Research And Implementation Of Generative Question Answering System Technology

With the rapid development of the Internet and the continuous improvement of information retrieval technology,people are eager to obtain the information and services they need,anytime,anywhere,in a fast and intelligent way.Unlike traditional search engines that return to a variety of related web pages,the question answering system directly returns the answers to user questions,saving the cost of the user's secondary screening of answers from relevant web pages.Technically,the question answering system can be divided into extractive question answering and generative question answering.The latter is attracting more and more researchers' attention because it can fuse multiple answers distributed in different segments.At present,the model of the generative question answering has the following two problems:1)The framework of "extraction-then-synthesis"is usually adopted,and the extraction module and the synthesis module are separately trained to cause errors before and after propagation,and the information between the modules before and after is difficult to share;2)It is difficult to accurately generate key named entities during the answer generation process from sequence to sequence model,and cannot generate out-of-vocabulary words.Aiming at these two problems,this paper proposes an end-to-end generative question answering system based on the framework of"encoding-interaction-prediction".Specifically,the question and the passage are encoded in the encoding layer through the bidirectional recurrent neural network;the two-way attention mechanism is used in the interaction layer to capture the matching relationship between the question and the passage,and the context information of the passage is merged through a bidirectional recurrent neural network to obtain the question-aware passage represent;the prediction layer fuses the copy and generates two mechanisms to obtain the answer text.At each step of the decoding,either a word can be copied from the passage or a new word can be generated from the vocabulary.Through a series of comparative experiments and case studies,this paper proves that the hybrid model is better than the copy model and the generation model on the generative question answering data.Based on the above model,a web-oriented generative question answering system was built.The question answering system receives the user's question,and can directly give the answer to the question according to the passage provided by the user;or can search for relevant information from the massive document of the Internet,and obtain the answer through understanding and synthesis.