In E-commerce, controllable text generation plays an essential role in generating attractive and suitable product descriptionsShao et al. (2021). These automatic description generation methods bring significant increases in writing efficiency and cost savings when facing billions of product data Zhang et al. (2019).
Most recent works depend on the single-pass paradigm to design the description generation manner Li et al. (2020a); Chan et al. (2019); Shao et al. (2021). In details, the generation model takes a set of inputs, which include the key selling points of a product (i.e., a set of attributes Shao et al. (2021)) and various forms of grounding, such as titles Zhang et al. (2019), customer reviews Zhan et al. (2021), or knowledge base Chen et al. (2019). Then, it outputs the final description, without considering user feedback (i.e., in a single pass).
Despite the success of the above studies, there is still a major limitation in this single-pass paradigm — it fails to interact with users and flexibly refine the generated description. Specifically, users’ expectations about the content might change after receiving the first version of description. For example, they may want to add or delete the content of a specific product attribute to obtain a more appropriate description. Unfortunately, it is inflexible for the single-pass paradigm to address this situation. On the one hand, existing models have to regenerate a description from scratch, even if needing a very few words changes from the previous version. On the other hand, utilizing manual post processing is time-consuming and often prohibitively expensive Green et al. (2013), since the edit operation includes finding the right place from the entire description, then rewriting the attribute-relevant content while fine-tuning context to keep readability.
To reach the ideal goal of generating descriptions interactively, we propose a new task to approximate the condition—controllable E-commerce description editing. In short, we allow users to flexibly modify the previous description (hereinafter known as draft) in a draft-command-edit manner. In particular, users input a specific command with an attribute (e.g., deleting or adding the attribute-relevant content in draft), the model then generates a new description (i.e., edit) based on the command, draft, and grounding (e.g., product title, product property). This paradigm would be easier and more practical to meet the new needs of users Faltings et al. (2020), while making the first attempt to high-volume processing on user-oriented description editing.
The key challenge of this task is that the rich supply of alignment data between draft and edit would be naturally inaccessible, which may cause big difficulties for these data-driven generation models. To overcome this low resource limitation, we propose a data augmentation method to automatically generate draft-edit data pairs. In more detail, we design two strategies to imitate the edit by humans. One of them uses a filling-in-the-blank generation model to approximate the human edit that removes the attribute-relevant tokens and slightly modifies the context to keep readability. Another one is based on the rules to imitate the edit that directly deletes the attribute tokens in the content.
Besides, we introduce a new draft-command-edit dataset called E-cEdits and a novel metric “Attribute Edit” to evaluate our method. E-cEdits is created by humans via crowdsourcing in Anonymity E-commerce scenario.111For the anonymous submission, we have temporarily hidden the specific real-world E-commerce platform name. As Figure 1 shown, we first crawl data from the platform, then the human annotators are asked to edit the description until it excludes the content about a pre-specified attribute. In the end, there are 9,000 <attribute, command, grounding, draft, edit> 5-tuples from 733 product categories. In addition, “Attribute Edit” aims to examine whether an attribute has been edited. In details, it computes a fuzzy matching score between the input attribute and the model output, evaluating whether the content of the user-specified attribute appears (for adding commands) / disappears (for deleting commands) in final results. It is simple but effective, and our experiments demonstrate that “Attribute Edit” metric is significantly correlated with human evaluation.
The main contributions of this work could be summarized as follows:
We propose a challenging new text generation task in E-commerce, which allows the model to flexibly modify the description constrained by users’ dynamical requirements. This paradigm could provide a novel insight to design a user-oriented generation manner in controllable text generation.
Responding to the key challenge of low resource in this task, we propose a new automatic data augmentation method to approximate human edit, which automatically generates pseudo data. Experiments on the E-cEdits dataset show that our system significantly outperforms baselines in automatic and human evaluations.
