PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction

Introduction

Narratives and news texts often reflect societal biases and stereotypes, such as the traditional gender role that women are passive and submissive Lakoff (1973); Fiske (1993); Fast et al. (2016). The task of controllable text revision, i.e., rephrasing text to a targeted style or framing, can help correct for these biases by altering and equalizing the way people are described. For example, automatically rewriting “Mey daydreamed about being a doctor” as “Mey pursued her dream to be a doctor” portrays Mey with more authority and decisiveness (Figure 1). Such controllable revision methods could be used to help reshape how gender roles are portrayed in media (e.g., through machine-in-the-loop writing systems; Clark et al., 2018).

To edit such biases out of text, a controllable rewriting model faces three key challenges. First, a model should be able to make edits beyond surface-level paraphrasing, as simple paraphrasing will often not adequately debias the underlying events described. For example, Mey’s portrayal in Figure 1 carries both overt bias (the choice of action) and subtle bias (the framing of the action), both of which require rewriting to be adequately debiased. Second, a model’s debiasing revisions should be purposeful and precise and should not make unnecessary changes to the underlying meaning of the original text. Lastly, since parallel data does not exist, models must learn to revise and debias text without supervised data, thereby preventing straightforward machine translation-style modelling.

We formulate Controllable Debiasing as a new controllable text revision task that aims to correct the implicit and possibly unwanted bias against or towards a specific character portrayed in text (§2). As shown in Figure 1 (top), we study the portrayal biases through the lens of connotation frames of power and agency Sap et al. (2017), which provide pragmatic knowledge about implied power and agency levels projected onto characters by a predicate.

We create PowerTransformer, an encoder-decoder model that rewrites sentences with a desired portrayal using agency connotation frames (§3). We combine a reconstruction and paraphrase objective into our model to overcome the lack of parallel supervised data, building off of the denoising autoencoder setup from Li et al. (2018a). To steer the revisions, we endow the model with connotation frame knowledge both at training time using control tokens, and at generation time using agency-based vocab boosting.

Our findings show that PowerTransformer is effective at rewriting sentences with desired agency connotations while only making minimal changes to their meaning, as measured through both human and automatic evaluations (§4). We also show that PowerTransformer significantly outperforms existing stylistic rewriting methods Prabhumoye et al. (2018); Dathathri et al. (2020) on those metrics. Additionally, through ablations studies, we establish the usefulness of each component of the model, finding benefits from both the joint objective (47% gain in accuracy) and the agency scaling (12% gain in accuracy).

Finally, in §5, we apply Controllable Debiasing to a corpus of modern English movies Gorinski and Lapata (2015) as a step towards removing gender bias in character portrayal established by prior work Sap et al. (2017). Using PowerTransformer, we revise the movie scripts and significantly increase the agency levels of female characters, thereby reducing the gender bias. Our findings show promise for using modern NLP tools to help mitigate societal biases in text. We release our preprocessed data and code at http://maartensap.com/controllable-debiasing.

Controllable Debiasing

Controllable Debiasing is a novel formalization of stylistic rewriting that aims to debias the portrayal of characters through controllable revision. To achieve the desired character portrayal, a system must be able to change the underlying meaning of events, unlike certain formalizations (e.g., politeness transfer; Rao and Tetreault, 2018) where full meaning preservation is required. Without this, systems run the risk of merely paraphrasing the biases in text. However, revisions must be precise and avoid unnecessary meaning changes, which can often occur in stylistic rewriting (e.g., reversing the sentiment of a review drastically changes its underlying meaning).

For our new rewriting task of changing portrayal bias, we focus on connotation frames that measure the power and agency ascribed to characters through the actions they take. Connotation frames Rashkin et al. (2016); Sap et al. (2017) distill implicit relations between a verb, its agent, and its theme. In this work, we use the positive, neutral, and negative agency dimensions, where agency is defined as the capacity to intentionally make changes or act upon one’s environment Dennett (1989). For example, illustrated in Figure 1, “X pursued Y” implies that X has positive agency.Future work could explore using the power dimension instead of agency, or alternative operationalizations of biases, e.g., Social Bias Frames (Sap et al., 2020) or regard towards minorities as introduced by Sheng et al. (2019). Using machine-in-the-loop writing systems (e.g., Ghazvininejad et al., 2016, 2017; Clark et al., 2018, Textiohttps://textio.com/), models trained on this task could help authors write news, stories, or movies that portray characters in less biased ways, and thereby help mitigate the negative effects of stereotypical portrayals in media Behm-Morawitz and Mastro (2008); Field et al. (2019).

