Tiny-StyleWizard: Unleashing the Potential of Small Language Models in Complex Style Transfer

Firstly, we want to sincerely thank you for the time and effort you've put into reviewing our paper and offering insightful feedback. We have thoroughly reviewed each point you raised and have responded to the key issues in the sections below. We trust that our replies will resolve any uncertainties and further highlight the significance and contributions of our research.

• The small language models we refer to are those with fewer than 100 million parameters, such as T5-tiny (31M) and T5-small (60M). As can be seen from Figure 5 in the paper, both of these small language models outperform the existing best models in terms of style transfer accuracy and semantic preservation.
• The complex text style transfer we claim originates from the BTTS paper[1]. In that work, they provide a preliminary definition of complex text style, which refers to styles that are difficult for non-experts to discern and categorize. Such styles might include personality, domain-specific jargon, or other highly specialized terminology. The complexity of these styles makes it challenging to use crowdsourcing for labeling, as only experts in the relevant fields can accurately differentiate between styles. Therefore, the challenge in complex text style transfer is to develop effective models capable of capturing and transferring these nuanced styles without relying heavily on labeled data.
• Details on how the models were fine-tuned: we fine-tuned the pretrained language models (BART, T5, LongLM etc.) with the huggingface coding framework on single A100 GPU. The initial learning rate is 2e-5. The batch size is 4.
• We will correct the citation issues in subsequent versions;
• We will also modify the layout of the tables according to your suggestions.
• Reading Table 1, the main takeaways are two-fold: First, using a few-shot classifier based on ChatGPT to calculate the accuracy of style transfer is effective, and it performs better than the model fine-tuned on the training set using DeBERTa. Second, using LLM-score to assess the degree of semantic preservation in style transfer is more reliable than other metrics. Our reasoning is that ChatGPT, trained on massive corpora and reinforced through feedback alignment, naturally preserves the semantics of the original text in its stylized output. A good assessment metric should, therefore, yield a high similarity when evaluating the original text against the text generated by ChatGPT.
• Regarding the differences in baselines: Table 4 shows the experimental results on the Genshin dataset, while Table 5 presents the results on the Shakespeare dataset. The former involves shorter text lengths (under 30 words), and the latter longer text lengths (average length of about 70 words). Both represent complex styles and have some foundational work. We established the baselines based on their respective foundational works.
• In manual evaluations, we used the fine-tuned BART-large model. We will bold the performance of the best models in all tables.
• As for diversity you mentioned, our game designers hope to express similar semantics within the same style using diverse wordings. According to their rich experience, players do not appreciate encountering repetitive and identical text throughout the game, as it can lead to aesthetic fatigue and diminish their gaming experience. Moreover, we hope that the stylized content viewed by different players will also vary in terms of wrod choice. Therefore, we believe that this metric is extremely important for the gaming business. Thus, we take this diversity metric into accout as well as the conventional style transfer accuracy and semantic preservation.

References

[1] Specializing Small Language Models towards Complex Style Transfer via Latent Attribute Pre-Training

We greatly appreciate the time and effort you dedicated to reviewing our paper and offering insightful feedback. We have thoughtfully reviewed each of your remarks and have responded to the key concerns in the sections below. It is our hope that these clarifications will resolve any outstanding questions and further highlight the significance and contributions of our research.

• Our research not only demonstrates that large language models can generate high-quality datasets for training smaller models with comparable performance in complex style transfer tasks, but also investigates the minimal size of language models required to effectively execute complex style transfer. Furthermore, it explores how different dimensions of complex style—specifically, the accuracy of style transfer and the degree of semantic preservation—place varying demands on the size of the language model.

• The calculation method of LLM-score is distinct from that of BERT-score. LLM-score assesses the similarity between model-generated text and reference text at the level of the whole sentence (cosine similarity). In contrast, BERT-score requires calculating the precision and recall at the word/token level for the generated text in comparison to the reference text. We believe that complex text styles are manifested across entire sentences, rather than being uniformly distributed across each word.

• Thank you for pointing out the issue with inappropriate citation. We will address this in subsequent versions of our paper.

• This traditional perspective has been discussed in the previous paper[1]. At that time, large language models, such as ChatGPT, had not yet achieved remarkable success in various natural language processing tasks. Following the emergence of ChatGPT, our view, similar to yours, is that introducing high-quality parallel corpora generated by large language models can break this trade-off. Therefore, we conducted this study. The experimental results in our paper demonstrate the validity of our idea.