To accompany this task, we release an E-commerce text editing dataset E-cEdits and design a novel metric “Attribute Edit”.
|shape: column||brass magnetic clasps, column, silver size:about 8mm wide, …, just add to the end of your diy bracelets crimp in the hole.||brass magnetic clasps, silver size:about 8mm wide, …, just add to the end of your diy bracelets crimp in the hole.|
|flavor: green-apple flavor||you will receive green-apple flavor a must have for braces, …, package including : 10 boxes ortho wax.||you will receive a must have for braces, …, package including : 10 boxes ortho wax.|
|club type: hybrids||brand new aftermarket adapter for taylormade hybrids, …, they have 1.5 degrees of loft adjustment.||brand new aftermarket adapter, …, they have 1.5 degrees of loft adjustment.|
|nintendo model: nintendo switch||charging data cable for nintendo switch type c usb charger, …, 1pcs x charging cable.||charging data cable for the one that uses type c usb charger, …, 1pcs x charging cable.|
2 Related Work
In this section, we first analyze the related works on text generation in E-commerce. Then, we review some representative works which introduce text editing into some text generation scenarios.
2.1 Text Generation in E-commerce
Recently, various attempts have been made in E-commerce text generation, such as title generation Mane et al. (2020), summarization Li et al. (2020b), dialogue Zhang et al. (2020a), and answer generation Gao et al. (2021). The most related task to us is product description generation. Apex Shao et al. (2021)
based on a Conditional Variational Autoencoder generates a description from a set of attributes. FPDGChan et al. (2019) considers the entity label of each word and increases the fidelity of the descriptions. KOBE Chen et al. (2019) takes a variety of factors into account while generating descriptions, including product aspects, user categories, and knowledge base. The previous methods have their applicative advantages and insurmountable disadvantages: for example, they consider the generation of description under the one-pass setting from scratch. However, the users’ needs for constraining the description could be dynamic. In contrast, we provide a new generation paradigm regarding the process in a draft-command-edit manner, which significantly drops the difficulty of the description generating by taking advantage of previous versions.
2.2 Text Editing
Dating back to the period of rule-based postediting Knight and Chander (1994), text editing has long been investigated for text generation. According to the differences in the ultimate goal, it can be roughly divided into two types: (1) Refining the sentences to be more fluency and factually grounded; (2) Adding or modifying the content. The first type has a set of different settings, such as post-editing Herbig et al. (2020); Mallinson et al. (2020), grammatical error correction Zhao and Wang (2020); Wan et al. (2020), and paraphrasing Goyal and Durrett (2020); Siddique et al. (2020). Our new task is more relevant to the second type, in which editing text based on prototypes has achieved promising performance. For example, Guu et al. (2018) sample prototype from the training corpus. FACTEDITOR Iso et al. (2020) creates a fact-based draft by rules as the model input in two data-to-text tasks. Faltings et al. (2020) crawl data from Wikipedia’s revision histories to form a Draft-Edit pair, and generate text according to the command in a progressively adding manner. However, directly incorporating the methods above into E-commerce is less portable. It is because the text editing manner of these methods is based on continuing writing (i.e., generating new sentences after the current text), which might be hard to modify previous content according to users’ wishes. In comparison, our task provides more flexible operations—adding and deleting, which takes modifying and adding content of the description both into account. Meanwhile, the editing object product attribute is the central theme of a product description, which promotes our task more adaptable to the application requirements in E-commerce Petrovski and Bizer (2017).
In this section, we introduce our new task and elaborate the benchmark E-cEdits. To ease of presentation, we start from formalize the new task in § 3.1. Then we give a detailed creation of the presented dataset E-cEdits in § 3.2.
3.1 Task Definition
The controllable E-commerce description editing task is defined as follows: given an attribute , a command and the specified forms of grounding , the system generates a new description based on the previous version . In our specific settings to instantiate this task, the command is defined as adding or deleting the content of attribute from the description while keeping readability. Meanwhile, the type of attribute-relevant content is various. It could be a word, a phrase, or a clause when appearing in the description. In addition, grounding is the supplementary information about the product, such as titles. In sum, the editing process is , given , , and .
