Controlling Linguistic Style Aspects in Neural Language Generation - - PowerPoint PPT Presentation

Controlling Linguistic Style Aspects in Neural Language Generation

Jessica Ficler and Yoav Goldberg

ISCOL 2017

Controlling Linguistic Style Aspects in Neural Language Generation

Jessica Ficler and Yoav Goldberg

ISCOL 2017

Our goal is to generate text… …while allowing control of its style.

Style

The same message (e.g. expressing a positive sentiment towards a movie) can be conveyed in different ways.

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style Personal voice

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style Personal voice Few adjectives

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style Personal voice Few adjectives “A genuinely unique, full-on sensory experience that treads its own path between narrative clarity and pure visual expression.”

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style Personal voice Few adjectives “A genuinely unique, full-on sensory experience that treads its own path between narrative clarity and pure visual expression.” Professional critic

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style Personal voice Few adjectives “A genuinely unique, full-on sensory experience that treads its own path between narrative clarity and pure visual expression.” Professional critic Impersonal voice

Style Aspects (Example)

“OMG... This movie actually made me cry a little bit because I laughed so hard at some parts.“ Colloquial style Personal voice Few adjectives “A genuinely unique, full-on sensory experience that treads its own path between narrative clarity and pure visual expression.” Professional critic Impersonal voice Many adjectives

Style Aspects (Example)

The challenge

Generate text that conforms to a set of content-based and stylistic requirements.

The challenge

Generate text that conforms to a set of content-based and stylistic requirements. full length, natural sentences

The challenge

Generate text that conforms to a set of content-based and stylistic requirements. more than 2 full length, natural sentences

Example

Theme: Acting Descriptive: True

Example

Theme: Acting Descriptive: True “A wholly original, well-acted, romantic comedy that's elevated by the modest talents

• f a lesser known cast.”

Example

Theme: Acting Descriptive: True “A wholly original, well-acted, romantic comedy that's elevated by the modest talents

• f a lesser known cast.”

Example

Theme: Acting Descriptive: True “A wholly original, well-acted, romantic comedy that's elevated by the modest talents

• f a lesser known cast.”

Example

Theme: Acting Descriptive: True “A wholly original, well-acted, romantic comedy that's elevated by the modest talents

• f a lesser known cast.”

Theme: Plot Descriptive: False “I think the poor writing and script are what caused this movie to bomb.”

Formal Definition

• We assume a set of k parameters 𝑞" … 𝑞%, each parameter 𝑞& with a

set of possible values 𝑊()

Formal Definition

• We assume a set of k parameters 𝑞" … 𝑞%, each parameter 𝑞& with a

set of possible values 𝑊()

• Input: specific assignment to these parameters

e.g.

Value Parameter False Professional True Personal ≤ 10 Length False Descriptive Other Theme Positive Sentiment

Formal Definition

• We assume a set of k parameters 𝑞" … 𝑞%, each parameter 𝑞& with a

set of possible values 𝑊()

• Input: specific assignment to these parameters

e.g. Output: a text that is compatible with the parameters values

Value Parameter False Professional True Personal ≤ 10 Length False Descriptive Other Theme Positive Sentiment

e.g. “I don't understand why it is rated so poorly.”

This work

We consider 6 parameters and values from the movie reviews domain

Content Style Sentiment Theme Professional Personal Descriptive Length

Content Parameters

Task Description – Content Parameters

Sentiment - The score that the reviewer gave the movie

Positive Neutral

“While the film doesn't quite reach the level of sugar fluctuations, it's beautifully animated.” “This movie is so much to keep you on the edge of your seat.”

Negative

“It’s a very low-budget movie that just seems to be a bunch

• f fluff.”

Task Description – Content Parameters

Theme - Whether the sentence's content is about the Plot, Acting, Production, Effects or none

• f these (Other)

Task Description – Content Parameters

Theme - Whether the sentence's content is about the Plot, Acting, Production, Effects or none

• f these (Other)

Plot - “The storyline had me laughing out loud.”