PowerTransformer

We present a new approach for Controllable Debiasing called PowerTransformer, which addresses two key challenges: the paucity of parallel supervised data for training and the difficulty of incorporating fine-grained control for steering the agency of the output. Our approach (Figure 2) jointly learns to reconstruct partially masked story sentences while also learning to paraphrase from an external corpus of paraphrases (§3.2). At generation time, we also include a boosting method for fine-grained steering towards the desired agency level as described in §3.3.

PowerTransformer is an encoder-decoder style model with an OpenAI-GPT transformer model Radford et al. (2018) as the base. The input sentence $\mathbf{x}$ is converted to a sequence of byte pair encodings (BPE) $\{x_{1},...,x_{n}\}$, and given to the encoder after being scrubbed of its agency markers as described below. To steer the model, we also give the encoder the target agency $t$, which we represent as one of three special tokens {$<$Pos$>$,$<$Equal$>$,$<$Neg$>$}.In earlier experiments, we also provided the original agency as an input to the model during training and decoding, but found that it made little difference in performance.

2 Joint Objective

We train our model on both a reconstruction and a paraphrasing task, for which inputs are masked and paraphrased versions of the output, respectively.

Inspired by the delete-retrieve-generate model from Li et al. (2018a), this objective teaches the model to recover masked out agency-associated verbs in sentences. We first assign an agency level to an input sentence by counting verbs in the agency lexicon from Sap et al. (2017).For sentences that have multiple verbs, we assign the agency level that the most verbs in the sentence have (e.g., a sentence with two positive agency verbs and one negative agency verb will be assigned positive agency). Then, we mask out all verbs indicative of the agency level, replacing them with a special $<$VERB$>$ token. In this setup, the target output is the original sentence $\mathbf{x}=\{x_{1},...,x_{n}\}$, with the masked sentence $\mathbf{\hat{x}}$ and the target agency level $t$ as inputs. During training, we minimize the cross entropy of the target output sentence given the inputs:

To go beyond reconstructing sentences, we add a paraphrasing objective using an out-of-domain paraphrase corpus (§4.1). We extract agency levels for each sentence and its paraphrase and mask out the agency verbs in the input, using the same methods as described above. Here, the inputs are the masked sentence $\mathbf{\hat{x}}$ and the target agency $t$, while the target output $\mathbf{y}=\{y_{1},...,y_{m}\}$ is the paraphrase. As with reconstruction, we minimize the cross entropy of the target output given the inputs:

3 Controlled Decoding with Vocab Boosting

Controllable Debiasing Experiments

In this section, we describe three experiments for investigating PowerTransformer performance. First, we evaluate performance of our full model and ablated baselines, using automatic metrics to quantify the effectiveness of each modelling component (§4.4). Next, we compare our full model to baselines from related work (§4.5). Lastly, given the limitations of automated metrics for evaluating generations (Liu et al., 2016; Mir et al., 2019), we obtain human judgments of model performance through crowdsourcing (§4.6). We additionally include examples of generations in Table 4.

In our experiments, we use a dataset of short stories for the reconstruction task and a parallel corpus of paraphrases for both paraphrase and reconstruction tasks. We show data statistics in Table 1, with additional preprocessing details in Appendix A.

The main focus of our study is controllable revision of story sentences; therefore, we select sentences from the ROC story corpus (ROC Mostafazadeh et al., 2016). After extracting agency levels for all sentences from the training stories, we sample roughly equal amounts of all three agency levels, and randomly split sentences into training, development, and test sets.We use a 80:13:7 train, development, test ratio.

As additional training data, we use the corpus of automatically aligned paraphrases of TV subtitles (Creutz, 2018, Para.). As with the ROC story corpus, we extract agency levels for each sentence and its paraphrase, then sample roughly equal amounts of pairs with all different sentence-paraphrase agency combinations (further details in §A.2). We randomly split the data into 45k train and 10k dev. instances (Table 1).Since this is just additional training data, we do not test our models on this corpus, but do use the dev. set for selecting some hyperparameters.

2 Metrics

In addition to human evaluations, we also use a variety of automated evaluation metrics to characterize different aspects of performance. We measure the accuracy of the change in agency by comparing the target agency level with that of the output (extracted using the connotation frames lexicon). As a measure of meaning preservation, we use BERT-score F1 metrics (Zhang et al., 2020) to compare the semantic similarity of the input sentence with the machine output.

