Research On Key Technologies Of Automatic Question Generation And Rewriting

Automatically asking questions is one of the important research topics in natural language processing.The methods are mainly divided into two types,question generation and question rewriting.Question generation methods aim to directly generate relevant questions based on the given input material.Question rewriting aims to generate semantically consistent but differently expressed paraphrase questions from existing original questions.Automatic question generation task has attracted the attention of many researchers in the field of artificial intelligence in recent years due to its wide range of application scenarios.For example,in the field of education,obtaining questions related to course or research content through generation or rewriting methods will help students or researchers better understand and master knowledge points and discover research priorities;In the field of business applications,this task is used in intelligent customer service to effectively help customers interact with enterprises and solve problems more conveniently.However,there are still many challenges in building a complete question generation and rewriting system,such as:(1)Diversity: Generating a large number of high-quality questions requires the system to have the ability to generate diversity.Existing automatic question generation techniques are limited by pre-defined templates or the lack of Beam Search algorithm,it may only focus on a specific fragment of the material and generate several similar content,which limits the scope of the content being examined.(2)Accuracy: Generating a large number of high-quality questions requires the system to have the ability to accurately generate relevant questions according to human-given constraints and materials.Existing automatic question generation techniques may not accurately capture the relationship between constraints and material,especially if the two belong to different modalities.(3)Scalability: Generating a large number of high-quality questions requires the system to have the ability to expand content.Existing automatic question generation or rewriting techniques generate or rewrite questions only related to input material or original questions,but the content contained in the input material or original questions is too limited to expand and include richer information.(4)Consistency: Generating a large number of high-quality questions requires the system to maintain the semantic consistency of the questions before and after rewriting.The generated paraphrase questions from existing automatic question rewriting techniques may modify key information in the original question and result in generating questions with similar descriptions but different semantics.Exploring and solving the above challenges has important theoretical value and guiding significance for building an efficient automatic question generation and rewriting system.This paper focuses on the four challenges faced in automatic question generation and rewriting tasks,analyzes the root causes of the four challenges,and conducts research on specific issues corresponding to the four challenges.Specifically,the main research contents and contributions of this paper are as follows:(1)Research on generated question diversity.Multiple-choice questions in reading comprehension tests are one of the common question types,and they have been a important research due to their widespread use.The quality of the distractors usually determines the quality of the questions.To ensure that each distractor has discriminability from one another,this paper proposes a research method based on multi-expert systems for generating diverse distractors.Each expert in the mechanism uses a attention mechanism to extract different parts of the reading comprehension material for corresponding distractor generation.In addition,the steps of extracting multiple sentences for generating distractors in this method can be used as a plug-and-play module for existing text generation models,which has strong scalability.(2)Research on multimodal features alignment.Constrained multimodal question generation focuses on asking questions related to fine-grained image regions related to given answers,which aim at generating questions that are more human-like.Image information and text information belong to two different modal representation spaces.Cross-modal information cannot achieve semantic alignment,meaning that text information cannot capture specific image features,which leads to the inability to accurately generate questions that are consistent with the constraints or image content.Therefore,this paper studies object-level fine-grained image representations and further proposes an attention mechanism to achieve semantic alignment between object content and text constraint information,thereby extracting target objects related to text constraints to generate more accurate question content.(3)Research on integrating knowledge in generating and rewriting.Existing methods for generating visual questions are limited by their reliance on image information and cannot generate questions that are more realistic.This study explores the fusion of three modal information sources: textual answer information,image information,and knowledge information.An innovative algorithm is proposed for integrating image and knowledge base information to enable the model to generate questions that go beyond visual content in images.This effectively avoids the question generation research being limited to a single source of information,providing researchers with more inspiration to generate questions with richer content.In addition,for the task of question rewriting,existing methods only rely on content replacement and word order transformation based on the original sentence,which leads to the problem of high similarity between the generated sentence and the original sentence.This paper studies the integration of additional knowledge and proposes a knowledge attention-based rewriting model to ensure that the generated sentence goes beyond the content contained in the original input statement.Our method can improve the results of the existing model on BLEU-1 by at least 2.93 on the Quora dataset.(4)Research on Maintaining Key Object for Rewriting.This paper explores the task of visual paraphrase generation for the first time,which analyzes the importance of image information and the challenges of the task.We propose an object-level paraphrase generation model that focuses on fine-grained object features,which can help maintain the information of key objects contained in the original sentence to ensure semantic consistency in the generated paraphrases.To account for variations in expression,we employ an object reordering module and a sequence generation module to adjust the arrangement of key objects and modify their associated descriptions.Additionally,we construct a dataset for the task of visual paraphrase generation,which serves as an appropriate benchmark for evaluating our proposed model and future research.