# nonparallel_text_style_transfer_with_selfparallel_supervision__e4ff3f1e.pdf

Published as a conference paper at ICLR 2022

NON-PARALLEL TEXT STYLE TRANSFER WITH SELF-PARALLEL SUPERVISION

Ruibo Liu , Chongyang Gao , Chenyan Jia , Guangxuan Xu , Soroush Vosoughi

Dartmouth College, Northwestern University, University of Texas, Austin University of California, Los Angeles {ruibo.liu.gr, soroush.vosoughi}@dartmouth.edu cygao@u.northwestern.edu chenyanjia@utexas.edu gxu21@cs.ucla.edu

The performance of existing text style transfer models is severely limited by the non-parallel datasets on which the models are trained. In non-parallel datasets, no direct mapping exists between sentences of the source and target style; the style transfer models thus only receive weak supervision of the target sentences during training, which often leads the model to discard too much style-independent information, or utterly fail to transfer the style. In this work, we propose La Mer, a novel text style transfer framework based on large-scale language models. La Mer first mines the roughly parallel expressions in the non-parallel datasets with scene graphs, and then employs MLE training, followed by imitation learning refinement, to leverage the intrinsic parallelism within the data. On two benchmark tasks (sentiment & formality transfer) and a newly proposed challenging task (political stance transfer), our model achieves qualitative advances in transfer accuracy, content preservation, and fluency. Further empirical and human evaluations demonstrate that our model not only makes training more efficient, but also generates more readable and diverse expressions than previous models.

1 INTRODUCTION

LM MLE Training 2

Imitating Learning 3 Mining 1

(a) Exsiting Methods

(b) Our Method: La Mer

Non-Parallel, Unsupervised (Roughly) Parallel, Self-Supervised

Figure 1: Red and blue circles represent the source and target texts respectively. (a) Existing methods crucially ignore the inherit parallelism within the data. (b) Our method first mines (roughly) parallel expressions, then learns how to transfer style with the self-supervision from the parallel expressions.

Text style transfer (TST) models learn how to transfer the style of text from source to target while preserving the style-independent content (John et al., 2019; Fu et al., 2018). Existing TST methods perform well when transferring simple styles, such as sentiment; however, they tend to do a poor job on more abstract and subtle styles, such as formality and political stance (Lee et al., 2021; Fu et al., 2019b).

The lack of parallel datasets is one of the main bottlenecks for text style transfer tasks. Parallel datasets contain pairwise sentences that only differ in specific styles, while non-parallel datasets typically contain a large corpora of sentences with different styles, without direct one-to-one mapping between sentences (Fu et al., 2019a; Krishna et al., 2020). In real-world settings, it is much easier to construct non-parallel datasets by collecting large quantities of non-parallel sentences with different styles; conversely, carefully-aligned parallel TST datasets are much more costly to construct, usually requiring human annotation, and are thus scarce.

Published as a conference paper at ICLR 2022

The lack of parallel datasets for style transfer tasks prompts researchers to develop unsupervised TST methods trained on non-parallel datasets (Dai et al., 2019; Luo et al., 2019; Shen et al., 2017). Current methods either rely on surface-level split of styleand content-bearing tokens (Li et al., 2018a; Castro et al., 2017), or separate content and style vectors in latent space (John et al., 2019; Liu et al., 2020a; Wang et al., 2019) once style and content have been separated, they then swap styles leaving the content unchanged. However, surface-level split is not reliable when the style cannot be extracted at the token level (Sudhakar et al., 2019); latent-space disentanglement could also fail due to a lack of pairwise supervision and results in discarding too much style-independent content (Reif et al., 2021). However, these methods are less than ideal. As demonstrated in Figure 1 (a), they tend to learn the mapping between source and target styles on the original randomly mapped sentence pairs, but fail to take into account the inherent parallelism within the data (Krishna et al., 2020).

This paper presents La Mer, a novel self-supervised text style transfer framework built upon recent text-to-text language models (LMs). As illustrated in Figure 1 (b), rather than accepting the random mapping in the given non-parallel dataset, we first align sentences from source and target styles based on their scene graphs, and then run conditional MLE training on LM to maximize the probability of generating (roughly) parallel style sentences. As a final refinement step, we deploy reinforced imitation learning with contrastive loss on the aligned data to further reward the best transfer demonstration and penalize the others. In comparison with existing methods, La Mer fully exploits the intrinsic parallelism within non-parallel data and innovatively applies the transfer learning power of large-scale LMs to boost text-style transfer performance.

Our contributions can be outlined as follows: First, We present a language model based text style transfer framework, LAMER, that achieves significant advances in two existing benchmark TST tasks and one newly proposed challenging TST task. Human evaluations further confirm that La Mer generation is not only more readable but also richer in diverse expressions than previous methods. Second, We propose a new challenging text style transfer task, Political Stance Transfer, and release a benchmark dataset, which includes news sentence-pairs describing the same fact but from opposite political stances (i.e., liberal and conservative). The stances of the sentences are labeled by expert human editors. Third, As part of La Mer, we introduce a simple but powerful method to create parallel corpora from non-parallel data. We show that this method can be used as a general treatment to improve existing TST models. Given the abundance of non-parallel corpora for different TST tasks, this method has the potential to greatly impact the field of TST.

Code for La Mer is available at https://github.com/Dapang Liu/La Mer.

La Mer is comprised of three main steps. First, we mine (roughly) parallels target style expressions for each of the source style sentences based on their LM-based sentence embeddings and scene graphs ( 2.2). Second, we deploy MLE training on the text-to-text LM BART (Lewis et al., 2020b) to maximize the probability of generating such roughly parallel sentences in target style ( 2.3). Finally, we use imitation learning to further refine the generation with the self-supervision from the mined parallel demonstrations ( 2.4).

2.1 NOTATION AND PROBLEM STATEMENT

For the task of binary text style transfer (src tgt), We denote the source and target style datasets as Dsrc and Dtgt, which contain M and N sentences expressing two different styles respectively (i.e., Dsrc = {ssrc 1 , ssrc 2 , ...ssrc M}, Dtgt = {s tgt 1 , s tgt 2 , ...s tgt N }). We also denote the tokens in source and target sentences as {x1, x2, ..., xt} ssrc i and {y1, y2, ..., yt} s tgt j , for given source and target sentences ssrc i and s tgt j . Given an input sentence ssrc of the source style, an ideal style transfer model should be able to generate an appropriate, fluent target sentence ˆstgt, which is of the target style while maintaining the style-independent content.

Existing methods learn a mapping frandom Dsrc Dtgt without considering the self-parallelism between Dsrc and Dtgt; La Mer, however, works on the assumption that m( M) sentences in Dsrc

have n parallels in Dtgt. The number of parallels n for each of the m sentences is variant but no more

Published as a conference paper at ICLR 2022

than N. Therefore, the goal of La Mer can be framed as first identifying a parallel subset of the original non-parallel dataset and learning a mapping on this dataset: fparallel Dsrc m Dtgt n (m M, n N).

2.2 MINING (ROUGHLY) PARALLEL SENTENCES

they had such an extensive menu and the beef was delicious .

overpriced menu, and mediocre beef .

nothing great on their menu

[they, menu, beef]

[they, menu, nothing, beef]

S-Emb. + SAS Alignment

Figure 2: An example of aligned parallels by our scene graph based method. We underline the scene entities and annotate how we compute d Order and d Exist with ˆstgt+ and ˆstgt-.

In this work, we argue that the non-parallel text style transfer datasets have inherent property of parallelism. We outline in the following three major strategies to mine roughly parallel sentences:

Random (RD). For each sentence ssrc Dsrc, randomly pick k sentences from Dtgt as parallels. This method mainly serves as an ablation to show the effectiveness of other strategies.

Sentence Embedding (S-Emb.). Sentences that only differ in linguistic style should have higher semantic similarity than other unrelated sentences. As a more advanced strategy, we leverage off-theshelf LMs (e.g., Ro BERTa (Liu et al., 2019)) to mine semantically similar sentences. We first encode every sentence in Dsrc and Dtgt with their embeddings extracted from an LM. Then, for each sentence ssrc Dsrc, we pick k most similar sentences from Dtgt through cosine similarity as parallels.

S-Emb. + Scene Alignment Score (S-Emb. + SAS). We add a scene graph based alignment step to the sentence embedding method to leverage shared entities between source and target sentences. Scene graph extraction has been successfully used in computer vision (Venugopalan et al., 2015; Chen et al., 2020b) to guide (Pan et al., 2020) or judge (Anderson et al., 2016) diverse caption generation. Inspired by the CV example and the observation that scene graphs of sentences with similar content but different styles tend to have overlapping nodes, our algorithm highlights the shared entities (nodes) of the sentence scene graph by running the scene graph parser released by Wu et al. (2019)1, which outputs mined scenes graphs in subject-relation-object triplets. In our case, we only consider the end nodes of these triplets (which are subj. and obj.) as the scenes entities. Figure 2 shows an example of mined sentences and corresponding scenes entities.