It is tough to obtain for from the E-commerce platform, since the history version for attributes editing is difficult to collect. To accompany this task and evaluate the effectiveness of the proposed method, we present a high-quality dataset containing 9,000 draft-command-edit tuples in the English E-commerce domain, which is wholly written by humans via crowdsourcing in the Anonymity E-commerce platform. To be concrete, we ask each annotator to remove the content of a pre-specified attribute from the complete description, which is consistent with the deleting editing in the task definition.222In fact, each description has multiple attributes (2.61 on average), and we randomly select one of them as the pre-specified hint for description editing. When considering the adding editing, data tuples can be easily obtained by exchanging the deleted editing samples’ source and target descriptions. It is worth mentioning that the editing operation of humans follows the “minimum modification principle”. This principle means that the annotator should remove the relevant content, may add punctuation or a few words to keep the readability. In addition, the third-party inspectors examine 200 random samples from the editing data for data quality assurance and make sure the pass rate is more than 95%. Finally, each sample of E-cEdits contains 5-tuple <attribute, command, grounding, draft, edit>. In the implementation, the command includes two signals, “[ADD]” and “[DEL]”, which denote adding and deleting commands, respectively. Meanwhile, we choose product title and category as the grounding following Zhang et al. (2019).
Statistical Analysis Table 2 gives a statistical overview of our dataset. More concretely, the mean length difference between draft (i.e., ) and edit (i.e., ) suggests that the words changing in editing is slightly. Meanwhile, descriptions come from 733 categories, and various types mean our dataset could approximate the real condition. Finally, the mean length of the attribute implies that it usually includes only a single word or a short phrase. As a result, the minimal information offered by attributes may bring algorithms difficulties to generate a description from scratch.
Examples The typical edit examples of E-cEdits are illustrated in Table 1. The phenomena can be mainly divided into two categories: 1) Deleting the attribute-relevant content. For instance, the annotator may delete the keywords (row 2), the phrase (row 3), or the clause (row 1); 2) Replacing the attribute words with attribute-free ones (row 4) to keep the text flowing. We can see that the editing contains various forms and appears in different positions, which will present the editing system in the low-resource settings with a great challenge.
In this section, we elaborate our data augmentation method and the description editing model. To ease of presentation, we start from toy examples to illustrate the overview of the data augmentation method in § 4.1, and give a detailed explanation of the implementation. Then, we present the editing model in § 4.2.
4.1 Automatic Data Augmentation
To remedy the low resource challenge in this task, we design a data augmentation method to imitate the edit by humans. Although draft-command-edit data pairs are difficult to obtain naturally, a large number of descriptions and corresponding attributes can be easily collected from E-commerce platforms. Thus, we consider strategically removing the content of a pre-specified attribute from the description (i.e., the deleting editing), which is similar to the dataset building method of humans editing. After that, we obtain the adding editing sample by exchanging the source and target descriptions in deleting, as mentioned in § 3.2.
Model-based Strategy Our basic idea is to mask the content of a pre-specified attribute in description with a signal “[FILL]” then use a filling-in-the-blank model to generate attribute-free content on that position. Therefore, we get a draft for each description while keeping readability. The workflow is shown in the blue flow of Figure 2. For each description, we extract a word or a phrase, then replace it with a mask token “[FILL]”. After that, we fine-tune a pre-trained Seq2Seq model (e.g., ProphetNet Qi et al. (2020)) to reconstruct the description (i.e., the “Pretrained model” part in Figure 2). In the inference, we use a phrase fuzzy matching tool based on the Levenshtein Distance to mask the content of a pre-specified attribute with “[FILL]”.333This tool can be accessed via https://pypi.org/project/fuzzywuzzy/. Then, we constrain the decoding space by removing the attribute tokens in vocabulary while generating. As a result, the model can generate an attribute-free description.
