Research And Implementation Of Topic-to-essay Generation Technology Based On Deep Learning

The demand for natural language generation has updated owning to rapid evolvement of technology.People expect the output to be higher quality and richer information through simpler inputs.The importance of research on semantically controllable text generation technology is growing.The topic-to-essay generation task as a branch of natural language generation is emerged adapted to the trend.It generates topic-related and coherence text based on defined topic.Through studying the existing topic-to-essay generation technology,we find that current topic-to-essay generation model is mainly implemented based on the encoder-decoder model with attention mechanism.The model is capable to generate topic-related text but the generation quality needs to be improved.The main problem is that the input semantic information is not adequate to guide the model.To solve this problem,we propose a corpus-based topic-to-essay generation model.A background network is built to calculate additional topic information using semantic information from the corpus.The empirical results show that this approach improves the quality of the generated text.The model is superior to the model that does not contain the corpus-based background network in both subjective and objective evaluation.Through research on the corpus-based topic-to-essay generation model,we find that although the model improves the results,it still has two problems.The first is that the model shows poor understanding about common-sense topic due to the corpus lack of such information.The second is that the additional topic information is affected by word segmentation,which may cause a biased result.To solve the first problem,we introducing sememes from the HowNet knowledge graph to expands the topic information.To solve the second problem,we measure the similarity between the sememes and the average word vectors of the non-current topic words.The expanded topic information is used to guide the model.The encoder-decoder framework with attention mechanism is used based on the previous research.The experiments demonstrates that our approach effectively solves the two problem and further improves the integrity and coherence of the text.The proposed model obtains best evaluation results in both subjective and objective metrics compared with the model based on the corpus background network,model based on the other knowledge graph,model based on HowNet without sememe selection mechanism and model dose not introduce additional information.