Task Description – Content Parameters

Theme - Whether the sentence's content is about the Plot, Acting, Production, Effects or none

• f these (Other)

Acting - “The cast are all excellent.” Plot - “The storyline had me laughing out loud.”

Task Description – Content Parameters

Theme - Whether the sentence's content is about the Plot, Acting, Production, Effects or none

• f these (Other)

Production - “The director's magical.” Acting - “The cast are all excellent.” Plot - “The storyline had me laughing out loud.”

Task Description – Content Parameters

Theme - Whether the sentence's content is about the Plot, Acting, Production, Effects or none

• f these (Other)

Effects - “Only saving grace is the sound effects.” Production - “The director's magical.” Acting - “The cast are all excellent.” Plot - “The storyline had me laughing out loud.”

Task Description – Content Parameters

Theme - Whether the sentence's content is about the Plot, Acting, Production, Effects or none

• f these (Other)

Effects - “Only saving grace is the sound effects.” Other - “I'm afraid that the movie is aimed at kids and adults weren't sure what to say about it.” Production - “The director's magical.” Acting - “The cast are all excellent.” Plot - “The storyline had me laughing out loud.”

Style Parameters

Task Description – Style Parameters

Length – Number of words

≤ 10 words 11-20 words 21-40 words > 40 words

Task Description – Style Parameters

Professional - Whether the review is written in the style of a professional critic or not

Task Description – Style Parameters

Professional - Whether the review is written in the style of a professional critic or not

True

“This is a breath of fresh air, it's a welcome return to the franchise's brand of satirical humor.”

Task Description – Style Parameters

Professional - Whether the review is written in the style of a professional critic or not

True False

“So glad to see this movie !!” “This is a breath of fresh air, it's a welcome return to the franchise's brand of satirical humor.”

Task Description – Style Parameters

Personal - Whether the review describes subjective experience (written in personal voice) or not

Task Description – Style Parameters

Personal - Whether the review describes subjective experience (written in personal voice) or not

True

“I could see the movie again”

Task Description – Style Parameters

Personal - Whether the review describes subjective experience (written in personal voice) or not

True False

“Very similar to the book.” “I could see the movie again”

Task Description – Style Parameters

Descriptive - Whether the review is in descriptive (contains a high ratio of adjectives) style or not

Task Description – Style Parameters

True

“Such a hilarious and funny romantic comedy.”

Descriptive - Whether the review is in descriptive (contains a high ratio of adjectives) style or not

Task Description – Style Parameters

True False

“A definite must see for fans of anime fans, pop culture references and animation with a good laugh too.” “Such a hilarious and funny romantic comedy.”

Descriptive - Whether the review is in descriptive (contains a high ratio of adjectives) style or not

And we would like to control for all these aspects si simul ultani niousl usly

Value Parameter Type False Professional Style True Personal Style ≤ 10 Length Style False Descriptive Style Other Theme Content Positive Sentiment Content

“I don't understand why it is rated so poorly.”

Model

Model

a conditioned language model:

𝑄 𝑥" … 𝑥- 𝑑 = 0 𝑄(𝑥2|𝑥", … , 𝑥25" , 𝑑)

• 27"

Model

a conditioned language model: Condition each word on the history, as well as on a context c.

𝑄 𝑥" … 𝑥- 𝑑 = 0 𝑄(𝑥2|𝑥", … , 𝑥25" , 𝑑)

• 27"

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start An

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start An An

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start An An

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start An An entertaining

Model

In our case, c is a concatenation of the parameters values embedding vectors

c:

Theme:Plot Proffesional:True Descriptive:True Length:≤10 Sentiment:Positive Personal:False

start An An entertaining entertaining and and visually visually attractive attractive family- friendly family-friendly story story .

The model is simple, but…

we need training data annotated with the appropriate values.

Text Parameters extract

Text Parameters extract

Meta data Heuristics

Text Parameters extract train

Meta data Heuristics

Text Parameters extract train

Meta data Heuristics

Rotten-Tomatoeswebsite. 7,500 movies. 1,002,625 movie reviews.

