Pashto poetry generation: deep learning with pre-trained transformers for low-resource languages

Development of dataset

We rolled up our sleeves to explore Pashto poetry and put together a bunch of beautiful poems ourselves. We gathered these poems from more than 20 different books, written by famous Pashto poets like Abaseen Yousafzai, Shaukat Ali Shaukat, and Hussain Ahmad Sadiq. In our collection, there is a grand total of 1,800 heartfelt ghazals. These ghazals are like emotional paintings, reflecting the diversity and depth of Pashto poetry.

Imagine each poem as a little story, told in Pashto. Some are about love, while others talk about important things happening in society and history. In these lines, there’s a mix of old traditions and new ideas. Through our collection, we aim to understand how Pashto poetry has changed and what parts of its history it holds onto.

Our collection isn’t just a pile of poems—it is a way to celebrate the talented poets who’ve given Pashto its poetic colors. Each ghazal is like a piece of art, reminding us of the language, culture, and feelings that make Pashto poetry special. By sharing these poems, we’re inviting everyone to join us on a journey into the world of Pashto poetic wonders.

As shown in Fig. 1, the Pashto poetry example highlights the profound cultural significance of Pashto literature and the richness of its poetic tradition. This particular verse exemplifies the theme of love and devotion, commonly found in Pashto poetry. Additionally, it underscores the timeless nature of these poetic expressions, as they continue to resonate with Pashto-speaking audiences across generations.

As shown in Fig. 2, the following methodology to generate praise poems consists of four phases. In this section, these phases are discussed in more detail.

Figure 2 visually represents these phases, demonstrating the flow of the research methodology and its integral components. This structured approach ensures the validity and effectiveness of the research process, ultimately leading to valuable outcomes.

Data pre-processing

In this phase, data preprocessing is performed by removing or modifying data that is incorrect, incomplete, irrelevant, or duplicate. We also removed unrelated characters or symbols, such as punctuation marks, line breaks, $, #, brackets, parentheses, and other unrelated characters from the dataset. Table 2 shows an example of the input before and after processing. During the data cleaning process for Pashto poetry, we used regular expressions (regex) in Python to identify and remove specific elements that did not conform to the desired format of the poetry. Here are a few examples of regular expressions commonly used in cleaning Pashto poetry data.

Table 2:

Example of the pre-processing step.

Before processing	After processing

DOI: 10.7717/peerj-cs.2163/table-2

Removing non-Pashto characters: A regular expression can be used to match and eliminate any non-Pashto characters, such as punctuation marks, numbers, or special symbols. For example, the expression ‘[^\u0600-\u06FF\s]’ matches any character outside the range of Pashto Unicode characters and whitespace. By replacing the matched characters with an empty string, the data can be cleaned by removing the unwanted elements.

Handling line breaks and extra whitespace: Pashto poetry often contains line breaks and extra whitespace that can affect the readability and format. Regular expressions can be used to address these issues. For example, the expression ‘\n+’ matches one or more consecutive line breaks, which can be replaced with a single line break or a whitespace character. Similarly, the expression ‘\s+’ matches one or more consecutive whitespace characters, which can be replaced with a single space.

By applying appropriate regular expressions, we were able to successfully clean the data and convert it into a specific format suitable for Pashto poetry. These expressions helped remove non-Pashto elements, address line breaks and whitespace issues, and ensure that the data adhered to the desired format of poetry. The cleaning process ensured that the resulting dataset was more coherent, standardized, and ready for subsequent analysis, such as training machine learning models for Pashto poetry generation.

Table 2 illustrates an example of the crucial data pre-processing step in the research methodology. In this phase, the raw text data undergoes several transformations to make it suitable for analysis. The ‘Before processing’ column displays the original, unprocessed text, which may contain various noise elements such as special characters, punctuation, and extra spaces.

In contrast, the ‘After processing’ column presents the text after undergoing pre-processing. During this step, unnecessary characters, symbols, and irregularities are removed or corrected, resulting in clean and structured data that can be effectively used for subsequent analysis tasks. Data pre-processing is a fundamental step in ensuring the quality and accuracy of research outcomes.

Proposed methods

In this work, we explore two approaches for Pashto poetry generation using pre-trained models. Both approaches are instruction-based, meaning that the models are given instructions on how to generate poetry. These two pre-trained transformer models are, LaMini-Cerebras-590M and Bloomz-560m. Each model has unique capabilities and characteristics that contribute to the creative generation of Pashto poetry. After an extensive evaluation, we selected two models to proceed with further fine-tuning and finalization for our project.

Table 3 provides a detailed comparison of two proposed models, BLOOMZ-560M and LaMini-Cerebras-590M, with respect to various key characteristics.