It is worth mentioning that we develop multiple policies for deciding which tokens or phrases should be masked with “[FILL]”. First, we use the TF-IDF score at the token level to choose and mask an essential word in each description (except for stop words and punctuation). The TF-IDF score could provide the uniqueness and local importance of a word at the corpus level Zhang et al. (2020b). Second, we design three masking approaches to increase the filling types’ variety at the phrase level, aiming to approximate human editing situations. Concretely, it includes random, conjunction-based, and attribute-based masking (the blue block “Phrase Level”). In randomly masking, we randomly choose 2-5 word pieces to mask each description, referring to the length statistic results. Meanwhile, we consider masking attribute-relevant words with phrase fuzzy matching tool thus models can enjoy benefits to better deal with attributes. In addition, we randomly mask the clauses connected with coordinating conjunctions as it is a typical form when multiple attributes appear in one sentence. Finally, we collect 1.2 million data pairs in total for model training. The proportion of each type in the training data set is: 50% for token level and 50% for phrase-level (50% attribute-based masking, 30% conjunction-based masking, and 20% randomly masking).
Rule-based Strategy We design two ways to imitate the editing type of directly deleting attribute-relevant content in a description. On the one hand, we directly remove the sentence in a description if it only contains one attribute (the “Sentence Level” block in the figure). On the other hand, as removing adjectives does not affect the sentence integrity, we also fall the attributes of adjectives into this category (the “For adjective” block).
For the description edit task, we use the standard auto-regressive sequence to sequence models as test beds, thus various generation models can be easily adapted to this task. Given an attribute , an command , groundings , and draft , the model generates Edit by:
where is the model parameters.
In implementation, we use the pre-trained Seq2Seq model ProphetNet Qi et al. (2020) as the backbone, which is effective on several text generation tasks Liu et al. (2020); Yang et al. (2022). To adapt the ProphetNet in our task, we follow Gururangan et al. (2020) to continue training it in three steps. Firstly, to adapt ProphetNet to the E-commerce domain, we collect 24 million grounding E-commerce titles to continue pre-train ProphetNet in a denoising sequence-to-sequence task. After that, we obtain a pre-trained E-commerce domain Seq2Seq model called ProphetNet-E. Secondly, we train ProphetNet-E in text editing task using our automatic augmentation dataset as described in § 4.1, which contains 600 thousand samples. Finally, we train the model on the E-cEdits dataset. As shown in Figure 3, in the editing task, it has to be mentioned that we concatenate all of the model’s input (, , , and ) by the separator signal “[SEP]” to adapt the models for our task.
|Baselines (row 1-3)||Transformer||56.32||55.57||0.37||14.23||2.54||13.05||3.50|
|Ablations (row 4-6)||only Grounding||61.72||85.77||-12.03||89.16||86.00||89.03||79.97|
|no Data Augmentation||61.24||85.97||-12.36||89.53||86.45||89.42||80.69|
Human Evaluation for Mass and our system. Note that the higher of “DEL” score represents the better quality as opposed to the automatic evaluation. The significance test is carried out for each group by the Two-Sample t-Test, all of the p values are less than 0.05.
5.1 Experimental Settings
For the E-cEdits dataset, we randomly sample 4,000/1,000 pairs for training/validation, and the remaining 4000 for testing. All of the models are implemented based on Fairseq Ott et al. (2019), and the specific parameters setting for each model can be found in § 5.2
. We set the max training epoch to 10 for each model. In the inference step, the beam size and length penalty are set to 4 and 1.2 respectively, to calculate the main results without post-processing.
is the most commonly used sequence to sequence model. We follow the hyperparameters of standard TransformerVaswani et al. (2017).