Text Parameters extract train

Meta data Heuristics

Rotten-Tomatoeswebsite. 7,500 movies. 1,002,625 movie reviews.

Professional

Professional

In rottentomatoes the critic reviews are separated from the audience review

Professional

In rottentomatoes the critic reviews are separated from the audience review

Professional Non Professional

Some of the non-professional reviewers are considered as “super reviewers”

Also professional

Sentiment

Sentiment

Sentiment scores

Sentiment

Sentiment

We normalized the critics scores to be on 0-5 scale Negative 0-2 Neutral 3 Positive 4-5

Text Parameters extract train

Meta data Heuristics

Rotten-Tomatoeswebsite. 7,500 movies. 1,002,625 movie reviews.

Text Parameters extract train

Meta data Heuristics

Rotten-Tomatoeswebsite. 7,500 movies. 1,002,625 movie reviews.

Content words Function words POS tags

Theme

Content words

Effects Production

Director Directed Production co-production

Acting

Acting Cast Performance Play Role Miscasting Actor

Plot

Story Storytelling Plot Script Manuscript Tale Scene Effects Song Music Voice Visual Soundtrack Shot

To determine the value for the theme parameter we searched for words that are related to the 4 topics and are common in our data set

Theme

Theme

Content words

Effects Production

Director Directed Production co-production

Acting

Acting Cast Performance Play Role Miscasting Actor

Plot

Story Storytelling Plot Script Manuscript Tale Scene Effects Song Music Voice Visual Soundtrack Shot

Each sentence was labeled with the category that has the most words in the sentence. Sentences that do not include any words from our lists are labeled as other To determine the value for the theme parameter we searched for words that are related to the 4 topics and are common in our data set

Theme

Personal Voice

Personal

True I My False Other cases To determine weather a review is written in personal voice we search for words that express subjectivity

Personal Pronouns

Descriptiveness

We assume that descriptive texts make heavy use of adjectives

True % JJ ≥35 False Other cases

Distribution of part-of-speech tags Descriptive

Length

Length

≤ 10 words 11-20 words 21-40 words > 40 words

Dataset Statistics

Our final data-set includes 2,773,435 sentences We divided the data set to training (~2.7M), development (~2K) and test (~2K) sets Each sentence is labeled with the 6 parameters

Parameters Values Text

easy

Parameters Values Text

easy

Text Parameters Values

easy hard

Text Parameters Values

extract hard

Text Conditioned Language Model Parameters Values

extract

Text Conditioned Language Model Does this work? Parameters Values

extract

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.”

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.” “The film’s simple, and a refreshing take on the complex family drama of the regions of human intelligence.” Value Parameter True Professional False Personal 11-20 Length False Descriptive Other Theme Positive Sentiment

Examples of Generated Sentences

Value Parameter False Professional True Personal 11-20 Length True Descriptive Other Theme Negative Sentiment “Ultimately, I can honestly say that this movie is full of stupid stupid and stupid stupid stupid stupid stupid.” “The film’s simple, and a refreshing take on the complex family drama of the regions of human intelligence.” Value Parameter True Professional False Personal 11-20 Length False Descriptive Other Theme Positive Sentiment

We would like to quantitatively measure our model capabilities.

Evaluation

• Evaluating LM Quality (Perplexity)
• Evaluating the Generated Sentences

Evaluating LM Quality

Sanity Check

• 1. Conditioned vs. Unconditioned

Does knowing the parameters indeed helps in achieving better language modeling results?

Sanity Check

• 1. Conditioned vs. Unconditioned

Does knowing the parameters indeed helps in achieving better language modeling results?

Test Dev 24.4 25.8 Not-conditioned 23.3 24.8 Conditioned

Knowing the correct parameter values indeed results in better perplexity!

Baseline

• 2. Conditioned vs. Dedicated LMs

Is our model effective comparing to train a separate unconditioned LM

• n subset of the data (dedicated LM)?