Let e(ssrc) and e(stgt) be the number of scene entities extracted from the source and target style sentence, and e(ssrc tgt) be the number of their overlaps, we compute Scene Alignment Score (SAS) as:

SAS(ssrc stgt) = 1 stgt (1 + β2) precision recall

β2 precision + recall , (1)

where precision (= e(ssrc tgt)

e(stgt) ) and recall (= e(ssrc tgt)

e(ssrc) ) describe the portion of target and source entities that are part of the overlap, and positive real factor β is chosen such that recall is considered β times as important as precision in a specific domain. We also normalize SAS by the length of stgt, since longer target sentence candidates are more likely to have higher SAS.

SAS serves as a complementary alignment metric to further filter the parallel sentences picked by the previous sentence-embedding-based metric. The complete procedure of S-Emb. + SAS alignment is:

1. Encode each sentence in Dsrc and Dtgt as sentence embeddings via a pre-trained LM.

2. Compute cosine similarity between all sentence pairs (si, sj) where si Dsrc and sj Dtgt.

Pick the k most similar sentences s (k) j from Dtgt as parallel candidates for si.

1The scene graph parser can be found here: https://github.com/vacancy/Scene Graph Parser

Published as a conference paper at ICLR 2022

3. For each sentence si Dsrc and its parallel candidates s (k) j , compute pairwise SAS by Equation 1. Pick those sj whose SAS scores (with respect to si) are above a threshold p as the final parallel sentences to si.

From empirical observations, we found that the selection of p and k is task-specific and has crucial effects on the performance of La Mer. We explore the outcomes of different choices of k and p in 3.3. Though SAS alignment La Mer relies on entity alignment for mining parallel corpora, other alignment mechanisms, such as (multilingual) LM alignment (i.e., the S-Emb. method) can potentially be used for transfer cases where entities may not be shared across styles (such as ancient v.s. modern English (Xu et al., 2012), or even multilingual text style transfer (Krishna et al., 2021)).

2.3 CONDITIONAL TOKEN-LEVEL MLE TRAINING

After mining parallel sentences for a source sentence ssrc, from our corpus, we take the source sentence ssrc as the condition, and preform a conditional token-level MLE training on the mined parallels in the manner of standard autoregressive pre-training. Assuming for a certain source sentence ssrc we have n target parallels mined in the previous step (i.e., ssrc {s tgt 1 , s tgt 2 , ...stgt n }, n N), the MLE loss can be computed as:

t=1 y (i) t log LM(y1 t 1 ssrc), (2)

where yi t s tgt i denotes the i-th step ground truth token belonging to the i-th target parallel sentence. LM(y1 t 1 ssrc) is the LM output at t-th step (a distribution on the vocabulary) given the source text and previous generation. We choose BART Lewis et al. (2020b) as the LM, which is already pre-trained on several text-to-text tasks, and concatenate the source and target text with </s><s> tokens pre-defined by BART as input data. At each step, with the supervision of ground truth target tokens y (i) t , we minimize the MLE loss only at the target part. The LM is thus trained to maximize the probability of generating tokens ˆyt approaching to y (i) t .

Though the results produced are reasonable, solely relying on MLE training has several defects. First, MLE treats all parallels of a certain source equally without penalizing low-quality demonstrations, which not only makes the model less robust on noisy data, but also possible to be misled by negative samples during local optimization. Second, token-level MLE does not pay special attention to scene entities shared between the source and target, which we found take up the main portion in the human written references (as demonstrated in A.3). We are thus motivated to further improve the performance using the self-supervision from the parallels.

2.4 REINFORCED IMITATION LEARNING BY CONTRAST

For all the mined target style parallels stgt of ssrc, we denote the one that has the highest similarity score (with S-Emb.) or has the highest SAS (with S-Emb. + SAS) as the expert demonstration stgt+, while the others are set as stgt-.

In the formulation of reinforcement learning, at each step t, the policy πθ observes the state y1 t 1 (partially generated text), takes an action at = yt (pick a token), and transits to the next state y1 t. In the context of TST tasks, the policy can be interpreted as the LM-based generation model: ˆπ LM(at y1 t 1, ssrc), and the set of stgt+ can be seen as the trajectory derived from an expert policy π+, while stgtis produced by an amateur policy π . The goal of this step imitation learning can thus be framed as learning a policy that has more probability to behave like π+ rather than π , which can be achieved by minimizing the contrastive loss:

JIL = max(ψ [π+, ˆπ] ψ [π , ˆπ] + , 0), (3)

where ψ ( ) is the distance measurement of two policies in terms of semantic coherence d SEM( ) and scene preservation d PSV( ), such that ψ ( ) = d SEM( ) + d PSV( ); and is a hyperparameter

Published as a conference paper at ICLR 2022

that controls the contrasting margin. The general form is inspired by the triplet loss (Schroff et al., 2015) and recent in-batch constrastive learning (Gao et al., 2021): we encourage the learned policy ˆπ to be more similar to the expert one π+ than the amateur one π by at least margin . d SEM( ) and d PSV( ) are detailed below.

Sequence-level Semantic Coherence. For sequence ˆs generated under current learning policy ˆπ and a certain target style demonstration stgt, we encode them using bi-directional LM Ro BERTa (Liu et al., 2019), which is widely used as sentence encoder in related work (Reimers & Gurevych, 2019; Li et al., 2020). Then, we take their negative cosine similarity as the distance for semantic coherence:

d SEM = emb T ˆs embstgt embˆs embstgt . Lower value in d SEM means the learned policy ˆπ can produce similar quality sequences in terms of semantics as target style demonstrations. It is a sequence-level measurement since the encoding is based on the whole sentence.

Token-level Scene Preservation. In our context, legitimate target style parallels should preserve most scene entities that appear in source style sentences. To explicitly promote the content preservation performance of the current policy, we introduce a token-level distance d PSV. We first extract scene entities from stgt+ and each of the stgt , and then measure the differences in both the order of appearance of the entities and the existence of the entities. Taking stgt+ as an example (similar calculation is done on each of the stgt-), we compute: d PSV+ = αd Order+ + (1 α)d Exist+ = α e(stgt+), e(ssrc) Lev. +(1 α) e(stgt+) e(stgt+ src) H

2 where e( ) means the extracted scene entities, e( ) is its one-hot encoded vector, x, y Lev. measures Levenshtein distance (Yujian & Bo, 2007), one type of edit distance, and for vectors e(ssrc) and e(stgt), we compute their hamming distance3 (Norouzi et al., 2012). Lower d Order+ means the appearing order of e(ssrc) resemble e(stgt+). Similarly, lower d Exist+ means e(ssrc) and e(stgt+) are mostly overlapped. Similar procedure applies to d PSV as well. We present more details about how we compute these distances in A.4.

2.4.1 POLICY GRADIENT BY REINFORCE

We choose the REINFORCE algorithm (Williams, 1992) to optimize the current policy πθ. As shown in Equation 4, we use greedy decoding results as the baseline (Rennie et al., 2017), and sampling decoding for the exploration step, which mimics the actor-critic manner in reinforcement learning (Fujimoto et al., 2018; Haarnoja et al., 2018). The policy gradient can thus be presented as:

θ J(θ) = Eπθ [ θ log πθ (J sample IL (stgt+, stgt-) J greedy IL (stgt+, stgt-))] , (4)

where J sample IL and J greedy IL can be both calculated by Equation 3. We take the difference of two the mode rewards as the advantage, feeding into the REINFORCE algorithm to optimize our policy πθ. Finally, the output policy πθ is able to generate style transferred sequences resembling the expert demonstrations stgt+. More details about the complete loop of policy gradient can be found in A.5

Discussion: Imitation Learning, Reward-based RL, and MLE. Framing token-by-token text generation as a sequential decision-making process, reward-based reinforcement learning (RL) has achieved success in text generation by directly setting indifferentiable evaluation metrics such as BLEU (Papineni et al., 2002) or ROUGE (Lin, 2004) as the reward function (Gong et al., 2019; Liu et al., 2020b). In the context of TST tasks, however, instead of reward-based RL, we argue imitation learning (IL) is a better choice, since: 1) it is challenging to design a uniform reward fitting a variety of source style text (Chen et al., 2020a), and 2) it is also unnecessary to calculate an elaborately designed reward. The roughly parallel target style sentences mined in the alignment step can naturally serve as expert demonstrations (Li et al., 2018b). They are high-quality and cost-free, which can greatly improve the sampling and learning efficiency of RL optimization.

A common obstacle of IL is known as covariate shift (Ross & Bagnell, 2010), which means IL can provide a relatively good performance for the test cases similar to the demonstrations (used for policy training), but it could still suffer from bad generalization for cases it never meets during training, as

2α controls the weights assigned to d Order and d Exist; set by running repeated experiments ranging the α from 0 to 1 by 0.1, and picking the best-performing α with respect to GM: α = {0.4, 0.3, 0.1} for the three tasks.

3Hamming norm v H is defined as the number of non-zero entries of vector v.

