This project concerns the generation of Shakespeare-like text sequences using mainly LSTMs and how to evaluate these texts sequences as well as enhancing the text generation process. We experimented with different extensions to improve the model, such as data augmentation, and different sampling methods such as beam search, temperature sampling, top k and top p. We also evaluated the model using qualitative and quantitative metrics. The qualitative metrics that were used were coherence, grammaticality, and Shakespeare similarity. The quantitative metrics used were spelling percentage, term token ratio, perplexity, BLEU score, and BERTScore.
Project report can be found here.