{ "paper_id": "2020", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:28:10.652586Z" }, "title": "Stable Style Transformer: Delete and Generate Approach with Encoder-Decoder for Text Style Transfer", "authors": [ { "first": "Joosung", "middle": [], "last": "Lee", "suffix": "", "affiliation": { "laboratory": "", "institution": "Kakao Enterprise Corp", "location": { "country": "South Korea" } }, "email": "rung.joo@kakaoenterprise.com" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Text style transfer is the task that generates a sentence by preserving the content of the input sentence and transferring the style. Most existing studies are progressing on non-parallel datasets because parallel datasets are limited and hard to construct. In this work, we introduce a method that follows two stages in nonparallel datasets. The first stage is to delete attribute markers of a sentence directly through a classifier. The second stage is to generate a transferred sentence by combining the content tokens and the target style. We experiment on two benchmark datasets and evaluate context, style, fluency, and semantic. It is difficult to select the best system using only these automatic metrics, but it is possible to select stable systems. We consider only robust systems in all automatic evaluation metrics to be the minimum conditions that can be used in real applications. Many previous systems are difficult to use in certain situations because performance is significantly lower in several evaluation metrics. However, our system is stable in all automatic evaluation metrics and has results comparable to other models. Also, we compare the performance results of our system and the unstable system through human evaluation. Our code and data are available at the link 1 .", "pdf_parse": { "paper_id": "2020", "_pdf_hash": "", "abstract": [ { "text": "Text style transfer is the task that generates a sentence by preserving the content of the input sentence and transferring the style. Most existing studies are progressing on non-parallel datasets because parallel datasets are limited and hard to construct. In this work, we introduce a method that follows two stages in nonparallel datasets. The first stage is to delete attribute markers of a sentence directly through a classifier. The second stage is to generate a transferred sentence by combining the content tokens and the target style. We experiment on two benchmark datasets and evaluate context, style, fluency, and semantic. It is difficult to select the best system using only these automatic metrics, but it is possible to select stable systems. We consider only robust systems in all automatic evaluation metrics to be the minimum conditions that can be used in real applications. Many previous systems are difficult to use in certain situations because performance is significantly lower in several evaluation metrics. However, our system is stable in all automatic evaluation metrics and has results comparable to other models. Also, we compare the performance results of our system and the unstable system through human evaluation. Our code and data are available at the link 1 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Text style transfer is a task that generates a sentence while preserving the content in a given sentence but changing the source style. The style of the sentence refers to a predefined class (e.g. sentiment, formality, tense) and the content refers to the rest of the sentence except for the style. Lack of parallel data makes text style transfer task difficult. This problem cannot be solved by supervised learning because there are no right sentences.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "1 https://github.com/rungjoo/Stable-Style-Transformer One previous method (Hu et al., 2017; Shen et al., 2017; Fu et al., 2018; Prabhumoye et al., 2018a; Logeswaran et al., 2018 ) of text style transfer is to learn latent representations to separate style and content from sentences. First, these approaches try adversarial training to learn a disentangled latent representation of the content and style. Secondly, a transferred sentence is generated from the decoder by combining the disentangled latent representation and the target style. However, the experimental results of (Lample et al., 2019) report that disentangled latent representation through adversarial training is hard to get and not necessary. Also, adversarial training is not effective to encode a sentence of various lengths into a vector representation of fixed length. Other methods of text style transfer do not depend on disentanglement. Dai et al. (2019a) ; Lample et al. (2019) ; Luo et al. (2019) do not attempt to find the disentangled latent representation in the sentence. Therefore, sentences with different styles are mapped to the same space. Xu et al. (2018a) ; ; Sudhakar et al. (2019) ; neutralize sentences by deleting style-dependent attribute markers. Remaining tokens resulting from the deletion of attribute markers are style independent, and then the content tokens and a style attribute are combined to generate the transferred sentence.", "cite_spans": [ { "start": 74, "end": 91, "text": "(Hu et al., 2017;", "ref_id": "BIBREF4" }, { "start": 92, "end": 110, "text": "Shen et al., 2017;", "ref_id": "BIBREF15" }, { "start": 111, "end": 127, "text": "Fu et al., 2018;", "ref_id": "BIBREF3" }, { "start": 128, "end": 153, "text": "Prabhumoye et al., 2018a;", "ref_id": "BIBREF11" }, { "start": 154, "end": 177, "text": "Logeswaran et al., 2018", "ref_id": "BIBREF9" }, { "start": 579, "end": 600, "text": "(Lample et al., 2019)", "ref_id": "BIBREF6" }, { "start": 912, "end": 930, "text": "Dai et al. (2019a)", "ref_id": "BIBREF0" }, { "start": 933, "end": 953, "text": "Lample et al. (2019)", "ref_id": "BIBREF6" }, { "start": 956, "end": 973, "text": "Luo et al. (2019)", "ref_id": "BIBREF10" }, { "start": 1126, "end": 1143, "text": "Xu et al. (2018a)", "ref_id": "BIBREF19" }, { "start": 1148, "end": 1170, "text": "Sudhakar et al. (2019)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We propose an approach with two stages using Delete and Generate without adversarial training for disentanglement. (1) Attribute markers of a sentence are extracted by using a pre-trained classifier as a Delete model. Our method is model-agnostic and is not affected by the design of the classifier. Attribute markers found in a sentence are deleted.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "(2) A transferred sentence is generated by combining the target attribute and the content tokens after stage-1. The Generate model consists of an encoder and decoder with the Transformer structure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In the method of deleting attribute markers, deletes attribute markers via a statistical manner using a frequency ratio and Sudhakar et al. (2019) ; Xu et al. (2018b) delete attribute markers using attention weights of a classifier. deletes attribute markers by fusion of the frequency ratio and the attention weights. We introduce an intuitive delete method that uses a change in classifier probability. If a change in classifier probability is significant when limiting certain tokens in a sentence, the token is considered an attribute marker. Our method does not need to build attribute dictionaries or define attention weights like previous methods and easily control the tradeoff between content and style. We test our methods on two text style transfer datasets: sentiment of Yelp reviews and Amazon reviews. Evaluation metrics are conducted in terms of content, fluency, style accuracy, and semantic. The content and style accuracy are measured