# multipleattribute_text_rewriting__e29cdba2.pdf

Published as a conference paper at ICLR 2019

MULTIPLE-ATTRIBUTE TEXT REWRITING

Guillaume Lample 1,3, Sandeep Subramanian 1,2, Eric Michael Smith1, Ludovic Denoyer1,3, Marc Aurelio Ranzato1, Y-Lan Boureau1

1Facebook AI Research, 2MILA, Universit e de Montr eal 3Sorbonne Universit es, UPMC Univ Paris 06 sandeep.subramanian.1@umontreal.ca {glample,ems,denoyer,ranzato,ylan}@fb.com

The dominant approach to unsupervised style transfer in text is based on the idea of learning a latent representation, which is independent of the attributes specifying its style . In this paper, we show that this condition is not necessary and is not always met in practice, even with domain adversarial training that explicitly aims at learning such disentangled representations. We thus propose a new model that controls several factors of variation in textual data where this condition on disentanglement is replaced with a simpler mechanism based on back-translation. Our method allows control over multiple attributes, like gender, sentiment, product type, etc., and a more fine-grained control on the trade-off between content preservation and change of style with a pooling operator in the latent space. Our experiments demonstrate that the fully entangled model produces better generations, even when tested on new and more challenging benchmarks comprising reviews with multiple sentences and multiple attributes.

1 INTRODUCTION

One of the objectives of unsupervised learning is to learn representations of data that enable fine control over the underlying latent factors of variation, e.g., pose and viewpoint of objects in images, or writer style and sentiment of a product review. In conditional generative modeling, these latent factors are given (Sohn et al., 2015; Mirza & Osindero, 2014; Ficler & Goldberg, 2017), or automatically inferred via observation of samples from the data distribution (Chen et al., 2017; 2016; Higgins et al., 2017).

More recently, several studies have focused on learning unsupervised mappings between two data domains such as images (Taigman et al., 2016; Isola et al., 2017; Zhu et al., 2017), words or sentences from different languages (Conneau et al., 2017; Lample et al., 2018).

In this problem setting, the generative model is conditioned not only on the desired attribute values, but also on a initial input, which it must transform. Generations should retain as many of the original input characteristics as possible, provided the attribute constraint is not violated. This learning task is typically unsupervised because no example of an input and its corresponding output with the specified attribute is available during training. The model only sees random examples and their attribute values.

The dominant approach to learn such a mapping in text is via an explicit constraint on disentanglement (Hu et al., 2017; Fu et al., 2017; Shen et al., 2017): the learned representation should be invariant to the specified attribute, and retain only attribute-agnostic information about the content . Changing the style of an input at test time then amounts to generating an output based on the disentangled latent representation computed from the input and the desired attributes. Disentanglement is often achieved through an adversarial term in the training objective that aims at making the attribute value unrecoverable from the latent representation.

This paper aims to extend previous studies on style transfer along three axes. (i) First, we seek to gain a better understanding of what is necessary to make things work, and in particular, whether

Authors contributed equally to this work

Published as a conference paper at ICLR 2019

Relaxed Annoyed

Relaxed Sitting by the Christmas tree and watching Star Wars after cooking dinner. What a nice night

Annoyed Sitting by the computer and watching The Voice for the second time tonight. What a horrible way to start the weekend

Annoyed Getting a speeding ticket 50 feet in front of work is not how I wanted to start this month

Relaxed Getting a haircut followed by a cold foot massage in the morning is how I wanted to start this month

Male Female

Male Gotta say that beard makes you look like a Viking... Female Gotta say that hair makes you look like a Mermaid...

Female Awww he s so gorgeous

can t wait for a cuddle. Well done

xxx Male Bro he s so f***ing dope can t wait for a cuddle. Well done bro

Age 18-24 65+

18-24 You cheated on me but now I know nothing about loyalty

ok 65+ You cheated on America but now I know nothing about patriotism. So ok.

65+ Ah! Sweet photo of the sisters. So happy to see them together today . 18-24 Ah

happy to see them together today

Table 1: Our approach can be applied to many different domains beyond sentiment flipping, as illustrated here with example re-writes by our model on public social media content. The first line in each box is an input provided to the model with the original attribute, followed by its rewrite when given a different attribute value.

disentanglement is key, or even actually achieved by an adversarial loss in practice. In Sec. 3.1 we provide strong empirical evidence that disentanglement is not necessary to enable control over the factors of variation, and that even a method using adversarial loss to disentangle (Fu et al., 2017) does not actually learn representations that are disentangled. (ii) Second, we introduce a model which replaces the adversarial term with a back-translation (Sennrich et al., 2015a) objective which exposes the model to a pseudo-supervised setting, where the model s outputs act as supervised training data for the ultimate task at hand. The resulting model is similar to recently proposed methods for unsupervised machine translation (Lample et al., 2017a; 2018; Artetxe et al., 2018; Zhang et al., 2018b), but with two major differences: (a) we use a pooling operator which is used to control the trade-off between style transfer and content preservation; and (b) we extend this model to support multiple attribute control. (iii) Finally, in Sec. 4.1 we point out that current style transfer benchmarks based on collections of user reviews have severe limitations, as they only consider a single attribute control (sentiment), and very small sentences in isolation with noisy labels. To address this issue, we propose a new set of benchmarks based on existing review datasets, which comprise full reviews, where multiple attributes are extracted from each review.

The contributions of this paper are thus: (1) a deeper understanding of the necessary components of style transfer through extensive experiments, resulting in (2) a generic and simple learning framework based on mixing a denoising auto-encoding loss with an online back-translation technique and a novel neural architecture combining a pooling operator and support for multiple attributes, and (3) a new, more challenging and realistic version of existing benchmarks which uses full reviews and multiple attributes per review, as well as a comparison of our approach w.r.t. baselines using both new metrics and human evaluations. We will open-source our code and release the new benchmark datasets used in this work, as well as our pre-trained classifiers and language models for reproducibility. This will also enable fair empirical comparisons on automatic evaluation metrics in future work on this problem.

2 RELATED WORK

There is substantial literature on the task of unsupervised image translation. While initial approaches required supervised data of the form (input, transformation, output), e.g., different images of the same object rendered with different viewpoints or/and different lighting conditions (Hinton et al., 2011; Yang et al., 2015; Kulkarni et al., 2015), current techniques are capable of learning completely unsupervised domain mappings. Given images from two different domains X and Y (where X could be the domain of paintings and Y the domain of realistic photographs), and the task is to learn two mappings F : X Y and G : Y X, without supervision, i.e., just based on images sampled from the two domains (Liu & Tuzel, 2016; Taigman et al., 2016; Isola et al., 2017). For instance, Zhu et al. (2017) used a cycle consistency loss to enforce F(G(y)) y and G(F(x)) x. This loss is minimized along with an adversarial loss on the generated outputs to constrain the model to generate realistic images. In Fader Networks (Lample et al., 2017b), a discriminator is applied on

Published as a conference paper at ICLR 2019

the latent representation of an image autoencoder to remove the information about specific attributes. The attribute values are instead given explicitly to the decoder at training time, and can be tuned at inference to generate different realistic versions of an input image with varying attribute values.

Different approaches have been proposed for textual data, mainly aiming at controlling the writing style of sentences. Unfortunately, datasets of parallel sentences written in a different style are hard to come by. Carlson et al. (2017) collected a dataset of 33 English versions of the Bible written in different styles on which they trained a supervised style transfer model. Li et al. (2018) released a small crowdsourced subset of 1,000 Yelp reviews for evaluation purposes, where the sentiment had been swapped (between positive and negative) while preserving the content. Controlled text generation from unsupervised data is thus the focus of more and more research.