Baseline

• 2. Conditioned vs. Dedicated LMs

Is our model effective comparing to train a separate unconditioned LM

• n subset of the data (dedicated LM)?

Data Set

Baseline

• 2. Conditioned vs. Dedicated LMs

Is our model effective comparing to train a separate unconditioned LM

• n subset of the data (dedicated LM)?

Data Set

Sentiment:Positive Sentiment:Neutral Sentiment:Negative

when generating text, we would choose the model that corresponds to the requested

When generating text, we would choose the model that corresponds to the requested value

Baseline

• 2. Conditioned vs. Dedicated LMs

Is our model effective comparing to train a separate unconditioned LM

• n subset of the data (dedicated LM)?

Data Set

Sentiment:Positive; Proffesional:True Sentiment:Positive; Proffesional:False Sentiment:Neutral; Proffesional:True Sentiment:Neutral; Proffesional:False Sentiment:Negative; Proffesional:True Sentiment:Negative; Proffesional:False The number of models that need to be trained depends on the number of parameters and the possible values

Baseline

• 2. Conditioned vs. Dedicated LMs

Is our model effective comparing to train a separate unconditioned LM

• n subset of the data (dedicated LM)?

Data Set

Sentiment:Positive; Proffesional:True; Theme:Other; Personal:False; Length:21-40; Descriptive:False Sentiment:Negative; Proffesional:False; Theme:Other; Personal:True; Length:21-40; Descriptive:False

. . . .

240

Evaluation (Language Model Quality)

We hypothesize that the conditioned LM will be able to:

Evaluation (Language Model Quality)

We hypothesize that the conditioned LM will be able to:

• Generalize across properties-combinations

Evaluation (Language Model Quality)

We hypothesize that the conditioned LM will be able to:

• Generalize across properties-combinations
• Share data between the different settings

Evaluation (Language Model Quality)

We hypothesize that the conditioned LM will be able to:

• Generalize across properties-combinations
• Share data between the different settings

And thus will be more effective than a dedicated LM

Evaluation (Language Model Quality)

We verify this hypothesis by training dedicated models and compare their results on the corresponding data to the results achieved by our model

Baseline

For a set of parameters and values 𝑞" … 𝑞-, we train n sub-models Each sub model 𝑛& is trained on the subset of sentences that match parameters 𝑞" … 𝑞&

Baseline

For a set of parameters and values 𝑞" … 𝑞-, we train n sub-models Each sub model 𝑛& is trained on the subset of sentences that match parameters 𝑞" … 𝑞& Example - given the set of parameters values: personal:false, sentiment:pos, professional:false, theme:other and length:≤10 we train 5 sub-models:

1. personal:false 2. persoal:false and sentiment:positive 3. persoal:false, sentiment:positive and professional:false 4. persoal:false, sentiment:positive, professional:false and theme:other 5. persoal:false, sentiment:positive, professional:false, theme:other and length:≤10

Baseline

For a set of parameters and values 𝑞" … 𝑞-, we train n sub-models Each sub model 𝑛& is trained on the subset of sentences that match parameters 𝑞" … 𝑞& Example - given the set of parameters values: personal:false, sentiment:pos, professional:false, theme:other and length:≤10 we train 5 sub-models:

1. personal:false 2. persoal:false and sentiment:positive 3. persoal:false, sentiment:positive and professional:false 4. persoal:false, sentiment:positive, professional:false and theme:other 5. persoal:false, sentiment:positive, professional:false, theme:other and length:≤10

As we add parameters, the size of the training set of the sub-model decreases.

Baseline

We measure the perplexity of the dedicated models on the test-set sentences that match the criteria and compare it to our conditioned LM and to an unconditioned language model. We do this for 4 different parameter-sets.

