Automatic Generation and Evaluation of French-Style Chinese Modern Poetry

Abstract

Literature has a strong cultural imprint and regional color, including poetry. Natural language itself is part of the poetry style. It is interesting to attempt to use one language to present poetry in another language style. Therefore, in this study, we propose a method to fine-tune a pre-trained model in a targeted manner to automatically generate French-style modern Chinese poetry and conduct a multi-faceted evaluation of the generated results. In a five-point scale based on human evaluation, judges assigned scores between 3.29 and 3.93 in seven dimensions, which reached 80.8–93.6% of the scores of the Chinese versions of real French poetry in these dimensions. In terms of the high-frequency poetic imagery, the consistency of the top 30–50 high-frequency poetic images between the poetry generated by the fine-tuned model and the French poetry reached 50–60%. In terms of the syntactic features, compared with the poems generated by the baseline model, the distribution frequencies of three special types of words that appear relatively frequently in French poetry increased by 12.95%, 15.81%, and 284.44% per 1000 Chinese characters in the poetry generated by the fine-tuned model. The human evaluation, poetic image distribution, and syntactic feature statistics show that the targeted fine-tuned model is helpful for the spread of language style. This fine-tuned model can successfully generate modern Chinese poetry in a French style.

1. Introduction

In the 19th century, philosopher Georg Wilhelm Friedrich Hegel listed architecture, sculpture, painting, music, and poetry as the five canonical arts of humankind in his book Lectures on Aesthetics (Vorlesungen über die Ästhetik). As the fifth art, poetry is the quintessence of language and is a pearl that will never be absent from the literature of various countries. Different linguistic and cultural backgrounds create different styles of poetry. When reading an exotic poem, we can feel the charming exotic atmosphere. Even if the foreign poem has been translated into the reader’s mother tongue, the reader can still feel the difference from their native culture. Even without knowing whether a poem has been translated, the reader may guess the original language and nationality of the author through the poem’s linguistic style—vocabulary, syntax, imagery, and rhetoric. All these elements constitute the language style of a poem. In the past, readers who wanted to read exotic poems could only rely on human translators to translate foreign poems into their own languages, or local poets could have imitated the foreign language style when they were inspired to create poetry. Today, with the emergence and development of large language models (LLMs) [1,2,3], readers may not have to wait long. If one can transfer a specific “language style” by fine-tuning the language model, one can generate poetry with a specific language style.

The greatest difference between human beings and other intelligent creatures on Earth is that human beings can articulate, comprehend, and disseminate knowledge through structured language, expressing complex logical concepts. The development and manifestation of all aspects of human civilization, from science and engineering to humanities and arts, are inseparable from language. Language abstracts objectively existing things into concepts, and, in turn, the concepts in language can help us understand the real world. Philosopher Ludwig Wittgenstein once said, “The limits of my language mean the limits of my world”. In ancient societies, humans knew nothing about unnamed plants, and a polymath living in the third century BCE could not understand what a portable phone meant; hence, the boundary of languages is to some extent equal to the boundary of human civilization. The poems created by human poets are rooted in historical soil, and their expression is often limited by the author’s common sense, culture, imagination, and background of the times. However, when the language style is disseminated through the targeted fine-tuning of large models, the output of poetry is not a simple reproduction and imitation. It is no longer limited by the shackles of the poet. It is freer and easier to let the imagination run wild and generate breakthrough language that did not exist before. The “emergence” of its language broadens the boundary of human language.

In the field of natural language processing (NLP), the culturally interesting study of poetry generation, like machine translation, has moved from generation based on rules and templates [4,5] to generation based on statistical machine learning [6,7] and then to generation based on neural networks [8,9]. Currently, poetry generation based on pre-trained large language models [10,11,12,13] has greatly improved in quality. France and China are both countries with profound cultural heritage. There are two reasons for choosing French poetry as the source of style and Chinese as the form of poetry presentation: on the one hand, many poetry schools have been born throughout the history of French literature, and a large number of outstanding poets have emerged who are deeply loved by Chinese readers. On the other hand, French is an inflectional language belonging to the Latin part of the Indo-European language family, and Chinese is an analytic language belonging to the Chinese part of the Sino-Tibetan language family. French and Chinese have distinct differences both in language and cultural background, and people can obviously feel a foreign atmosphere when reading poetry in other language styles. Therefore, in this work, we generate Chinese modern poetry based on a pre-trained large language model, fine-tune it for the French language style, and apply human feedback reinforcement learning for further training to generate Chinese modern poetry with a distinctive French style.

The remainder of this paper is organized as follows: Section 2 reviews the related work on poetry generation and reinforcement learning and explains why this research content and the related methods were chosen. Section 3 introduces our proposed method for generating French-style Chinese modern poetry. Section 4 introduces the experiments. Section 5 offers the experimental and evaluation results from various perspectives. Finally, Section 6 concludes the paper and proposes future work.