Published as a conference paper at ICLR 2022

the number of demonstrations is finite. Similar problems called exposure bias (Bengio et al., 2015; Ranzato et al., 2016) also exists in MLE when the supervised data is scarce. We claim our token-level MLE training on LMs can somewhat mitigate such problems, since the large-scale LMs are already pre-trained on huge quantities of data across many domains as a prior.

3 EXPERIMENTS

3.1 DATASETS AND EXPERIMENTAL SETUP

Sentiment Transfer. We use the Yelp reviews dataset collected by Shen et al. (2017) which contains 250k negative sentences and 380k positive sentences, organized in non-parallel fashion. Li et al. (2018a) released a small test set that has 500 human written references for both target sentiments, and they are already one-on-one aligned with the source sentiment sentences.

Formality Transfer. A more challenging TST task is to modify the formality of a given sentence. We use the GYAFC dataset (Rao & Tetreault, 2018), which contains formal and informal sentences from two domains. In this paper, we choose the Family Relationship domain, which consists of about 52k training sentences, 5k development sentences, and 2.5k test sentences. Different from the Yelp dataset, the test set of GYAFC has four human written references for each source style test case. Our evaluation results for this task are reported as the average against the four references.

Political Stance Transfer. News outlets reporting the same event naturally meet the requirement of style transfer data, which is aligned in content but different in style. We leverage Allsides.com4 as an aggregator, where outlets reporting the same events with different political stances are already organized and labeled by expert human editors. In total, we collect 2,298 full-length news articles from 6/1/2012 to 4/1/2021 in pairs of liberal-conservative leaning. Though the data seems parallel at the article level, in this work we focus on sentence level TST. We further split the articles into about 56k sentence pairs as a non-parallel dataset for training (detailed in A.2). These titles are written by expert writers from liberalor conservative-learning media outlets, and aligned by the Allsides editors. For ethical consideration, we manually examined the articles to removing those containing sensitive topics (such as hate speech). To adhere to the outlets TOS, we will release our dataset as links to the articles (instead of the text), with the appropriate code to extract the content for training purposes. A closer comparison of our proposed novel dataset with another political stance TST dataset (Voigt et al., 2018) can be found in A.1.

The detailed statistics of all three benchmark datasets and our experimental setup can be found in A.3. We have discussed the hyperparameter selection for the best performing La Mer in 3.3.

3.2 MAIN RESULTS ON TEXT STYLE TRANSFER PERFORMANCE

TST models should be able to transfer source style sentence to a target style, while maintaining the style-independent content as much as possible. Most existing works evaluate their methods with transfer accuracy (ACC; classification accuracy) and content-preservation (BLEU-ref; token overlaps with reference). In this work, besides using ACC and BLEU-ref, we also report more advanced metrics SIM and FL introduced by Krishna et al. (2020) to further evaluate our method.

La Mer has competitive performance. Table 1 shows the main results comparing La Mer with eight popular TST models on the three benchmark datasets. In addition to the independent ACC and BLEU score5, we also report their geometric mean (GM) for direct comparison with the Dual RL and STRAP baselines; however, it should be noted that GM has not been well studied as an evaluation metric for TST and is only included for comparison purposes. We also consider inverse PINC scores (Chen & Dolan, 2011) i-PINC as a complementary metric of BLEU. i-PINC counts the n-gram overlaps after removing the tokens shared between source and target style sentences, and thus properly penalizes the Input Copy method (i-PINC = 0).

Compared with existing methods, La Mer performs better in terms of overall metric GM, and shows particularly strong net transfer ability (since i-PINC does not count duplicated tokens between ssrc

4https://www.allsides.com/story/admin 5We use BLEU-ref, which computes average n-gram BLEU (n ranges from 1 to 4).

Published as a conference paper at ICLR 2022

Table 1: La Mer achieves competitive performance in three TST tasks. CAE (Shen et al., 2017), Del Retr Gen (Li et al., 2018a), Dual RL (Luo et al., 2019), Style Transformer (Dai et al., 2019), Deep Latent Seq (He et al., 2020) are popular TST models not built on LMs. TSF-Del Ret Gen (Sudhakar et al., 2019), IMa T (Jin et al., 2019), and STRAP (Krishna et al., 2020) are recent LM-based TST models. Input Copy directly copies the source style sentence as the transfer output. GM: Geometric Mean of Transfer Accuracy (ACC) and BLEU. i-PINC measures net transfer ability by not counting words that can be directly copied from the source style sentence, as discussed in 3.2. We color (

) the best results, and underline the second best, not including the Input Copy reference.

Sentiment SAS = 0.73

Formality SAS = 0.29

Political Stance SAS = 0.30

Existing Methods ACC BLEU GM i-PINC ACC BLEU GM i-PINC ACC BLEU GM i-PINC

Input Copy (ref.) 7.4 61.7 21.4 0 1.9 54.6 10.2 0 7.5 17.8 11.6 0 CAE 70.5 6.0 20.6 1.2 39.5 35.8 37.6 5.9 31.5 5.1 12.7 1.4 Del Retr Gen 65.4 11.8 27.8 4.8 61.9 14.8 30.3 6.6 93.8 3.5 18.1 6.2 Dual RL 85.6 21.2 42.5 4.0 71.1 31.9 47.9 7.2 91.8 8.1 27.3 5.7 Style Transformer 93.6 17.1 40.0 5.2 72.2 32.7 48.6 4.1 48.6 7.3 18.8 1.1 Deep Latent Seq 87.9 18.7 40.5 7.3 80.5 18.5 38.6 6.1 85.8 20.8 42.2 3.0

LM-based TSF-Del Ret Gen 89.5 31.7 53.3 4.5 79.3 17.4 37.1 7.3 65.8 23.7 39.5 2.9 IMa T 95.9 22.5 46.5 3.3 72.1 28.2 45.1 12.7 82.9 19.5 40.2 5.3 STRAP 92.3 32.5 54.8 6.4 75.4 32.6 49.6 9.7 84.3 21.3 42.4 5.0

Ours: La Mer w/. RD 97.1 19.8 43.8 2.5 64.2 20.4 36.2 5.5 75.9 10.1 27.7 6.1 w/. S-Emb. 90.1 40.6 60.5 5.6 65.1 33.8 46.9 10.9 78.2 22.1 41.6 9.2 w/. S-Emb. + SAS 97.0 34.1 57.5 9.6 76.5 39.2 54.8 13.3 82.7 30.5 50.2 13.6

and stgt). Among the three mining strategies we proposed in 2.2, the S-Emb. + SAS seems to offer the best performance boosting in terms of content preservation, but shows no clear advantage on style control accuracy. When comparing La Mer with the three LM-based models, we find the advantage of La Mer mainly comes from the gain in content preservation (BLEU), which demonstrates the effectiveness of our scene graph alignment. We also find some of the other TST models produce low-fluency (i.e., high perplexity) transferred sentences (see Table 10 in A.7), though we must note that recent studies have shown perplexity to be a poor measure of fluency (Krishna et al., 2020; Mir et al., 2019; Lee et al., 2021).

Table 2: Few-shot performance on Formality TST (only 1% training data available) of La Mer and other LM-based baselines. We also include a zero-shot TST method built on GPT-3 (Reif et al., 2021) for reference. We annotate standard deviation next to the results as well. Training Data (%) 1% (0% for GPT3) 100%

Method ACC SIM FL JA,S,F ACC SIM FL JA,S,F TSF-Del Ret Gen 16.51.4 5.51.9 90.30.1 2.61.1 54.11.4 30.50.8 98.90.1 29.70.5 IMa T 35.90.4 22.10.2 23.00.4 3.20.7 59.51.2 39.71.1 99.60.1 37.40.2 STRAP 37.70.7 12.50.9 70.40.2 5.30.4 67.71.2 72.51.1 90.40.7 45.51.2 Ours: La Mer 48.80.4 27.40.6 89.80.1 13.50.3 66.40.5 74.30.3 97.80.1 53.20.1 GPT-3prompt TST 10.20.6 17.11.1 98.00.2 6.50.4

Table 3: Compared with other baselines, La Mer requires neither extra data nor training additional models.