MASS Song et al. (2019) uses the encoder-decoder framework to reconstruct a sentence fragment given the remaining part of the sentence. We use the largest released pretraining model “MASS-middle-uncased” trained on Wikipekia and BookCorpus. It contains six layers (embedding/hidden size 1024 and 16 head for each attention layer) for both encoder and decoder.444https://github.com/microsoft/MASS
ProphetNet Qi et al. (2020)
introduces an n-stream self-attention mechanism and a self-supervised objective named future n-gram prediction. Two versions of ProphetNet are used, and the main difference is the source of the training dataset. “ProphetNet” in Table3 is trained on English Wikipekia and BookCorpus (16GB in total), while “ProphetNet-E” is continue trained on E-commerce text. Both of two models include a 12-layer encoder (decoder) with 1024 embedding/hidden size and 4096 feed-forward filter size.555https://github.com/microsoft/ProphetNet
5.3 Evaluation Metrics
Automatic Evaluation Following the automatic evaluation methods in both text editing Faltings et al. (2020); Iso et al. (2020) and description generation in E-commerce Chen et al. (2019); Chan et al. (2019), we first use BLEU Papineni et al. (2002) and ROUGE Lin (2004) to keep in line with previous works. The BLEU score is calculated by the built-in function in Fairseq, while the ROUGE score is calculated by the “files2rouge” tool.666https://github.com/pltrdy/files2rouge
While ROUGE and BLUE could evaluate fluency of the generated description, there is no way to explicitly examine whether an attribute has been edited, which is one of the key points in this task. To tackle this problem, we propose a new evaluation indicator “Attribute Edit”, which computes the fuzzy matching score between the input attribute and the model output (the matching tool is introduced in § 4.1, and score range 0-100). It is worth mentioning that this indicator is significantly correlated with human evaluation (details can be found in § 5.5). Note that we use “ADD” and “DEL” to represent the evaluation scores on adding commands and deleting commands, while “ALL” denotes the average score of them. In deleting operations, the lower matching score indicates better model performance. Therefore, we convert “DEL” scores into negative ones when computing the overall score (i.e., “ALL” in Table 3).
Human Evaluation We also conduct a human evaluation to compare our system with baselines. 10 human graders are asked to evaluate the fluency and attribute relevance (for deleting command, we evaluate the attribute irrelevance) across 50 randomly selected examples from our test set, which include 25 deleting editing samples and 25 adding editing samples. Following Genie Khashabi et al. (2021), we use a discrete Likert scale for 5 categories: exceptionally bad, bad, just OK, good, and perfect instead of a continuous one. Finally, we convert categories into scores (1-5 from exceptionally bad to perfect) and get the average score.
|Item||Attribute Edit||Character-level LD|
|ADD||0.92 (p-value<0.0001)||-0.01 (p-value=0.96)|
|DEL||0.89 (p-value<0.0001)||0.17 (p-value=0.23)|
5.4 Main Results
We evaluate the performance of our system and baselines on E-cEdits dataset and further provide ablations. We report the main result on Table 3 and the human evaluation can be found in Table 4, from which we can make the following conclusions:
1. Our system consistently outperforms baselines both in automatic evaluation and human evaluation. For automatic evaluation (as shown in Table 3), our system outperforms all baselines both in fluency (BLEU and ROUGE) and attribute relevance (Attribute Edit). For example, comparing with MASS that gets the highest 89.30 BLUE and 94.72/92.05/94.69 ROUGE scores in text editing baselines (row 2), our system still beats MASS with 91.78 BLEU and 96.52/94.01/96.28 ROUGE score (row 7). Especially in Attribute Edit, using the proposed data augmentation method can significantly improve the editing model compared with other baselines. We choose MASS to compare with our system in human evaluations because it is the best baseline in both ROUGE and BLUE. As shown in Table 4, our system also outperforms MASS in both fluency and attribute relevance. Consistent with the automatic evaluation, these two models show significant differences in “Attribute-relevant” scores (2.05 vs. 3.95).
