories Make Sense?
• We crawl Wikipedia and SparkNotes for summaries
• Removing uninformative summaries results in 125 character pairs to evaluate
• Workers prefer the RMN to the HTMM for 87 out of the 125 relationships (69.6%, Fleiss κ=0.32)
49
In this task, you will be comparing two timelines of how a relationship between a pair of literary characters changes over time. We will provide you with a summary of the relationship, and your job is to select which of the two timelines (A or B) better captures the content of the summary.
![Page 61: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/61.jpg)
50
Summary: Govinda is Siddhartha’s best friend and sometimes his follower. Like Siddhartha, Govinda devotes his life to the quest for understanding and enlightenment. He leaves his village with Siddhartha to join the Samanas, then leaves the Samanas to follow Gotama. He searches for enlightenment independently of Siddhartha but persists in looking for teachers who can show him the way. In the end, he is able to achieve enlightenment only because of Siddhartha’s love for him.
![Page 62: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/62.jpg)
Qualitative Analysis: Good and Bad Trajectories
51
![Page 63: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/63.jpg)
52
TIM
EHTMMRMN
Storm Island: David and Lucy
HTMMRMN
A Tale of Two Cities: Darnay and Lucie
HTMMRMN
Dracula: Arthur and Lucy
love love
sadness
joy
love
fantasy
love
fantasy
sickness
death
sadness
death
marriage
sickness
murder
passage of time
I love him more than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
poor girl, there is peace for her at last. It is the end!
Arthur placed the stake over her heart … he struck with all his might. The Thing in the coffin writhed …
Arthur and Lucy “ground-truth”: marriage -> sickness -> death -> murder
![Page 64: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/64.jpg)
53
TIM
E
HTMMRMN
Storm Island: David and Lucy
HTMMRMN
A Tale of Two Cities: Darnay and Lucie
HTMMRMN
Dracula: Arthur and Lucy
love love
sadness
joy
love
fantasy
love
fantasy
sickness
death
sadness
death
marriage
sickness
murder
passage of time
I love him more than ever. We are to be married on 28 September.
I feel so weak and worn out … looked quite grieved … I hadn't the spirit
poor girl, there is peace for her at last. It is the end!
Arthur placed the stake over her heart … he struck with all his might. The Thing in the coffin writhed …
Arthur and Lucy “ground-truth”: marriage -> sickness -> death -> murder
learned trajectories:
![Page 65: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/65.jpg)
54
TIM
E
HTMMRMN
Storm Island: David and Lucy
HTMMRMN
A Tale of Two Cities: Darnay and Lucie
HTMMRMN
Dracula: Arthur and Lucy
Event-based similarities between the two models
The RMN is led astray by the novel’s sad tone
![Page 66: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/66.jpg)
55
TIM
E
HTMMRMN
Storm Island: David and Lucy
HTMMRMN
A Tale of Two Cities: Darnay and Lucie
HTMMRMN
Dracula: Arthur and Lucy
Event-based similarities between the two models
The RMN is led astray by the novel’s sad tone
![Page 67: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/67.jpg)
Qualitative Analysis: Using Existing Datasets
56
![Page 68: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/68.jpg)
What Makes a Good Relationship?• Dataset of Massey et al. (2015) has affinity
annotations for relationships in Project Gutenberg • 120 non-neutral relationships are also present in our dataset
57
![Page 69: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/69.jpg)
What Makes a Good Relationship?• Dataset of Massey et al. (2015) has affinity
annotations for relationships in Project Gutenberg • 120 non-neutral relationships are also present in our dataset
58
love death sadness0.9 0.05 0.050.8 0.1 0.10.6 0.3 0.10.7 0.1 0.20.8 0.1 0.1
love death sadness0.1 0.7 0.20.2 0.3 0.50.15 0.25 0.60.05 0.65 0.30.1 0.2 0.7
positive negative
Don Quixote & Sancho PanzaCandide & CunégondeAnna Karenina & Vronsky…
Dracula & Jonathan HarkerDr. Jekyll & Mr. HydeHester Prynne & Chillingworth…
![Page 70: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/70.jpg)
What Makes a Good Relationship?• Dataset of Massey et al. (2015) has affinity
annotations for relationships in Project Gutenberg • 120 non-neutral relationships are also present in our dataset
59
love death sadness0.9 0.05 0.050.8 0.1 0.10.6 0.3 0.10.7 0.1 0.20.8 0.1 0.1
0.76 0.13 0.11
love death sadness0.1 0.7 0.20.2 0.3 0.50.15 0.25 0.60.05 0.65 0.30.1 0.2 0.70.12 0.42 0.46
positive negative
average the positive and negative trajectories
![Page 71: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/71.jpg)
What Makes a Good Relationship?• Dataset of Massey et al. (2015) has affinity
annotations for relationships in Project Gutenberg • 120 non-neutral relationships are also present in our dataset
60
love death sadness0.9 0.05 0.050.8 0.1 0.10.6 0.3 0.10.7 0.1 0.20.8 0.1 0.1
0.76 0.13 0.11
love death sadness0.1 0.7 0.20.2 0.3 0.50.15 0.25 0.60.05 0.65 0.30.1 0.2 0.70.12 0.42 0.46
positive negative
1. love 2. death 3. sadness
1. sadness 2. death 3. love
![Page 72: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/72.jpg)
61
RMNeducation
love religion
sex
HTMMlove
parental business outdoors
Most Positive Descriptors
![Page 73: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/73.jpg)
62
RMNeducation
love religion
sex
RMNpolitics murder sadness royalty
HTMMlove
parental business outdoors
HTMMlove
politics violence
crime
Most Positive Descriptors
Most Negative Descriptors
![Page 74: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/74.jpg)
63
RMNeducation
love religion
sex
RMNpolitics murder sadness royalty
HTMMlove
parental business outdoors
HTMMlove
politics violence
crime
Most PositiveDescriptors
Most Negative Descriptors
![Page 75: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/75.jpg)
Why is Politics Negative?• Both models rank politics as highly negative • The affinity data we look at comes primarily from
Victorian-era authors (e.g., Charles Dickens and George Eliot)
64
Victorian-era authors are “obsessed with otherness… of antiquated social and legal institutions, and of autocratic and/or dictatorial abusive government” (Zarifopol-Johnston,1995)
![Page 76: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/76.jpg)
Areas for Improvement• Difficult to evaluate unsupervised relationship
modeling, requires considerable human effort
• Our data processing leaves out a lot of information • e.g., spans of text in which one but not both characters in a
relationship are mentioned • only considers undirected relationships between pairs
• Model performance is directly tied to the quality of character disambiguation and coreference resolution • e.g., first person pronouns
65
![Page 77: Feuding Families and Former Friends: Unsupervised Learning ...](https://reader033.fdocuments.in/reader033/viewer/2022053012/629134557b0f9b726c573785/html5/thumbnails/77.jpg)
Recap• Introduced the task of unsupervised relationship
modeling as well as an interpretable neural network architecture, the RMN, for this task
• Found that the RMN generates higher quality descriptors and more interpretable trajectories than topic model baselines
• Future work: collaborate with humanities researchers to help answer literary questions with the RMN
66