Ext. Resource? Data Model

TSF-Del Ret Gen ! ! IMa T ! % STRAP ! ! Ours: La Mer % %

GPT-3prompt TST % %

La Mer is robust in few-shot settings. La Mer learns the inherent parallelism within the unsupervised data, which could be challenging in data scarce cases. Recent works have explored the limits of TST in few-shot settings (Riley et al., 2021; Krishna et al., 2021), we conduct similar studies by limiting the available training data (to 1%), aiming to answer whether or not La Mer can still provide comparable performance as others. We consider LM-based baselines which show strong results in Table 1. We also compare with a zero-shot prompt-based TST baseline built upon GPT-3 (Reif et al., 2021) for reference. For evaluation metrics, we use the recently proposed J-score framework (including ACCuracy, SIMilarity, and FLuency) (Krishna et al., 2020), which solves many of the

Published as a conference paper at ICLR 2022

33.00 33.25 33.50 33.75 34.00 34.25 34.50

(k=200, p=0.6)

(k=200, p=0.8)

(k=300, p=0.6)

(k=200, p=0.7)

(k=100, p=0.6)

30 33 36 39 42 45 65

85 (k=500, p=0.4)

(k=200, p=0.3)

(k=200, p=0.5)

(k=200, p=0.4)

(k=50, p=0.4)

20 22 24 26 28 30 32 70

(k=50, p=0.4)

(k=500, p=0.3)

(k=500, p=0.4)

(k=500, p=0.5)

(k=1000, p=0.4)

Political Stance

Figure 3: Effect on performance when choosing different k and p for S-Emb. + SAS alignment. We also annotate (with a star) the recommended settings which reaches best balance between style control accuracy and content preservation BLEU score.

issues in existing metrics 6. As shown in Table 2, we find that La Mer achieves great performance (J-score) in few-shot cases, partially because other baselines rely on external systems trained on ample data;7 In extreme data hungry cases, the compromised external modules would propagate errors to following parts as they are coupled. La Mer, instead, requires neither extra data nor additional training on external systems (See Table 3), but fully leverages the self-supervision hidden in the unsupervised data, which achieves robust performance even in few-shot scenarios.

IL refinement is crucial for La Mer content preservation. As shown in Table 4, the BLEU score for La Mer drops substantially when the imitation learning refinement stage is totally removed (No IL, MLE only), and we find this drop is higher for challenging style tasks (such as political stance transfer), partially because the quality variance of their demonstrations is higher, i.e., a further guidance towards good demonstrations will make a big difference. We also conducted further ablation studies on components of IL: In general, removing d Order seems to cause greater performance drop in formality TST, while d Exist seems important to successful political stance TST. Removing all scene preservation distance sub-components (i.e., IL (d SEM)) leads to noticeable performance deterioration across all three TST tasks. We take these ablation study results as evidence that the technologies we propose for La Mer are all crucial.

Table 4: Further ablation on components of La Mer imitation learning (IL). d SEM: semantic coherence distance, d Order and d Exist are sub-components of scene preservation distance d PSV. Removing all IL modules (MLE only) will cause severe drop in content preservation (BLEU), especially for challenging TST tasks (Formality and Political Stance TST).

Sentiment Formality Political Stance

Ablation on La Mer IL ACC BLEU ACC BLEU ACC BLEU

IL (d SEM + d Order + d Exist) 97.0 34.1 76.5 39.2 93.4 18.5 IL (d SEM + d Order) 95.5 ( 1.5) 34.4 ( 0.3) 74.7 ( 1.8) 37.7 ( 1.5) 93.2 ( 0.2) 15.8 ( 2.7) IL (d SEM + d Exist) 96.5 ( 0.5) 33.9 ( 0.2) 72.6 ( 3.9) 34.1 ( 5.1) 92.7 ( 0.7) 17.1 ( 1.4) IL (d SEM) 95.1 ( 1.9) 33.3 ( 0.8) 71.9 ( 4.6) 32.5 ( 6.7) 89.9 ( 0.5) 13.9 ( 4.6)

No IL (MLE Only) 93.1 ( 3.9) 33.2 ( 0.9) 68.7 ( 7.8) 27.2 ( 12.0) 89.8 ( 0.6) 8.8 ( 9.7)

3.3 CONFIGURING LAMER

The filtering parameter p and k are hyperparameters that are crucial for the construction of roughly parallel datasets. Thus, for all three TST tasks, we evaluate La Mer performance given different combination of p and k, as shown in Figure 3. We find there is a approximate trade-off relationship between style control accuracy and content preservation BLEU score: In general, higher k (more parallels) seems to improve accuracy but normally has lower BLEU, while for a fixed k, a proper

6E.g., BLEU score can be artificially high when directly copying source style sentence as the target style generation, and perplexity is unbounded and discourages creative TST generation. 7E.g., TSF-Del Ret Gen requires training a classifier to locate style-carrying spans of text. STRAP fine-tunes a GPT-2 model to learn how to paraphrase. IMa T trains a plain Seq2Seq model from scratch as an LM.

Published as a conference paper at ICLR 2022

p can be reached empirically. We annotate the recommended settings of p and k for the three TST tasks we studied in this work.

3.4 HUMAN EVALUATION

We conducted human evaluation to answer: 1) How good is La Mer s SAS alignment? , and 2) How is the quality of the style-transferred sentences? (The exact questions and procedures are details in A.6.). We recruited 212 participants from MTurk. Most annotators agreed that SAS-aligned pairs share similar style-independent content (Table 5). We conducted several one-way ANOVAs to test the difference among the seven baselines and La Mer (political stance TST results shown in Table 6; see Table 8 for sentiment TST and Table 9 for formality TST in A.6). La Mer was significantly better than next-best in all three perspectives (p < .05), and had the most ratings that where 5.

Table 5: Effectiveness of SAS alignment for mining weakly parallel sentences. Similarity: How much the mined pairs are similar in styleindependent content? SAS Align: How well SAS scores reflect such similarity? Ratings are on a 7 point scale. All 4: The ratio for which the ratings are 4 for both questions.

Similarity SAS

TST Tasks Mean SD Mean SD All 4 Sentiment 5.421.36 5.561.17 88% Formality 5.581.14 5.741.09 82% Political 5.451.17 5.431.16 74%

Table 6: Human evaluations on the political stance transfer task (on a 7 point scale). We bold the highest average rating. La Mer was rated significantly higher than the next-best method (p < .05). All 5: The ratio for which the ratings are 5 for all three perspectives.

Political Stance Style Content Readability

Methods Mean SD Mean SD Mean SD All 5 CAE 4.91 1.59 4.85 1.58 5.04 1.68 12% Del Retr Gen 5.14 1.60 4.93 1.63 4.96 1.60 11% Dual RL 5.15 1.37 4.87 1.59 5.15 1.53 27% Style Transformer 5.31 1.19 5.34 1.36 5.18 1.40 28% Deep Latent Seq 5.42 0.57 5.29 1.20 5.30 1.26 25% TSF-Del Ret Gen 4.87 0.65 5.23 1.14 5.20 1.13 15% STARP 5.33 1.25 5.31 1.40 5.43 1.07 21% Ours: La Mer 5.59 1.17 5.44 1.11 5.56 1.05 37%

4 RELATED WORK

Besides the baselines we have studies in the paper, we notice there are other TST methods focusing on either latent vector manipulation (Liu et al., 2020a; Wang et al., 2019; John et al., 2019) or surface level editing (Castro et al., 2017; Karadzhov et al., 2017; Mansoorizadeh et al.). La Mer differs from them as it is learning to transfer with self-supervision mined from the unsupervised data, without complicated manipulation or editing. The idea of mining roughly parallel demonstrations in La Mer seems to echo the recent findings of retrieval-based generation systems (Lewis et al., 2020a; Lee et al., 2019) retrieving relevant documents from external resources (e.g., Wikipedia) as grounding can improve the quality of generation (Lewis et al., 2020c; Guu et al., 2020). La Mer does not rely on any external resources, but mines parallel demonstrations from the training data itself, and thus is easier to deploy and alleviates potential bias.

5 CONCLUSION

We have proposed La Mer, a self-supervised text style transfer method built on large-scale LMs. La Mer demonstrates the potential of leveraging self-parallelism within large non-parallel corpora, which can guide efficient TST models development or even LM design. On three diverse TST tasks, we show that La Mer performs better than eight existing methods while being robust to data scarcity. Human evaluations further validate the advantage of our model compared to existing baselines.

Future work of La Mer could study how to extend to multilingual scenarios, or transfer the style that essentially changes the form of language (e.g., ancient v.s. modern English). One could also investigate combining the alignment module of La Mer with the text2text LM in one model, and jointly train to further boost the performance. Another direction is to further study the integration of La Mer with larger size foundation models, like GPT-3 175B to understand La Mers potential in few-shot settings.

Published as a conference paper at ICLR 2022

6 ETHICS AND REPRODUCIBILITY STATEMENT

The goal of La Mer is to provide a simple but efficient general-purpose text style transfer framework by leveraging large-scale pre-trained LMs. La Mer can generate text in a specified target style given source style input, which can serve as a useful tool for social good. For example, La Mer can assist language formalization (formality transfer), enable data anonymization (writing style transfer), or depolarize politically biased text (Liu et al., 2022; 2021a), via transferring the identifiable style of the input text.

The performance of La Mer depends on the pre-trained LM, which means that the generation of La Mer can be affected by certain societal biases found in the pre-trained LM it is based on (e.g., BART) (Blodgett et al., 2020). Other negatives such as abusive language (Hancock et al., 2020), hate speech (Kennedy et al., 2018), and dehumanization (Mendelsohn et al., 2020) are also prevalent in the language data for LM pre-training. When deployed in real-world settings, La Mer may increase the risk of reinforcing and propagating polarizing and toxic language in public space through people s dissemination (Waseem et al., 2017). Moreover, the generated text of La Mer could potentially be included in the next iteration of large-scale LM training data collection, which may amplify offensive stereotypes and overtly abusive views in the future LMs.

