Research On Question Generation Task Based On LSTM Model

Question Generation task refers to the reverse generation of multi-angle questions given a paragraph and a specified answer.The task of question generation has a wide range of applications and has become one of the research hotspots in the field of Natural Language Processing.However,there are many problems with traditional question generation models.For one thing,the model is insufficient in extracting the semantic information of long texts and cannot obtain rich information features;and for another,the model lacks of the interaction between the context and the answer,which causes the semantics of the generated question to deviate from the context and the answer.In view of the above problems,our paper improves the question generation model based on LSTM.The encoder and decoder of our model are both LSTM models,and the model utilizes a maximum output-based copy mechanism and a pointer generation network to avoid the situation where the input sequence is repeated and the output sequence duplicates repeated words.Based on the above model architecture,our paper proposes a question generation model that integrates rich language features and bidirectional attention layers.The model extracts deep semantic features by augmenting the input text embedding layer information;in addition,it uses a bidirectional attention layer to fuse text and answer information.The main contents of this study are divided into the following two aspects:(1)We propose a word vector representation method based on rich language features.That is,in the embedding layer,the method of integrating the external knowledge of sememe is used to capture the semantic knowledge with a smaller granularity than the word vector,thereby enhancing the semantic representation of the text itself,and solving the problem of insufficient semantic extraction of long texts.In addition,the recommended sememe set is obtained through the cosine similarity algorithm.supplements the problem of "missing of word meaning" caused by no corresponding sememe words in the vocabulary,and also solves the problem of "mixing of word meanings" caused by irrelevant sememe sets in the sememe set.At the paragraph of the embedding layer,the model can obtain an answer-oriented context representation by adding the answer position information.(2)We propose a question generation method based on variant bidirectional attention layers.That is,after the text and the answer are encoded separately,the answer-aware context representation vector is obtained through the bidirectional attention layer network.In addition,the model first adds a self-attention mechanism to the context to capture the semantic correlation within the long sequence of text data,obtain the key information in the context,and then fuse with the answer.The variant bidirectional attention layer mechanism proposed by our model can aggregate the answer information and effectively avoid the problem of information loss caused by the early weighting of the model.Finally,extensive experiments show that our model outperforms the baseline model on the SQu AD1.1 dataset.And,our model can generate higher-quality questions that are more in line with human language habits.