Sememe-based Topic-to-Essay Generation with Neural Networks

Topic-to-Essay generation task aims to generate topic related and coherence text based on user input. Previous research generates text solely based on the input words and the generation results in not satisfactory. The traditional methods using semantic relationships of words from corpus to guide the model, which made the model rely heavily on the corpus. In this paper, we propose a sememe-based topic-to-essay generation model(S-TEG), which integrates sememes from external knowledge graph HowNet with input topic words to guide the model. In order to prevent introducing noise, we elaborately devise measuring the similarity of the non-current topic words to filter sememes information. The experiment results demonstrate that our approach achieved 4.10 average score in subjective evaluation and a 3.60 BLEU score, which shows that our model is able to generate text that is more coherence, topic-related and fits the daily logic.