Though La Mer was developed as an neutral tool mainly for research purposes, it can also be used in malicious ways such as generating sophisticated spam in a particular style disguise , which we strongly discourage (Rae et al., 2021; Liu et al., 2021b). Due to the lack of fact checking systems as a safety measure, La Mer is at the risk of generating seemingly coherent but counterfactual text in a certain target style, which could further undermine social trust in computer-mediated forms of communication (Liu et al., 2021c; Petroni et al., 2019). We thus recommend the users of La Mer to integrate necessary safety checks into the actual systems.

In addition, our experiments and analysis are done when source and target style sentences are both in English, and therefore we do not claim that our findings will generalize across all languages, although our framework has the potential to be extended to other languages with necessary modifications.

Though it s beyond the scope of this paper and our individual ability to address all above-mentioned complications, we strongly encourage users of our system (and other TST models and systems based on pre-trained language models) to acknowledge and beware of these issues.

Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pp. 382 398. Springer, 2016.

Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pp. 1171 1179, 2015. URL https://proceedings.neurips.cc/paper/2015/hash/ e995f98d56967d946471af29d7bf99f1-Abstract.html.

Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. Language (technology) is power: A critical survey of bias in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 5454 5476, Online, 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.485. URL https://aclanthology.org/ 2020.acl-main.485.

Daniel Castro, Reynier Ortega, and Rafael Muñoz. Author masking by sentence transformation notebook for pan at clef 2017. In CLEF 2017 Evaluation Labs and Workshop Working Notes Papers, pp. 11 14, 2017.

David Chen and William Dolan. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human

Published as a conference paper at ICLR 2022

Language Technologies, pp. 190 200, Portland, Oregon, USA, 2011. Association for Computational Linguistics. URL https://aclanthology.org/P11-1020.

Liqun Chen, Ke Bai, Chenyang Tao, Yizhe Zhang, Guoyin Wang, Wenlin Wang, Ricardo Henao, and Lawrence Carin. Sequence generation with optimal-transport-enhanced reinforcement learning. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pp. 7512 7520. AAAI Press, 2020a. URL https: //aaai.org/ojs/index.php/AAAI/article/view/6249.

Shizhe Chen, Qin Jin, Peng Wang, and Qi Wu. Say as you wish: Fine-grained control of image caption generation with abstract scene graphs. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pp. 9959 9968. IEEE, 2020b. doi: 10.1109/CVPR42600.2020.00998. URL https://doi.org/10.1109/ CVPR42600.2020.00998.

Ning Dai, Jianze Liang, Xipeng Qiu, and Xuanjing Huang. Style transformer: Unpaired text style transfer without disentangled latent representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 5997 6007, Florence, Italy, 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1601. URL https://aclanthology. org/P19-1601.

Yao Fu, Yansong Feng, and John P. Cunningham. Paraphrase generation with latent bag of words. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, Neur IPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pp. 13623 13634, 2019a. URL https://proceedings.neurips.cc/ paper/2019/hash/5e2b66750529d8ae895ad2591118466f-Abstract.html.

Yao Fu, Hao Zhou, Jiaze Chen, and Lei Li. Rethinking text attribute transfer: A lexical analysis. In Proceedings of the 12th International Conference on Natural Language Generation, pp. 24 33, Tokyo, Japan, 2019b. Association for Computational Linguistics. doi: 10.18653/v1/W19-8604. URL https://aclanthology.org/W19-8604.

Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. Style transfer in text: Exploration and evaluation. In Sheila A. Mc Ilraith and Kilian Q. Weinberger (eds.), Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pp. 663 670. AAAI Press, 2018. URL https://www.aaai.org/ocs/index.php/AAAI/AAAI18/ paper/view/17015.

Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in actorcritic methods. In Jennifer G. Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15, 2018, volume 80 of Proceedings of Machine Learning Research, pp. 1582 1591. PMLR, 2018. URL http://proceedings.mlr.press/v80/fujimoto18a.html.

Tianyu Gao, Xingcheng Yao, and Danqi Chen. Sim CSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 6894 6910, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.emnlp-main.552. URL https://aclanthology.org/2021.emnlp-main.552.

Hongyu Gong, Suma Bhat, Lingfei Wu, Jin Jun Xiong, and Wen-mei Hwu. Reinforcement learning based text style transfer without parallel training corpus. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 3168 3180, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1320. URL https:// aclanthology.org/N19-1320.

Published as a conference paper at ICLR 2022

Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pp. 3929 3938. PMLR, 2020. URL http://proceedings. mlr.press/v119/guu20a.html.

Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In Jennifer G. Dy and Andreas Krause (eds.), Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15, 2018, volume 80 of Proceedings of Machine Learning Research, pp. 1856 1865. PMLR, 2018. URL http://proceedings. mlr.press/v80/haarnoja18b.html.

Jeffrey T Hancock, Mor Naaman, and Karen Levy. Ai-mediated communication: definition, research agenda, and ethical considerations. Journal of Computer-Mediated Communication, 25(1):89 100, 2020.

Junxian He, Xinyi Wang, Graham Neubig, and Taylor Berg-Kirkpatrick. A probabilistic formulation of unsupervised text style transfer. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. Open Review.net, 2020. URL https: //openreview.net/forum?id=HJl A0C4t PS.

Zhijing Jin, Di Jin, Jonas Mueller, Nicholas Matthews, and Enrico Santus. IMa T: Unsupervised text attribute transfer via iterative matching and translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 3097 3109, Hong Kong, China, 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1306. URL https:// aclanthology.org/D19-1306.

Vineet John, Lili Mou, Hareesh Bahuleyan, and Olga Vechtomova. Disentangled representation learning for non-parallel text style transfer. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 424 434, Florence, Italy, 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1041. URL https://aclanthology. org/P19-1041.

Georgi Karadzhov, Tsvetomila Mihaylova, Yasen Kiprov, Georgi Georgiev, Ivan Koychev, and Preslav Nakov. The case for being average: A mediocrity approach to style masking and author obfuscation. In International Conference of the Cross-Language Evaluation Forum for European Languages, pp. 173 185. Springer, 2017.

Brendan Kennedy, Drew Kogon, Kris Coombs, Joseph Hoover, Christina Park, Gwenyth Portillo Wightman, Aida Mostafazadeh Davani, Mohammad Atari, and Morteza Dehghani. A typology and coding manual for the study of hate-based rhetoric. Psy Ar Xiv. July, 18, 2018.

Kalpesh Krishna, John Wieting, and Mohit Iyyer. Reformulating unsupervised style transfer as paraphrase generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 737 762, Online, 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.emnlp-main.55. URL https://aclanthology.org/ 2020.emnlp-main.55.

Kalpesh Krishna, Deepak Nathani, Xavier Garcia, Bidisha Samanta, and Partha Talukdar. Few-shot controllable style transfer for low-resource settings: A study in indian languages. Ar Xiv preprint, abs/2110.07385, 2021. URL https://arxiv.org/abs/2110.07385.

Dongkyu Lee, Zhiliang Tian, Lanqing Xue, and Nevin L. Zhang. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 93 102, Online, 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.8. URL https: //aclanthology.org/2021.acl-long.8.

Published as a conference paper at ICLR 2022

Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 6086 6096, Florence, Italy, 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1612. URL https://aclanthology.org/P19-1612.

Mike Lewis, Marjan Ghazvininejad, Gargi Ghosh, Armen Aghajanyan, Sida I. Wang, and Luke Zettlemoyer. Pre-training via paraphrasing. In Hugo Larochelle, Marc Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, Neur IPS 2020, December 6-12, 2020, virtual, 2020a. URL https://proceedings.neurips.cc/ paper/2020/hash/d6f1dd034aabde7657e6680444ceff62-Abstract.html.

Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 7871 7880, Online, 2020b. Association for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.703. URL https://aclanthology.org/2020.acl-main.703.

Patrick S. H. Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Hugo Larochelle, Marc Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan Tien Lin (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, Neur IPS 2020, December 6-12, 2020, virtual, 2020c. URL https://proceedings.neurips.cc/paper/2020/hash/ 6b493230205f780e1bc26945df7481e5-Abstract.html.

Jingjing Li, Zichao Li, Lili Mou, Xin Jiang, Michael R. Lyu, and Irwin King. Unsupervised text generation by learning from search. In Hugo Larochelle, Marc Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin (eds.), Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, Neur IPS 2020, December 6-12, 2020, virtual, 2020. URL https://proceedings.neurips.cc/paper/ 2020/hash/7a677bb4477ae2dd371add568dd19e23-Abstract.html.

Juncen Li, Robin Jia, He He, and Percy Liang. Delete, retrieve, generate: a simple approach to sentiment and style transfer. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 1865 1874, New Orleans, Louisiana, 2018a. Association for Computational Linguistics. doi: 10.18653/v1/N18-1169. URL https://aclanthology.org/N18-1169.

Zichao Li, Xin Jiang, Lifeng Shang, and Hang Li. Paraphrase generation with deep reinforcement learning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 3865 3878, Brussels, Belgium, 2018b. Association for Computational Linguistics. doi: 10.18653/v1/D18-1421. URL https://aclanthology.org/D18-1421.

Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pp. 74 81, Barcelona, Spain, 2004. Association for Computational Linguistics. URL https://aclanthology.org/W04-1013.

Dayiheng Liu, Jie Fu, Yidan Zhang, Chris Pal, and Jiancheng Lv. Revision in continuous space: Unsupervised text style transfer without adversarial learning. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pp. 8376 8383. AAAI Press, 2020a. URL https://aaai.org/ojs/index.php/AAAI/article/view/6355.

Ruibo Liu, Guangxuan Xu, Chenyan Jia, Weicheng Ma, Lili Wang, and Soroush Vosoughi. Data boost: Text data augmentation through reinforcement learning guided conditional generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 9031 9041, Online, 2020b. Association for Computational Linguistics. doi: 10.18653/v1/2020. emnlp-main.726. URL https://aclanthology.org/2020.emnlp-main.726.

Published as a conference paper at ICLR 2022

Ruibo Liu, Chenyan Jia, Jason Wei, Guangxuan Xu, Lili Wang, and Soroush Vosoughi. Mitigating political bias in language models through reinforced calibration. Proceedings of the AAAI Conference on Artificial Intelligence, 35(17):14857 14866, 2021a. URL https://ojs.aaai.org/ index.php/AAAI/article/view/17744.

Ruibo Liu, Jason Wei, and Soroush Vosoughi. Language model augmented relevance score. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 6677 6690, Online, 2021b. Association for Computational Linguistics. doi: 10.18653/v1/2021. acl-long.521. URL https://aclanthology.org/2021.acl-long.521.

Ruibo Liu, Guoqing Zheng, Shashank Gupta, Radhika Gaonkar, Chongyang Gao, Soroush Vosoughi, Milad Shokouhi, and Ahmed Hassan Awadallah. Knowledge infused decoding. In International Conference on Learning Representations, 2021c. URL https://openreview.net/pdf? id=upn DJ7itech.

Ruibo Liu, Chenyan Jia, Jason Wei, Guangxuan Xu, and Soroush Vosoughi. Quantifying and alleviating political bias in language models. Artificial Intelligence, 304:103654, 2022. ISSN 0004-3702. doi: https://doi.org/10.1016/j.artint.2021.103654. URL https://www.sciencedirect. com/science/article/pii/S0004370221002058.

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. Ar Xiv preprint, abs/1907.11692, 2019. URL https://arxiv.org/abs/1907.11692.

Fuli Luo, Peng Li, Jie Zhou, Pengcheng Yang, Baobao Chang, Xu Sun, and Zhifang Sui. A dual reinforcement learning framework for unsupervised text style transfer. In Sarit Kraus (ed.), Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pp. 5116 5122. ijcai.org, 2019. doi: 10.24963/ijcai. 2019/711. URL https://doi.org/10.24963/ijcai.2019/711.

Muharram Mansoorizadeh, Taher Rahgooy, Mohammad Aminiyan, and Mahdy Eskandari. Author obfuscation using wordnet and language models notebook for pan at clef 2016.

Julia Mendelsohn, Yulia Tsvetkov, and Dan Jurafsky. A framework for the computational linguistic analysis of dehumanization. Frontiers in artificial intelligence, 3:55, 2020.

Remi Mir, Bjarke Felbo, Nick Obradovich, and Iyad Rahwan. Evaluating style transfer for text. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 495 504, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1049. URL https://aclanthology.org/N19-1049.

Mohammad Norouzi, David J. Fleet, and Ruslan Salakhutdinov. Hamming distance metric learning. In Peter L. Bartlett, Fernando C. N. Pereira, Christopher J. C. Burges, Léon Bottou, and Kilian Q. Weinberger (eds.), Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems 2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States, pp. 1070 1078, 2012. URL https://proceedings.neurips.cc/paper/2012/hash/ 59b90e1005a220e2ebc542eb9d950b1e-Abstract.html.

Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pp. 10867 10876. IEEE, 2020. doi: 10.1109/CVPR42600.2020.01088. URL https://doi.org/10.1109/CVPR42600.2020.01088.

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 of the Association for Computational Linguistics, pp. 311 318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. doi: 10.3115/1073083.1073135. URL https://aclanthology. org/P02-1040.

Published as a conference paper at ICLR 2022

Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 2463 2473, Hong Kong, China, 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1250. URL https://aclanthology.org/D19-1250.

Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. Scaling language models: Methods, analysis & insights from training gopher. Ar Xiv preprint, abs/2112.11446, 2021. URL https://arxiv.org/abs/2112.11446.

Marc Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba. Sequence level training with recurrent neural networks. In Yoshua Bengio and Yann Le Cun (eds.), 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016. URL http://arxiv.org/abs/1511.06732.

Sudha Rao and Joel Tetreault. Dear sir or madam, may I introduce the GYAFC dataset: Corpus, benchmarks and metrics for formality style transfer. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 129 140, New Orleans, Louisiana, 2018. Association for Computational Linguistics. doi: 10.18653/v1/N18-1012. URL https://aclanthology. org/N18-1012.

Emily Reif, Daphne Ippolito, Ann Yuan, Andy Coenen, Chris Callison-Burch, and Jason Wei. A recipe for arbitrary text style transfer with large language models. Ar Xiv preprint, abs/2109.03910, 2021. URL https://arxiv.org/abs/2109.03910.

Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERTnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pp. 3982 3992, Hong Kong, China, 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1410. URL https://aclanthology.org/D19-1410.

Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pp. 1179 1195. IEEE Computer Society, 2017. doi: 10.1109/CVPR.2017.131. URL https://doi.org/10.1109/CVPR. 2017.131.

Parker Riley, Noah Constant, Mandy Guo, Girish Kumar, David Uthus, and Zarana Parekh. Text SETTR: Few-shot text style extraction and tunable targeted restyling. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 3786 3800, Online, 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.293. URL https://aclanthology.org/2021.acl-long.293.

Stéphane Ross and Drew Bagnell. Efficient reductions for imitation learning. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pp. 661 668. JMLR Workshop and Conference Proceedings, 2010.

Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pp. 815 823. IEEE Computer Society, 2015. doi: 10.1109/CVPR.2015.7298682. URL https://doi.org/10.1109/CVPR.2015. 7298682.

Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. Style transfer from nonparallel text by cross-alignment. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pp.

Published as a conference paper at ICLR 2022

6830 6841, 2017. URL https://proceedings.neurips.cc/paper/2017/hash/ 2d2c8394e31101a261abf1784302bf75-Abstract.html.

Akhilesh Sudhakar, Bhargav Upadhyay, and Arjun Maheswaran. transforming delete, retrieve, generate approach for controlled text style transfer. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 3269 3279, Hong Kong, China, 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1322. URL https:// aclanthology.org/D19-1322.

Subhashini Venugopalan, Marcus Rohrbach, Jeffrey Donahue, Raymond J. Mooney, Trevor Darrell, and Kate Saenko. Sequence to sequence - video to text. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pp. 4534 4542. IEEE Computer Society, 2015. doi: 10.1109/ICCV.2015.515. URL https://doi.org/10.1109/ ICCV.2015.515.

Rob Voigt, David Jurgens, Vinodkumar Prabhakaran, Dan Jurafsky, and Yulia Tsvetkov. Rt Gender: A corpus for studying differential responses to gender. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan, 2018. European Language Resources Association (ELRA). URL https://aclanthology.org/ L18-1445.

Ke Wang, Hang Hua, and Xiaojun Wan. Controllable unsupervised text attribute transfer via editing entangled latent representation. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d Alché-Buc, Emily B. Fox, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, Neur IPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pp. 11034 11044, 2019. URL https://proceedings.neurips.cc/paper/2019/hash/ 8804f94e16ba5b680e239a554a08f7d2-Abstract.html.

Zeerak Waseem, Thomas Davidson, Dana Warmsley, and Ingmar Weber. Understanding abuse: A typology of abusive language detection subtasks. In Proceedings of the First Workshop on Abusive Language Online, pp. 78 84, Vancouver, BC, Canada, 2017. Association for Computational Linguistics. doi: 10.18653/v1/W17-3012. URL https://aclanthology.org/W17-3012.

Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229 256, 1992.

Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pp. 6609 6618. Computer Vision Foundation / IEEE, 2019. doi: 10.1109/CVPR.2019.00677. URL http://openaccess.thecvf.com/content_CVPR_ 2019/html/Wu_Unified_Visual-Semantic_Embeddings_Bridging_Vision_ and_Language_With_Structured_Meaning_CVPR_2019_paper.html.

Wei Xu, Alan Ritter, Bill Dolan, Ralph Grishman, and Colin Cherry. Paraphrasing for style. In Proceedings of COLING 2012, pp. 2899 2914, Mumbai, India, 2012. The COLING 2012 Organizing Committee. URL https://aclanthology.org/C12-1177.

Li Yujian and Liu Bo. A normalized levenshtein distance metric. IEEE transactions on pattern analysis and machine intelligence, 29(6):1091 1095, 2007.