• Our prompt for style transfer accuracy classifier:

  ${CHARACTER_PROFILE}
  
  You are an excellent style classifier, please determine whether the style presented in the input sentence is consistent with ${YOUR_CHARACTER}'s speaking style. Output either `yes` (consistent) or `no` (inconsistent). 
  
  ${DEMO}
  
  input sentence: ${text}
  label: 
  

• Our prompt for style transfer in Table 6:

  The prompt to generate parallel corpus for LuXun-style transfer is shown beneath. The initial parallel sample(s) used for demonstration is obtained by zero-shot prompting. The prompt to generate parallel corpus for JinYong-style transfer is similar.

  给定一段童话文本，在保持原文语义不变的情况下，将其转换为带有鲁迅风格的文本。
  
  ${DEMO}
  
  童话风格的文本：${text}
  鲁迅风格的文本：
  
  Given a piece of fairy tale text, it is transformed into a text with the style of Lu Xun while maintaining the original meaning.
  
  ${DEMO}
  
  Fairy-tale style text: ${text}
  LuXun-style text:
  

[1] Storytrans: Non-parallel story author-style transfer with discourse representations and content enhancing

At the outset, we wish to extend our heartfelt thanks for the time and effort you invested in reviewing our paper and offering constructive feedback. We have diligently examined each of your observations and have tackled the primary issues in the subsequent sections. Our aim is that our detailed responses will clear up any uncertainties and more effectively showcase the importance and impact of our study.

• We will revise the terms that are difficult to understand. Some modifications are listed as follows:

  • unleash the capability of these models -> fully utilize the potential of these models
  • Small language models specialized for the magician capabilities of style transfer -> Small language models specialized for style transfer

• We will present the prompts we used and the related results here. We used ChatGPT as a few-shot classifier for categorizing text styles. The prompt we used is shown beneath. [Prompt text here]. The term 'fine-tuned model' refers to the model we obtained by fine-tuning DeBERTa on the training set containing a variety of different style corpora. In our work, the training dataset is the training part of genshin, Shakespeare, LuXun and JinYong.

  The prompt used for evaluating the style transfer accuracy:

  ${CHARACTER_PROFILE}
  
  You are an excellent style classifier, please determine whether the style presented in the input sentence is consistent with ${YOUR_CHARACTER}'s speaking style. Output either `yes` (consistent) or `no` (inconsistent). 
  
  ${DEMO}
  
  input sentence: ${text}
  label: 
  

• We will adjust the layout of the figures and references in the paper based on your suggestions.

• We invited 10 designers from the gaming industry to evaluate the texts generated by different models. These designers have played multiple open-world games, including 'Genshin Impact.' They are capable of discerning and judging the quality of the stylized texts generated by the models.

• Regarding the differences in baselines: Table 4 shows the experimental results on the Genshin dataset, while Table 5 presents the results on the Shakespeare dataset. The former has shorter text lengths (under 30 words), and the latter has longer text lengths (average length of about 70 words). Both represent complex styles and have some preliminary work. We set up the baselines based on their respective preliminary works.

First and foremost, we would like to express our sincere gratitude for your time and effort in reviewing our paper and providing valuable feedback. We have carefully considered each of your comments and addressed the main concerns in the following sections. We hope that our responses will clarify any issues and further demonstrate the value and contribution of our work.

In our initial experiments, we employed the GST model[1] as a representative of strong supervised baselines which leverage parallel dataset (either synthetic or mined from unsupervised ground truth). The results demonstrated that the performance of the GST model on the Genshin dataset was inferior to that of our proposed method. Similar experimental outcomes have been observed in previous studies[2]. The last paper[3] you mentioned directly employs a large language model combined with specialized prompting techniques for text style transfer. We believe that this approach is fundamentally analogous to using ChatGPT for text style transfer, which is why we did not include it as a baseline in our study. Our experimental findings indicate that our method achieves superior results in text style transfer tasks compared to ChatGPT.

The primary reasons for employing small language models in complex style transfer are attributable to the proprietary nature of closed-source large language models such as OpenAI's ChatGPT, the substantial deployment costs associated with open-source large language models as well as the significant inference latency.

We also conducted a manual evaluation on the Genshin dataset, with the results presented in Table 3. The participants in the manual evaluation were professionals from the gaming industry, specifically game designers. These designers have experience playing multiple open-world games, including Genshin. They possess the expertise to accurately identify whether the texts generated by the model align with the characters' established settings and personalities.

[1] Transforming Delete, Retrieve, Generate Approach for Controlled Text Style Transfer

[2] Specializing Small Language Models towards Complex Style Transfer via Latent Attribute Pre-Training

[3] A Recipe for Arbitrary Text Style Transfer with Large Language Models