similarly to previous studies. Fluency is measured in two ways: general-fluency using pre-trained GPT-2 (Radford et al., 2019) and data-fluency using finetuned GPT-1 (Radford, 2018) . Semantic is newly evaluated using BERTscore (Zhang* et al., 2020) in this paper. The goal of BERTscore is to evaluate semantic equivalence between two sentences. In this paper, we use a pre-trained model GPT and BERT (Devlin et al., 2019) that perform well in natural language processing/generation to evaluate transferred sentences with various automatic evaluations. Since automatic evaluations are not perfect evaluations of generated sentences, it is hard to know which system is the best, but we can determine which system has a problem. Comparative models are unstable in some evaluation metrics. But our proposed model has stable results for all automatic evaluations and is called SST (Stable Style Transformer). In addition, we first observe a point that can enhance the style controlling ability by generating sentences through latent space walking in the vector space of the style attribute token.", "cite_spans": [ { "start": 124, "end": 146, "text": "Sudhakar et al. (2019)", "ref_id": "BIBREF16" }, { "start": 149, "end": 166, "text": "Xu et al. (2018b)", "ref_id": "BIBREF20" }, { "start": 1057, "end": 1079, "text": "(Radford et al., 2019)", "ref_id": "BIBREF14" }, { "start": 1119, "end": 1134, "text": "(Radford, 2018)", "ref_id": "BIBREF13" }, { "start": 1181, "end": 1202, "text": "(Zhang* et al., 2020)", "ref_id": null }, { "start": 1354, "end": 1375, "text": "(Devlin et al., 2019)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "One line of text style transfer research (Shen et al., 2017; Fu et al., 2018; Hu et al., 2017; Prabhumoye et al., 2018b; Logeswaran et al., 2018) is to separate content and style from sentences through disentangled learning. Hu et al. (2017) uses the VAE model to derive the disentanglement of the content between the generated sentence and the original sen-tence through KL loss. Shen et al. (2017) introduce the aligned auto-encoder and the cross aligned autoencoder using learning discriminators. Fu et al. (2018) propose a multi-decoder and StyleEmbedding model. The multi-decoder model has decoders for each style, and the style embedding model uses only one decoder by inserting style embedding into the decoder. The methods of Prabhumoye et al. 2018b; Logeswaran et al. (2018) used backtranslation to learn latent representations.", "cite_spans": [ { "start": 41, "end": 60, "text": "(Shen et al., 2017;", "ref_id": "BIBREF15" }, { "start": 61, "end": 77, "text": "Fu et al., 2018;", "ref_id": "BIBREF3" }, { "start": 78, "end": 94, "text": "Hu et al., 2017;", "ref_id": "BIBREF4" }, { "start": 95, "end": 120, "text": "Prabhumoye et al., 2018b;", "ref_id": "BIBREF12" }, { "start": 121, "end": 145, "text": "Logeswaran et al., 2018)", "ref_id": "BIBREF9" }, { "start": 225, "end": 241, "text": "Hu et al. (2017)", "ref_id": "BIBREF4" }, { "start": 381, "end": 399, "text": "Shen et al. (2017)", "ref_id": "BIBREF15" }, { "start": 500, "end": 516, "text": "Fu et al. (2018)", "ref_id": "BIBREF3" }, { "start": 759, "end": 783, "text": "Logeswaran et al. (2018)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "The second line of text style transfer research is not to rely on learning for latent representation. The first approach (Xu et al., 2018b; Sudhakar et al., 2019; is to find and delete tokens called attribute markers that are highly related to style. uses the delete method of attribute markers as a statistical method based on frequency ratio, and Sudhakar et al. 2019; Xu et al. (2018b) use the attention scores of the Transformer classifier and LSTM classifier, respectively. deletes attribute markers by fusion of the frequency ratio and attention scores. The second approach (Dai et al., 2019b; Lample et al., 2019; Luo et al., 2019) does not attempt to control content and style separately. Therefore, sentences with different styles are encoded to gather in the same latent representation space. Dai et al. (2019b) ; Lample et al. (2019) are based on the learning method using cycle reconstruction loss. Lample et al. (2019) reported that disentanglement is not easy and that latent representations learned through adversarial training are unnecessary because learned latent representations depend on style. Unlike the previous models, (Luo et al., 2019) learns dual models in two directions: style1 (e.g. negative) to style2 (e.g. positive) and style2 (e.g. positive) to style1 (e.g. negative) by reinforcement learning.", "cite_spans": [ { "start": 121, "end": 139, "text": "(Xu et al., 2018b;", "ref_id": "BIBREF20" }, { "start": 140, "end": 162, "text": "Sudhakar et al., 2019;", "ref_id": "BIBREF16" }, { "start": 371, "end": 388, "text": "Xu et al. (2018b)", "ref_id": "BIBREF20" }, { "start": 580, "end": 599, "text": "(Dai et al., 2019b;", "ref_id": "BIBREF1" }, { "start": 600, "end": 620, "text": "Lample et al., 2019;", "ref_id": "BIBREF6" }, { "start": 621, "end": 638, "text": "Luo et al., 2019)", "ref_id": "BIBREF10" }, { "start": 803, "end": 821, "text": "Dai et al. (2019b)", "ref_id": "BIBREF1" }, { "start": 824, "end": 844, "text": "Lample et al. (2019)", "ref_id": "BIBREF6" }, { "start": 911, "end": 931, "text": "Lample et al. (2019)", "ref_id": "BIBREF6" }, { "start": 1143, "end": 1161, "text": "(Luo et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "In language model research, the RNN-based language model is weak in long dependency. Therefore, the recent study of text style transfer (Dai et al., 2019b; Sudhakar et al., 2019; has been conducted with Transformer (Vaswani et al., 2017) which is known to have good performance in language modeling. Dai et al. (2019b) uses a method of using the encoder and the decoder of the Transformer, and Sudhakar et al. (2019) uses a method of fine-tuning the decoder to the style transfer datasets with the pre-trained GPT-1 as an initial state. solved the problem of text style transfer in a similar way to Text Infill- Figure 1 : The proposed model framework consists of Delete and Generate process. Delete process is a method using a pre-trained classifier, and the Generate process consists of an encoder and a decoder. In the training time, our model receives feedback from the classifier's probability of the generated sentence.", "cite_spans": [ { "start": 136, "end": 155, "text": "(Dai et al., 2019b;", "ref_id": "BIBREF1" }, { "start": 156, "end": 178, "text": "Sudhakar et al., 2019;", "ref_id": "BIBREF16" }, { "start": 215, "end": 237, "text": "(Vaswani et al., 2017)", "ref_id": "BIBREF17" }, { "start": 300, "end": 318, "text": "Dai et al. (2019b)", "ref_id": "BIBREF1" } ], "ref_spans": [ { "start": 612, "end": 620, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "ing or Cloze by presenting Attribute Conditional Masked Language Model (AC-MLM) using pretrained BERT.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "In this paper, we chose the first approach (Delete and Generate) that does not rely on latent representations in the second research line, referring to the results of Lample et al. (2019) . Our system has a Transformer encoder and decoder because the style transfer task is given input text. If the system uses only a decoder such as Sudhakar et al. (2019) , there is a disadvantage that it cannot include bidirectional encoding of the content token. Or, if only bidirectional encoders are used, such as AC-MLM, the position and length of the masking tokens to be filled in a sentence is not flexible.", "cite_spans": [ { "start": 167, "end": 187, "text": "Lample et al. (2019)", "ref_id": "BIBREF6" }, { "start": 334, "end": 356, "text": "Sudhakar et al. (2019)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "2" }, { "text": "In this section, we introduce our proposed method. The style transfer problem definition is described in Section 3.1. An overview of the model is shown in Section 3.2. The proposed generation process is introduced in Sections 3.3 and 3.4. The learning mechanism is described in Section 3.5.