2. Related Work

2.1. Poetry Generation

Researchers from all over the world have explored and studied poetry generation in Chinese [14,15], English [16,17,18], French [18,19], Spanish [4,20], Portuguese [21], Arabic [12,13,22], and many other languages. According to the different styles of generated poetry, it can be divided into “modern poetry generation” and “classical poetry generation”. According to the content of the generated poetry, it can be divided into “generating poems on specified themes” and “generating poems in the style of a specific poet”. There are quite detailed divisions in the forms of Chinese poetry: the forms of pre-Tang poetry (古体诗) and regulated verse (近体诗) belong to classical poetry, written in ancient Chinese, with restrictions on rhyme, the tones of Chinese characters, and the number of characters; modern poetry (现代诗) is written in modern Chinese, and there are no specific requirements for rhyme, tone, and word count. Researchers have conducted many studies on the generation of Chinese classical poetry [7,9,14,15,23,24,25] but have paid relatively little attention to the generation of modern Chinese poetry [26]. Most of the research content focuses on the embodiment of Chinese internal poetic images and emotions. To our knowledge, however, there has been little research on using Chinese to generate poetry in other language and cultural styles. Therefore, we decided to make this literary attempt using fine-tuning methods for the pre-trained model to “transplant” the style of French poetry to Chinese poetry, hoping to use style propagation to obtain some creative results.

French was chosen as the source of language style in this study because French has a rich literary heritage, and the cultural background it carries is different from Chinese and is quite recognizable. In terms of the choice of poetry form, modern Chinese has a richer vocabulary than ancient Chinese and has more words that can express modern and foreign concepts. Therefore, “Chinese Modern Poetry” written in modern Chinese was our first choice for presenting French-style Chinese poetry.

2.2. Background and Operation of Large Language Models (LLMs)

Large language models have revolutionized natural language processing by leveraging vast amounts of data and computational power to generate human-like text. These models, such as GPT-3, are built using transformer architectures, which allow the efficient processing and generation of text by capturing long-range dependencies in language [27]. The development and fine-tuning of LLMs involve several critical steps, each contributing to the model’s ability to understand and generate coherent and contextually appropriate text.

LLMs, such as GPT-3, are pre-trained on diverse and extensive datasets containing text from a wide range of sources, including books, articles, and websites. This pre-training phase enables the model to learn the statistical properties of a language, including grammar, syntax, semantics, and even some level of common-sense reasoning. The pre-trained model can then be fine-tuned for specific tasks, such as poetry generation, by training it on domain-specific datasets [28]. Fine-tuning adjusts the model’s parameters to better suit the nuances and styles required for the target application [29].

The operation of LLMs involves several key components and steps: (1) Pre-training: During the pre-training phase, the model is trained on a large corpus of text using unsupervised learning techniques. The primary objective is to minimize the difference between the predicted and actual next words in a sequence, a process known as language modeling. This step equips the model with a broad understanding of language [1,30]. (2) Fine-tuning: After pre-training, the model undergoes fine-tuning on a more specific dataset relevant to the desired application. For instance, to generate French-style modern Chinese poetry, the model is fine-tuned using a dataset comprising modern Chinese translations of works by French poets. This step allows for the model to capture specific styles, themes, and linguistic patterns [31]. (3) Human Feedback Collection: This step presents multiple samples generated by the model to human evaluators who label these samples based on their quality and adherence to the desired style. This feedback is used to create a reward model that predicts human preferences, guiding the model to generate higher-quality text [32]. (4) Reward Model Training: The reward model is trained using the feedback collected from human evaluators. This model assigns a score to each generated text sample, indicating its quality. The reward model helps in refining the text generation process by providing a quantitative measure of text quality [33]. (5) Reinforcement Learning: Using the reward model, the LLM is further fine-tuned. This step involves adjusting the model’s parameters to maximize the reward, ensuring that the generated text aligns closely with human preferences and desired stylistic attributes. By following these steps, LLMs can be optimized to generate high-quality and stylistically consistent text [34].

2.3. Reinforcement Learning from Human Feedback

Reinforcement Learning (RL) [35] is a widely used machine learning method in the field of artificial intelligence. Since it was proposed in 1961, in addition to poetry generation [14], reinforcement learning has been used in many fields such as medical care [36,37], industry [38], robots [39], and games [40,41]. Reinforcement learning does not rely on explicit instructions but learns through repeated interactions with the environment and can better cope with various uncertain environments. However, traditional reinforcement learning also has its limitations. First, the real environment contains complex changes, making it difficult to accurately define reward or penalty functions in advance. Second, in some scenarios for human users, the results learned by the computer may deviate from the user’s true intention. The new reinforcement learning method, Reinforcement Learning from Human Feedback (RLHF) [34], proposed in recent years introduces manual guidance into machine learning, which effectively overcomes the limitations of traditional reinforcement learning. RLHF learns reward models from human feedback such as ratings, corrections, and preferences in order to optimize reinforcement learning. Integrating human feedback into reinforcement learning is like a ship having a navigation satellite. The satellite can correct the ship that deviates from the route and provide an appropriate follow-up route, which can improve both efficiency and safety.

The connection between natural language and human beings is close. Once RLHF was proposed, it quickly became popular in the field of natural language processing. In recent years, RLHF has been used to fine-tune large language models [42,43] and generate book summaries [44] and has performed well in improving the quality of the output results and reducing dirty data and inappropriate content.