An theme that is common to most recent studies is that style transfer can be achieved by disentangling sentence representations in a shared latent space. Most solutions use an adversarial approach to learn latent representations agnostic to the style of input sentences (Fu et al., 2017; Hu et al., 2017; Shen et al., 2017; Zhang et al., 2018a; Xu et al., 2018; John et al., 2018; Zhao et al., 2018). A decoder is then fed with the latent representation along with attribute labels to generate a variation of the input sentence with different attributes.

Unfortunately, the discrete nature of the sentence generation process makes it difficult to apply to text techniques such as cycle consistency or adversarial training. For instance, the latter (Shen et al., 2017; dos Santos et al., 2018; Zhang et al., 2018c) requires methods such as REINFORCE (He et al., 2016) or approximating the output softmax layer with a tunable temperature (Hu et al., 2017; Prabhumoye et al., 2018; Yang et al., 2018), all of which tend to be slow, unstable and hard to tune in practice. Moreover, all these studies control a single attribute (e.g. swapping positive and negative sentiment).

The most relevant work to ours is Zhang et al. (2018b), which also builds on recent advances in unsupervised machine translation. Their approach first consists of learning cross-domain word embeddings in order to build an initial phrase-table. They use this phrase-table to bootstrap an iterative back-translation pipeline containing both phrase-based and neural machine translation systems. Overall, their approach is significantly more complicated than ours, which is end-to-end and does not require any pre-training. Moreover, this iterative back-translation approach has been shown to be less effective than on-the-fly back-translation which is end-to-end trainable (Lample et al., 2018).

3 CONTROLLABLE TEXT REWRITING

This section briefly introduces notation, the task, and our empirical procedure for evaluating disentanglement before presenting our approach.

We consider a training set D = (xi, yi)i [1,n] of n sentences xi X paired with attribute values yi. y Y is a set of m attribute values y = (y1, ..., ym). Each attribute value yk is a discrete value in the set Yk of possible values for attribute k, e.g. Yk = {bad, neutral, good} if yk represents the overall rating of a restaurant review.

Our task is to learn a model F : X Y X that maps any pair (x, y) of an input sentence x (whose actual set of attributes are y) and a new set of m attribute values y to a new sentence x that has the specified attribute values y, subject to retaining as much as possible of the original content from x, where content is defined as anything in x which does not depend on the attributes.

The architecture we consider performs this mapping through a sequence-to-sequence auto-encoder that first encodes x into a latent representation z = e(x), then decodes (z, y) into x = d(z, y), where e and d are functions parameterized by the vector of trainable parameters θ. Before giving more detail on the architecture, let us look at disentanglement.

3.1 ARE ADVERSARIAL MODELS REALLY DOING DISENTANGLEMENT?

Almost all the existing methods are based on the common idea to learn a latent representation z that is disentangled from y. We consider z to be disentangled from y if it is impossible to recover y from z. While failure to recover y from z could mean either that z was disentangled or that the classifier

Published as a conference paper at ICLR 2019

chosen to recover y was either not powerful enough or poorly trained, success of any classifier in recovering y demonstrates that z was in fact not invariant to y.

λadv Discriminator Acc (Train) Post-fit Classifier Acc (Test)

0 89.45% 93.8% 0.001 85.04% 92.6% 0.01 75.47% 91.3% 0.03 61.16% 93.5% 0.1 57.63% 94.5% 1.0 52.75% 86.1% 10 51.89% 85.2%

fast Text - 97.7%

Table 2: Recovering the sentiment of the input from the encoder s representations of a domain adversarially-trained Fader model (Fu et al., 2017). During training, the discriminator, which was trained adversarially and jointly with the model, gets worse at predicting the sentiment of the input when the coefficient of the adversarial loss λadv increases. However, a classifier that is separately trained on the resulting encoder representations has an easy time recovering the sentiment. We also report the baseline accuracy of a fast Text classifier trained on the actual inputs. As a preliminary study, we gauge the degree of disentanglement of the latent representation. Table 2 shows that the value of the attribute can be well recovered from the latent representation of a Faderlike (Fu et al., 2017) model even when the model is trained adversarially. A classifier fit post-hoc and trained from scratch, parameterized identically to the discriminator(see paragraph on model architecture in Section 3.3 for details), is able to recover attribute information from the distengeled content representation learned via adversarial training. This suggests that disentanglement may not be achieved in practice, even though the discriminator is unable to recover attribute information well during training. We do not assert that disentangled representations are undesirable but simply that it isn t mandatory in the goal of controllable text rewriting. This is our focus in the following sections.

3.2 OUR APPROACH

Evaluation of controlled text generation can inform the design of a more streamlined approach: generated sentences should (1) be fluent, (2) make use of the specified attribute values, and (3) preserve the rest of the content of the input.

Denoising auto-encoding (DAE) (Fu et al., 2017) is a natural way to learn a generator that is both fluent and that can reconstruct the input, both the content and the attributes. Moreover, DAE is a weak way to learn about how to change the style, or in other words, it is a way to force the decoder to also leverage the externally provided attribute information. Since the noise applied to the encoder input x may corrupt words conveying the values of the input attribute y, the decoder has to learn to use the additional attribute input values in order to perform a better reconstruction. We use the noise function described in Lample et al. (2017a) that corrupts the input sentence by performing word drops and word order shuffling. We denote by xc a corrupted version of the sentence x.

As discussed in Sec. 3.1, disentanglement is not necessary nor easily achievable, and therefore, we do not seek disentanglement and do not include any adversarial term in the loss. Instead, we consider a more natural constraint which encourages the model to perform well at the task we are ultimately interested in - controlled generation via externally provided attributes. We take an input (x, y) and encode x it into z, but then decode using another set of attribute values, y, yielding the reconstruction x. We now use x as input of the encoder and decode it using the original y to ideally obtain the original x, and we train the model to map ( x, y) into x. This technique, called back-translation (BT) (Sennrich et al., 2015a; Lample et al., 2017a; 2018; Artetxe et al., 2018), has a two-fold benefit. Initially when the DAE is not well trained and x has lost most of the content present in x, the only useful information provided to the decoder is the desired attribute y. This encourages the decoder to leverage the provided attributes. Later on during training when DAE is better, BT helps training the sequence-to-sequence for the desired task. Overall, we minimize:

(x,y) D log pd x|e(xc), y + λBT X

(x,y) D, y Y log pd

x|e d e(x), y , y (1)

Published as a conference paper at ICLR 2019

where pd is the probability distribution over sequences x induced by the decoder, e(xc) is the encoder output when fed with a corrupted version xc of the input x, and d(e(x), y) is a variation of the input sentence x written with a randomly sampled set of attributes y. In practice, we generate sentences during back-translation by sampling from the multinomial distribution over words defined by the decoder at each time step using a temperature T.

3.3 IMPLEMENTATION

So far, the model is the same as the model used for unsupervised machine translation by Lample et al. (2018), albeit with a different interpretation of its inner workings, no longer based on disentanglement. Instead, the latent representation z can very well be entangled, but we only require the decoder to eventually overwrite the original attribute information with the desired attributes. Unfortunately, this system may be limited to swapping a single binary attribute and may not give us enough control on the trade-off between content preservation and change of attributes. To address this limitations, we introduce the following components:

Attribute conditioning In order to handle multiple attributes, we separately embed each target attribute value and then average their embeddings. We then feed the averaged embeddings to the decoder as a start-of-sequence symbol.