|Attribute: Chain style: handmade chains|
|Draft: Beads for necklaces bracelets making, with silver iron eye pins, mixed stone…|
|Model output: Handmade stone chips beads chains for necklaces bracelets making, with silver iron eye pins, mixed stone…|
|Attribute: Item name: original rear camera module replacement part|
|Draft: For sony xperia xa ultra compatible: for sony xperia xa ultra (only fit for xa ultra, not fit for xa1 ultra) features: original high…|
|Model output: Original rear camera module replacement part for sony xperia xa ultra (only fit for xperia ultra), not fit for xa1 ultra, features: original high…|
|Attribute: Suitable: abs systems|
|Draft: Feature : automatic brake… and magnetic reservoir base to prevent accidental spillage. Suitable for use with abs systems. Supplied with range of…|
|Model output: Feature : automatic brake… and magnetic reservoir base to prevent accidental spillage . Supplied with range…|
|Attribute: Product type: stopper|
|Draft: We have our own button factory… bulk order have more discount. Item description: plastic stoppers: 24mm width * 23mm height.|
|Model output: We have our own button factory… bulk order have more discount.|
2. Both command and grounding play important roles in the controllable text editing task. We also explore the impact of grounding and command on model performance. As shown in ablations (row 4 and row 5), using explicit command may improve performance in fluency, while using the grounding benefits model in better attribute editing. Meanwhile, ProphetNet-E outperforms all baselines (row 6) when in the “no Data Augmentation” condition. That is, using E-commerce text for pretraining could better adapt ProphetNet to this task (compared with row 3).
5.5 Further Analysis
Attribute Relevance Evaluation We further compute Pearson correlation coefficient between the human score and our “Attribute Edit” score, to verify it can effectively evaluate whether models carry out an editing operation. Meanwhile, we also choose the character-level Levenshtein distance as the baseline, which is widely used in judging the two sentences’ similarity Snover et al. (2006).777https://pypi.org/project/python-Levenshtein/0.11.2/ Table 5 illustrates that there is a significant statistical correlation between the proposed “Attribute Edit” score and human evaluation. In addition, the character-level Levenshtein distance is irrelevant with human evaluation as all of the p-values greater than 0.05.
Case Analyze Table 6 illustrates four examples of description editing results with our edit system. Especially, we can see that the operation of adding and deleting is not just simply copying or removing all the words in the pre-specified attribute, as finding an appropriate position in draft for editing operation is one of the challenges in this task. For example, with deleting command, sample 4 needs to remove the attribute “product type: stopper”. Our model not only removes the word “stopper”, but also the relevant content “24mm width…”, which keeps the readability of the Edit version.
In this paper, we propose a new controllable text editing task allowing users flexibility to constrain the attribute-relevant content of the product description by commands in a draft-command-edit manner, and introduce a high-quality draft-command-edit dataset E-cEdits written by humans. Meanwhile, in response to the low resource condition—the key challenge in this task, we design a data augmentation method that contains two strategies to generate pseud data pairs. Experiments demonstrate that our method significantly and consistently outperforms baselines both in automatic evaluation and human evaluation. In sum, as a new attempt, we tentatively give a simple but effective implementation of product description editing, successfully approximate the ideal goal of generating descriptions interactively. Thus in the future, such a paradigm deserves a closer and more detailed exploration. Therefore, we will investigate to design this interactive generation manner in a more superior way.
Stick to the facts: learning towards a fidelity-oriented e-commerce product description generation.
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. 4960–4969. Cited by: §1, §2.1, §5.3.
- Towards knowledge-based personalized product description generation in e-commerce. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 3040–3050. Cited by: §1, §2.1, §5.3.
- Text editing by command. arXiv preprint arXiv:2010.12826. Cited by: §1, §2.2, §5.3.
- Meaningful answer generation of e-commerce question-answering. ACM Transactions on Information Systems (TOIS) 39 (2), pp. 1–26. Cited by: §2.1.
- Neural syntactic preordering for controlled paraphrase generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pp. 238–252. Cited by: §2.2.
- The efficacy of human post-editing for language translation. In Proceedings of the SIGCHI conference on human factors in computing systems, pp. 439–448. Cited by: §1.
