Research On Opinion-question Answering For Chinese Reading Comprehension

Reading comprehension is to read a given document and answer a series of questions.It is a research focus in natural language processing.Set against the backdrop of the Beijing College Entrance Examinations,this thesis studies the opinion-question answering for reading comprehension in Chinese literature documents.The main works was as follows:(1)The difficulty of opinion-question is analyzed in reading comprehension in college entrance examination and puts forward three key techniques: question extension,sort learning and answer extraction.(2)Multi-topic answer extraction.A new method was proposed for extracting answers in reading comprehension to overcome the shortcomings that used key-words.First,through the LDA(Latent Dirichlet Allocation)model,the similarity of the distribution of the topics between each sentence of the article and the question were calculated.Then,the sentences with high similarity of the topics distribution for obtaining the candidate sentences with the characteristics of opinion were classified.Finally,the similarity between the sentences in article and the question were calculated and sorted to choose the first ? sentence as the answer.This method achieved 47.8% and 38.5% accuracy in the answer extraction of the small-scale and large-scale opinion question corpus and can automatically sort the answer sentence to the front.(3)Contextual-aware based problem extension.To increase the recall rate of the answer extraction,the model needs to have a deeper understanding of the original problem,namely to find the relevant semantics of the problem under the reading material.We use a context-aware model based on a hierarchical recurrent network to establish the interdependence among the reading materials,questions and answers,and semantically extend the original problem based on this model.The experimental results show that the accuracy of the semantic extension of neural network to the original problem reaches 68%,which can extend the deep semantics of the problem to some extent.At the same time,the use of the extended problem is 4% more accurate than the use of the original problem when using the multi-topic answer extraction method.(4)Multi-feature-fusion-based sort learning.To reduce the instability caused by incorrect extension information and artificial selection of feature weights,we apply the learning-to-rank model to the questions of the college entrance examination perspective.The extended questions and the original questions are respectively calculated with the reading material topic distribution similarity,content similarity,the opinion sentence features,the BM25 score,and the PageRank score.These features are used as features and merged to construct an L2R-ordered learning model.Finally,the trained model is used to extract the answer sentence,and the top ? sentence is selected as the answer sentence.The experimental results show that the multi-feature-fusion-based sort learning method increases 6.5% over the multi-topic answer extraction method on a larger-scale test set,achieving an accuracy of 45.0%.(5)Answering system.Combining the problem extension,sort learning and answer extraction techniques,it is applied to the solution of the questions of the college entrance examination.By manually scoring the answers obtained by the machine,the answer frame proposed in this thesis can achieve an accuracy of 42.1%.