Published as a conference paper at ICLR 2022

A.1 COMPARING WITH ANOTHER POLITICAL STANCE DATASET

Transferring political stance has also been studied in Voigt et al. (2018), where the author also collected a dataset for political stance TST. Comparing with their dataset, we found ours has following differences:

Language Formality: Their dataset is composed of comments on Facebook posts from members of the United States Senate and House, while our dataset consists of sentences from news articles written by professional reporters belonging to well-known media outlets (CNN, Fox News, etc.). The text in the dataset of Voigt et al. (2018) tends to be more casual and oral (e.g., congrats , lee ! ). Conversely, the text in our dataset is generally more formal and well-formed.

Completeness: Their dataset does not contain the original post to which the comments are responding, which causes some potential ambiguity on its political stance labels especially when the comment is generic (e.g., happy birthday jerry : - ) ). Most of the sentences in our dataset are complete statements of certain facts. We also conducted an extra distillation step (described in Appendix A in the supplementary document) to only keep the sentences that are salient to the article s stance label.

Coverage: The comments in Voigt et al. (2018) are mainly from 2017-2018, while our dataset collects nearly 10-year of articles from 6/1/2012 to 4/1/2021, which covers a wider range of societal issues and events. In the released version of our data we will add the time stamps to the sentences for customized filtering.

A.2 SENTENCE DISTILLATION FOR CONSTRUCTING POLITICAL STANCE DATASET

As mentioned in Section 3.1 of the paper, the articles collected from Allsides.com are already manually aligned (based on their political ideology) by Allsides editors. We collected a total of 2,298 full-length news articles from opposite ideologies (i.e., liberal and conservative) grouped by events. To create a sentence-level TST dataset for political stance transfer, we split the articles into sentences and took their article-level ideologies (as assigned by the editors) as their sentence-level stance labels. However, we found that not all sentences explicitly exhibit the stance label inherited from their corresponding articles, so we implemented an additional distillation step to select those sentences that are salient with respect to their assigned political stance.

Assuming we have an article with N sentences X = {x1, x2, ..., x N} and the title of the article xtitle, we aim to find the subset of sentences Ddistill X which all explicitly exhibit the assigned political stance. To better describe our approach, we first introduce two concepts: (1) stance saliency, and (2) distillation cost.

Stance Saliency. We compute the stance saliency vi for each sentence xi X\Ddistill, which is a measure of how well xi exhibits a certain political stance. Ddistill is the set of sentences already picked our distilled set. We assume that the stance exhibited in xtitle matches that of the article (as assigned by the editors), Thus, we define stance saliency vi as the similarity between xi and the union of xtitle and Ddistill, which is:

vi = LCS(xi, xtitle Ddistill), (5)

where LCS( ) is longest common sequence between the two input sequences. However, this metric is biased towards longer sentences as they can have more tokens matched with the title. We address this issue through the intriduction of a distilation cost, described below.

Distillation Cost. We assign a distillation cost wi to each sentence xi in an article. The cost is simply the length of the sentence:

wi = xi . (6)

Published as a conference paper at ICLR 2022

Note that for both concepts our computation is based on lower-cased and lemmatized sentence tokens. We use λ to denote the ratio of the sentences to be picked as distilled sentences in an article consisting of N sentences, and define Wmax = λ N i=0 wi as the maximum distillation cost allowed.

Now that for each sentence we have a stance saliency and a distillation cost, we aim to choose a set of sentences as the distillation which have the highest possible sum of saliency for which the sum of distillation costs is not greater than Wmax. Given a function ϕ(xi) = {1, 0} where 1 corresponds to sentence xi being picked for the distilled set and 0 means otherwise, the objective of distillation can be expressed as:

i=1 vi ϕ(xi) , s.t.

i=1 wi ϕ(xi) Wmax . (7)

This is actually a NP-complete combinatorial optimization problem, called the Knapsack problem, which we solve using dynamic-programming (DP). In general, the distillation strategy aggressively picks sentences of high stance saliency while keeping the cost of distillation from exceeding the mask cost limitation Wmax. The procedure is demonstrated in Algorithm 1 below.

Algorithm 1: Sentence Distillation for Political Stance Dataset Input: article title xtitle, article sentences X, distillation ratio λ.

Initialization Init current distillation set Ddistill as ; Init current distillation cost Wdistill with xtitle ;

Multiple Rounds of Distillation while Wdistill Wmax do

Compute vi for each xi X\Ddistill with xtitle and Ddistill (by Equation 1); Compute wi for each xi (by Equation 2); Find best distillation strategy {ϕ(xi)} via DP with vi, wi and Wdistill (see Alg. 2);

Update Distillation Results Every Step Update current distillation set Ddistill = Ddistill {xi} whose ϕ(xi) = 1; Update current distillation cost Wdistill in terms of updated Ddistill; end return distillation sentences set Ddistill;

Note that we initialize Wdistill with the length of title. At every iteration of the loop we solve the DP problem with the limit set by current Wdistill, which is a variable that estimates how many sentences are already in the distillation set (no more than Wmax). The DP-based algorithm used in Algorithm 1 for finding the best distillation strategy is shown is Algorithm 2 below.

A.3 STATISTICS OUR ALL THREE DATASETS, AND EXPERIMENTAL SETUP

We present more statistical details of all three datasets in Table 7. All of our experiments were run on a single RTX-2080 GPU, with batch size 4 and 2/3/2 epochs for La Mer in the above three TST tasks. Following previous work, we choose style control accuracy (ACC), BLEU score as our main metrics. ACCs are judged by a pre-trained Ro BERTa (Liu et al., 2019) classifier with F1 = {0.97, 0.86, 0.93} using the training data of the three TST tasks. About the J-score we calculate in few-shot experiments, we use the models released by the author8.

A.4 ADDITIONAL DETAILS ABOUT TOKEN-LEVEL SCENE PRESERVATION

In this section, we provide additional details regarding how we compute the token-level scene preservation d PSV. Figure 4 shows an example sentiment transfer task. For a given positive (source style) sentence ssrc and its corresponding negative (target style) parallels stgt, we first extract their

8The evaluation scripts, and pre-trained models can be found here: http://style.cs.umass.edu/

Published as a conference paper at ICLR 2022

Sentiment Formality Political Stance

Training Data (S-Emb. + SAS) positive negative formal informal liberal conservative

# Sent Pairs 553,430 93,157 36,307 # Scene Entities 742,136 777,839 145,731 133,779 194,032 272,071 Avg. Sent Len 7.32 7.58 11.52 11.05 19.39 24.75 Avg. # Scene Ent. / Sent 1.34 1.41 1.56 1.44 5.34 7.49 Avg. # Scene Ent.src tgt (vs. RD) 0.95 (0.73) 0.49 (0.02) 0.96 (0.10) Avg. SAS (vs. RD) 0.72 (0.55) 0.29 (0.01) 0.25 (0.01)

Human Reference positive negative formal informal liberal conservative

# Sent Pairs 500 1020 525 # Scene Entities 964 905 1,467 1,388 2,915 3,012 Avg. Sent Len 9.90 9.80 10.9 9.35 10.30 11.19 Avg. # Scene Ent. / Sent 1.93 1.81 1.44 1.36 5.58 5.76 Avg. # Scene Ent.src tgt 1.53 0.55 1.78 Avg. SAS 0.73 0.29 0.30

Table 7: Data statistics for the three TST tasks in this work: Sentiment, Formality, and Political Stance Transfer. We compare the S-Emb. + SAS aligned data (the upper part), with the human-annotated reference file (the bottom part). Notice that after alignment, the SAS of training data becomes much closer to that of human reference, which demonstrates we successfully align the non-parallel data into parallel form (comparable to the human reference). Among the three TST datasets, we find the political stance TST has the longest sentence length (see Avg. Sent Len), and more scene entities involved on average (see Avg. # Scene Ent. / Sent). Human written references also reveal similar statistics (see the lower part of the table), which will potentially increase the difficulty for entity-matched transfer. We thus consider the political stance TST as a more challenging task than the other two.

Algorithm 2: DP-based Best Distillation Strategy Search Input: stance saliency vi, distill cost wi, and current cost limit Wdistill. Record the number of stance saliency M = vi ; Init DP-table T[M + 1][Wdistill + 1] with all 0; for i = 1, 2, . . . , M do

for j = 1, 2, . . . , Wdistill do

if j wi 1 < 0 then

T[i][j] = T[i 1][j]; Record distillation picking choice ϕ(xi); else

T[i][j] = max(T[i 1][j], T[i 1][j wi 1] + vi 1); Record distillation picking choice ϕ(xi); end end end {ϕ(xi)M i=1} backtracking via records; return best distillation strategy {ϕ(xi)M i=1};

scene entities using the method by Wu et al. (2019) (Figure 4 (a)), and based on these entities we compute two distances d Order and d Exist as sub-components of d PSV.

