Blending the Powers of BERT and Neural Style Transfer for Artistic Text Generation in Poetry
Manish Dakhore, Muktha Eti, Mandar Pramod Diwakar, A. Sivanantham, Lokesh Verma, M. Shyam · 2024
Developing original content that prioritizes both semantic elegance and truth requires the collaboration of two individuals. The visually distinctive appearance of this piece is achieved through the integration of BERT (Bidirectional Encoder Representations from Transformers) capabilities with the sophisticated aesthetics of Neural Style Transfer (NST). Three distinct methods are merged in the proposed method: decoding and generation (D&G), semantic encoding (BERT), and style embedding (NST). The initial methodology, "Semantic Encoding with BERT," guarantees that the resultant poem remains faithful to its initial intent through the utilization of contextual embeddings in BERT. Second, you may create a unique poetry by inserting style components into an existing reference text using Style Embedding with NST. The third algorithm deciphers the style embeddings using an RNN-based generative model. Poetry that is both original and contextualized is the result. Scientific evaluation and comparison of the proposed technique with more conventional approaches is based on a number of criteria. Among these considerations are creative variety, originality, continuity, diversity, meaning accuracy, cohesiveness, and style preservation. The results demonstrate that the proposed approach outperforms conventional ones in generating aesthetically original and easily comprehensible prose.