Big science/bloomz-560m

Big science/bloomz-560m is a large language model developed by Big Science, a distributed research project to train and deploy large language models. The model was trained on a massive dataset of text and code and can follow human instructions in dozens of languages zero-shot. To train the model, Bigscience used a process called multitask finetuning (xP3). This involves training the model in a variety of tasks, such as translation, text generation summarization, and question answering. This helps the model to learn to generalize to new tasks and languages. The resulting model has 560 million parameters and is one of the largest and most powerful language models in the world.

Figure 3 depicts a high-level overview of BLOOM, a large language model (LLM) released in 2023 by Hugging Face. BLOOM is a 176B parameter model, making it one of the largest and most powerful LLMs available to the public.

Brief explanation of the key components of Fig. 3.

1. Tokenizer: The tokenizer is responsible for breaking down text into individual tokens, which can be words, subwords, or other units.

2. Encoder: The encoder takes the tokens from the tokenizer and converts them into a sequence of hidden states. Each hidden state represents the model’s understanding of the token and its context.

3. Decoder: The decoder takes the hidden states from the encoder and generates text one token at a time.

4. Attention: Attention is a mechanism that allows the model to focus on specific parts of the input sequence when generating output.

The arrows in the image represent the flow of data through the model. Text is first tokenized and then passed to the encoder. The encoder generates a sequence of hidden states, which are then passed to the decoder. The decoder generates text one token at a time, using attention to focus on specific parts of the input sequence.

The Big science/bloomz-560m model was trained for Pashto poetry using Google Collab. The first step was to format the dataset into JSON format. The model was then trained using a set of parameters, including BATCH_SIZE = 128: This is the number of examples that are processed at a time. warmup_steps = 100: This is the number of steps that are used to gradually increase the learning rate. max_steps=300: This is the maximum number of steps that the model is trained for. learning_rate = 3e−4: This is the learning rate, which is a measure of how quickly the model is updated. fp16 = True: This flag tells the model to use 16-bit floats instead of 32-bit floats. This can help to improve performance. logging_steps = 10: This parameter specifies how often the model’s progress is logged. After the model was trained, it was fine-tuned using a different set of parameters.

The fine-tuning parameters included: num_train_epochs = 10: This is the number of epochs that the model is fine-tuned for. weight_decay = 0.01: This is a regularization parameter that helps to prevent the model from overfitting. The fine-tuning process helped the model to learn the specific features of Pashto poetry.

As a result, the model was able to generate more creative and accurate Pashto poetry. Here is a brief explanation of the terminology used in the training process: JSON: A format for storing data in text files. Transformer: A neural network architecture that has been shown to be very effective for natural language processing tasks. Epoch: A complete pass through the training data. Batch size: The number of examples that are processed at a time. Learning rate: A measure of how quickly the model is updated. Regularization: A technique used to prevent the model from overfitting. Overfitting: A problem that occurs when the model learns the training data too well and is unable to generalize to new data. Fine-tuning: A process of adjusting the model’s parameters to improve its performance on a specific task. After training the model it takes almost 40 min to train and fine-tuned for our data, they output of the train model is TrainOutput (global_step = 300, training_loss = 1.5561199736595155, metri = {‘train_runtime’: 2365.3552, ‘train_smple_per_second’: 8.117, ‘train_steps_per_second’: 0.127, ‘total_flos’: 8910943530516480.0, ‘train_loss’: 1.5561199736595155, ‘epoch’: 10.85}) parameters provided for generating poetry: max_length = 300: This parameter sets the maximum length of the generated poetry, in terms of the number of tokens (words or characters, depending on the model). do_sample = True: This parameter indicates whether to use sampling when generating poetry. Sampling refers to the process of randomly selecting the next token from a probability distribution, rather than always selecting the most likely token. top_k = 200: This parameter limits the number of possible next tokens to consider during sampling to the top k most likely tokens, where k is the value specified. top_p = 0.91: This parameter limits the number of possible next tokens to consider during sampling based on their cumulative probability, such that the sum of probabilities of all tokens considered is less than or equal top_p, where p is the value specified. temperature = 0.7: This parameter controls the “creativity” of the generated poetry by adjusting the randomness of the sampling process. A higher temperature value will result in more random and potentially creative output, while a lower temperature value will result in more conservative and predictable output.

Figure 4 displays the training loss of the Big Science/BLOOMZ-560M model during the training process. The y-axis represents the loss, which begins at four and gradually decreases as the model learns and converges. This loss reduction is indicative of the model’s improvement in minimizing prediction errors over time.

The x-axis corresponds to the max_steps parameter, a key training argument in the Transformers library, set to a value of 300. The max_steps parameter determines the maximum number of training steps or iterations the model undergoes during the training process. It is a crucial parameter in fine-tuning models like Big Science/BLOOMZ-560M, influencing the training duration and performance.

Analyzing the training loss curve in Fig. 4 provides insights into the model’s convergence and learning dynamics. A decreasing loss indicates that the model is progressively improving its ability to make accurate predictions and understand the underlying patterns in the training data.