We also tried an approach similar to Michel & Neubig (2018), where the output layer of the decoder uses a different bias for each attribute label. We observed that the learned biases tend to reflect the labels of the attributes they represent. Examples of learned biases can be found in Table 14. However, this approach alone did not work as well as using attribute-specific start symbols, nor did it improve results when combined with them.

Latent representation pooling To control the amount of content preservation, we use pooling. The motivating observation is that models that compute one latent vector representation per input word usually perform individual word replacement, while models without attention are much less literal and tend to lose content but have an easier time changing the input sentence with the desired set of attributes. Therefore, we propose to gain finer control by adding a temporal max-pooling layer on top of the encoder, with non-overlapping windows of width w. Setting w = 1 results in a standard model with attention, while setting w to the length of the input sequence boils down to a sequence-to-sequence model without attention. Intermediate values of w allow for different tradeoffs between preserving information about the input sentence and making the decoder less prone to copying words one by one.

The hyper-parameters of our model are: λAE and λBT trading off the denoising auto-encoder term versus the back-translation term (the smaller the λBT /λAE ratio the more the content is preserved and the less well the attributes are swapped), the temperature T used to produce unbiased generations (Edunov et al., 2018) and to control the amount of content preservation, and the pooling window size w. We optimize this loss by stochastic gradient descent without back-propagating through the back-translation generation process; back-translated sentences are generated on-the-fly once a new mini-batch arrives.

Model Architecture We use an encoder parameterized by a 2-layer bidirectional LSTM and a 2layer decoder LSTM augmented with an attention mechanism (Bahdanau et al., 2014). Both LSTMs and our word embedding lookup tables, trained from scratch, have 512 hidden units. Another embedding lookup table with 512 hidden units is used to embed each attribute value. The decoder conditions on two different sources of information: 1) attribute embedding information that presented that it as the first token, similar to Lample et al. (2018) and at the softmax output as an attribute conditional bias following Michel & Neubig (2018). When controlling multiple attributes, we average the embeddings and bias vectors that correspond to the different attribute values. 2) The decoder also conditions on a temporally downsampled representation of the encoder via an attention mechanism. The representations are downsampled by temporal max-pooling with a non-overlapping window of size 5. Although our best models do not use adversarial training, in ablations and experiments that study disentanglement, we used a discriminator paramaeterized as 3 layer MLP with 128 hidden units and Leaky Re LU acivations.

Published as a conference paper at ICLR 2019

4 EXPERIMENTS

4.1 DATASETS

We use data from publicly available Yelp restaurant and Amazon product reviews following previous work in the area (Shen et al., 2017; Li et al., 2018) and build on them in three ways to make the task more challenging and realistic. Firstly, while previous approaches operate at the sentence level by assuming that every sentence of a review carries the same sentiment as the whole of review, we operate at the granularity of entire reviews. The sentiment, gender1 of the author and product/restaurant labels are therefore more reliable. Secondly, we relax constraints enforced in prior works that discard reviews with more than 15 words and only consider the 10k most frequent words. In our case, we consider full reviews with up to 100 words, and we consider byte-pair encodings (BPE) Sennrich et al. (2015b) with 60k BPE codes, eliminating the presence of unknown words. Finally, we leverage available meta-data about restaurant and product categories to collect annotations for two additional controllable factors: the gender of the review author and the category of the product or restaurant being reviewed. A small overview of the corresponding datasets is presented below with some statistics presented in Table 3. Following Li et al. (2018), we also collect human reference edits for sentiment and restaurant/product categories to serve as a reference for automatic metrics as well as an upper bound on human evaluations (examples in Appendix Table 12).

Yelp Reviews This dataset consists of restaurant and business reviews provided by the Yelp Dataset Challenge2. We pre-process this data to remove reviews that are either 1) not written in English according to a fast Text (Joulin et al., 2016) classifier, 2) not about restaurants, 3) rated 3/5 stars as they tend to be neutral in sentiment (following Shen et al. (2017)), or 4) where the gender is not identifiable by the same method as in Reddy & Knight (2016); Prabhumoye et al. (2018). We then binarize both sentiment and gender labels. Five coarse-grained restaurant category labels, Asian, American, Mexican, Bars & Dessert, are obtained from the associated meta-data. Since a review can be written about a restaurant that has multiple categories (ex: an Asian restaurant that serves desserts), we train a multi-label fast Text classifier to the original data that has multiple labels per example. We then re-label the entire dataset with this classifier to pick the most likely category to be able to model the category factor as a categorical random variable. (See Appendix section A.2 for more details.) Since there now exists two variants of the Yelp dataset, we refer to the one used by previous work (Shen et al., 2017; Fu et al., 2017; Li et al., 2018) as SYelp and our created version with full reviews along with gender and category information as FYelp henceforth.

Amazon Reviews The amazon product review dataset (He & Mc Auley, 2016) is comprised of reviews written by consumers of Amazon products. We followed the same pre-processing steps as in the Yelp dataset with the exception of collecting gender labels, since a very large fraction of amazon usernames were not present in a list of gender-annotated names. We labeled reviews with the following product categories based on the meta-data: Books, Clothing, Electronics, Movies, Music. We followed the same protocol as in FYelp to re-label product categories. In this work, we do not experiment with the version of the Amazon dataset used by previous work, and so we refer to our created version with full reviews along with product category information as just Amazon henceforth. Statistics about the dataset can be found in Table 3.

Public social media content We also used an unreleased dataset of public social media content written by English speakers to illustrate the approach with examples from a more diverse set of categories3. We used 3 independent pieces of available information about that content: 1) gender (male or female) 2) age group (18-24 or 65+), and 3) writer-annotated feeling (relaxed or annoyed).

1Note that using gender (or any other attribute for that matter) as a differentiating attribute between several bodies of text implies that there are indeed signatures of gender in the data. These signatures could be as innocuous as some first names like Mary being usually associated with women, or disheartening like biases and stereotypes exposed by statistical methods, (e.g., man is to computer programmer as woman is to homemaker (Bolukbasi et al., 2016)). We certainly do not condone those stereotypes, and on the contrary, we hope that showing that our models can uncover these biases might down the line turn them into powerful tools for researchers who study fairness and debiasing (Reddy & Knight, 2016). 2https://www.yelp.com/dataset/challenge 3See footnote 1.

Published as a conference paper at ICLR 2019

Sentiment Gender Category

SYelp Positive Negative Male Female American Asian Bar Dessert Mexican 266,041 177,218 - - - - - - -

FYelp Positive Negative Male Female American Asian Bar Dessert Mexican 2,056,132 639,272 1,218,068 1,477,336 904,026 518,370 595,681 431,225 246,102

Amazon Positive Negative - - Book Clothing Electronics Movies Music 64,251,073 10,944,310 - - 26,208,872 14,192,554 25,894,877 4,324,913 4,574,167

Social Media Content Relaxed Annoyed Male Female 18-24 65+ 7,682,688 17,823,468 14,501,958 18,463,789 12,628,250 7,629,505

Table 3: The number of reviews for each attribute for different datasets. The SYelp, FYelp and the Amazon datasets are composed of 443k, 2.7M and 75.2M sentences respectively. Public social media content is collected from 3 different data sources with 25.5M, 33.0M and 20.2M sentences for the Feeling, Gender and Age attributes respectively.

To make the data less noisy, we trained a fast Text classifier (Joulin et al., 2016) for each attribute and only kept the data above a certain confidence threshold.

4.2 EVALUATION

Automatic evaluation of generative models of text is still an open research problem. In this work, we use a combination of multiple automatic evaluation criteria informed by our desiderata. We would like our systems to simultaneously 1) produce sentences that conform to the set of pre-specified attribute(s), 2) preserve the structure and content of the input, and 3) generate fluent language. We therefore evaluate samples from different models along three different dimensions:

Attribute control: We measure the extent to which attributes are controlled using fast Text classifiers, trained on our datasets, to predict different attributes.

