Automatic Question Generation For Texts

Text-based question generation generates related questions from a given sentence or paragraph.With the continuous development of natural language generation technology,question generation has been focused in the field of artificial intelligence.This task requires that the system understands the input texts and has some reasoning ability.At present,sequence-to-sequence neural network model has been used to generate questions for sentences containing answers.However,these methods have major problems :（1）the generated interrogatives do not match the answer types;（2）the relevance between the question and the answer is not strong;（3）the question is not clearly stated,etc.For problems,the main work of this paper is as follows:（1）A question generation model based on the answer and contextual information is adopted to solve the problem that the generated interrogatives do not match the answer type and the relevance of questions and the answer is low.The main idea of this model is: firstly,the interrogative matching with the answer type is determined according to the relation between the answer and contextual information.Then use the answer and contextual information to determine the word related to the question and make the question use the words in the original sentence as much as possible.Finally,combining the original sentence as input to come into question.Related experimental results show: this model increases the BLEU₄ value by 3%-5% and the ROUGE_L value by 3%-6%.（2）For the problem of unclear expressions of the questions,this paper uses a question generation model based on reinforcement learning and integrated multi-sentence information to solve it.The main idea of this model is as follows: first,the Transformer model encodes the information of sentences and the answer respectively.Secondly,the matching information between sentences and the answer can be obtained through Co-Attention mechanism.Finally,the model based on the idea of reinforcement learningand fine-tuned with the strategy gradient technique to maximize several rewards for assessing question quality.Related test results show: compared with the baseline system,BLEU₄ and ROUGE_L in this model have improved by 2%-5% and 3%-5%,respectively.（3）A question generation prototype system is implemented.The system mainly includes two core modules: the determination of question objects and the generation of questions,which are used to realize the function of generating a series of related questions based on a given text.The work of this paper are effective in improving the performance of the question generation system,but further research on inter-clause relation and extraction of words related to questions are needed in the future.