As additional metrics, we measure the fluency, the repetitiveness, and diversity of the output. Following previous work Dai et al. (2019), we measure fluency as perplexity (PPL) of the output sentence using a pre-trained GPT model that has not been fine-tuned for this task. As an additional metric of potential text degeneration, we compute the fraction of output sentences that have a bigram that is repeated two or more times (w/ rep). Finally, we compute the fraction of generations that are unique with respect to the rest of the output, to ensure diverse, input-specific generations (unique).

3 Experimental Setup

We randomize ROC story and paraphrase data, and use OpenAI GPT LM as our pretrained model. For decoding, we use top-$p$=0.4 nucleus sampling Holtzman et al. (2020), and a boosting strength of $\beta$=5 (hyperparameters and details in §B.1).

4 Investigating Effectiveness of Approach

We first establish our model’s effectiveness at Controllable Debiasing on our dev. set, and investigate the importance of various components in our approach through ablation analyses. For qualitative analyses, we also show example revisions in Table 4 (and Table 6 in the appendix).

We first investigate the importance of the reconstruction objective, by comparing our joint objective model (Joint) with a model trained with just the paraphrasing objective (without masking, ParaOnly). Then, to quantify the effect of boosting, we compare models with (Boost) and without (noBoost) agency-specific vocab boosting. Note that ParaOnly+noBoost is equivalent to a GPT-based encoder-decoder model, similar to seq2seq frameworks commonly used in paraphrasing tasks (Cao et al., 2017; Li et al., 2018b; Prakash et al., 2016).

As a final comparison, we implement a model variant that more closely mirrors the delete-retrieve-generate paradigm (Li et al., 2018a) by adding a “retrieve” step in which we concatenate transformer input with a verb retrieved from the verb agency lexicon that is most similar to the masked out verb (SupplyVerb).We retrieve a verb from the Sap et al. (2017) lexicon that has the target agency and is most similar to the masked out verb, where similarity is defined as cosine distance between word embeddings using GloVe 300-d embeddings (Pennington et al., 2014).

4.2 Results

In Table 2, our results show that the full model (Joint+Boost) yields text revisions with the most accurate target agency and the most meaning preservation. In general, we find that both the joint objective and vocab boosting (Boost) substantially increase the target agency accuracy, as also illustrated in examples (d) and (e) in Table 4. However, unsurprisingly, vocab boosting also slightly lowers fluency, yielding higher perplexities than models’ non-boosted counterparts. Our results also show that using the joint objective with boosting increases the diversity of output, but causes marginally more repetition of bigrams.

Counterintuitively, our ablations show that supplying a verb to the model as an explicit retrieval step (SupplyVerb) does not improve the agency or meaning metrics and actually hurts the fluency of the output (as measured by higher perplexities). Upon qualitative investigation (Table 6 in the appendix), the retrieved verb is often related to a different word sense of the masked verb, breaking the grammaticality of the sentence.

5 Comparison with External Approaches

To further validate our approach, we compare against two baselines from related style transfer and stylistic generation tasks. As these models were designed for binary style transfer, we only report our baseline and model results on the positive and negative agency portions of our data.

We compare to the backtranslation style transfer model from Prabhumoye et al. (2018). This model first translates input sentences to a pivot language (preserving the meaning but losing language-specific style), then relies on style-specific decoder-translators for generating the output sentence. We include set-up details in §B.3.

Recent work in controllable generation has introduced PPLM, a new plug-and-play technique with promising results for decoding stylistic text (Dathathri et al., 2020). This method operates on an underlying neural language model at decoding time. It uses backpropagation from a stylistic discriminator to update the past and present hidden representations to be more consistent with the targeted style or domain. We adapt the approach to controllable revision by replacing the base language model with an autoencoder trained on a reconstruction objective, described in detail in §B.2.

5.2 Results

We present results in Table 3. Our experiments show that PowerTransformer performs better than the baselines overall. Specifically, while the BST revisions obtain slightly higher accuracy on the output agency levels, these revisions have the both the lowest diversity and meaning preservation, suggesting the model ignores the input (Table 4). PPLM shows opposite trends, yielding the lowest accuracy with high meaning preservation and high diversity of generations. Illustrated in Table 4, this model often makes less purposeful and less concise alterations.

6 Evaluating with Human Judgements

To validate our automatic evaluations, we collect human judgments of the controllable revisions from several baselines and PowerTransformer (Joint+Boost).