Fluency: Fluency is measured by the perplexity assigned to generated text sequences by a pre-trained Kneser Ney smooth 5-gram language model using Ken LM (Heafield, 2011).

Content preservation: We measure the extent to which a model preserves the content present of a given input using n-gram statistics, by measuring the BLEU score between generated text and the input itself, which we refer to as self-BLEU. When a human reference is provided instead, we compute the BLEU score with respect to it, instead of the input, which we will refer to as just BLEU (Papineni et al., 2002). BLEU scores in this paper correspond to the BLEU score with respect to human references averaged across generations conditioned on all possible attribute values except for that of the input. However, when reporting self-BLEU scores, we also average across cases where generations are also conditioned on the same attribute value as the input.

A combination of these metrics, however, only provides a rough understanding of the quality of a particular model. Ultimately, we rely on human evaluations collected via a public crowd-sourcing platform. We carried out two types of evaluations to compare different models. 1) Following a protocol similar to Li et al. (2018), we ask crowd workers to annotate generated sentences along the three dimensions above. Fluency and content preservation are measured on a likert-scale from 1 to 5 and attribute control is evaluated by asking the worker to predict the attribute present in the generated text. 2) We take a pair of generations from two different models, and ask workers to pick the generation they prefer on the overall task, accounting for all the dimensions simultaneously. They are also presented with a no preference option to discard equally good or bad generations from both models.

4.3 MODEL SELECTION

Since our automatic evaluation metrics are only weak proxies for the quality of a model, we set minimum thresholds on the content preservation and attribute control criteria and only consider models above a certain threshold.

The few models that met the specified threshold on the validation set were evaluated by humans on the same validation set and the best model was selected to be run on the test set.

Published as a conference paper at ICLR 2019

Model Accuracy BLEU PPL

Fader/Style Embedding (Fu et al., 2017) 18% 16.7 56.1 Multi Decoder (Fu et al., 2017) 52% 11.3 90.1 Controllable Text (Hu et al., 2017) 85% 20.6 232.0 CAE (Shen et al., 2017) 72% 6.8 53.0 Retrieval (Li et al., 2018) 81% 1.3 7.4 Rule-based (Li et al., 2018) 73% 22.3 118.7 Delete Only (Li et al., 2018) 77% 14.5 67.1 Delete And Retrieve (Li et al., 2018) 79% 16.0 66.6

Fader (Ours w/o backtranslation & attention) 71% 15.7 35.1 Ours 87% 14.6 26.2 Ours 85% 24.2 26.5 Ours 74% 31.2 49.8

Input copy 13% 30.6 40.6

Table 4: Automatic evaluation of models on the SYelp test set from Li et al. (2018). The test set is composed of sentences that have been manually written by humans, which we use to compute the BLEU score. Samples for previous models were made available by Li et al. (2018). For our model, we report different results corresponding to different choices of hyper-parameters (pooling kernel width and back-translation temperature) to demonstrate our model s ability to control the trade-off between attribute transfer and content preservation.

4.4 COMPARISONS TO PRIOR WORK

Our first set of experiments aims at comparing our approach with different models recently proposed, on the SYelp dataset. Results using automatic metrics are presented in Table 4. We compare the same set of models as in Li et al. (2018) with the addition of our model and our own implementation of the Fader network (Lample et al., 2017b), which corresponds to our model without back-translation and without attention mechanism, but uses domain adversarial training (Ganin et al., 2016) to remove information about sentiment from the encoder s representation. This is also similar to the Style Embedding model presented by Fu et al. (2017). For our approach, we were able to control the trade-off between BLEU and accuracy based on different hyper-parameter choices. We demonstrate that our approach is able to outperform all previous approaches on the three desired criteria simultaneously, while our implementation of the fader is competitive with the previous best work.

Fluency Content Sentiment

DAR (Li et al. (2018)) 3.33 (1.39) 3.16 (1.43) 64.05% Ours 4.07 (1.12) 3.67 (1.41) 69.66% Human (Li et al. (2018)) 4.56 (0.78) 4.01 (1.25) 81.35%

Our Model No Preference DAR

DAR vs Our Fader 37.6% 32.7% 29.7% DAR vs Ours 54.4% 24.7% 20.8%

Table 5: Top: Results from human evaluation to evaluate the fluency / content preservation and successful sentiment control on the Li et al. (2018) SYelp test set. The mean and standard deviation of Fluency and Content are measured on a likert scale from 1-5 while sentiment is measured by fraction of times that the controlled sentiment of model matches the judge s evaluation of the sentiment (when also presented with a neutral option). Bottom: Results from human A/B testing of different pairs of models. Each cell indicates the fraction of times that a judge preferred one of the models or neither of them on the overall task.)

Since our automatic evaluation metrics are not ideal, we carried out human evaluations using the protocol described in Section 4.2. Table 5 (top) shows the fluency, content preservation and attribute control (sentiment) scores obtained by our model, Delete And Retrieve (DAR) and turkers from Li

Published as a conference paper at ICLR 2019

et al. (2018) 4 on the SYelp dataset. While humans clearly outperform both models, our model is better than Delete And Retrieve on all 3 dimensions. We further demonstrate our model s strength over Delete And Retrieve in Table 5 (bottom) in an A/B test between two the models, where crowd workers prefer our model 54.4% compared to theirs 20.8%. Interestingly, our baseline Fader model is also able to do better (37.6% vs 29.7%), suggesting limitations in our automatic metrics, since Fader does not do as well in Table 4.

4.5 EVALUATING MULTIPLE ATTRIBUTE CONTROL

Dataset (Model) Attributes Sentiment Category Gender Accuracy self-BLEU Accuracy self-BLEU Accuracy self-BLEU

Yelp (DAR) Sentiment 78.7% 42.1 - - - -

Yelp (Our Fader) Sentiment 85.5% 31.3 - - - - Sentiment + Category 85.1% 20.6 46.1% 22.6 - - Sentiment + Category + Gender 86.6% 20.4 47.7% 22.5 58.5% 23.3

Yelp (Ours)

Sentiment 87.4% 54.5 - - - - Sentiment + Category 87.1% 38.8 64.9% 44.0 - - Sentiment + Category + Gender 88.5% 31.6 64.1% 36.5 59.0% 37.4 Gender - - - - 59.1% 47.0

Amazon (Ours) Sentiment 82.6% 54.8 - - - - Sentiment + Category 82.5% 48.9 81.4% 41.8 - -

Input Copy - 50.0% 100.0 20.0% 100.0 50.0% 100.0

Table 6: Results using automatic evaluation metrics on the FYelp and Amazon test sets. Different rows correspond to the set of attributes being controlled by the model.

Table 6 presents the quantitative results obtained by the FYelp and Amazon datasets when controlling single and multiple attributes. For this table, unlike for Table 4, the Delete And Retrieve (DAR) results were obtained by re-training the model of Li et al. (2018). We use our implementation of the Fader model since we found it to be better than previous work, by human evaluation (Table 5). While we control all attributes simultaneously during training, at test time, for the sake of quantitative evaluations, we change the values only of a single attribute while keeping the others constant. Our model clearly outperforms the baseline Fader model. We also demonstrate that our model does not suffer significant drops in performance when controlling multiple attributes over a single one.