- Don’t stop pretraining: adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, D. Jurafsky, J. Chai, N. Schluter, and J. R. Tetreault (Eds.), pp. 8342–8360. External Links: Cited by: §4.2.
- Prototypes generating sentences by editing prototypes. Transactions of the Association for Computational Linguistics 6, pp. 437–450. Cited by: §2.2.
- MMPE: a multi-modal interface for post-editing machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 1691–1702. Cited by: §2.2.
- Fact-based text editing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pp. 171–182. Cited by: §2.2, §5.3.
- Genie: a leaderboard for human-in-the-loop evaluation of text generation. arXiv preprint arXiv:2101.06561. Cited by: §5.3.
- Automated postediting of documents. In AAAI, Vol. 94, pp. 779–784. Cited by: §2.2.
Aspect-aware multimodal summarization for chinese e-commerce products.
Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34, pp. 8188–8195. Cited by: §1.
- Aspect-aware multimodal summarization for chinese e-commerce products. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34, pp. 8188–8195. Cited by: §2.1.
- ROUGE: a package for automatic evaluation of summaries. In Text Summarization Branches Out, Barcelona, Spain, pp. 74–81. Cited by: §5.3.
- Glge: a new general language generation evaluation benchmark. arXiv preprint arXiv:2011.11928. Cited by: §4.2.
- FELIX: flexible text editing through tagging and insertion. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, EMNLP 2020, Online Event, 16-20 November 2020, pp. 1244–1255. Cited by: §2.2.
- Product title generation for conversational systems using bert. arXiv preprint arXiv:2007.11768. Cited by: §2.1.
- Fairseq: a fast, extensible toolkit for sequence modeling. In Proceedings of NAACL-HLT 2019: Demonstrations, Cited by: §5.1.
- Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pp. 311–318. Cited by: §5.3.
- Extracting attribute-value pairs from product specifications on the web. In Proceedings of the International Conference on Web Intelligence, pp. 558–565. Cited by: §2.2.
- ProphetNet: predicting future n-gram for sequence-to-sequence pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pp. 2401–2410. Cited by: §4.1, §4.2, §5.2.
- Controllable and diverse text generation in e-commerce. In The World Wide Web Conference, Cited by: §1, §1, §2.1.
Unsupervised paraphrasing via deep reinforcement learning. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 1800–1809. Cited by: §2.2.
- A study of translation edit rate with targeted human annotation. In Proceedings of association for machine translation in the Americas, Vol. 200. Cited by: §5.5.
MASS: masked sequence to sequence pre-training for language generation.
Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, K. Chaudhuri and R. Salakhutdinov (Eds.), Proceedings of Machine Learning Research, Vol. 97, pp. 5926–5936. External Links: Cited by: §5.2.
- Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V. N. Vishwanathan, and R. Garnett (Eds.), pp. 5998–6008. External Links: Cited by: §5.2.
- Improving grammatical error correction with data augmentation by editing latent representation. In Proceedings of the 28th International Conference on Computational Linguistics, pp. 2202–2212. Cited by: §2.2.
- GCPG: A general framework for controllable paraphrase generation. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, S. Muresan, P. Nakov, and A. Villavicencio (Eds.), pp. 4035–4047. External Links: Cited by: §4.2.
- Probing product description generation via posterior distillation. Cited by: §1.
- Automatic generation of pattern-controlled product description in e-commerce. In The World Wide Web Conference, pp. 2355–2365. Cited by: §1, §1, §3.2.
- Multi-turn dialogue generation in e-commerce platform with the context of historical dialogue. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pp. 1981–1990. Cited by: §2.1.
- POINTER: constrained progressive text generation via insertion-based generative pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, B. Webber, T. Cohn, Y. He, and Y. Liu (Eds.), pp. 8649–8670. External Links: Cited by: §4.1.
- MaskGEC: improving neural grammatical error correction via dynamic masking. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34, pp. 1226–1233. Cited by: §2.2.