Unlike news, weather forecasts, or instructions, which have fixed formats and vocabulary categories, poetry, as an art form full of romance and imagination, has all-encompassing themes, rich content, and complex rhetoric and is full of symbols, metaphors, and allusions. Regardless of whether one uses Unsupervised Learning (UL), Supervised Learning (SL), or traditional reinforcement learning, it is difficult to achieve good performance in this highly open language generation. The new reinforcement learning method, RLHF, has excellent performance on tasks without fixed standards; so, we applied this method when fine-tuning the pre-trained model to enhance the expression of the French style in modern Chinese poetry.

3. Method

3.1. Model Architecture

Our model primarily operates in three steps: collecting human feedback, training reward models, and refining policy, as depicted in Figure 1. Initially, we begin with a pre-trained instance of GPT-3 and further enhance the model through supervised learning using a dataset containing modern Chinese translations of works by various French poets.

Figure 1. The model architeture. First, we collect dataset samples through human feedback and rank candidate outputs. In the second step, we train the reward model r using the defined loss. Lastly, the model utilizes Proximal Policy Optimization (PPO) to train policy $\pi$ with rewards from r.

3.2. Collecting Human Feedback

In the first stage, we presented four samples $\{y_0,y_1,y_2,y_3\}$ with respect to topic x to the labelers for selection. One sample $y_0$ was drawn from a dataset of modern Chinese translations of works by French poets, while the other three $\{y_1,y_2,y_3\}$ were generated using a language model as a strategy for acquiring these samples.

There were several reasons for using three y samples. First, having multiple generated samples helps capture the diversity of the model’s output. By presenting multiple samples $\{y_1,y_2,y_3\}$ generated by the language model, we could more comprehensively evaluate the model’s performance under different conditions. This diversity allowed for evaluators to see variations in the model’s output, thus making more reliable quality judgments. Second, multiple samples provide richer data and enhance the training effect of the model. In training reward models, more samples mean more comparison pairs and ranking information, which are crucial for model optimization. By comparing multiple samples, the model can better learn what kind of output meets expectations, thus improving the overall generation quality. Lastly, comparing multiple samples can reduce the bias caused by the randomness of a single sample.

We chose three y samples to strike a balance between diversity and manageability. Including more samples (e.g., four or five) can increase the complexity and cognitive load for evaluators, potentially leading to fatigue and less reliable judgments. Additionally, more samples would require significantly more computational resources and time for both the generation and evaluation processes. Therefore, the use of three samples ensured that we captured sufficient diversity without overwhelming the evaluators or overextending computational resources.

Human evaluators ranked these samples based on a set of guidelines, specifying the relative quality of each sample. These guidelines included aspects such as language fluency, poetic expression, grammatical accuracy, and adherence to the original French style. The evaluators’ feedback was stored in a human feedback dataset $S=\{x,y_0,y_1,y_2,y_3\}$, providing references for subsequent model optimization.

A graphical representation of the process is shown in Step 1 of Figure 1. In this way, we systematically collected high-quality human feedback, helping the model to progressively improve the quality and stylistic consistency of the generated poetry. The key to this stage was the subjective judgment of human evaluators, which filtered out the most fitting samples, thus providing valuable data for training the reward models.

3.3. Training the Reward Models

Next, we proceeded to train a reward model to anticipate the human-preferred outcomes. Specifically, samples gathered during the human feedback phase were utilized for training the reward model, aiming to forecast whether the new poetry samples resonated with the preferences of human labelers. An illustrative overview of this process is depicted in Step 2 of Figure 1.

In particular, we defined a reward function, denoted as $r:X\times Y\to \mathbb{R}$. This function was expected to assign higher values to completions ranked higher by human evaluators and lower values to those ranked lower, with respect to prompt x. Consequently, we optimized the following objective function:

$$loss\left(r\right)=-E_{\left(x,y_w,y_l\right)\sim S}\left[log\left(\sigma \left(r\left(x,y_w\right)-r\left(x,y_l\right)\right)\right)\right].$$

(1)

Here, $y_w$ represents the preferred completion out of the pair of $y_w$ and $y_l$, and S is the dataset of human comparisons. The loss function aims to maximize the probability that the reward model correctly identifies the higher-ranked completion $y_w$ over the lower-ranked completion $y_l$.

To train the reward model, we employed dataset S composed of pairs $(x,y_w,y_l)$ where each pair reflected human judgments on the relative quality of two samples. Reward function $r(x,y)$ was designed to predict the quality of sample y given topic x. Sigmoid function $\sigma$ ensured that the output was a probability, indicating how much more likely it was that $y_w$ was preferred over $y_l$.

The optimization process involved iteratively adjusting the parameters of the reward model to minimize the loss function defined above. The gradient descent algorithm is typically used for this purpose, allowing for us to find the parameter values that best align the model’s predictions with human preferences. The gradients of the loss function were computed, and this gradient was used to update the parameters of the reward model, thereby refining its ability to predict the human-preferred outcomes.

Through this training process, the reward model became increasingly adept at discerning the qualities that made the poetry samples resonate with human evaluators. This trained reward model was then used in the subsequent step to refine the policy of the language model, ensuring that the generated poetry aligned more closely with human preferences.