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Approach", "sec_num": "3" }, { "text": "Given a dataset consist of sentence and label:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "3.1" }, { "text": "D = {(x 1 , s 1 ), \u2022 \u2022 \u2022 , (x N , s N )}", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "3.1" }, { "text": "where x i is a sentence and s i is a style attribute (e.g. sentiment) and N is the number of sentences in the dataset. Our goal is to train the model to generate a sentence y i with a different style while preserving the content of the sentence x i . For example, if x i is \"The food is salty and tasteless\" and s i is \"negative\" attribute, then y i is generated to mean \"The food is not salty and delicious\" which has a \"positive\" attribute. However the dataset is non-parallel, so the model cannot access y i aligned with x i .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Statement", "sec_num": "3.1" }, { "text": "Our approach consists of two stages: Delete and Generate framework in Fig. 1 . The first stage is the Delete process with a pre-trained style classifier. The pre-trained style classifier finds and deletes tokens that contain a lot of style attributes. The second stage is encoding the content tokens and combine them with a target style to generate a sentence. Both the encoder and the decoder have the Transformer structure, which is better than RNN and robust to long dependency.", "cite_spans": [], "ref_spans": [ { "start": 70, "end": 76, "text": "Fig. 1", "ref_id": null } ], "eq_spans": [], "section": "Model Overview", "sec_num": "3.2" }, { "text": "The stage-1 is the process of finding and deleting tokens for a given sentence and style attribute. In the previous studies, the strategies of deleting attribute markers are the frequency-ratio method and the classifier's attention score (or fusion of both). However, the frequency ratio method requires a pre-built vocabulary for the training dataset and it is difficult to understand contextual information. The attention score method has a limitation on the structure of the classifier, because it must learn the style classifier using self-attention regardless of accuracy. It is also unclear whether the attention score is directly proportional to the attribute.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "We propose a novel method of removing attribute markers using a pre-trained classifier without a pre-built dictionary and attention scores. Our method is a model-agnostic method and it is more intuitive to find attribute markers than the previous method. Given an input sentence x, the style probability follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "p x = p \u03b8 C (s|x)", "eq_num": "(1)" } ], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "where p is a probability predicted by the classifier and s is style label. If we delete token t i from the sentence x, the style probability changes as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "p x,t i = p \u03b8 C (s|x, t i )", "eq_num": "(2)" } ], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "where", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "x = (t 1 , t 2 , \u2022 \u2022 \u2022 , t n )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "and n is the number of tokens in x. The probability difference between Eq. 1 and Eq. 2 is defined as Important Score(IS) of the token(t i ):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "IS k t i = p x k \u2212 p x k ,t i (3)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "where x k is the remained tokens after k tokens are deleted. The value of IS k t i determines how much the token t i affects the style classifier. The token is deleted in order of the largest IS, and the Delete process ends if only one of the following two conditions: (1) p x k is less than \u03b1, or (2) the ratio of content tokens is less than \u03b2. \u03b1 is a hyperparameter that determines that a sentence no longer has a source style attribute. \u03b2 is a hyperparameter that determines how much of the content is preserved. The two hyperparameters make it easy to control the trade-off of content and style, and the experimental results are explained in Section 4.8.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-1: Delete process", "sec_num": "3.3" }, { "text": "Our model generates a transferred sentence with the encoder and the decoder of the Transformer.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Stage-2: Generate process", "sec_num": "3.4" }, { "text": "All content tokens given as a result of Delete process are input to a bidirectional self-attention the Transformer encoder. Explicitly, the Transformer encoder maps content tokens", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder", "sec_num": "3.4.1" }, { "text": "x c = (t 1 , \u2022 \u2022 \u2022 , t m ) to the continuous representation z = (z 1 , \u2022 \u2022 \u2022 , z m ) as follow: (z 1 , \u2022 \u2022 \u2022 , z m ) = Encoder(t 1 , \u2022 \u2022 \u2022 , t m ; \u03b8 E ) (4)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Encoder", "sec_num": "3.4.1" }, { "text": "In order to generate a sentence with the desired style, two special tokens, style and start, are initially input to the decoder in Fig. 1 . The position of special tokens is always fixed in front, so we did not add positional embedding. We use teacher-forcing at training time and no teacher-forcing at test time to generate sentences. If the generated token is the special token end, the Generate process ends. The decoder auto-regressively predicts the conditional probability of the next step token as follows:", "cite_spans": [], "ref_spans": [ { "start": 131, "end": 137, "text": "Fig. 1", "ref_id": null } ], "eq_spans": [], "section": "Decoder", "sec_num": "3.4.2" }, { "text": "softmax(y j ) = p \u03b8 D (t j |t 1 , \u2022 \u2022 \u2022 , t j\u22121 ,s, z) (5)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder", "sec_num": "3.4.2" }, { "text": "where y j is the logit vector of the decoder,s is a desired style and t j is the predicted token in j step.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoder", "sec_num": "3.4.2" }, { "text": "Since we only have non-parallel datasets, we can't do supervised learning about transferred sentences. Therefore, we train SST to minimize two losses according to style conditions: s (source style) or\u015d (target style).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Training", "sec_num": "3.5" }, { "text": "SST reconstructs the original sentence x conditioned on x c and source style s. Reconstruction loss follows the equation:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reconstruction loss", "sec_num": "3.5.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "L rec = \u2212log p \u03b8 E ,\u03b8 G (x|x c , s)", "eq_num": "(6)" } ], "section": "Reconstruction loss", "sec_num": "3.5.1" }, { "text": "In non-parallel datasets, the reconstruction loss cannot be calculated if the style of the generated sentence is\u015d.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reconstruction loss", "sec_num": "3.5.1" }, { "text": "If the model is only trained with reconstruction loss, the decoder will not see how to transform the style. Therefore, a discrepancy occurs between training time and test time. To learn how to generate sentencex with a target style\u015d, we introduce style loss as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Style loss", "sec_num": "3.5.