Research And Application Of Deep Text Matching Algorithm In Question Answering System

The automatic question-answering system can understand the user's question and return the answer to the user in a short and clear language.Now automatic question answering has become an important research hotspot in the field of artificial intelligence.In the automatic question-answering system,the part related to text matching is very critical,which directly determines the accuracy of the question answering system.Therefore,this thesis studies the two tasks of answer selection and similar question recognition in the automatic question answering system.This thesis focuses on the two tasks of algorithm optimization and implementation,the main research work includes:1.For the answer selection task,the question and the answer may not have the same vocabulary.Therefore,the answer selection model based on lexical features cannot obtain the matching relationship between the two sentences.The model should analyze the relationship between the sentences more from a semantic perspective,so this thesis uses the BERT pre-training model to implement the answer selection task,and proposes two answer selection models using feature-based and fine-tuning strategies.In the featurebased method,the pre-training BERT model is used to obtain the context word vector,and the multi-head self-attention weighting vector and bi-directional attention weighting vector are compared and aggregated to obtain the matching score of the question and answer pair.In the fine-tuning-based method,the BERT model is extended by MVLSTM(Multi View Long Short Term Memory),so that the question and answer pairs can directly interact,and the semantic information of sentence matching can be fully obtained.Finally,the performance of the proposed model is better than that of the mainstream model on the open datasets Trec QA and Wiki QA.2.For the task of the identification of similar question,in the current model based on text semantic representation,neural network can efficiently obtain the vector representation of sentences,but these methods do not fully obtain the matching features of two sentences,so the performance of the model needs to be improved.Recently,the use of the BERT pre-training language model to fine-tune the target task has made significant progress in some sub-fields of natural language processing.Therefore,this thesis will use the BERT language model to fine-tune the identification of similar question task,and combine it with the Bi LSTM-CNN((Bi-directional LSTMConvolutional Neural Networks))network model to calculate the semantic similarity between two questions.The experimental results show that the performance of the BERT fine-tuning model based on Bi LSTM-CNN on the financial data sets is better than other algorithms.3.An automatic question answering prototype system based on the identification of similar question and answer selection is constructed.The system is composed of a data acquisition and processing module,a similar question recognition module,an answer selection module and a human-computer interaction layer.In order to demonstrate the function of the identification of similar question module,this thesis designs a similar sentence recognition interface.Two sentences to be recognized can be input on the interface,and the semantic relationship between the two sentences can be obtained through system processing.