3.4. Refining the Policy with Reinforcement Learning

Finally, our aim was to fine-tune a large model (policy) denoted as $\pi$. Initially, this policy was set to $\pi =\rho$. We refined $\pi$ through reinforcement learning, utilizing the reward model from the preceding step to ensure its effectiveness in generating Chinese poems resembling the style of French poetry. To prevent the significant deviation of learned policy $\pi$ from initial policy $\rho$, we introduced regularization term $\beta KL(\pi ,\rho )$ as penalty. Thus, we conducted reinforcement learning on the reward model using the following equation:

$$R(x,y)=r(x,y)-\beta log\frac{\pi \left(y\right|x)}{\rho \left(y\right|x)}.$$

(2)

Here, hyperparameter $\beta$ controls the extent of deviation allowed for the large model. It serves as an entropy bonus, preventing the policy from straying too far from the effective range of the initial policy $\rho$, which is crucial to our style transfer task. We relied on the KL term to encourage coherence and thematic consistency, as we relied on the human evaluators to assess the style.

To optimize our policy, we utilized Proximal Policy Optimization (PPO), a popular algorithm in reinforcement learning known for its stability and efficiency. PPO works by iteratively improving the policy through gradient updates, ensuring that each update does not move the policy too far from the previous one, thus maintaining stability. The process of refining the policy involves several critical steps.

Policy ${\pi }_{\theta }$ is parameterized by $\theta$, determining the probability of taking action y given state x, denoted as ${\pi }_{\theta }\left(y\right|x)$. PPO uses advantage function ${\widehat{A}}_t$ to evaluate how much better taking action y in state x is compared to the average action. Advantage function ${\widehat{A}}_t$ plays a crucial role in this process. It measures the relative value of a particular action taken in a given state compared to the average value of all possible actions in that state. By using the advantage function, we provided a clearer signal for learning, emphasizing actions that led to better outcomes than expected. In our work, the advantage function is defined as

$${\widehat{A}}_t=R_t-V\left(s_t\right),$$

(3)

where $R_t$ is the actual reward and $V\left(s_t\right)$ is the state value function, representing the expected cumulative reward from state $s_t$.

Next, the probability ratio $r_t\left(\theta \right)$ was calculated as

$$r_t\left(\theta \right)=\frac{{\pi }_{\theta }\left(y\right|x)}{{\pi }_{{\theta }_{\mathrm{old}}}\left(y\right|x)}$$

(4)

Here, ${\pi }_{\theta }$ is the current policy, and ${\pi }_{{\theta }_{\mathrm{old}}}$ is the old policy. This ratio compares the probability of the action under the current policy to that under the old policy.

Then, PPO was designed to prevent large destabilizing updates by introducing mechanisms, including clipped probability ratios. To ensure stable policy updates, we define the clipped objective function as

$$L^{CLIP}\left(\theta \right)=E_t\left[min\left(r_t\left(\theta \right){\widehat{A}}_t,\mathrm{clip}\left(r_t\left(\theta \right),1-ϵ,1+ϵ\right){\widehat{A}}_t\right)\right],$$

(5)

where $\mathrm{clip}(r_t\left(\theta \right),1-ϵ,1+ϵ)$ clips probability ratio $r_t\left(\theta \right)$ to range $[1-ϵ,1+ϵ]$, and $ϵ$ is a small constant that controls the extent of the policy update.

Policy parameters $\theta$ were updated by maximizing the clipped objective function, typically using gradient ascent methods. This objective function balances between making significant updates when the advantage is large and making conservative updates when the change is too drastic.

The connection between the PPO process and the minimization of $R(x,y)$ can be understood as follows. By optimizing advantage function ${\widehat{A}}_t$, PPO indirectly works towards maximizing the expected cumulative reward, which corresponds to minimizing negative reward function $R(x,y)$. The clipped objective function ensures that the policy updates are stable and do not result in drastic changes. This stability is crucial for effectively minimizing $R(x,y)$, as it prevents the policy from oscillating or diverging. By using the advantage function and probability ratio, PPO balances exploration and exploitation, ensuring that the policy not only seeks immediate rewards but also explores better actions to minimize $R(x,y)$ in the long run. Thus, PPO optimizes the policy by using a clipped objective function to make stable updates, leveraging the advantage function to guide these updates; it effectively works towards minimizing reward function $R(x,y)$, aligning the policy with the goal of achieving optimal performance.

In our poetry generation task, PPO was used to optimize the model’s policy, ensuring it generated poems that aligned with the desired style. By repeatedly updating the policy using the PPO algorithm, guided by the reward model and regularization terms, we ensured the final policy produced high-quality stylistically consistent poetry.

In summary, the overall training process of our model is as follows:

• Collect dataset samples $(x,y_0,y_1,y_2,y_3)$ by gathering human feedback and candidate outputs $y_i$. Human labelers are asked to rank all four $y_i$.
• Train the reward model r using the loss defined in Equation (1) on the dataset samples obtained in the previous step.
• Utilize Proximal Policy Optimization (PPO) to train $\pi$, where the reward is defined by Equation (2).