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "L style = \u2212log p \u03b8 C (x =\u015d|x c ,\u015d)", "eq_num": "(7)" } ], "section": "Style loss", "sec_num": "3.5.2" }, { "text": "Style loss is measured by a pre-trained classifier to determine whether the transferred sentence has a\u015d. Since the generated sentence is a discrete space, we utilize soft-embedding of predicted tokens to optimize through style loss. When the SST is trained, the parameters of the classifier are not finetuned.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Style loss", "sec_num": "3.5.2" }, { "text": "The Transformer encoder and decoder consist of 3 layers, and each layer has 4 heads. The style classifier consists of 5 convolution filters based on Kim (2014) . Text is tokenized using Byte-Pair-Encoding, and (word, style, position) embeddings ", "cite_spans": [ { "start": 149, "end": 159, "text": "Kim (2014)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Model Details", "sec_num": "3.6" }, { "text": "In this paper, we test our model on two datasets, YELP and AMAZON, which are provided in . The Yelp dataset is for business reviews, and the Amazon dataset is product reviews. Both datasets are labeled negative and positive and statistics are shown in Table 1 .", "cite_spans": [], "ref_spans": [ { "start": 252, "end": 259, "text": "Table 1", "ref_id": null } ], "eq_spans": [], "section": "Experiments 4.1 Dataset", "sec_num": "4" }, { "text": "Human references are used to measure human-BLEU and BERTscore. We used 2 Yelp human references and 1 Amazon human reference. Yelp: ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Human References", "sec_num": "4.2" }, { "text": "We compare the previous models with three approaches. The first comparisons are CrossAligned (Shen et al., 2017) , [StyleEmbedding, multi-decoder] (Fu et al., 2018) , and BackTranslation (Prabhumoye et al., 2018b), which attempt to separate content and style through latent representation learning.", "cite_spans": [ { "start": 93, "end": 112, "text": "(Shen et al., 2017)", "ref_id": "BIBREF15" }, { "start": 147, "end": 164, "text": "(Fu et al., 2018)", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Previous Method", "sec_num": "4.3" }, { "text": "The second comparisons are [DeleteOnly, DeleteAn-dRetrieve] , UnpariedRL (Xu et al., 2018b) and [B-GST, G-GST] (Sudhakar et al., 2019) , which delete attribute markers and then generate the sentence. [TemplateBased, RetrieveOnly] return the target sentence through retrieve without generating. The final comparison is DualRL (Luo et al., 2019) , which does not distinguish between content and style.", "cite_spans": [ { "start": 73, "end": 91, "text": "(Xu et al., 2018b)", "ref_id": "BIBREF20" }, { "start": 111, "end": 134, "text": "(Sudhakar et al., 2019)", "ref_id": "BIBREF16" }, { "start": 200, "end": 229, "text": "[TemplateBased, RetrieveOnly]", "ref_id": null }, { "start": 325, "end": 343, "text": "(Luo et al., 2019)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Previous Method", "sec_num": "4.3" }, { "text": "We evaluated the systems in 4 ways and results are shown in Table 2 and 3.", "cite_spans": [], "ref_spans": [ { "start": 60, "end": 67, "text": "Table 2", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Automatic Evaluation", "sec_num": "4.4" }, { "text": "Content preserving intensity is measured by G-BLEU, the geometric mean of self-BLEU and human-BLEU, as in previous works. A high BLEU score indicates that the model is good at content preservation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Content", "sec_num": "4.4.1" }, { "text": "In the Yelp dataset, RetireveOnly and BackTranslation are considered unstable models because G-BELU score is too low compared to other systems. In the Amazon datasets, CrossAligned and Re-trieveOnly are too low compared to other systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Content", "sec_num": "4.4.1" }, { "text": "Most style transfer studies measure style accuracy using a classifier. We also evaluate style accuracy with a classifier (note that this is different from the one used in training).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attribute", "sec_num": "4.4.2" }, { "text": "In the Yelp dataset, StyleEmbedding, multidecoder, and UnpairedRL have quite a low accuracy. In the Amazon datasets, StyleEmbedding, DeleteOnly, and DeleteAndRetrieve are unstable in style transfer.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attribute", "sec_num": "4.4.2" }, { "text": "Fluency is considered the perplexity of the transferred sentence. We use GPT-1 and GPT-2, which is known to perform well as a language model. General-fluency (g-PPL) is measured using pretrained GPT-2 and data-fluency (d-PPL) is measured using GPT-1 (instead of GPT-2 due to GPU memory) finetuned to the dataset. General-Fluency is a general view because the language model is not fitted to the data, and data-fluency is an evaluation metric in terms of the specific data of style transfer tasks. The total-fluency (t-PPL) is the geometric mean of d-PPL and g-PPL, and lower values indicate better fluency.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Fluency", "sec_num": "4.4.3" }, { "text": "In the Yelp dataset, TemplateBased is unstable because t-PPL is much larger than other systems. In the Amazon dataset, it is determined that the fluency of B-GST and G-GST is unstable.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Fluency", "sec_num": "4.4.3" }, { "text": "Semantic is measured using BERTscore. Unlike BLEU and ROUGE, BERTscore is an evaluation metric defined in continuous space. Pretrained model is used to calculate cosine similarity by extracting the contextual token embed-dings from a human reference and a transferred sentence. BERTscore solves the limitations of previous metrics and measures a better correlation between the reference and the candidate. The original BERTscore ranged from 0 to 1, but we rescale it from 0 to 100 to clearly see the difference.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic", "sec_num": "4.4.4" }, { "text": "We set the unstable threshold as a margin point lower than the mean of all systems. The margin point is a gap between an average and a lower bound with 95% confidence considering all systems as the gaussian distribution in the BERTscore evaluation. CrossAligned, multi-decoder, Re-trieveOnly, and BackTranslation have limitations on Yelp datasets. CrossAligned, multi-decoder, and RetrieveOnly have limitations on Amazon datasets.