Demonstrations of our model s ability to control single and multiple attributes are presented in Table 8 and Table 9 respectively. What is interesting to observe is that our model does not just alter single words in the input to control an attribute, but often changes larger fragments to maintain grammaticality and fluency. Examples of re-writes by our model on social media content in Table 1 show that our model tends to retain the overall structure of input sentences, including punctuation and emojis. Additional examples of re-writes can be found in Table 10 and Table 11 in Appendix.

4.6 ABLATION STUDY

Model Test (FYelp) Test (Li et al., 2018) Accuracy self-BLEU Accuracy BLEU

Our model 87% 54.5 80% 25.8 -pooling 89% 47.9 - - -temperature 86% 45.2 80% 21.3 -attention 93% 25.4 80% 22.1 -back-translation 86% 32.8 69% 16.4 +adversarial 86% 45.5 78% 25.1 -attention -back-translation 90% 26.0 71% 15.7

Table 7: Model ablations on 5 model components on the FYelp dataset (Left) and SYelp (Right).

4https://github.com/lijuncen/Sentiment-and-Style-Transfer/tree/master/ evaluation/outputs/yelp

Published as a conference paper at ICLR 2019

In Table 7, we report results from an ablation study on the SYelp and FYelp datasets to understand the impact of the different model components on overall performance. The different components are: 1) pooling, 2) temperature based multinomial sampling when back-translating, 3) attention, 4) back-translation, 5) the use of domain adversarial training and 6) attention and back-translation in conjunction. We find that a model with all of these components, except for domain adversarial training, performs the best, further validating our hypothesis in Section 3.1 that it is possible to control attributes of text without disentangled representations. The absence of pooling or softmax temperature when back-translating also has a small negative impact on performance, while the attention and back-translation have much bigger impacts.

0 20 40 60 80 Epoch

Accuracy / Self-BLEU

No Pooling - Accuracy No Pooling - Self-BLEU Pooling size = 3 - Accuracy Pooling size = 3 - Self-BLEU Pooling size = 8 - Accuracy Pooling size = 8 - Self-BLEU

Figure 1: Accuracy and self-BLEU curves on the FYelp dataset for different pooling operator configurations. Without pooling, the model tends to converge to a copy mode very quickly, with a high self-BLEU score and a poor accuracy. The pooling operator alleviates this behaviour and provides models with a different trade-off accuracy / content preservation.

Table 13 shows examples of reviews re-written by different models at different checkpoints, showing the trade-off between properly modifying the attribute and preserving the original content. Figure 1 shows how the trade-off between content preservation (self-BLEU) and attribute control (accuracy) evolves over the course of training and as a function of the pooling kernel width.

5 CONCLUSION

We present a model that is capable of re-writing sentences conditioned on given attributes, that is not based on a disentanglement criterion as often used in the literature. We demonstrate our model s ability to generalize to a realistic setting of restaurant/product reviews consisting of several sentences per review. We also present model components that allow fine-grained control over the trade-off between attribute control versus preserving the content in the input. Experiments with automatic and human-based metrics show that our model significantly outperforms the current state of the art not only on existing datasets, but also on the large-scale datasets we created. The source code and benchmarks will be made available to the research community after the reviewing process.

Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. Unsupervised neural machine translation. In International Conference on Learning Representations (ICLR), 2018.

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ar Xiv preprint ar Xiv:1409.0473, 2014.

Published as a conference paper at ICLR 2019

Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, pp. 4349 4357, 2016.

Keith Carlson, Allen Riddell, and Daniel Rockmore. Zero-shot style transfer in text using recurrent neural networks. ar Xiv preprint ar Xiv:1711.04731, 2017.

Micka el Chen, Ludovic Denoyer, and Thierry Arti eres. Multi-view data generation without view supervision. In ICLR 2018, 2017.

Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In Advances in neural information processing systems, pp. 2172 2180, 2016.

Alexis Conneau, Guillaume Lample, Marc Aurelio Ranzato, Ludovic Denoyer, and Herv e J egou. Word translation without parallel data. ar Xiv preprint ar Xiv:1710.04087, 2017.

Cicero Nogueira dos Santos, Igor Melnyk, and Inkit Padhi. Fighting offensive language on social media with unsupervised text style transfer. ar Xiv preprint ar Xiv:1805.07685, 2018.

Sergey Edunov, Myle Ott, Michael Auli, and David Grangier. Understanding back-translation at scale. ar Xiv preprint ar Xiv:1808.09381, 2018.

Jessica Ficler and Yoav Goldberg. Controlling linguistic style aspects in neural language generation.

ar Xiv preprint ar Xiv:1707.02633, 2017.

Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. Style transfer in text: Exploration and evaluation. ar Xiv preprint ar Xiv:1711.06861, 2017.

Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Franc ois Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096 2030, 2016.

Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, Tieyan Liu, and Wei-Ying Ma. Dual learning for machine translation. In Advances in Neural Information Processing Systems, pp. 820 828, 2016.

Ruining He and Julian Mc Auley. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In proceedings of the 25th international conference on world wide web, pp. 507 517. International World Wide Web Conferences Steering Committee, 2016.

Kenneth Heafield. Kenlm: Faster and smaller language model queries. In Proceedings of the Sixth

Workshop on Statistical Machine Translation, pp. 187 197. Association for Computational Linguistics, 2011.

Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representations (ICLR), 2017.

Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Transforming auto-encoders. In International Conference on Artificial Neural Networks, pp. 44 51. Springer, 2011.

Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. Toward controlled generation of text. ar Xiv preprint ar Xiv:1703.00955, 2017.

Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017.

Vineet John, Lili Mou, Hareesh Bahuleyan, and Olga Vechtomova. Disentangled representation learning for non-parallel text style transfer. ar Xiv preprint ar Xiv:1808.04339, 2018.

Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. ar Xiv preprint ar Xiv:1607.01759, 2016.

Published as a conference paper at ICLR 2019

Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ar Xiv preprint

ar Xiv:1412.6980, 2014.

Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Ondrej Bojar Chris Dyer, Alexandra Constantin, and Evan Herbst. Moses: Open source toolkit for statistical machine translation. In Annual Meeting of the Association for Computational Linguistics (ACL), demo session, 2007.

Tejas D Kulkarni, William F Whitney, Pushmeet Kohli, and Josh Tenenbaum. Deep convolutional inverse graphics network. In Advances in neural information processing systems, pp. 2539 2547, 2015.

Guillaume Lample, Ludovic Denoyer, and Marc Aurelio Ranzato. Unsupervised machine translation using monolingual corpora only. ar Xiv preprint ar Xiv:1711.00043, 2017a.

Guillaume Lample, Neil Zeghidour, Nicolas Usunier, Antoine Bordes, Ludovic Denoyer, et al. Fader networks: Manipulating images by sliding attributes. In Advances in Neural Information Processing Systems, pp. 5967 5976, 2017b.

Guillaume Lample, Myle Ott, Alexis Conneau, Ludovic Denoyer, and Marc Aurelio Ranzato. Phrase-based & neural unsupervised machine translation. ar Xiv preprint ar Xiv:1804.07755, 2018.

Juncen Li, Robin Jia, He He, and Percy Liang. Delete, retrieve, generate: A simple approach to sentiment and style transfer. ar Xiv preprint ar Xiv:1804.06437, 2018.

Ming-Yu Liu and Oncel Tuzel. Coupled generative adversarial networks. In Advances in neural

information processing systems, pp. 469 477, 2016.

Paul Michel and Graham Neubig. Extreme adaptation for personalized neural machine translation.

ar Xiv preprint ar Xiv:1805.01817, 2018.

Alexander H Miller, Will Feng, Adam Fisch, Jiasen Lu, Dhruv Batra, Antoine Bordes, Devi Parikh, and Jason Weston. Parlai: A dialog research software platform. ar Xiv preprint ar Xiv:1705.06476, 2017.

Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. ar Xiv preprint ar Xiv:1411.1784, 2014.

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pp. 311 318. Association for Computational Linguistics, 2002.

Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhutdinov, and Alan W Black. Style transfer through back-translation. ar Xiv preprint ar Xiv:1804.09000, 2018.

Sravana Reddy and Kevin Knight. Obfuscating gender in social media writing. In Proceedings of

the First Workshop on NLP and Computational Social Science, pp. 17 26, 2016.

Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pp. 86 96, 2015a.

Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pp. 1715 1725, 2015b.

Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. Style transfer from non-parallel text by cross-alignment. In Advances in Neural Information Processing Systems, pp. 6830 6841, 2017.

Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems, pp. 3483 3491, 2015.

Published as a conference paper at ICLR 2019

Yaniv Taigman, Adam Polyak, and Lior Wolf. Unsupervised cross-domain image generation. ar Xiv

preprint ar Xiv:1611.02200, 2016.

Jingjing Xu, Xu Sun, Qi Zeng, Xuancheng Ren, Xiaodong Zhang, Houfeng Wang, and Wenjie Li. Unpaired sentiment-to-sentiment translation: A cycled reinforcement learning approach. ar Xiv preprint ar Xiv:1805.05181, 2018.

Jimei Yang, Scott E Reed, Ming-Hsuan Yang, and Honglak Lee. Weakly-supervised disentangling with recurrent transformations for 3d view synthesis. In Advances in Neural Information Processing Systems, pp. 1099 1107, 2015.

Zichao Yang, Zhiting Hu, Chris Dyer, Eric P Xing, and Taylor Berg-Kirkpatrick. Unsupervised text style transfer using language models as discriminators. ar Xiv preprint ar Xiv:1805.11749, 2018.

Ye Zhang, Nan Ding, and Radu Soricut. Shaped: Shared-private encoder-decoder for text style adaptation. ar Xiv preprint ar Xiv:1804.04093, 2018a.

Zhirui Zhang, Shuo Ren, Shujie Liu, Jianyong Wang, Peng Chen, Mu Li, Ming Zhou, and Enhong Chen. Style transfer as unsupervised machine translation. ar Xiv preprint ar Xiv:1808.07894, 2018b.

Zhirui Zhang, Shuo Ren, Shujie Liu, Jianyong Wang, Peng Chen, Mu Li, Ming Zhou, and Enhong Chen. Style transfer as unsupervised machine translation. ar Xiv preprint ar Xiv:1808.07894, 2018c.

Yanpeng Zhao, Wei Bi, Deng Cai, Xiaojiang Liu, Kewei Tu, and Shuming Shi. Language style transfer from sentences with arbitrary unknown styles. ar Xiv preprint ar Xiv:1808.04071, 2018.

Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. ar Xiv preprint, 2017.

Published as a conference paper at ICLR 2019

A SUPPLEMENTARY MATERIAL

A.1 TRAINING DETAILS

We used the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 10 4, β1 = 0.5, and a batch size of 32. As in Lample et al. (2018), we fix λBT = 1, and set λAE to 1 at the beginning of the experiment, and linearly decrease it to 0 over the first 300, 000 iterations. We use greedy decoding at inference. When generating pseudo-parallel data via back-translation, we found that increasing the temperature over the course of training from greedy generation to multinomial sampling with a temperature of 0.5 linearly over 300,000 steps was useful (Edunov et al., 2018). Since the class distribution for different attributes in both the Yelp and Amazon datasets are skewed, we train with balanced minibatches when there is only a single attribute being controlled and with independent and uniformly sampled attribute values otherwise. The synthetic target attributes during back-translation y are also balanced by uniform sampling.

A.2 DATASET CREATION DETAILS

In addition to the details presented in Section 4.1, we present additional details on the creation of the FYelp and Amazon datasets.

FYelp: Reviews, their rating, user information and restaurant/business categories are obtained from the available metadata. We construct sentiment labels by grouping 1/2 star ratings into the negative category and 4/5 into the positive category while discarding 3 star reviews. To determine the gender of the person writing a review, we obtain their name from the available user information and then look it up in a list of gendered names5 following Prabhumoye et al. (2018); Reddy & Knight (2016). We discard reviews for which we were unable to obtain gender information with this technique. Restaurant/business category meta-data is available for each review, from which we discard all reviews that were not written about restaurants. Amongst restaurant reviews, we manually group restaurant categories into parent categories to cover a significant fraction of the dataset. The grouping is as follows:

Asian - Japanese, Thai, Ramen, Sushi, Sushi Bar, Chinese, Asian Fusion, Vietnamese, Korean, Noodles, Dim Sum, Cantonese, Filipino, Taiwanese American - American (New), American (Traditional), Canadian (New), Southern Mexican/Latin American - New Mexican Cuisine, Mexican, Tacos, Tex-Mex, Tapas Bars, Latin American Bars - Brasseries, Nightlife, Bars, Pubs, Wine Bars, Sports Bars, Beer, Cocktail Bars Desserts - Desserts, Bakeries, Ice Cream & Frozen Yogurt, Juice Bars & Smoothies Donuts, Cupcakes, Chocolatiers & Shops

As described in Section 4.1, we train a classifier on these parent categories and relabel the entire dataset using this.

Amazon: Reviews, their rating, user information and restaurant/business categories are obtained from the metadata made available by He & Mc Auley (2016). We construct sentiment labels in the same manner as in FYelp. We did not experiment with gender labels, since we found that Amazon usernames seldom use real names. We group Amazon product categories into parent categories manually, similar to FYelp as follows:

Books - Books, Books & Comics, Children s Books, Literature & Fiction, Comic Books, Kindle e Books etc. Electronics - Car Electronics, Cell Phones, Electrical & Electronics, Electronics, Electronics & Gadgets, Mobiles, Tablets, Headphones etc. Movies - Movies, Movies & TV, Movies & Video, TV & Film Clothing - Clothing, Shoes & Jewelry, Baby Clothing, Fashion

5https://www.ssa.gov/oact/babynames/names.zip

Published as a conference paper at ICLR 2019

Music - CDs & Vinyl, Music, Digital Music, Children s Music, World Music, Electronic Music

We relabel reviews with a trained product category classifier similar to FYelp.

For the FYelp and Amazon datasets, we normalize, lowercase and tokenize reviews using the moses (Koehn et al., 2007) tokenizer. With social media content, we do not lowercase data in order to exploit interesting capitalization patterns inherent in the data, but we still run other pre-processing steps. We use byte-pair encodings (BPE) (Sennrich et al., 2015b) with 60k replacements, on all 3 datasets, to deal with large vocabulary sizes.

Human Annotated References Li et al. (2018) released a set of human reference edits when controlling the sentiment of a review, on a test set of 500 examples on the SYelp dataset. We follow suit by collecting a similar dataset of 500 human reference edits, which will be made publicly available, for both sentiment and product categories on the FYelp and Amazon datasets. When collecting such data, we use pre-trained sentiment/category classifiers to interactively guide crowd workers using the Parl AI (Miller et al., 2017) platform, to produce edits with the desired attribute value as well as significant content overlap with the input.

Published as a conference paper at ICLR 2019

A.3 ADDITIONAL QUALITATIVE EXAMPLES

Positive Negative (Yelp)

Positive frozen hot chocolate with peanut butter cups = amazing. i ll be back for some food next time! Negative frozen hot chocolate with peanut butter ? horrible. i ll stick with the coffee shop next door!

Negative one word: underwhelming. save your money and find the many restaurants in vegas that offers a real experience. Positive one word: delicious. save room for the best and most authentic indian food in vegas.