We design a head-to-headWe use head-to-head evaluations as those have been shown to be more reliable than scale-rating evaluations Kiritchenko and Mohammad (2017). crowdsourcing task on Amazon Mechanical Turk where we ask raters to compare two outputs from different models given the same input sentence and target agency (see Figure 5 in the appendix). We first ask them to judge whether either output is gibberish, then, in two questions, choose which revision has better targeted agency and which better preserves the meaning of the original sentence. For consistency, each pair is rated by three judges. To ensure the quality of our evaluations, we selected workers who could reliably distinguish high from low agency sentences in a qualification task (see Figure 6 in the appendix).

For this evaluation, we generate three revisions–one for each target agency level–for a random subset of 100 test examples. We compare the output of our full PowerTransformer model with two external baselines (PPLM and BST). For further comparison, we also include the most competitive ablated baseline from Table 2 (i.e., Joint+noBoost).

6.2 Results

In Figure 3, we show the percentages of times in which PowerTransformer was preferred over the three baseline models.Judgments in our evaluation task had an average pairwise agreement of 75% (Krippendorf’s $\alpha$=.52). Percentages $>$50% indicate a preference towards PowerTransformer.

Overall, the sentence revisions by PowerTransformer are preferred over all of the baselines in obtaining the desired agency level. For meaning preservation, our model is always selected over BST, mirroring BertScores in Table 3. The difference is less stark when comparing to PPLM which sometimes makes no changes or irrelevant changes to the input sentence, and reversed when comparing to the ablated noBoost.

Additionally, BST revisions were marked as gibberish substantially more than those by other models ($63\%$ vs. 3-7$\%$). While this seemingly contradicts BST’s low perplexity scores, this is in line with previous work showing automatic fluency metrics can favor degenerate, bland, or repetitive language (Holtzman et al., 2020).

Gender Bias in Movies

As a proof-of-concept of Controllable Debiasing, we investigate whether gender biases in portrayals of movie characters can be mitigated using PowerTransformer.

We draw our data from the 767 modern English movie scripts by Gorinski and Lapata (2015), focusing on the narrations which describe characters and their actions (as opposed to the character’s dialogue utterances). Described in further detail in §C in the appendix, we automatically extract characters and assign them a binaryNote that gender is a social construct that goes beyond the man-woman binary Lorber et al. (1991), however more inclusive analyses (e.g., with non-binary genders) are not possible given the limited information about the individuals mentioned in our data. gender (man, woman) using a list of highly gendered names (e.g., “Sarah”, “William”) and a list of gendered words (e.g., “waiter,” “waitress”). Following previous work Ramakrishna et al. (2017); Sap et al. (2017), we assign narration sentences to characters if their name appears in them.

Our corpus contains 16,763 characters from 767 different English movies. Of those characters, 68% are inferred to be men and only 32% to be women,There were 2597 characters for which the gender could not be inferred. consistent with known gender skews in movie characters Google (2017). This bias in representation is also present at the narrative level. Specifically, female characters are only mentioned in $n_{narr,f}=$27 narrations on average, compared to $n_{narr,m}=$34 narrations for male characters (Cohen’s $|d|=0.13$, $p<0.001$). Similarly, compared to their male counterparts, female characters are described in significantly fewer words ($n_{words,f}=329$, $n_{words,m}=435$, $|d|=0.14$, $p<0.001$) and with fewer verbs ($n_{verbs,f}=41$, $n_{verbs,m}=54$, $|d|=0.13$, $p<0.001$).

2 Debiasing Portrayal in Movies

Given the known bias that female characters are portrayed with less agency (Sap et al., 2017), our goal is to re-balance their agency levels to be more on par with those of male characters. Therefore, we revise only the sentences describing female characters to have higher agency, using PowerTransformer. Then we extract connotation frames of agency for revised script sentences, and aggregate per character. Shown in Figure 4, revisions successfully increase the instances of positive agency of female characters, and decrease their negative agency or passiveness.

We further examine the change in gender association of positive and negative agency, to verify the effectiveness of Controllable Debiasing. We first count all the positive and negative agency verbs used to describe characters (in original or rewritten sentences). Following Sap et al. (2017), we then fit a logistic regression model to quantify the association between character’s gender with their agency levels, controlling for their number of words, verbs, and narrations. For better interpretation of the $\beta$ coefficients, we $z$-score all the continuous variables.