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Semantic", "sec_num": "4.4.4" }, { "text": "SST : For comparison with other systems, we select the \u03b1 and \u03b2 of the appropriate trade-off points for style transfer and content preservation. When experimenting with the Yelp datasets, SST model is evaluated in two cases where (\u03b1, \u03b2) is (0.7, 0.5) and (0.7, 0.75). SST(0.7, 0.5) changes styles better with style accuracy of 79.5%, but SST(0.7, 0.75) has better performance on other metrics. In the Amazon datasets, SST model is evaluated when (\u03b1, \u03b2) is (0.6, 0.5). The effects of \u03b1 and \u03b2 are discussed in detail in Section 4.8. Table 4 shows human evaluation results for content, fluecy, and style. Comparison models, StyleEmbedding and TemplatedBased, each have weaknesses in attribute and fluency. BackTranslation has weaknesses in content and semantic in automatic evaluation. In the yelp test set, we randomly sampled 250 and gave 6 people hired through the Amazon mechanical turk 2 evaluate content, fluency, and style between 1 and 5 points. As a result, BackTranslation and StyleEmbedding show the worst results for content, fluency, and style, respectively. Since humans evaluate fluency from a general point of view, the fluency performance of BackTranslation, which is poor in overall performance, and TemplateBased, which has poor t-PPL performance, are similarly bad. We confirm that our system has adequate performance in human evaluation as well as automatic evaluation.", "cite_spans": [], "ref_spans": [ { "start": 530, "end": 537, "text": "Table 4", "ref_id": null } ], "eq_spans": [], "section": "Semantic", "sec_num": "4.4.4" }, { "text": "Human systems do not obtain the highest performance scores except for human-BLEU and BERTscore, which are calculated using human references. But which of the sentences in human and machines is actually realistic? Probably human. It is difficult to determine the best system with only automatic evaluation, but it is possible to determine which system is stable or unstable. If a system has significantly lower performance during the evaluation, it is considered unstable. The stable systems in the Yelp dataset are SST, DeleteOnly, DeleteAn-dRetireve, DualRL, B-GST, and G-GST. For the Amazon dataset, the stable systems are SST and TemplateBased. For all the metrics in both datasets, the stable systems are SST and DualRL. In automatic evaluation, DualRL outperforms SST, but DualRL does not share the model parameters of positive to negative and negative to positive tasks. Therefore, direct comparison is difficult because DualRL is regarded as two models.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Result Analysis", "sec_num": "4.6" }, { "text": "We trained SST by changing the random seed of the model initialization several times and found that SST can always yield similar and comparable results. SST can be inferred as a stable system for the following reasons: (1) G-BLEU: Delete and Generate approaches show the stable performance of G-BLEU because the methods generate a sentence based on content tokens. There is no guarantee that content tokens will always be maintained, but content tokens help the generator. (2) Attribute: Our delete process is a method of determining whether certain tokens are deleted with Important Score. The direct and model-agnostic deletion is effective for neutralizing sentences. SST also improves a style accuracy by adding style control loss. (3) Fluency: TemplatedBased, B-GST, and G-GST show non-ideal fluency in d-PPL. TemplatedBased is considered unstable because it simply inserts attribute tokens of training data when generating test sentences. Since B-GST and G-GST use pre-trained GPT, they also have the ability to predict the distribution of tokens that are not in training data. The ability to predict generalized tokens is usually helpful, but can sometimes be harmful to d-PPL. SST, the Transformer encoder-decoder structure, learns only the distribution of given data and therefore has a stable d-PPL. (4) Semantic: Transformer language modeling is known to perform better on various tasks than RNN. Even in the style transfer task, the Transformer-based structures seem to reflect the linguistic characteristics.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Result Analysis", "sec_num": "4.6" }, { "text": "We observed that unstable systems performed poorly in human evaluation as well in automatic Table 4 : Human evaluation results. The higher the number, the better. Red means the worst result in the corresponding evaluation term.", "cite_spans": [], "ref_spans": [ { "start": 92, "end": 99, "text": "Table 4", "ref_id": null } ], "eq_spans": [], "section": "Result Analysis", "sec_num": "4.6" }, { "text": "evaluation. However, since performing human evaluation every time is expensive, choosing a stable system with automatic evaluation can be helpful. Table 5 shows the samples of the generation of the models, which shows the lack of comparison models. In Yelp's negative to positive example, there are only SST and DualRL models that change the style while preserving content that includes taste and price of the food. In Yelp's positive to negative example, the professionals word contains a combination of style and content. In this case, the deletion and generation framework has the disadvantage of corrupting content information.", "cite_spans": [], "ref_spans": [ { "start": 147, "end": 154, "text": "Table 5", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Result Analysis", "sec_num": "4.6" }, { "text": "If we use style loss for SST training, Table 6 shows that the style accuracy has 4 point gain. Fluency", "cite_spans": [], "ref_spans": [ { "start": 39, "end": 46, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Ablation Study", "sec_num": "4.7" }, { "text": "Yelp (negative to positive) Yelp (positive to negative) Input (source) the food was so-so and very over priced for what you get . these two women are professionals . SST the service is so-so and very reasonably priced for what you get . these two women are rude . CrossAligned the food was fantastic and very very nice for what you . these two dogs are hard down . StyleEmbedding the food was so-so and very over priced for what you get . these two pot everywhere was . DeleteOnly the food was so-so and very over priced for what you get . i would n't like these two women are professionals .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Study", "sec_num": "4.7" }, { "text": "DeleteAndRetrieve the service is fantastic and the food was so-so and the food is very priced for what you get . these two scam women are professionals .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Study", "sec_num": "4.7" }, { "text": "Back-translation the food is delicious and the staff are very good for me . this place is just not good . UnpariredRL the food was so-so and very over priced for what great qualities . these two women are great . DualRL the food was surprising and very reasonably priced for what you get . these two women are unprofessional . B-GST the food was amazing -so fresh and very good for what you get . these two women are terrible liars . G-GST the food was priced right -so nice and very good for what you get . these two women are condescending . Human DRG the food was great and perfectly priced these two women are not professionals . Human DualRL the food was good and the price is low . these two women are not professionals at all Amazon (negative to positive)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Study", "sec_num": "4.7" }, { "text": "Amazon (positive to negative) Input (source) i have to lower the rating another notch . it seems to be of very good quality in its build . SST love the rating another one , it seems to be of very poor quality in its build . CrossAligned i would recommend this for the price . it s not be for a good game for my phone . StyleEmbedding i have to get by a one market . it seems to be the num extend is good nice high cases .