Asian Mexican (Yelp)

Asian best thai food i ve ever had in the us. great duck specials on monday.. best yellow curry fried rice.. Mexican best mexican food i ve ever had in my life. great guacamole on the side.. best carnitas tacos i have ever had..

Mexican awesome carne asada! try the papa verde with steak! it s delicious and the portions are great! Asian awesome orange chicken! try the orange chicken with the spicy sauce! it s delicious and the portions are great!

Male Female (Yelp)

Male good food. my wife and i always enjoy coming here for dinner. i recommend india garden. Female good food. my husband and i always stop by here for lunch. i recommend the veggie burrito.

Female we are regulars here... me n my husband just gorge on these freaking amazing donuts!!!!! loved it Male we are regulars here... every time we come here she loves the new york style pizza!!!!!!

Positive Negative (Amazon)

Positive i love this game. takes patience and strategy, go online and look for hints and cheats, they help alot. great game!!! Negative i don t like this game. it takes a lot of time to figure out how to play, and it doesn t work. i would not recommend this game.

Negative i did not like the conflicting historical data. what was real and what was not. i prefer fiction with facts intact. Positive i enjoyed the historical references. what a great read and i loved it. i highly recommend this book.

Movies Books (Amazon)

Movies very good movie with outstanding special effects i recommend this to all shi filovers. good acting great plot lots of action Books very good book with outstanding character development i recommend this to all the readers. good job great plot twists and turns

Books definitely not a stone barrington book, but a story told that keeps you wanting more. great read! Movies definitely not a film noir, but a story that keeps you on the edge of your seat. great acting and a great story.

Clothing Electronics (Amazon)

Clothing got this cause it said it would help with tennis elbow and guess what my tennis elbow still bothering me Electronics got this cause it said it would help with windows xp and guess what my windows xp still crashed

Electronics i have no choice. this is the only black ink that works with my printer. Clothing i have no choice. this is the only black color that works with my dress.

Table 8: Example re-writes by our model on the FYelp and Amazon datasets when controlling a single attribute. The first line in every box is the pre-specified input with its attribute on the left, and the subsequent line is our model s re-write conditioned on a different attribute value.

Sentiment Category Input / Generations

Positive Movies exciting new show. john malkovich is superb as always. great supporting cast. hope it survives beyond season 1 Positive Books exciting new book. john grisham is one of the best. great read. hope he continues to write more. Negative Books nothing new. john grisham is not as good as his first book. not a good read. Positive Clothing awesome new watch. fits perfectly. great price. great quality. hope it lasts for a long time. Negative Clothing horrible. the color is not as pictured. not what i expected. it is not a good quality. Positive Electronics works great. the price is unbeatable. great price. great price. hope it lasts for a long time. Negative Electronics worthless. the picture is not as clear as the picture. not sure why it is not compatible with the samsung galaxy s2. Positive Movies exciting new show. john goodman is great as always. great supporting cast. hope it continues to end. Negative Movies horrible. the acting is terrible. not worth the time. it s not worth the time. Positive Music awesome new album. john mayer is one of the best. great album. hope he continues to release this album. Negative Music horrible. the songs are not as good as the original. not worth the price.

Negative Dessert the bread here is crummy, half baked and stale even when fresh. i won t be back. Positive American the burgers here are juicy, juicy and full of flavor! i highly recommend this place. Negative American the bread here is stale, dry and over cooked even though the bread is hard. i won t be back. Positive Asian the sushi here is fresh, tasty and even better than the last. i highly recommend this place. Negative Asian the noodles here are dry, dry and over cooked even though they are supposed to be fresh. i won t be back. Positive Bar the pizza here is delicious, thin crust and even better cheese (in my opinion). i highly recommend it. Negative Bar the pizza here is bland, thin crust and even worse than the pizza, so i won t be back. Positive Dessert the ice cream here is delicious, soft and fluffy with all the toppings you want. i highly recommend it. Negative Dessert the bread here is stale, stale and old when you ask for a fresh sandwich. i won t be back. Positive Mexican the tacos here are delicious, full of flavor and even better hot sauce. i highly recommend this place. Negative Mexican the beans here are dry, dry and over cooked even though they are supposed to be fresh. i won t be back.

Table 9: Demonstrations of our model s ability to control multiple attributes simultaneously on the Amazon dataset (top) and FYelp dataset (bottom). The first two columns indicate the combination of attributes that are being controlled, with the first row indicating a pre-specified input

Published as a conference paper at ICLR 2019

Relaxed Annoyed

Relaxed Wow! Sitting on my sister s patio, a glass of wine, and glorious music! Hmmmm! Annoyed Wow! Sitting on my sister s patio, a glass of wine, and the neighbors are out! Geez!

Relaxed Nothing like a few

after a long productive day!! Annoyed Nothing like a flat tire

after a long day at work!!

Relaxed I decided it was time for a little chill time with my people tonight, I Missed them! Plus I need a break.

Annoyed I thought I was sleeping for a little while at the end of the week, I m done! Plus I need a nap.

Relaxed Rain = Sleepy! Love the sound of rain on my tin roof.

FYI: Tomorrow is FRIDAY! Annoyed Rain = Mad! The sound of rain on my tin roof.

Rain is overrated!

Relaxed Yay!! A quick pedicure before I pick up the little angel from school!

Annoyed Yay!! A week before I pick up the little bastard from school! FML

Relaxed Had an amazing day driving around. The sea, the woods, just great.

Annoyed Had an amazing day driving around. The weather, the roads are delayed, and the traffic is closed.

Relaxed Happiness is watching movie in a rainy day cuddling with your kids in a cozy bed

Annoyed Yet again watching tv in a rainy day cuddling with my kids in a cozy house

Relaxed Love sitting in the tree listening to the woods wake up!! Last day of 6 day

Annoyed Love sitting in the tree listening to the wake up call!! Last day of my holiday was just ruined

Relaxed A cup of hot creamy #coffee is all i need

#relaxed #focused Annoyed A cup of hot coffee is the worst feeling ever

Relaxed I am in bed, listening to my music. 50 60 70 2000s.....

Annoyed I am in bed, trying to sleep. This 50 + degree weather.....

Relaxed A nice end to a busy day, dinner & drinks with the hubby

Annoyed A bad end to a busy day, dinner & drinks are just ruined

Annoyed Relaxed

Annoyed Looks like my shovelling is done. Just broke my shovel. Relaxed Looks like my adjustable couch is done. Just chilling.

Annoyed Nothing makes me more mad on Thursday get paid and the ATM at back is also broke. Thanks for the great service Relaxed Nothing better than a good massage on the beach and watching the sunset at the pool. Thanks for the great company

Annoyed Who cares that Golden State won... im so over it... ready for football Relaxed Out of town by the Sea, and... im so excited... ready for vacation

Annoyed Some people just can t be themself! #actthesamearoundeveryone Relaxed Having a great time with the bestie! #momanddaughterday #muchneeded

Annoyed When gmail is isn t connecting and you tried to get stuff submitted before midnight Relaxed When the day is just over and you finally get to enjoy some rays before bed

Annoyed I ve never been this mad in my life!!!

Relaxed I ve never been this relaxed in my life!!!

Annoyed I hate it when I can t get to sleep. Stop it brain!! Relaxed I love it when I can t get up. Beautiful weather!!

Annoyed When u get one of those phone calls that messes up ur day

.. Relaxed When u get one of those massage chairs that helps ur body

Annoyed Don t u just love it when people don t respond to ur texts

Relaxed Don t u just love it when it s time to pamper yourself

Annoyed 5 of battery and no charger since it just broke a min ago ugh!!

