Text Filling Method Based on Prediction Network and Semantic Fusion Loss

Aiming at the problem of semantic loss caused by poor semantic coherence in the text filling process, a text filling method based on prediction network and semantic fusion loss is designed to reduce semantic loss through fusion of cross entropy and similarity. First, the missing text is input into the bidirectional long-short-term memory network code after the word embedding end to extract the context semantic features; Then, the decoder uses the attention mechanism to capture contextual semantic information to fill in the missing blanks one by one; Finally, a semantic similarity fusion loss method is proposed, which calculates the distance between the semantic vectors of coding layer and prediction layer, and adjusts the prediction network together with cross entropy loss function to iteratively reduce the semantic loss and make the filled text semantic more coherent and smooth. The experimental results show that using different missing rates and screening strategies for testing on Yelp, Grimm and NBA news datasets has greatly improved accuracy and fluency. Indicating that the proposed method is better than other methods in text filling task.