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Ablation Study", "sec_num": "4.7" }, { "text": "DeleteOnly i have to lower the rating and it fits into another notch . i have previously charged num num different bt headsets that last num num hours longer . DeleteAndRetrieve i have to lower the rating another notch and i love it . initially it was very good quality in its build . B-GST i have lower levels for the other notch . it seems to be of very good quality in taste . G-GST i have lower the steel another notch . it seems to be of very good value in return . Human DRG i have to raise the rating another notch . it seems to be of very poor quality in its build Table 6 : Ablation result of style loss in the Yelp dataset. (Con: content, Attr: attribute, Flu: Fluency, Sem: Semantic) and semantic are slightly better. It is observed that style loss improves the data-fluency, resulting in better total fluency. However, style loss decreases G-BLUE slightly by allowing the transferred sentence to change the attribute better.", "cite_spans": [], "ref_spans": [ { "start": 573, "end": 580, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Ablation Study", "sec_num": "4.7" }, { "text": "With \u03b1 and \u03b2 we can simply adjust the trade-off of content and style. The results of Yelp are shown in Source(negative) when i was finally there , i was very disappointed . after deletion when i finally , i very . style: negative when i finally left , i was very disappointed .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Trade-off between Content and Style", "sec_num": "4.8" }, { "text": "\u2193 when i finally left , i was very disappointed. when i finally walked in , i was very disappointed . when i finally got , i was very happy . style: positive when i finally got , i was very happy . Table 7 : One sample of the Yelp dataset. SST generates a sentence from style vector space to negative to positive Fig. 2 . Smaller \u03b1 and \u03b2 allow the model to focus on style changes, while larger \u03b1 and \u03b2 allow the model to focus on content preserving. The tradeoff of content and style changes linearly with \u03b1 and is sensitive to \u03b2. The appropriate \u03b1 and \u03b2 depend on datasets.", "cite_spans": [], "ref_spans": [ { "start": 198, "end": 205, "text": "Table 7", "ref_id": null }, { "start": 313, "end": 319, "text": "Fig. 2", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Trade-off between Content and Style", "sec_num": "4.8" }, { "text": "In this section we observe the transferred sentences according to the weight of positive and negative in the continuous style vector space. Ideally, a neutral sentence should be generated when the style attribute has the same weight for negative and positive. An example is shown in Table 7 . A lot of data, like this example, don't show a neutral sentence even if the style has the same weight for the negative and positive. If we train our model to reflect this problem, we can expect better style control.", "cite_spans": [], "ref_spans": [ { "start": 283, "end": 290, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Latent Space Walking", "sec_num": "4.9" }, { "text": "We propose Stable Style Transformer (SST) that rewrites the sentences with Delete and Generate. SST is a system that can be used in the real world with overall stable results compared to other comparable systems. We show that filtering out unstable systems through human evaluation is expensive, so selecting a stable system through automatic evaluation can be helpful. The proposed direct and model-agnostic deletion method allows the classifier to intuitively delete attribute markers and easily handle the trade-off of content and style. In future work, we will study solutions for the case where attribute markers also contain content in the deletion and generation framework.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "5" }, { "text": "https://www.mturk.com/", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Style transformer: Unpaired text style transfer without disentangled latent representation", "authors": [ { "first": "Ning", "middle": [], "last": "Dai", "suffix": "" }, { "first": "Jianze", "middle": [], "last": "Liang", "suffix": "" }, { "first": "Xipeng", "middle": [], "last": "Qiu", "suffix": "" }, { "first": "Xuanjing", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "5997--6007", "other_ids": { "DOI": [ "10.18653/v1/P19-1601" ] }, "num": null, "urls": [], "raw_text": "Ning Dai, Jianze Liang, Xipeng Qiu, and Xuanjing Huang. 2019a. Style transformer: Unpaired text style transfer without disentangled latent represen- tation. In Proceedings of the 57th Annual Meet- ing of the Association for Computational Linguis- tics, pages 5997-6007, Florence, Italy. Association for Computational Linguistics.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Style transformer: Unpaired text style transfer without disentangled latent representation", "authors": [ { "first": "Ning", "middle": [], "last": "Dai", "suffix": "" }, { "first": "Jianze", "middle": [], "last": "Liang", "suffix": "" }, { "first": "Xipeng", "middle": [], "last": "Qiu", "suffix": "" }, { "first": "Xuanjing", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "5997--6007", "other_ids": { "DOI": [ "10.18653/v1/P19-1601" ] }, "num": null, "urls": [], "raw_text": "Ning Dai, Jianze Liang, Xipeng Qiu, and Xuanjing Huang. 2019b. Style transformer: Unpaired text style transfer without disentangled latent represen- tation. In Proceedings of the 57th Annual Meet- ing of the Association for Computational Linguis- tics, pages 5997-6007, Florence, Italy. Association for Computational Linguistics.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "BERT: Pre-training of deep bidirectional transformers for language understanding", "authors": [ { "first": "Jacob", "middle": [], "last": "Devlin", "suffix": "" }, { "first": "Ming-Wei", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Kenton", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Kristina", "middle": [], "last": "Toutanova", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "4171--4186", "other_ids": { "DOI": [ "10.18653/v1/N19-1423" ] }, "num": null, "urls": [], "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. 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), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Style transfer in text: Exploration and evaluation", "authors": [ { "first": "Zhenxin", "middle": [], "last": "Fu", "suffix": "" }, { "first": "Xiaoye", "middle": [], "last": "Tan", "suffix": "" }, { "first": "Nanyun", "middle": [], "last": "Peng", "suffix": "" }, { "first": "Dongyan", "middle": [], "last": "Zhao", "suffix": "" }, { "first": "Rui", "middle": [], "last": "Yan", "suffix": "" } ], "year": 2018, "venue": "Thirty-Second AAAI Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Explo- ration and evaluation. In Thirty-Second AAAI Con- ference on Artificial Intelligence.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Toward controlled generation of text", "authors": [ { "first": "Zhiting", "middle": [], "last": "Hu", "suffix": "" }, { "first": "Zichao", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Xiaodan", "middle": [], "last": "Liang", "suffix": "" }, { "first": "Ruslan", "middle": [], "last": "Salakhutdinov", "suffix": "" }, { "first": "Eric", "middle": [ "P" ], "last": "Xing", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 34th International Conference on Machine Learning", "volume": "70", "issue": "", "pages": "1587--1596", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1587-1596. JMLR. org.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Convolutional neural networks for sentence classification", "authors": [ { "first": "Yoon", "middle": [], "last": "Kim", "suffix": "" } ], "year": 2014, "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)", "volume": "", "issue": "", "pages": "1746--1751", "other_ids": { "DOI": [ "10.3115/v1/D14-1181" ] }, "num": null, "urls": [], "raw_text": "Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746-1751, Doha, Qatar. Association for Computational Lin- guistics.