Relaxed 5 minutes of sun and a bottle of beer now just watching a movie

Table 10: Examples of our model s ability to re-write sentences from public social media content when conditioned on information about the feeling expressed by the writer (Relaxed vs Annoyed)

Published as a conference paper at ICLR 2019

65+ Reality leaves a lot to the imagination. - John Lennon 18-24 Reality leaves a lot of memes - John Cena

65+ Photos are beautiful. It must be unreal in person

18-24 Cool pic bro - It must be fab in person!

65+ where did the time go ? LOVE the toothless smile! 18-24 lemme show u something

toothless smile

65+ so pretty I feed so many in my yard and just love seeing them everyday. 18-24 so pretty I wanna tag in my family and just love them everyday

65+ Just where are you. I am loving the pictures and just a bit envious. 18-24 Just wanna be you I m loving the pictures and just a bit excited

65+ Yes, so true! It is incredible what moms learn from daughters and grands! 18-24 Lmao so true! It s incredible what moms learn from daughters and kids x

65+ Fantastic picture, Peter you look younger all the time, must be the love! 18-24 Osm pic Peter you look younger all the time, must be the love

65+ What a sweet little girl. One can see the love

between you 2. 18-24 What a sweet little girl

can see the love

between you 2

65+ Congratulation young man. I am very proud of you, keep up the good work. 18-24 Congratulation sis I m very proud of you keep up the good work

65+ So good to hear you are doing well. Love seeing pictures of your precious granddaughter

18-24 So good to hear you re doing well

seeing pictures of your precious baby

18-24 Yeah ight bitch. We gon see who get the last laugh

65+ These ignorant idiots. We might see who get the last laugh!!!!!

18-24 Y all are cute tho and I m happy for you guys

65+ Hi, you are cute folks and I am happy for you guys.

i dont love boys

they re so urgh

65+ What a lovely group of boys! They are so fortunate to have an exceptional faces.

Table 11: Model re-writes of sentences from public social media content when conditioned on the age-group of the writer (18-24 vs 65+)

Published as a conference paper at ICLR 2019

Positive Negative (Yelp)

Positive happy to find this hidden gem near my office. great food and best of all, fast delivery. Negative the restaurant near my office was such a dump. late delivery and gross, cold food.

Negative omfg no. i ordered baklava and they nuked it on a styrofoam dish for me. the insides were still cold and it tasted like cancer :/ Positive yes! i ordered baklava and it was the perfect temperature on a fancy plate. the inside was great and it tasted excellent.

Mexican Asian (Yelp)

Mexican just awful. so called asada burrito was cold and bland... just a lump of plain carnitas and some iceberg lettuce in a cheap tortilla. reminiscent of taco bell in the 90s but worse. Asian just awful. so called spring roll was cold and bland....jump a lump of meat and vegetables in an egg noodle wrapper. reminded me of frozen chinese food but worse.

Asian American (Yelp)

Asian my new favorite curry house! definitely coming back. chicken katsu and takoyaki is soooo goood! cant wait to try the pork katsu. American my new favorite american bistro house! we are definitely coming back. fried chicken and waffles is soooo good. cant wait to try the new bbq rib sandwich.

Mexican Dessert (Yelp)

Mexican tacos were delicious. tried the carne asad, bbq pork, and chorizo. came with onion and cilantro topping and a house made choice of mild or hot salsa. Dessert cheesecake was delicious. tried the strawberry flavored one with chocolate drizzle. came with a fresh cherry on top and a house made choice of iced or hot coffee

Positive Negative (Amazon)

Negative too sci fifor me. characters not realistic. situation absurd. i abandoned it halfway through, unusual for me. simply not my taste in mysteries. Positive i love how sci fithis was! the characters are relatable, and the plot was great. i just had to finish it in one sitting, the story got me hooked. this is has to be one of my favorite mysteries.

Positive my mom love this case for her i-pod. she uses it a lot and she is one satisfied customer. she would recommended it. Negative my mom initially liked this case for her i-pod. she used it for a while and it broke. since it is not solid she would not recommend it.

Clothing Books (Amazon)

Clothing nice suit but i wear a size 8 and ordered at 12 and it was just a bit too small. Books great book but i can t read small text with my bad eyesight, unfortunately, this one happened to be printed rather small.

Books Clothing (Amazon)

Books great book about the realities of the american dream. well written with great character development. i would definitely recommend this book. Clothing great dress with american flags printed on it. well made with great materials. i would definitely recommend this dress.

Books Music (Amazon)

Books i loved the book an can t wait to read the sequel! i couldn t put the book down because the plot and characters were so interesting. Music i loved the music and can not wait for the next album! i couldn t stop listening because the music was so interesting.

Table 12: Examples of human edits from our FYelp and Amazon datasets. The first line in every box was the input presented to a crowd worker followed by their corresponding edit with the specified attribute.

Accuracy Self-BLEU Input / Swap

not a fan. food is not the best and the service was terrible the two times i have visited. 78.7% 73.8 great little place. food is great and the service was great the two times i have visited. 83.5% 51.5 best food in town. food is great and the service was the best two times i have visited. 92.8% 27.7 best thai food in town. the food is great and the service is excellent as always. 96.8% 13.1 best chinese food in town. great service and the food is the best i have had in a long time.

overpriced specialty food. also very crowded. service is slow. would not recommend at any time. 78.7% 73.8 great homemade food. also very crowded. service is fast. would recommend at least once. 83.5% 51.5 great specialty food. also very crowded. service is friendly. would recommend any time at the time. 92.8% 27.7 great variety of food. also very friendly staff. good service. would recommend at least once a week. 96.8% 13.1 great tasting food. very friendly staff. definitely recommend this place for a quick bite.

Table 13: Examples of controlling sentiment with different model checkpoints that exhibit different trade-offs between content preservation (self-BLEU) and attribute control (Accuracy). The first line in both examples is the input sentence, and subsequent lines are a model s outputs with decreasing content preservation with rows corresponding to model checkpoints at different epochs during training.

Published as a conference paper at ICLR 2019

Positive Negative Male Female American Asian Bar Dessert Mexican

pampered dishonest ammo manicure primanti guu promoter patisserie tortas relaxation fraud tenant pedi bobby chashu bouncers froyo fundido delightfully incompetence bachelor hubs flay izakaya bouncer buttercream arepas complemented unethical barbers bridesmaids lux tonkotsu hakkasan dunkin burritos cutest insulted wife bridal nacho khao postino groomers tostada plush audacity firestone pedicure bj soju bachi bakeries taquitos punctual confronted provider instructors gown tonkatsu brio bakery nacho housemade cockroach data mattresses applebee banchan cabanas custard fajita precision crooks plumber stylist burgr shabu films doughnuts guac masterpiece disrespect motor jacuzzi chilis teppanyaki trader pastries refried restored roaches contractor hubby mesa gai hooters gelato mexico comprehensive refunds hertz pregnancy bmw kbbq karaoke cheesecakes empanadas sublime shrugged hvac bachelorette denny pho irish donut tapas made liars qualified husbands mastro hotpot harry croissants queso tastefully rudest incompetence barre cellar soi darts doughnut cantina treasures accused transmission cutest bachi karaage nightclub danish salsas addicting inconsiderate summary lashes rubbed omakase applebee cheesecake pollo marvelous roach contractors sephora flatbread saigon whiskey fritter asada handsome rudely automotive bf skins panang cereal macarons barrio healing cancellation audio boyfriends grille cantonese perform oreos barbacoa

Table 14: Examples of the learned attribute biases for sentiment and restaurant categories on FYelp.