Topic-to-Essay Generation with Corpus-Based Background Information

Abstract

Abstract This study aims to generate more topic-related and coherence essays based on user-defined topic words. Existing research generates essays without considering the semantic information from the corpus. However, the corpus contains statistical relationships of words which can be used to guide the model to generate more coherent and fluent essays. To fill this gap, we propose a corpus-based topic-to-essay generation model (C-TEG). We elaborately devise a background network based on the co-occurrence relationships of words from the corpus. The empirical results demonstrate that our approach has achieved 4.14 average score in subjective evaluation and a better BLEU-2 score, which shows that our model is able to generate more topic-related and coherent text than existing models.