4. Experiments

4.1. Datasets

We fine-tuned the 175 billion-parameter version of GPT-3 for poetry generation and evaluation on a widely available French poetry dataset. Additionally, we gathered a new dataset, which we used as a quality assessment system. This dataset consisted of modern Chinese translations [45,46,47,48] of works by French poets (comprising 108,000 characters), meticulously proofread and corrected. The poems in the dataset were categorized into seven types based on content and annotated word by word according to 22 grammatical categories. We considered both the original Chinese word classes and syntactic structures and the specific word classes and structures in foreign languages to effectively incorporate foreign language styles during model training. The distribution of poem types in the dataset is presented in Table 1.

Table 1. Distribution of poetry types in the dataset.

Dataset	Types of Poetry	Number	Percentage
Modern Chinese versions of the works of French poets (108,000 Chinese characters, 302 poems)	Lyric Poetry	96	31.8%
	Narrative Poem	48	15.9%
	Love Poetry	92	30.4%
	Philosophical Poem	31	10.3%
	Mourning Poem	7	2.3%
	Satirical poem	13	4.3%
	Pastoral poetry	15	5.0%

4.2. Human Data Collection

Our prompt dataset primarily comprised text prompts submitted to the GPT API. Each natural language prompt typically defines the task directly through natural language instructions (e.g., “Write a Chinese lyrical poem with a French style”); however, it may also be indirectly specified through a few examples (e.g., providing a summary of modern Chinese translations of works by French poets and then prompting the model to generate a poem) or implicit continuation (e.g., providing the beginning of a French poem).

To create our comparison dataset and conduct primary evaluations, we assembled a contract annotation team of approximately 10 individuals, all native Chinese speakers with a deep understanding of French literature and a strong appreciation for poetry. Our objective was to select annotators proficient in identifying poems in the French style. In each case, our annotators endeavored to grasp the user’s intent behind the prompt and evaluated the result based on their best judgment and the guidelines provided by us, as follows.

Guidelines: We asked the annotators to select their favorite poem based on the following criteria:

• Relevance to the theme;
• Consistency of content (logical coherence);
• Creativity;
• Form, vocabulary, syntax, figurative language, idioms, and allusions.

To effectively incorporate foreign language styles during model training, the annotations considered both the original Chinese word classes and syntactic structures and the specific word classes and structures in foreign languages.

5. Results and Evaluation

We used the following three methods to evaluate the generated results: human evaluation, poetic image coincidence overlap statistics, and syntactic feature proximity.

5.1. Human Evaluation

A total of 26 human evaluators participated in the human evaluation, all of whom were native Chinese speakers familiar with French literature and who had a good appreciation of poetry. In order to ensure the objectivity of the evaluation results, we mixed the Chinese versions of the real French poet’s works into the generated poems for evaluation as a reference group. The selected works were all poems of a poet who is not well-known in China to ensure that the evaluators had not read these works before. Therefore, the human evaluators did not know in advance which of the poems were generated poems and which were works of human poets.

To ensure a comprehensive evaluation, we compared our model with the state-of-the-art poetry generation method, SPG [49]. This advanced method incorporates mutual information to enhance the fluency and coherence of generated poems. The 70 poems evaluated included 20 poems generated by the pre-trained model without fine-tuning, 20 poems generated by the model after fine-tuning, 20 poems generated by SPG, and 10 Chinese versions of works by French poets. The average length of these 70 poems was 17.34 lines, with an average of 11.33 Chinese characters per line. The evaluators independently scored the 70 poems. The comparison of verses generated by the baseline pre-trained model and the fine-tuned pre-trained model is shown in Table 2.

Table 2. Examples of generated verses and their translations.

Model	Generated Verses	Generated Verses in English
Baseline Pre-trained Model	它神秘的闪光吸引着我， 我迷失在它的深处， 就像诗人迷失在他的诗句中， 或画家迷失在他的颜料中。	Its mysterious gleam draws me, I am lost in its depths, Like a poet lost in his verses, Or a painter lost in his paints.
	但是，唉！时间不会停留， 生命疾驰向它的终极命运， 而你，美丽的玫瑰，终将凋零， 像一颗流星，消失在深渊之中。	But, alas! Time will not stand still, Life is galloping towards its ultimate destiny, And thou, the beautiful rose, will fade away, Like a shooting star, disappearing into the abyss.
	让我们的诗句像尖锐的剑一样， 削短谎言和鄙视态度， 让我们的文字像反射镜一样， 展现真相，毫不妥协。	Let our verses, like a sharp sword, Cut short lies and scornful attitudes, Let our words, like a mirror, Reveal the truth without compromise.
	我记得我们快乐的时光， 我们儿时的笑声和游戏， 但现在你不在了， 我的心充满了悲痛。	I remember our happy times, Our childhood laughs and games, But now that thou art gone, My heart is filled with sorrow.
	我是一位灵魂的探险家， 在火焰的海洋上航行， 寻找存在的本质， 和永恒的意识本质。	I am an explorer of the soul, Sailing on the sea of fire, In search of the essence of existence, And the eternal essence of consciousness.
Fine-tuned Pre-trained Model	这是一个美丽、优雅、充满诗意的世界， 在这里，人们可以逃离现实， 生活似乎是一首交响曲， 每一刻都是神的礼物。	This is a beautiful, elegant and poetic world, Here, people can escape from reality, Life seems like a symphony, Each moment is a gift of god.
	我望着远方的马赛， 海风拂面带来了花香， 想象着和你一起， 漫步在蓝色海岸的小路上。	I look at Marseilles in the distance, The sea breeze brings the scent of flowers, I imagine being with thee, Stroll along the path of the Blue Coast.
	我仍然记得你， 你温柔的声音和明亮的眼睛， 但时间过得如此之快， 现在你已经如此遥远。	I still remember thee, Thy soft voice and bright eyes, But time flies so fast, Now thou art so far away.
	他们沉迷于自己的虚荣， 认为自己是社会的国王， 但他们不知道真正的伟大， 在于谦卑和善良。	Indulging in their vanity, They think they are the kings of society, But they don’t know that the true greatness, Lies in humility and kindness.
	我爱你，如同风爱大海， 没有止境，没有边际，没有阻隔， 我爱你，如同烈火一样， 毫不畏惧地燃烧，摧毁一切。	I love thee, as the wind loves the sea, No limit, no border, no barrier, I love thee, like the raging flames Burn undauntedly and destroy everything.