Figure 5 provides an overview of the training process for the Big Science/BLOOMZ-560M model in terms of epochs and the max_steps parameter. The y-axis represents the epoch count, indicating how many complete passes the model has made through the training data. The x-axis represents the max_steps parameter, which determines the maximum number of training steps per epoch.

Training deep learning models, such as Big Science/BLOOMZ-560M, is often organized into epochs, where each epoch involves iterating over the entire training dataset once. The max_steps parameter influences the number of training steps within each epoch and, consequently, the training duration. Table 4 An Example of the Input-Output of Big Science/BLOOMZ-560M (Generated Pashto Poetry by Big Science/BLOOMZ-560M Model).

Table 4:

An example of the input-output of Bigscience/bloomz-560m.

An input sentence	An output sentence

DOI: 10.7717/peerj-cs.2163/table-4

Table 5 provides a summary of the key hyperparameters used during the fine-tuning process of the Big Science/BLOOMZ-560M model.

Table 6 provides an overview of the key specifications and architecture details of the Imran1/bloom_pg model.

MBZUAI/LaMini-Cerebras-590M

LaMini-Cerebras-590M is a large language model (LLM) developed by the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). It is a finetuned version of the Cerebras-GPT-590M model, which was trained on. The MBZUAI/LaMini-Cerebras-590M model is a large language model that was trained on a massive dataset of Pashto text. The model was trained using the Cerebras Wafer-Scale Engine, which is a supercomputer that is specifically designed for training large language models. The MBZUAI/LaMini-Cerebras-590M model can be used to generate Pashto poetry.

The Fig. 6 architecture, of MBZUAI/LaMini-Cerebras-590M model, which is a fine-tuned version of the Cerebras-GPT-590M model on the LaMini-instruction dataset. The LaMini-instruction dataset contains 2.58M samples for instruction fine-tuning, which allows the model to learn to perform a variety of tasks, such as generating different creative text formats of text content, translating languages, writing different kinds of creative content, and answering questions in an informative way.

The MBZUAI/LaMini-Cerebras-590M architecture is a decoder-only model, which means that it does not have an encoder. This is because the model is designed to be used for instruction following tasks, where the input is a natural language instruction, and the output is the desired response.

The architecture of the MBZUAI/LaMini-Cerebras-590M model is as follows:

1. Input: The input to the model is a natural language instruction.

2. Tokenizer: The tokenizer breaks down the instruction into individual tokens, which can be words, subwords, or other units.

3. Transformer decoder: The Transformer decoder takes the tokens from the tokenizer and generates a sequence of hidden states. Each hidden state represents the model’s understanding of the token and its context.

4. Output: The output of the model is the generated response to the instruction.

The Transformer decoder is a stack of self-attention layers. Self-attention is a mechanism that allows the model to focus on specific parts of the input sequence when generating output. This is important because it allows the model to understand the context of the instruction and generate a response that is relevant and informative.

The model can be used to generate a variety of Pashto poetry, including ghazals, quatrains, and sonnets. The model can also be used to generate Pashto poetry on a variety of topics, such as love, loss, and nature. The MBZUAI/LaMini-Cerebras-590M model is a valuable tool for generating Pashto poetry. The model can be used to generate creative and accurate Pashto poetry. The model is also available on the Hugging Face Hub, which makes it easy to use.

The MBZUAI/LaMini-Cerebras-590M model was fine-tuned for Pashto poetry generation using a JSON format dataset of Pashto poetry. The model was fine-tunegradient_accumulation_steps = GRADIENT_ACCUMULATION_STEPS: This is the number of times that the gradients are accumulated before they are used to update the model’s parameters. This can help to improve performance. warmup_steps = 100: This is the number of steps that are used to gradually increase the learning rate. This can help to prevent the model from overfitting. max_steps = 260: This is the maximum number of steps that the model is fine-tuned for. learning_rate = 2e−4: This is the learning rate, which is a measure of how quickly the model is updated. fp16 = True: This flag tells the model to use 16-bit floats instead of 32-bit floats. This can help to improve performance. logging_steps = 10: This parameter specifies how often the model’s progress is logged. The fine-tuning process took approximately 40 min to complete. The model was able to generate Pashto poetry that was both creative and accurate. After training the output of the train model shows that the model was trained for 260 steps, and that the loss on the training data was 0.8195405574945304. The training process took 1,708.3726 s, and the model processed 9.74 training samples per second. The model performed 0.152 training steps per second, and it performed a total of 1.3027304839053312e+16 floating-point operations during training. The model was trained for 9.4 epochs.

Figure 7 displays the training loss of the LaMini-Cerebras model during the training process. The y-axis represents the loss, which begins at 2.3 and gradually decreases as the model learns and converges. This loss reduction is indicative of the model’s improvement in minimizing prediction errors over time.

