Research On Algorithms Combining Attention Mechanism And Gate Mechanism In Question Answering System

In this century,services from all walks of life have been widely connected to the Internet.With the rapid growth of content production and user scale,massive data are generated all the time.These data lead to the inefficiency and overload of users receiving information.And users need to filter out valid information from them.Facing with such a huge amount of valuable data resources,how to search and extract information quickly and effectively has become a hot research field in academia and industry.Question answering system as a representative technology is highly concerned.Question answering system can help users read documents and questions and get answers quickly,thus replacing the tedious process of finding answers under a large number of related documents in the past,and significantly improving the user experience.In this thesis,in response to the current related research on the failure to use candidate answer information,the model ignores key text information,and matches the answer literally too stably,we introduce a customized gated mechanism and attention mechanism for further optimization and improvement.The relevant research results obtained are as follows:1.This thesis proposes a QANet-based candidate answer rethink model.This thesis proposes the model by simulating the behavior of selecting and referring to multiple candidate answers during reading comprehension,and comprehensively judging the behavior of giving the final answer.The model interacts and fuses the selected multiple candidate answers with the original text and the question,prompting the model to synthesize the three aspects of information to obtain a more accurate answer.Comparative experiments show that the model has achieved a certain performance improvement compared with the baseline model.At the same time,multiple groups of ablation experiments were conducted to compare the performance differences of the models under different numbers of candidate answers.2.This thesis proposes a multi-granularity gated multi-hop skimming and intensive model.This model adds skimming,intensive reading and gated mechanism modules to the pretraining model to simulate the behavior of humans reading multiple times and comprehensively screening information during question and answer tasks.At the same time,based on the idea of granular computing,a multi-granularity module for computing context granularity and sequence granularity is added to the model,and the word granularity obtained by the above modules forms a parallel structure.The model can simulate the behavior of human beings to understand the text from the perspective of words to sentences,from parts to the whole.Experiments show that the model has made new progress in indicators such as EM and F1.Experimental analysis shows that the model can better understand the question and context and give more accurate answers.At the same time,the attention mechanism is visualized to verify the effectiveness of the skimming module.