In the past, researchers often used the four-dimension evaluation method [8,11,22,23,24] (I. Fluency II. Coherence III. Meaning IV. Poeticness) and the six-dimension evaluation method [17,50,51] (I. How typical is the text as a poem? II. How understandable is it? III. How good is the language? IV. Does the text evoke mental images? V. Does the text evoke emotions? VI. How much does the subject like the text?) to evaluate generative poetry involving a single language. However, since the generated poetry in this experiment involved two languages—the language of the style source and the language used for the style presentation—we designed a seven-dimension evaluation method to evaluate the generated results in a more targeted manner. These seven dimensions were “grammar, logicality, rhetoric, creativity, depth, style, and human-likeness”.

The scoring of each perspective was presented on a five-point Likert Scale, and the standards are shown in Table 3.

Table 3. Evaluation Dimensions and Evaluation Criteria.

Grammar	Evaluation content	Fluency and grammatical correctness of the poem.
	Evaluation criteria	5 Excellent; 4 Good; 3 Average; 2 Below Average; 1 Poor.
Logicality	Evaluation content	Clarity of thinking, contextual coherence, lack of inconsistency.
	Evaluation criteria	5 Excellent; 4 Good; 3 Average; 2 Below Average; 1 Poor.
Rhetoric	Evaluation content	Language richness and rhetorical appropriateness.
	Evaluation criteria	5 Excellent; 4 Good; 3 Average; 2 Below Average; 1 Poor.
Creativity	Evaluation content	Uniqueness of expression, novelty of imagery, no clichés.
	Evaluation criteria	5 Excellent; 4 Good; 3 Average; 2 Below Average; 1 Poor.
Depth	Evaluation content	Inspiring and thought-provoking.
	Evaluation criteria	5 Excellent; 4 Good; 3 Average; 2 Below Average; 1 Poor.
Style	Evaluation content	Similarity to French poetry style.
	Evaluation criteria	5 Very High; 4 High; 3 Average; 2 Below Average; 1 Very Low.
Human- likeness	Evaluation content	Likelihood of the poem written by human.
	Evaluation criteria	5 Definitely; 4 Very Probably; 3 Probably; 2 Probably Not; 1 Definitely Not.

We obtained a total of 12,740 rating data from 26 judges on 70 poems in seven dimensions. The calculated average values of each dimension are shown in Figure 2. Among them, Human-written represents the score of the human poets’ works in this dimension, Model-2 represents the score of the poems generated by the original pre-trained model, and Model-1 represents the score of the poems generated by the fine-tuned pre-trained model.

Figure 2. Human evaluation results of poems generated by the baseline pre-trained model, poems generated by the fine-tuned pre-trained model, and human-written poems.

The results of Cronbach $\alpha$ and McDonald’s $\omega$ that are higher than 0.8 indicate that the reliability is very high, and results between 0.7 and 0.8 indicate that the reliability is high. As shown in Table 4, the Cronbach $\alpha$ index calculated from our human evaluation data was between 0.822 and 0.936, and the McDonald’s $\omega$ index was between 0.802 and 0.947, which shows that the human evaluation data were reliable.

Table 4. Reliability of human evaluation results.

Reliability	Grammar	Logicality	Rhetoric	Creativity	Depth	Style	Human-Likeness
Cronbach $\alpha$	0.936	0.876	0.886	0.896	0.895	0.822	0.916
McDonald’s $\omega$	0.947	0.891	0.890	0.883	0.895	0.802	0.915

Our comparison with the SPG demonstrated the effectiveness of our algorithm, showcasing superior performance across multiple dimensions. We can see from the average scores of the human evaluation that the poems generated by the fine-tuned pre-trained model reached above “average” (i.e., >3 points) in each evaluation dimension and were better than the baseline model in all evaluation dimensions. Among them, the “grammar” and “logicality” dimensions of increased by 0.05 points and 0.12 points, respectively. This may be because the baseline model already performed well on these basic aspects; so, there was less room for improvement. In the advanced dimensions “rhetoric” and “creativity”, the average scores increased by 0.29 points and 0.28 points, respectively. In terms of the high-order dimensions of “depth”, “style”, and “human-likeness”, the average scores increased by 0.35 points, 0.24 points, and 0.36 points, respectively. This shows that the fine-tuning of the pre-trained model helped to improve the performance of the generated poetry in terms of rhetoric, creativity, and depth. It significantly improved the approximation of the target language style and was closer to human-written poetry from the reader’s perspective. The fine-tuned pre-trained model generated poems whose human evaluation scores in the seven dimensions of grammar, logicality, rhetoric, creativity, depth, style, and human-likeness, respectively, reached 93.6%, 90.8%, 86.8%, 80.8%, 83.7%, 86.9%, and 88.7% of the scores of real poets’ works. It performed well in terms of grammar, logicality, rhetoric, style, and being human-like; however, it still lacked creativity and depth.