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Multiple-attribute text rewriting", "authors": [ { "first": "Guillaume", "middle": [], "last": "Lample", "suffix": "" }, { "first": "Sandeep", "middle": [], "last": "Subramanian", "suffix": "" }, { "first": "Eric", "middle": [], "last": "Smith", "suffix": "" }, { "first": "Ludovic", "middle": [], "last": "Denoyer", "suffix": "" }, { "first": "Marc'aurelio", "middle": [], "last": "Ranzato", "suffix": "" }, { "first": "Y-Lan", "middle": [], "last": "Boureau", "suffix": "" } ], "year": 2019, "venue": "International Conference on Learning Representations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, and Y- Lan Boureau. 2019. Multiple-attribute text rewrit- ing. In International Conference on Learning Rep- resentations.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Delete, retrieve, generate: a simple approach to sentiment and style transfer", "authors": [ { "first": "Juncen", "middle": [], "last": "Li", "suffix": "" }, { "first": "Robin", "middle": [], "last": "Jia", "suffix": "" }, { "first": "He", "middle": [], "last": "He", "suffix": "" }, { "first": "Percy", "middle": [], "last": "Liang", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "1865--1874", "other_ids": { "DOI": [ "10.18653/v1/N18-1169" ] }, "num": null, "urls": [], "raw_text": "Juncen Li, Robin Jia, He He, and Percy Liang. 2018a. Delete, retrieve, generate: a simple approach to sen- timent 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), pages 1865-1874, New Orleans, Louisiana. Associ- ation for Computational Linguistics.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Delete, retrieve, generate: a simple approach to sentiment and style transfer", "authors": [ { "first": "Juncen", "middle": [], "last": "Li", "suffix": "" }, { "first": "Robin", "middle": [], "last": "Jia", "suffix": "" }, { "first": "He", "middle": [], "last": "He", "suffix": "" }, { "first": "Percy", "middle": [], "last": "Liang", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "1865--1874", "other_ids": { "DOI": [ "10.18653/v1/N18-1169" ] }, "num": null, "urls": [], "raw_text": "Juncen Li, Robin Jia, He He, and Percy Liang. 2018b. Delete, retrieve, generate: a simple approach to sen- timent 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), pages 1865-1874, New Orleans, Louisiana. Associ- ation for Computational Linguistics.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Content preserving text generation with attribute controls", "authors": [ { "first": "Lajanugen", "middle": [], "last": "Logeswaran", "suffix": "" }, { "first": "Honglak", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Samy", "middle": [], "last": "Bengio", "suffix": "" } ], "year": 2018, "venue": "Advances in Neural Information Processing Systems", "volume": "", "issue": "", "pages": "5103--5113", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lajanugen Logeswaran, Honglak Lee, and Samy Ben- gio. 2018. Content preserving text generation with attribute controls. In Advances in Neural Informa- tion Processing Systems, pages 5103-5113.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "A dual reinforcement learning framework for unsupervised text style transfer", "authors": [ { "first": "Fuli", "middle": [], "last": "Luo", "suffix": "" }, { "first": "Peng", "middle": [], "last": "Li", "suffix": "" }, { "first": "Jie", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "Pengcheng", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Baobao", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Xu", "middle": [], "last": "Sun", "suffix": "" }, { "first": "Zhifang", "middle": [], "last": "Sui", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19", "volume": "", "issue": "", "pages": "5116--5122", "other_ids": { "DOI": [ "10.24963/ijcai.2019/711" ] }, "num": null, "urls": [], "raw_text": "Fuli Luo, Peng Li, Jie Zhou, Pengcheng Yang, Baobao Chang, Xu Sun, and Zhifang Sui. 2019. A dual rein- forcement learning framework for unsupervised text style transfer. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intel- ligence, IJCAI-19, pages 5116-5122. International Joint Conferences on Artificial Intelligence Organi- zation.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Style transfer through back-translation", "authors": [ { "first": "Yulia", "middle": [], "last": "Shrimai Prabhumoye", "suffix": "" }, { "first": "Ruslan", "middle": [], "last": "Tsvetkov", "suffix": "" }, { "first": "Alan", "middle": [ "W" ], "last": "Salakhutdinov", "suffix": "" }, { "first": "", "middle": [], "last": "Black", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "866--876", "other_ids": { "DOI": [ "10.18653/v1/P18-1080" ] }, "num": null, "urls": [], "raw_text": "Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhut- dinov, and Alan W Black. 2018a. Style transfer through back-translation. In Proceedings of the 56th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), pages 866-876, Melbourne, Australia. Association for Computational Linguistics.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Style transfer through back-translation", "authors": [ { "first": "Yulia", "middle": [], "last": "Shrimai Prabhumoye", "suffix": "" }, { "first": "Ruslan", "middle": [], "last": "Tsvetkov", "suffix": "" }, { "first": "Alan", "middle": [ "W" ], "last": "Salakhutdinov", "suffix": "" }, { "first": "", "middle": [], "last": "Black", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "866--876", "other_ids": { "DOI": [ "10.18653/v1/P18-1080" ] }, "num": null, "urls": [], "raw_text": "Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhut- dinov, and Alan W Black. 2018b. Style transfer through back-translation. In Proceedings of the 56th Annual Meeting of the Association for Com- putational Linguistics (Volume 1: Long Papers), pages 866-876, Melbourne, Australia. Association for Computational Linguistics.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Improving language understanding by generative pre-training", "authors": [ { "first": "Alec", "middle": [], "last": "Radford", "suffix": "" } ], "year": 2018, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Alec Radford. 2018. Improving language understand- ing by generative pre-training.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Language models are unsupervised multitask learners", "authors": [ { "first": "Alec", "middle": [], "last": "Radford", "suffix": "" }, { "first": "Jeffrey", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Rewon", "middle": [], "last": "Child", "suffix": "" }, { "first": "David", "middle": [], "last": "Luan", "suffix": "" }, { "first": "Dario", "middle": [], "last": "Amodei", "suffix": "" }, { "first": "Ilya", "middle": [], "last": "Sutskever", "suffix": "" } ], "year": 2019, "venue": "OpenAI Blog", "volume": "", "issue": "8", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8).", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Style transfer from non-parallel text by cross-alignment", "authors": [ { "first": "Tianxiao", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Tao", "middle": [], "last": "Lei", "suffix": "" }, { "first": "Regina", "middle": [], "last": "Barzilay", "suffix": "" }, { "first": "Tommi", "middle": [], "last": "Jaakkola", "suffix": "" } ], "year": 2017, "venue": "Advances in neural information processing systems", "volume": "", "issue": "", "pages": "6830--6841", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. In Advances in neural informa- tion processing systems, pages 6830-6841.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "transforming\" delete, retrieve, generate approach for controlled text style transfer", "authors": [ { "first": "Akhilesh", "middle": [], "last": "Sudhakar", "suffix": "" }, { "first": "Bhargav", "middle": [], "last": "Upadhyay", "suffix": "" }, { "first": "Arjun", "middle": [], "last": "Maheswaran", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Akhilesh Sudhakar, Bhargav Upadhyay, and Arjun Ma- heswaran. 