Writing Creative Poems With Deep Learning And NLP

“There is no rule on how to write. Sometimes it comes easily and perfectly: sometimes it’s like drilling rock and then blasting it out with charges” — Ernest Hemingway

In line with Ernest Hemingway's observation that writing can either flow smoothly or require immense effort, this project harnesses the power of Natural Language Processing (NLP) and Deep Learning to craft beautiful poems with precision and creativity. The system combines the elegance of human language with advanced neural network architectures, designed to understand linguistic patterns, structures, and emotions. By leveraging large-scale data and modern machine learning algorithms, it generates poetry that is both emotionally resonant and technically refined, embodying the art of writing while reducing the cognitive strain associated with it. This system not only automates the creative process but also captures the nuances of poetic expression, offering a seamless fusion of technology and artistry.

Libraries Used

• Tensorflow
• Keras
• Numpy
• Pandas
• Matplotlib
• Seaborn
• Sklearn
• Spacy
• BeautifulSoup
• Wordcloud

Model Details

For generating poetry, I have developed a sophisticated deep learning model that leverages the power of Long Short-Term Memory (LSTM) layers, which are particularly effective in capturing and retaining long-term dependencies within sequences of text. This is crucial for poetry generation, as it enables the model to understand and preserve the flow and meaning across different sentences or lines of a poem.

The architecture begins with an Embedding layer, which converts the input data into a dense representation, transforming words or tokens into vectors that can be more easily understood by the model. Following this, two LSTM layers are used sequentially. The first LSTM layer outputs a vector of size 100 from an input of 50 dimensions, capturing the contextual relationships between words in the sequence. The second LSTM further refines this representation, enabling the model to learn deeper patterns and dependencies.

After the LSTM layers, the architecture incorporates two Dense layers. The first Dense layer, using the ReLU (Rectified Linear Unit) activation function, helps in learning non-linear relationships, introducing complexity and ensuring that the model can handle intricate patterns in poetic structure. Batch Normalization is applied next, which normalizes the output from the Dense layer, improving the model’s learning efficiency and stability. Finally, the last Dense layer utilizes a softmax activation function, providing a probability distribution over the output space, which is crucial for selecting the next word or token in the poem.

This carefully designed architecture allows the model to balance creativity with structure, generating poems that are not only aesthetically pleasing but also coherent and contextually meaningful.

Model Traning

The model was trained over the course of 170 epochs, allowing it to gradually refine its ability to generate poetry. During the training process, the Adam optimizer was employed, which is well-regarded for its efficiency in handling sparse gradients and its adaptability in adjusting learning rates. This choice of optimizer ensures that the model converges more quickly and effectively toward minimizing the error, even in complex, high-dimensional spaces like those encountered in natural language generation.

Additionally, the model utilizes categorical cross-entropy as its loss function. This function is particularly suited for multi-class classification problems, such as predicting the next word in a sequence from a large vocabulary. By employing categorical cross-entropy, the model is penalized more severely when it makes incorrect predictions, particularly when the predicted probability distribution diverges significantly from the actual label. This encourages the model to become more precise in its predictions, thus improving its overall accuracy in generating coherent and meaningful poetic verses.

Model Analysis

After training the model for 170 epochs, it has demonstrated exceptional performance, achieving an accuracy of 95% and reducing the loss to a remarkable 0.1782. The learning curves illustrate steady and consistent progress throughout the training process.

In the loss graph, the initial value starts high at over 7, and we observe a sharp decline during the first 50 epochs, indicating effective learning. As training continues, the loss gradually plateaus, maintaining a very low level, which suggests that the model is not only learning well but also generalizing effectively with minimal overfitting.

The accuracy graph, on the other hand, shows a progressive increase, where accuracy quickly rises in the first few epochs, crossing the 60% mark by epoch 50. By epoch 170, the accuracy stabilizes near 95%, demonstrating the model's strong predictive capabilities. The gradual improvement in accuracy with minimal fluctuations further reinforces the stability and reliability of the model over time.

In summary, these metrics and visualizations reflect a well-optimized model, capable of achieving high accuracy while maintaining a very low loss, indicative of efficient learning and excellent generalization.

Beautiful Poems (Model's Ouput)

Title: Friendship

an worry as we send instead of my purpose an
irish trunk tshirts when my hand back i found my
cousin milton worked for a cable company the boy i
have become the person who says darling who says sugarpie
metaphor maybe nothing to do but brush back the tears

Title: Love

of us bear that have so near his message is
anything central orchards flung out on the land urban forests
or whose hollows was mean visited up there still sister
and you are both let me be no one and
yesterday something shattering happened not yesterday but several that is
anything central orchards flung out on the land urban forests

Title: Life

passes through places pj duffy landscapes of south ulster patrick
amount soul lady certainly for night give well historical the
cat has the chance to make the sunlight beautiful to
imagine the next weeping of the next ear you see
everything looped spiraled circular thought but the labyrinths not a

Title: Sadness

i believe for love my own name love and warn
of time and even whatever i know an grant you
are running away from everyone who loves you from your
husband is stretched out on the ground as if he
brings me chocolate from the pentagon dark chocolates shaped like

Conclusion

In conclusion, this project successfully integrates the artistry of poetry with the precision of deep learning, resulting in a system that generates poems with both creativity and coherence. By leveraging advanced NLP techniques and LSTM-based architectures, the model captures the nuances of language and poetic structure, producing emotionally resonant and technically refined verses. The model's training over 170 epochs, with the use of efficient optimizers and loss functions, has yielded impressive results, achieving 95% accuracy and a low loss of 0.1782. These metrics, coupled with the smooth learning curves, demonstrate the system's ability to balance artistic expression with algorithmic efficiency, ultimately automating the creative process while preserving the essence of poetry.