5.2. Poetic Image Evaluation

“Poetic imagery” has strong cultural and regional characteristics and has an important influence on poetry style. Common images in poetry vary from language to language. France has been deeply influenced by Christian culture in history, and religious and mythological images such as “Dieu (God)”, “paradis (heaven)”, “enfer (hell)”, “ange (angel)”, and “diable (devil)” often appear in French but are rarely encountered in Chinese poetry. Regional flora and fauna often serve as important images in poetry, presenting a unique cultural beauty. For example, the unique image expression “青纱帐 (green curtain)” appears in Chinese poetry, which is used to represent the large areas of tall and dense crops in the fields; as shown in Table 5, in the poem “L’Adieu” by French poet Guillaume Apollinaire, a kind of cold-resistant European flower “bruyère (heather)” is used as the main poetic image to express loneliness and sadness and to enhance the atmosphere of parting.

Table 5. Example of poetic images in French poetry.

Verses from “L’Adieu”	Translation in English
J’ai cueilli ce brin de bruyère	I picked this sprig of heather
L’automne est morte souviens-t’en	Autumn is dead remember it
Nous ne nous verrons plus sur terre	We will no longer see each other on earth
Odeur du temps brin de bruyère	Smell of time sprig of heather
Et souviens-toi que je t’attends	And remember that I’m waiting for you

Since poetic imagery is an important component of poetic style, we gathered statistics on the poetic image distribution of the generated poetry to evaluate the degree of presentation of the language style in the generated poetry. As a reference group, we counted the top 50 high-frequency poetic images in 200 works by 80 modern Chinese poets and the top 50 high-frequency poetic images in 302 works by 18 famous French poets. Then, we used the baseline model and the fine-tuned pre-trained model to generate 50 poems, respectively, and we counted the top 50 high-frequency poetic images. Finally, we calculated the consistency of the top 10 to top 50 high-frequency poetic images between Chinese poetry and French poetry, French poetry and the poetry generated by the baseline model, and French poetry and the poetry generated by the fine-tuned pre-trained model. The calculation results of the poetic image consistency are shown in Figure 3, in which the poetic image consistency between Chinese poetry and French poetry is represented by CR, the poetic image consistency between French poetry and the poetry generated by the baseline model is represented by GR1, and the poetic image consistency between French poetry and the poetry generated by the fine-tuned pre-trained model is represented by GR2.

Figure 3. Consistency of high-frequency poetic images between Chinese poetry and French poetry, French poetry and poetry generated by the baseline model, French poetry and poetry generated by the fine-tuned pre-trained model.

According to Figure 3, the CR was between 30% and 53.5%, the GR1 was between 46% and 55%, and the GR2 was between 40% and 60%. Among them, the GR1 was higher than the CR at three points: TOP10, TOP20, and TOP40, while the GR2 was higher than the CR at every point (i.e., TOP10, TOP20, TOP30, TOP40, and TOP50). In addition, except for the GR2 being lower than and equal to GR1 at the TOP10 and TOP20 points, respectively, it was higher than the GR1 at the TOP30-50 points. Compared with the baseline model, the GR2 improved in these by 20%, 4.8% and 8.7%, respectively. This shows that fine-tuning training can effectively enhance the presentation of the French style in generated poetry.

Figure 4a–c shows the distribution of the high-frequency images of Chinese poetry, French poetry, and French-style Chinese poetry generated by the fine-tuned pre-trained model.

Figure 4. Distribution of high-frequency poetic images in French poetry, Chinese poetry, and poetry generated by fine-tuned pre-trained model. (a) Distribution of top 30 high-frequency poetic images; (b) Distribution of top 40 high-frequency poetic images; (c) Distribution of top 50 high-frequency poetic images.

5.3. Syntactic Feature Evaluation

The style of language is not only defined by specific meanings, but it is also closely related to syntactic structure. When appreciating some translated literary, film, and television works, even if the content is ordinary and there is no semantic expression with clear content differences, we can still distinguish the original language of the sentence through sentence structure. The translation method that significantly transplants the syntactic structure of the foreign language is often called “translationese”. In the field of poetry translation, the “translationese” of excessively copying foreign syntax makes the sentences clumsy, eccentric, and difficult to understand. However, if the scale of the “translationese” can be grasped well, it can express the exotic style without damaging the meaning, the readability, and the coherence of the verses.

Based on the above, we identified the following three important characteristics of French that differ from Chinese in terms of the frequency of specific types of vocabulary.

I. In French, personal pronouns generally cannot be omitted; however, in Chinese, personal pronouns are often omitted, as shown in Table 6.

