The text generation models encounter challenges in learning the high-dimensional and abundance of data from the source domain, which leads to inaccurate text generation. In addition, generating the text sequence with contextual knowledge is challenging, even training the learning model with the domain knowledge. Hence, it is essential to effectively extract the knowledge from the source domain in the transfer learning model, which necessitates the contextual embedding of the input sequence and dimensionality reduction. Thus, this work develops the Deep Transfer Learning-based Text Generation (DTGEN) with the Embedding from Language Model (ELMo), Variational AutoEncoder (VAE), and Bidirectional-Long Short Term Memory (Bi-LSTM). The proposed DTGEN model focuses on the low-dimensional text feature representation and the deep transfer learning-based text generation. Initially, the proposed DTGEN approach contextually models the input text sequence with the help of the ELMo and generates the vectorized input data as the feature representation. Moreover, it generates the low-dimensional feature vector for the embedding input-based knowledge transfer by adopting the VAE model with the weighted regularization. In subsequence, the DTGEN approach learns the relevant knowledge from the source domain based on the contextual vector representation and analyzes the sequential patterns of the text data in the target domain. Consequently, the proposed approach predicts the terms based on the sequential relationships in the target text and the knowledge extracted from the source text with the assistance of the Bi-LSTM model. Thus, the experimental results demonstrate that the proposed DTGEN approach yields 2.89% higher text generation accuracy than the existing Deep Learning-based Next Word Prediction (DNWP) approach while testing on the NewsRoom dataset. Furthermore, the DTGEN approach outperforms the LSTM-based text generation by 10.11% higher accuracy for the samples in Cornell Movie-dialog Corpus (CMC) dataset.