The x-axis corresponds to the max_steps parameter, a key training argument in the Transformers library, set to a value of 250. The max_steps parameter determines the maximum number of training steps or iterations the model undergoes during the training process. It is a crucial parameter in fine-tuning models like LaMini-Cerebras, influencing the training duration and performance.

Figure 8 provides an overview of the training process for the LaMini-Cerebras model in terms of epochs and the max_steps parameter. The y-axis represents the epoch count, indicating how many complete passes the model has made through the training data. The x-axis represents the max_steps parameter, which determines the maximum number of training steps per epoch.

Training deep learning models, such as LaMini-Cerebras is often organized into epochs, where each epoch involves iterating over the entire training dataset once. The max_steps parameter influences the number of training steps within each epoch and, consequently, the training duration. An example of the input output and the generated Pashto poetry by LaMini-Cerebras Model are shown in Table 7 below.

Table 7:

An example of the input-output of LaMini-Cerebras model.

An input sentence	An output sentence

DOI: 10.7717/peerj-cs.2163/table-7

Table 8 provides a summary of the key hyperparameters used during the fine-tuning process of the LaMini-Cerebras 560M model.

Table 9 provides an overview of the key specifications and architecture details of the Imran1/cerebus_pg model.

Quantitative evaluation

In this study, we used BLEU scores to evaluate the quality of the generated Pashto poetry, using the default parameter of 4-grams. BLEU scores are a commonly used metric to measure the similarity between the generated text and the reference text, with a higher score indicating a greater level of similarity.

$$\mathrm{B}\mathrm{L}\mathrm{E}\mathrm{U}=\mathrm{B}\mathrm{P}\cdot (\sum _{n=1}^N{w}_n\log p_n)$$

In Tables 8 and 9, we present a sample of the outputs generated by both the bloom_pg and cerebras_pg models. Our experimental results showed that both models performed well in generating new Pashto poetry that adhered to the style and tone of the poetry in our training dataset. The generated poetry was presented in a poetic format, with proper structure and rhythm, and demonstrated the models’ ability to capture the underlying patterns and structures of Pashto poetry. These results are promising and demonstrate the potential for using machine and deep learning techniques to generate high-quality Pashto poetry.

Both the Imran1/bloom_pg and Imran1/cerebus_pg models have demonstrated their ability to generate high-quality Pashto poetry. As Table 10 illustrates, the generated Pashto poems exhibit a rich and expressive use of the language, capturing the essence of various themes and emotions. Furthermore, we have provided manual English translations for these poems, ensuring accurate and contextually meaningful renditions, allowing for a cross-linguistic appreciation of their poetic beauty and meaning. Table 11 presents the mean BLEU-Default scores for both models. Our observation indicates that both models performed well and were able to generate novel words that preserve the tone and style of Pashto poetry written by poets.

Table 10:

Models output.

Model name	The generated poetry in Pashto	The generated poetry in English
Imran1/bloom_pg		Round and round beauty captivates, but blood draws pain O beloved, I sacrifice myself for you, every wish captures hearts I have cherished everything, not leaving a single part behind You bought a gem, we would seize all the rivers Such are the wise ones who crafted explosives Such are the wise ones who captured bombs in flames My intention is not that you become a killer in vain But dying, your hands would grasp your home
Imran1/cerebus_pg		Neither through rank nor through gold I am a merchant of love This is my only capital That I am infamous in all lands With no boundaries I wander on such paths That soar up to the sky I fly on the wings of such thoughts

DOI: 10.7717/peerj-cs.2163/table-10

Qualitative evaluation

Table 12 shows the results of the human evaluation of the Imran1/cerebus_pg and Imran1/bloom_pg models based on various criteria, including meaning, coherence, rhyme, and rhythm. The scores reflect the assessment of these models in generating Pashto poetry that aligns with these criteria.

The BLEU scores alone are insufficient to evaluate the quality of generated poems, as they do not consider important factors such as meaning and coherence. Therefore, human evaluation is necessary in this field of research. However, human evaluation can be challenging due to different opinions and preferences in poetry. In this study, we randomly selected two poems from each model and asked two experts to evaluate them based on four criteria: meaning, coherence, rhyme, and rhythm. Each criterion was scored from zero to one, with zero being the worst and one the highest score. The results of the human evaluation are presented in Table 12. As shown in the table, the Bigscience/Bloomz-560m model received higher scores in terms of meaning, coherence, and rhyme compared to the other model that is LaMini-Cerebras-590M. These findings indicate that the Bigscience/Bloomz-560m model is better at generating Pashto poetry that is not only similar to the training dataset but also the generated poetry is meaningful, coherent and have interconnectivity. The comparison of the human evaluation of the two models is also shown as bar chart in Fig. 9.