2019. \"transforming\" delete, retrieve, generate approach for controlled text style transfer. In Proceedings of the 2019 Conference on Empiri- cal Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Attention is all you need", "authors": [ { "first": "Ashish", "middle": [], "last": "Vaswani", "suffix": "" }, { "first": "Noam", "middle": [], "last": "Shazeer", "suffix": "" }, { "first": "Niki", "middle": [], "last": "Parmar", "suffix": "" }, { "first": "Jakob", "middle": [], "last": "Uszkoreit", "suffix": "" }, { "first": "Llion", "middle": [], "last": "Jones", "suffix": "" }, { "first": "Aidan", "middle": [ "N" ], "last": "Gomez", "suffix": "" }, { "first": "Illia", "middle": [], "last": "Kaiser", "suffix": "" }, { "first": "", "middle": [], "last": "Polosukhin", "suffix": "" } ], "year": 2017, "venue": "Advances in Neural Information Processing Systems", "volume": "30", "issue": "", "pages": "5998--6008", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141 ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Pro- cessing Systems 30, pages 5998-6008. Curran Asso- ciates, Inc.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Mask and infill: Applying masked language model for sentiment transfer", "authors": [ { "first": "Xing", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Tao", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Liangjun", "middle": [], "last": "Zang", "suffix": "" }, { "first": "Jizhong", "middle": [], "last": "Han", "suffix": "" }, { "first": "Songlin", "middle": [], "last": "Hu", "suffix": "" } ], "year": 2019, "venue": "Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19", "volume": "", "issue": "", "pages": "5271--5277", "other_ids": { "DOI": [ "10.24963/ijcai.2019/732" ] }, "num": null, "urls": [], "raw_text": "Xing Wu, Tao Zhang, Liangjun Zang, Jizhong Han, and Songlin Hu. 2019. Mask and infill: Apply- ing masked language model for sentiment transfer. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI- 19, pages 5271-5277. International Joint Confer- ences on Artificial Intelligence Organization.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Unpaired sentiment-to-sentiment translation: A cycled reinforcement learning approach", "authors": [ { "first": "Jingjing", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Xu", "middle": [], "last": "Sun", "suffix": "" }, { "first": "Qi", "middle": [], "last": "Zeng", "suffix": "" }, { "first": "Xiaodong", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Xuancheng", "middle": [], "last": "Ren", "suffix": "" }, { "first": "Houfeng", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Wenjie", "middle": [], "last": "Li", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "979--988", "other_ids": { "DOI": [ "10.18653/v1/P18-1090" ] }, "num": null, "urls": [], "raw_text": "Jingjing Xu, Xu Sun, Qi Zeng, Xiaodong Zhang, Xu- ancheng Ren, Houfeng Wang, and Wenjie Li. 2018a. Unpaired sentiment-to-sentiment translation: A cy- cled reinforcement learning approach. In Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 979-988, Melbourne, Australia. Asso- ciation for Computational Linguistics.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Unpaired sentiment-to-sentiment translation: A cycled reinforcement learning approach", "authors": [ { "first": "Jingjing", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Xu", "middle": [], "last": "Sun", "suffix": "" }, { "first": "Qi", "middle": [], "last": "Zeng", "suffix": "" }, { "first": "Xiaodong", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Xuancheng", "middle": [], "last": "Ren", "suffix": "" }, { "first": "Houfeng", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Wenjie", "middle": [], "last": "Li", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "979--988", "other_ids": { "DOI": [ "10.18653/v1/P18-1090" ] }, "num": null, "urls": [], "raw_text": "Jingjing Xu, Xu Sun, Qi Zeng, Xiaodong Zhang, Xu- ancheng Ren, Houfeng Wang, and Wenjie Li. 2018b. Unpaired sentiment-to-sentiment translation: A cy- cled reinforcement learning approach. In Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 979-988, Melbourne, Australia. Asso- ciation for Computational Linguistics.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "2020. {BERTS}core: Evaluating text generation with {bert}", "authors": [ { "first": "Tianyi", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "*", "middle": [], "last": "", "suffix": "" }, { "first": "Varsha", "middle": [], "last": "Kishore", "suffix": "" }, { "first": "*", "middle": [], "last": "", "suffix": "" }, { "first": "Felix", "middle": [], "last": "Wu", "suffix": "" }, { "first": "*", "middle": [], "last": "", "suffix": "" }, { "first": "Kilian", "middle": [ "Q" ], "last": "Weinberger", "suffix": "" }, { "first": "Yoav", "middle": [], "last": "Artzi", "suffix": "" } ], "year": null, "venue": "International Conference on Learning Representations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. {BERTS}core: Evaluating text generation with {bert}. In Interna- tional Conference on Learning Representations.", "links": null } }, "ref_entries": { "FIGREF0": { "text": "provides 1 human reference and 3 additional human references inLuo et al. (2019). We used 2 human references, one from and one (the best performance in automatic evaluation) fromLuo et al. (2019), to increase reliability. Amazon: We used the human reference provided by.", "num": null, "type_str": "figure", "uris": null }, "FIGREF1": { "text": "(a) Trade-off curve of G-BLEU and Style accuracy according to \u03b1 (at \u03b2 = 0.5) in Yelp (b) Trade-off curve of G-BLEU and Style accuracy according to \u03b2 (at \u03b1 = 0.7) in Yelp.", "num": null, "type_str": "figure", "uris": null }, "TABREF2": { "num": null, "type_str": "table", "html": null, "text": "", "content": "
: Automatic evaluation results of the Yelp dataset (s: self, h: human, G: geometric mean, f: fine-tuned, p:
pre-trained). The red indicates that the evaluation score is significantly worse than other systems. Our model is
referred to as SST(\u03b1, \u03b2). The bold black indicates the better performance of our systems for the four metrics that
determine it is a stable system.
ContentAttributeFluencySemantic
Models-BLEU h-BLEU G-BLEU Classifier(%)d-PPLg-PPLt-PPLBERTscore
SST (0.6, 0.5)45.4720.3430.4166.54.51367.7340.7289.17
CrossAligned0.760.610.6874.81.11119.3711.5185.95
StyleEmbedding32.0312.9520.3742.43.42369.2435.5487.39
multi decoder16.486.6110.4470.31.39343.7221.8686.09
TemplateBased68.5433.7948.1264.85.36368.4144.4490.65
DeleteOnly57.4828.5640.52502.78251.2426.4390.55
DeleteAndRetrieve60.7530.8343.2852.42.43221.9223.2290.92
RetrieveOnly2.821.231.8682.35.65135.2227.6485.54
B GST58.2125.4738.559.112448.44 193.73 1552.9491.23
G GST51.0221.132.8157.318106458.93 2882.689.48
human: DRG47.6710069.0446.912.38132.1840.45100
input copy10047.668.9918.53.76188.3326.6193.77
" }, "TABREF3": { "num": null, "type_str": "table", "html": null, "text": "Automatic evaluation results of the Amazon dataset. Evaluation metrics are the same as Yelp, but Back-Translation, UnpariedRL, and DualRL do not provide results from Amazon datasets.", "content": "
ModelContent Fluency Style
SST(0.7, 0.75)3.323.373.3
BackTranslation2.693.152.99
TemplatedBased3.183.163.19
StyleEmbedding3.563.492.88
" }, "TABREF4": { "num": null, "type_str": "table", "html": null, "text": "Examples of comparison of generated sentences of AI systems. SST is our model. Attributes are colored. Red is negative and blue is positive.", "content": "
ConAttrFluSem
ModelG-BLEU Cls(%) t-PPL BERTscore
SST (0.7, 0)19.1182.2306.6589.96
-Style loss19.7878.2341.5189.84
" } } } }