We confirm that indeed, Controllable Debiasing using PowerTransformer can reverse the bias in portrayal in movies. In original scripts, male characters were portrayed with significantly higher positive agency ($\beta_{\text{pos}}=1.2$, $p<0.001$) and lower negative agency ($\beta_{\text{neg}}=-0.3$, $p<0.001$) than female characters. However, our model successfully reverses this gender bias, portraying women with significantly more positive agency ($\beta^{\prime}_{\text{pos}}=-62.6$, $p<0.001$) and significantly less negative agency ($\beta^{\prime}_{\text{neg}}=8.7$, $p<0.001$).

Our findings on movie scripts show the promise of using Controllable Debiasing to successfully mitigate gender biases in portrayal of characters, which could be extended to other domains (e.g., news or fiction, Field and Tsvetkov, 2019; Fast et al., 2016). Additionally, future work could consider alternative views of portrayal biases (e.g., “regard” or bias directed at different demographic groups; Sheng et al., 2019; Sap et al., 2020), or use more holistic views of gender roles (e.g., “masculine default” cultures; Cheryan and Markus, 2020).

Related Work

Controllable Debiasing is a new formalization of the unsupervised stylistic rewriting task, contrasting with supervised approaches which benefit from parallel corpora (e.g., Xu et al., 2012, 2015; Rao and Tetreault, 2018; Pryzant et al., 2020). In unsupervised settings, a majority of work has dealt with the dearth of parallel data by using encoder-decoder setups paired with discriminators to disentangle style from content and steer generations (e.g., Shen et al., 2017; Zhang et al., 2018; Fu et al., 2018; Yang et al., 2018; Niu and Bansal, 2018; Romanov et al., 2019; Dai et al., 2019; John et al., 2019) or backtranslation setups Prabhumoye et al. (2018); Lample et al. (2018). In contrast, Li et al. (2018a) introduce a modular approach (later adapted to transformer models by Sudhakar et al., 2019) that relies on drop-in replacement of attribute markers followed by language correction. PowerTransformer improves on this approach with an additional out-of-domain paraphrasing objective.

While a majority of related existing stylistic rewriting work defines style as sentiment (e.g., on reviews), a notable exception is Nogueira dos Santos et al. (2018), who use stylistic rewriting to make text less hateful or offensive. Similar in spirit, Controllable Debiasing is a novel formalization that aims to address and revise social biases expressed in text, but using the nuanced implications distilled in connotation frames of power and agency instead of binary offensiveness.

Our work also draws inspiration from controllable generation methods (e.g., Koncel-Kedziorski et al., 2016; Hu et al., 2017; Ficler and Goldberg, 2017). While those methods steer the generation output to contain desired attributes, controllable revision is constrained to revise an input sentence in addition to generating with desired attributes.

Conclusion

We introduce a new text revision task of Controllable Debiasing, to help debias the portrayal of characters through the lens of connotation frames of power and agency. To this end, we create PowerTransformer, a transformer-based encoder-decoder trained on a joint reconstruction and paraphrasing objective. Our approach demonstrates promising results to revise sentences with targeted power and agency, and outperforms ablations and baselines on both automatic and human evaluations. Finally, as a case study, we show the feasibility for Controllable Debiasing at debiasing the portrayal of characters in movie scripts. Our findings highlight the potential of neural models as a tool for editing out social biases in text.

Acknowledgements

The authors thank the anonymous reviewers and meta-reviewers for their helpful feedback. We also thank Aishwarya Nirmal and Kenta Takatsu for their preliminary exploration of this task. Additionally, we thank Ari Holtzman, Lucy Lin, Sofia Serrano, Elizabeth Clark, and other members of the UW NLP community for their thoughtful input. This research was supported in part by NSF (IIS-1524371, IIS-1714566), DARPA under the CwC program through the ARO (W911NF-15-1-0543), DARPA under the MCS program through NIWC Pacific (N66001-19-2-4031), and the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE-1256082.

References

Appendix A Additional data description

This English corpus originally contains 100,000 five-sentence stories written by crowdworkers about realistic everyday scenarios. We select the data for our task by first extracting agency levels for each sentence, filtering out those with indeterminable agency. Additionally, we filter out sentences with four or more verbs, to prevent the sentence masking from deleting too many content words.

A.2 Paraphrase corpus

This corpus contains paraphrases of spoken dialogue extracted from movie and TV subtitles.From http://www.opensubtitles.org OpusParcus was created by automatically aligning the subtitles sentences using several probabilistic metrics, including likelihood under a round-trip translation paraphrasing model Bannard and Callison-Burch (2005) and pointwise mutual information. For our paraphrasing dataset, we apply the same filtering as with the ROC story corpus to the English portion of the OpusParcus training corpus and select the top 10% highest scoring paraphrases using the PMI scoring from the original paper. We extract agency levels for each pair of paraphrases, and select pairs to obtain roughly equal number of agency-level pairs (i.e., 1/9th positive-neutral, 1/9th positive-negative, etc.) We preprocess the text by stripping any leading periods and commas.