For d Order, as shown in Figure 4 (b), we compare each of the stgt entities with the ssrc entities, and measure how much stgt needs to be edited to make the two sets of entities from opposite styles identical to each other. As mentioned in the paper, we use the Levenshtein distance as the edit distance; whose set of edits operation is: 1) insertion of a single token, 2) deletion of a single token, and 3) substitution of a single token. Each of these operations has unit cost, so the Levenshtein

Published as a conference paper at ICLR 2022

[they, menu, beef]

[they, menu, nothing, beef]

[menu, beef]

[nothing, menu]

[they, menu, beef]

[they, menu, beef]

[0, 1, 1, 0]

[1, 1, 0, 1]

they had an extensive menu and the beef

was delicious .

overpriced menu, and mediocre beef .

nothing great on their menu .

Figure 4: (a) extraction of scene entities (underlined) from source sentence and parallel target sentences. Sentences with a + correspond to stgt+, and those with a - correspond to stgt- (i.e., expert and amateur demonstrations for imitation learning); (b) entity appearing order (d Order, computed by edit distance); and (c) entity existence (d Exist, computed by hamming distance).

distance is equal to the minimum number of operations required to transform target style entities to source style entities.

As shown in Figure 4 (c), we calculate another distance component, d Exist, via hamming distance to emphasize the the difference between the sets of entities in source and target sentences. We first get the union of the entities from a pair of source and target sentences; based on the union set, we can obtain the corresponding binary vector representations of both style entities: 1 means the entity exists in a certain style sentence, and 0 otherwise. Note the binary vectors for each of the sentences will be the same length (since it corresponds to the union of entities). Hamming distance measures the difference between source and target sentences on the shared entities set (this was shown to be an upper bound on the Levenshtein distance). Though there are overlaps between this metric and d Order, we use this metric to emphasize the difference in scene entities.

A.5 MORE DETAILS ABOUT POLICY GRADIENT

We detail our policy gradient algorithm in Algorithm 3. Note that Jsafe IL is a safe-reward threshold. If the greedy policy can already reach this threshold the policy will not be updated. Empirically we set Jsafe IL to {0.8, 0.6, 0.4} for the three TST tasks (sentiment, formality, and political stance).

Algorithm 3: Imitating Learning of La Mer

Input: current policy πθ, expert policy trajectory stgt+, and amateur policy trajectory stgt-

Output: a learned policy πθ approaching to expert policy π+

Get baseline via Greedy Decoding ˆsgreedy arg max ˆπ LM(at y1 t 1)t [1,T ]; Compute J greedy IL with ˆsgreedy, stgt+, and stgtby Eq. 3;

if J greedy IL J safe IL then No Optimization on Already Good Policies Skip current step Imitation Learning;

Exploration by Sampling Decoding while i max exploration steps do

ˆs (i) sample sampling ˆπ LM(at y1 t 1)t [1,T ] ;

Compute J sample IL with ˆs (i) sample, stgt+, and stgtby Eq. 3;

Optimize Policy by REINFORCE Compute policy gradient as θJ(θ) = Eπθ [ θ log πθ (J sample IL J greedy IL )]; Update policy with policy gradient by taking K steps SGD (via Adam); return πθ π+

Published as a conference paper at ICLR 2022

We choose the REINFORCE algorithm (Williams, 1992) to optimize the current policy πθ. One known issue of the vanilla REINFORCE method is the instability of its optimization because of the high variance of the gradients. A common remedy is to add a baseline function (i.e., a parameterized value-network whose purpose is to estimate the unbiased future behavior of current policy (Ranzato et al., 2016)). Rennie et al. (2017) introduce a better approach that estimates the baseline at test time. They use the greedy decoding results as the baseline in their self-critic sequence training algorithm. We adopt their method to provide the baseline, and add an exploration step using sampling decoding. As shown in Algorithm 3, by iteratively switching between these two modes, we calculate corresponding contrastive losses defined by Equation 3, and take the difference of two modes losses as the advantage, feeding into the REINFORCE algorithm to optimize our policy πθ. Finally, the output policy πθ can generate style transferred sequences resemble the expert demonstrations stgt+.

A.6 ADDITIONAL DETAILS ABOUT HUMAN EVALUATION

We provide more details about our human evaluation in this section. The results on sentiment TST are shown in Table 8 and those of formality TST are shown in Table 9.

Details About Participants and Compensation. Participants (N=212) were all from the United States and above 18 years old. Participants were required to have a HIT approval rate greater than 95%. The average age of the participants was 34.95 years-old (SD=9.90, Median=32). 141 of the participants (66.5%) were male, and 71 female (33.5%). Participants received 14.41 years of education on average (SD=4.46, Median=16). When asked to self-report their party affiliation, 52.4% participants self-reported as Democratic, 35.3% participants were self-reported Republicans, and 12.3% participants identified as independent. Each participant was compensated 1 USD per task. The average completion time per questionnaire was about 7.6 minutes as reported by the platform, which results in a $7.9/hr wage, above the regional and federal minimum wage.

Table 8: Human judgement results (on a 7-point scale) of seven existing methods and La Mer on sentiment transfer task. We bold the highest average rating. SD: Standard Deviation.

Sentiment Style Content Readability

Methods Mean SD Mean SD Mean SD

CAE (Shen et al., 2017) 5.09 1.49 4.79 1.56 4.84 1.50 Del Retr Gen (Li et al., 2018a) 4.94 1.62 4.73 1.63 5.05 1.39 Dual RL (Luo et al., 2019) 5.43 1.40 4.89 1.73 5.10 1.40 Style Transformer (Dai et al., 2019) 5.09 1.55 4.71 1.56 5.04 1.50 Deep Latent Seq (He et al., 2020) 4.89 1.79 4.63 1.76 5.05 1.49 TSF-Del Ret Gen (Sudhakar et al., 2019) 4.66 1.35 4.73 1.24 5.30 1.13 STARP (Krishna et al., 2020) 5.50 1.20 4.65 1.40 5.33 1.20 Ours: La Mer 5.45 1.21 4.91 1.71 5.43 1.18

Table 9: Human judgement results (on a 7-point scale) of seven existing methods and La Mer on formality transfer task. We bold the highest average rating. SD: Standard Deviation.

Formality Style Content Readability

Methods Mean SD Mean SD Mean SD

CAE (Shen et al., 2017) 5.17 1.59 5.09 1.67 4.89 1.64 Del Retr Gen (Li et al., 2018a) 5.19 1.57 4.91 1.65 4.86 1.62 Dual RL (Luo et al., 2019) 5.49 1.27 5.13 1.29 5.36 1.31 Style Transformer (Dai et al., 2019) 5.25 1.38 5.22 1.37 5.26 1.37 Deep Latent Seq (He et al., 2020) 5.22 1.52 5.30 1.46 5.17 1.47 TSF-Del Ret Gen (Sudhakar et al., 2019) 4.86 1.25 5.23 1.24 5.20 1.49 STARP (Krishna et al., 2020) 5.33 1.21 4.65 1.37 5.39 1.20 Ours: La Mer 5.50 1.23 5.45 1.29 5.39 1.21

Details About Human Evaluation Procedure. We evaluate the style transferred sentences generated by baselines and our La Mer. The sentences are picked from the test set generations for the three

Published as a conference paper at ICLR 2022

tasks. Each participant was asked to read two sets of sentences (one set containing a source style sentence, corresponding human-written target style sentence as ground-truth, and eight versions of target style sentences generated by seven existing methods and La Mer for comparison). After reading each set, they were asked to rate how well the generated sentences did in style control, readability, and content preservation on a 7-point Likert scales. For example, for content preservation, they were asked How much do you agree with the following statement (1Strongly disagree to 7Strongly agree): The machine-generated target style text preserves the context of the source style text well (i.e. they talk about the same thing). . Higher ratings correspond to better performance.

A.7 PERPLEXITY MEASURE OF LAMER AND OTHER BASELINES

In Table 10 we present the perplexity measure of La Mer and other baselines as a proxy for fluency. As can be seen, certain methods, such as CAE and Style Transformer, have very high perplexity in more challenging TST tasks (e.g., Political Stance), potentially because the task requires longer generation and more entities involved in the sentences (see Table 7 for further discussions). However, it should also be noted that prior work have shown perplexity to be a poor measure of fluency (Krishna et al., 2020; Mir et al., 2019; Lee et al., 2021).

Table 10: Perplexity measure of La Mer and other baselines. We use GPT2-large as the judgement LM, and bold the lowest perplexity (the best), and underline the second best.

Sentiment Formality Political Stance

Existing Methods PPL PPL PPL

Input Copy (ref.) 54.0 16.2 28.0 CAE 53.4 135.4 1869.0 Del Retr Gen 18.5 30.3 86.0 Dual RL 17.0 31.9 141.2 Style Transformer 78.0 129.1 350.0 Deep Latent Seq 27.8 17.2 113.7

LM-based TSF-Del Retr Gen 52.1 24.5 102.3 IMa T 16.7 25.1 67.2 START 12.5 15.9 33.7

Ours: La Mer w/. RD 13.9 13.6 32.3 w/. S-Emb. 11.4 13.5 47.6 w/. S-Emb. + SAS 13.3 11.7 33.5