Evaluation (Language Model Quality)

The dedicated model achieves better perplexity than our model on data with personal:false

Evaluation (Language Model Quality)

The dedicated model achieves better perplexity than our model on data with personal:false The gap is getting smaller as the dedicated model includes more properties

Evaluation (Language Model Quality)

The dedicated model achieves better perplexity than our model on data with personal:false The gap is getting smaller as the dedicated model includes more properties Eventually the conditioned model result is better than the dedicated model result

Evaluation (Language Model Quality)

This is the case also in the other 3 sets that were experimented

Evaluation (Language Model Quality)

The dedicated LM scores are better than our model when:

• Only few conditioning parameters are needed
• The coverage of the parameter combination in the training set is large enough

Evaluation (Language Model Quality)

The dedicated LM scores are better than our model when:

• Only few conditioning parameters are needed
• The coverage of the parameter combination in the training set is large enough

We conclude that the conditioned model manages to generalize from sentences with different sets of properties, and is effective also with large number of conditioning factors.

Evaluation (Language Model Quality)

• 3. Conditioned vs. Flipped Conditioning

How effective are the conditioning parameters individually?

Evaluation (Language Model Quality)

• 3. Conditioned vs. Flipped Conditioning

How effective are the conditioning parameters individually? We compare the perplexity when using the correct conditioning values to the perplexity achieved when flipping the parameter value to an incorrect one.

Evaluation (Language Model Quality)

23.3 Correct Value 27.2 Replacing Descriptive with non-Descriptive 27.5 Replacing Personal 25 Replacing Professional 24.3 Replacing Sentiment Pos with Neg The model distinguishes descriptive text and personal voice better than it distinguishes sentiment and professional text.

Evaluating the Generated Sentences

Evaluation (Generated Sentences)

How well sentences generated by the model match the requested behavior (conditioning properties)?

Evaluation (Generated Sentences)

• 1. Capturing Individual Properties

For each parameter, we measure the correspondence of the sentences to the requested values.

Evaluation (Generated Sentences)

• 1. Capturing Individual Properties

Length

Max Min Avg Requested Length 21 1 7.6 <=10 25 5 20.6 11-20 49 7 34 21-40

Evaluation (Generated Sentences)

Descriptive descriptive:true – 85.7% descriptive descriptive:false – 96% non-descriptive

We measure the percentage of sentences that are considered as descriptive when requesting descriptive:true, and when requesting descriptive:false

Evaluation (Generated Sentences)

Personal personal:true – 100% personal personal:false – 99.85% non-personal

We measure the percentage of sentences that are considered as personal voice when requesting personal:true, and when requesting personal:false

Evaluation (Generated Sentences)

Theme

% Other % Effects % Prod % Acting % Plot Requested value 0.3 0.2 0.8 98.7 Plot 1.6 0.6 95.3 2.5 Acting 2.6 97.4 Production 2.4 91.7 5.9 Effects 99.9 0.03 0.03 0.04 Other

For each of the possible theme values, we compute the proportion of the sentences that were generated with the corresponding value. The confusion shows that the majority of sentences are generated according to the requested theme

Evaluation (Generated Sentences)

Professional The professional property could not be evaluated automatically We performed manual evaluation using Mechanical Turk

Evaluation (Generated Sentences)

Professional The professional property could not be evaluated automatically We performed manual evaluation using Mechanical Turk Can a person distinguish professional:true from professional:false?

Evaluation (Generated Sentences)

Professional The professional property could not be evaluated automatically We performed manual evaluation using Mechanical Turk Can a person distinguish professional:true from professional:false? We randomly created 1000 sentence-pairs:

• professional:true
• professional:false

Evaluation (Generated Sentences)

(t) “This film has a certain sense of imagination and a sobering look at the clandestine indictment.” (f) “I know it’s a little bit too long, but it’s a great movie to watch !!!!” Which of the sentences was written by a professional critic?

Evaluation (Generated Sentences)

Settings 5 different annotators. majority vote. Result The annotators were able to tell apart the professional from non- professional sentences generated sentences in 72.1% of the cases.