Appendix B Experimental details

We use the Hugging Face Wolf et al. (2019) implementation of OpenAI’s GPT model (117M parameters; Radford et al., 2018). our final setup uses AdamW Loshchilov and Hutter (2019) as our optimizer with a learning weight of 1e-5, batch size of 4 and maximum sequence length of 64. In preliminary results, we find that $\beta$=5 aptly steers the generation while avoiding repetition issues.

All the experiments are performed on NVIDIA TITAN card and use the model hyperparameters listed in Table 5.

We train this model for 10 epochs with each epoch taking approximately an hour. The learning rate is 1e-5 with AdamW optimizer, which is tuned manually in the [1e-6, 1e-3] range for 7 times. We use $p=0.4$ for nucleus sampling and $p$ is tuned manually in the [0.4, 0.9] range for 5 values.

B.1.2 PowerTParaOnly+Static

The PowerTParaOnly+Static loads the trained model from PowerTParaOnly+None and add re-scaling to the logits. The re-scaling factor, $\beta$ was tuned manually tuned in the range. We try 8 $\beta$s and use 5 in the final model. We use the same $p$ as PowerTParaOnly+None

B.1.3 PowerTJoint+None

Similar to PowerTParaOnly+None, we train this model for 10 epochs with each epoch taking approximately an hour. The learning rate is 1e-5 with AdamW optimizer, which is tuned manually in the [1e-6, 1e-3] range for 7 times.We use the same $p$ as PowerTParaOnly+None

B.1.4 PowerTJoint+Static

The PowerTJoint+Static loads the trained model from PowerTJoint+None and add re-scaling to the logits. The re-scaling factor, $\beta$ was tuned manually tuned in the range. We try 8 $\beta$s and use 5 in the final model. We use the same $p$ as PowerTParaOnly+None

B.2 PPLM details

The PPLM decoding method can be used on top of any model, but their original codebase is for use with a pre-trained language model rather than a model for paraphrasing or style transfer. We augment their techniques for this task by replacing the base model in their code with a denoising autoencoder that was trained to reconstruct the input sentence. The denoising autoencoder was implemented using the base GPT2 model (to fit with their code library and be similar size to our model). It was trained on our ROC only training data with a reconstruction objective. In order to denoise the autoencoder, we randomly “dropout” about 50% of the tokens from the context by replacing them with mask tokens. This autoencoder is trained to reconstruct input sentences, but when used with the PPLM decoding method, the input gets dynamically updated to decode a sentence that is similar in meaning but more likely to have a positive/negative agency according to a discriminator that is trained on top of the autoencoder. The PPLM decoding method also has hyperparameters that control the strength of the target label. If set too high, then the output could be degenerate. We manually set the hyperparameters to be as strong possible without producing degenerate text, using a subset of the dev. set as a guide.

B.3 Backtranslation details

We use the code provided by Prabhumoye et al. (2018) for running this baseline. After lowercasing all the negative and positive agency examples in our training data (ROC and OpusParcus), we translate to French using the machine translation model provided in the code base. This baseline requires training a style classifier (agency) and two decoders (one for each agency level). Since the classifier essentially re-learns the agency lexicon, we do not search for hyperparameters, and simply set a learning rate of 5, and 6 epochs. For training the decoders, we perform grid search to find the best hyperparameters. We experiment with a learning rates of {0.5, 1, 2, 5}, {2, 3, 5} epochs, a classification-loss weight of {0.5, 1, 2}, and a word-loss weight of {0.5, 1, 2}, and select the configuration with the best word-level accuracy on the dev. set. We use SGD with a batch size of 64 for all experiments, and refer the reader to the code base for other default parameters.

Appendix C Gender Bias in Movies

The movie scripts mention characters in all caps, making it easy to identify and extract them. We then cross reference the name (or, description for unnamed characters, e.g., “the doorman”) with a list of gendered nameshttp://www.cs.cmu.edu/Groups/AI/util/areas/nlp/corpora/names/0.html and gendered words (e.g., “waitress,” “policeman,” “police woman”). To allow for better rewriting using our model, we split the narratives into sentences (using NLTK’s sentence tokenizer Bird et al., 2009), and assign each sentence to a character if their name appears in the sentence.