Table 6. Examples of the occurrence and omission of personal pronouns in French poetry and Chinese poetry.

Language	Example	Translation in English
French	En vain j’ai voulu de l’espace	In vain I wanted space
	Trouver la fin et le milieu;	To find the end and the centre;
	Sous je ne sais quel oeil de feu	Under I don’t know what fiery eye
	Je sens mon aile qui se casse	I feel my wing breaking
	— Charles Baudelaire, Les Plaintes d’un Icare	— Charles Baudelaire, Lament of An Icarus
Chinese	为期待而绝望	(We’re) despairing for the sake of expectation
	为绝望而期待	(We’re) hoping for the sake of despair
	期待是最漫长的绝望	Expectation is the longest despair
	绝望是最完美的期待	Despair is the most perfect expectation
	—— 杨炼《偈子》	— Yang Lian, Gatha

II. Nouns in French usually appear together with articles. Most singular things are preceded by an indefinite article (French: un/une) indicating the number one, making “one” appear more frequently. In Chinese, in addition to numerals, there are often quantifiers involved in modifying the nouns. When the numeral before the noun is one (Chinese: 一), the numeral is often omitted, and only the quantifier is retained; so, the frequency of “one” is lower, as shown in Table 7.

Table 7. Examples of the occurrence and omission of numeral “one” in French poetry and Chinese poetry.

Language	Example	Translation in English
French	Un air chaud me frappait,	A hot air hit me,
	comme un souffle de forge,	Like a breath of forge,
	A chacun des soupirs qui soulevaient sa gorge.	With each sigh that raised his throat
	— Guy de Maupassant, Au bord de l’eau	— Guy de Maupassant, The Water’s Edge
Chinese	也许，人活到了这般年岁	Perhaps, when a person lives to this age
	就能够听到——时间	(He) can hear — time
	——它就像是个屠夫	— It is like the sound of (a) butcher
	在暗地里不停地磨刀子的声音	Sharpening his knife in the dark
	—— 芒克《晚年》	— Mang Ke, Twilight Years

III. The second-person singular honorific vous is used more frequently in French poetry. In comparison, the second-person singular honorific 您 and 君 appear less frequently in Chinese poems written in modern Chinese, as shown in Table 8.

Table 8. Examples of the use of second-person singular honorifics in French poetry and Chinese poetry.

Language	Example	Translation in English
French	Vous avez un regard singulier et charmant;	You have a singular and charming look;
	Comme la lune au fond du lac qui la reflète,	Like the moon at the bottom of the lake which reflects it.
	Votre prunelle, où brille une humide paillette,	Your pupil, where a moist glitter shines,
	Au coin de vos doux yeux roule languissamment	In the corner of your sweet eyes rolls languidly
	— Théophile Gautier, A deux beaux yeux	— Théophile Gautier, To Two Beautiful Eyes
Chinese	你站在桥上看风景，	Thou stand on the bridge and look at the scenery,
	看风景人在楼上看你。	People watching the scenery are watching thee upstairs.
	明月装饰了你的窗子，	The bright moon decorates thy window,
	你装饰了别人的梦。	Thou decorate someone else’s dream.
	——卞之琳《断章》	— Bian Zhilin, Fragments of Chapter

As can be seen in Figure 5, the poems generated by the fine-tuned model had an increase in the frequency of personal pronouns, the frequency of the numeral “一(a/an/one)”, and the frequency of second-person honorifics. Compared with the baseline model, the frequency of these three features increased by 12.95%, 15.81%, and 284.44% per thousand characters, respectively. Therefore, the fine-tuned poetry generation model was closer to the French language style in these details without affecting the semantic expression of modern Chinese.

Figure 5. Proportion of personal pronoun, numeral “一”, and second-person singular honorifics per 1000 Chinese characters in poems generated by the baseline pre-trained model and in poems generated by the fine-tuned pre-trained model.

6. Conclusions

In this paper, we proposed a method to fine-tune a pre-trained model in a targeted manner to automatically generate French-style modern Chinese poetry and conduct a multi-faceted evaluation of the generated results. In a human evaluation using a five-point scale, judges assigned scores between 3.29 and 3.93 in the seven dimensions of grammar, logicality, rhetoric, creativity, depth, style, and being human-like, which reached 80.8–93.6% of the scores of Chinese versions of real French poetry in these dimensions. In terms of the high-frequency poetic imagery, the consistency of the top 30–50 high-frequency poetic images between the poetry generated by the fine-tuned model and the French poetry reached 50–60%. In terms of the syntactic features, compared with the poems generated by the baseline model, the distribution frequencies of three special types of words that appear relatively frequently in French poetry increased by 12.95%, 15.81%, and 284.44% per 1000 Chinese characters in the poetry generated by the fine-tuned model. The human evaluation, poetic image distribution, and syntactic feature statistics show that the targeted fine-tuning of the pre-trained model is helpful for the spread of the language style. This fine-tuned model can successfully generate modern Chinese poetry in a French style.

Through human evaluation, we found that the model performed well in basic dimensions, such as grammar and logicality; however, it still had a large gap with real poets’ works in terms of creativity and depth. In the future, we will continue studying ways to improve the performance of the model in these weak aspects.