Evaluation (Generated Sentences)

Analysis In a few cases the sentence that was generated for professional:true was indeed not professional enough “Looking forward to the trailer.”

Evaluation (Generated Sentences)

In many cases, both sentences could indeed be considered as either professional or not Example: (t) “This is a cute movie with some funny moments, and some of the jokes are funny and entertaining.” (f) “Absolutely amazing story of bravery and dedication.”

Evaluation (Generated Sentences)

Sentiment Manual annotations using Mechanical Turk We randomly created 300 pairs of generated sentences for each of the following settings:

• positive/negative
• positive/neutral
• negative/neutral.

Evaluation (Generated Sentences)

Sentiment Manual annotations using Mechanical Turk We randomly created 300 pairs of generated sentences for each of the following settings:

• positive/negative
• positive/neutral
• negative/neutral.

Which of the reviewers liked the movie more than the other?

Evaluation (Generated Sentences)

Settings 5 different annotators. Majority vote. Results

86.3% Positive/Negative 63% Positive/Neutral 69.7% Negative/Neutral

Evaluation (Generated Sentences)

Examples where the intended sentiment was not recognized by the annotators:

(Pos) “It’s a shame that this film is not as good as the previous film, but it still delivers.” (Neg) “The premise is great, the acting is not bad, but the special effects are so bad.”

Evaluation (Generated Sentences)

• 2. Generalization Ability

Can the model generate sentences for parameter combinations it has not seen in training?

Evaluation (Generated Sentences)

theme:plot and personal:true

75,421

We removed from the training set about ~75K sentences which were labeled as theme:plot and personal:true

336,567 477,738

Evaluation (Generated Sentences)

theme:plot and personal:true

75,421

We removed from the training set about ~75K sentences which were labeled as theme:plot and personal:true

336,567 477,738

We don’t train

• n these

Evaluation (Generated Sentences)

We then asked the trained model to generate sentences with theme:plot and personal:true

Evaluation (Generated Sentences)

We then asked the trained model to generate sentences with theme:plot and personal:true Results 100% of the generated sentences indeed contained personal pronouns 82.4% of them fit the theme:plot criteria (The result achieved by the full model is 97.8%)

Evaluation (Generated Sentences)

Examples: “Some parts weren’t as good as I thought it would be and the acting and script were amazing.” “I really liked the story and the performances were likable and the chemistry between the two leads is great.”

Comparison to Previous Work

Comparison to Previous Work

Most work focus on content that need to be conveyed in the generated text

• Reviews generation conditioned on category and numeric rating

scores (Lipton et al., 2015; Tang et al., 2016)

• Dialog generation conditioned on a dialog act (“request”, “inform”)

and information to be conveyed (“price=low, food=italian, near=citycenter”) (Wen et al., 2015; Dusek and Jurcicek , 2016b,a)

Comparison to Previous Work

• Recipe generation conditioned on a list of ingredients (Kiddon et

al.,2016)

• Textual biographies generation conditioned on Wikipedia infoboxes

(Lebret et al.,2016)

Comparison to Previous Work

Generation conditioned on style:

• In dialog generation, Li et al. (2016) condition the text on the

speaker’s identity (age, gender, location) for improving the factual consistency of the utterances

• In Machine Translation, Sennrich et al. (2016a) model translates

English to German with a feature that encodes whether the generated text (in German) should express politeness.

Comparison to Previous Work

• Hu et al. (2017) that tackles the same problem as ours:

conditioning on multiple aspects of the generated text. Their model features a VAE based method coupled with a discriminator network. Hu et al. (2017) restrict themselves to sentences of up to length 16, and only two conditioning aspects (sentiment and tense).

Summary

• Most work on neural natural language generation focus on controlling

the content of the generated text

• We experiment with controlling several stylistic aspects of the

generated text, in addition to its content

• The method is based on conditioned RNN language model
• Simple but very effective!
• We demonstrate the approach on the movie reviews domain
• We show that it is successful in generating coherent sentences

corresponding to the required linguistic style and content