Search

In order to solve the problem of ineffective utilization of the semantic information between documents in the traditional multi-document extractive summarization method and the excessive redundant content in the summary result, a Khmer multi-document extractive summarization method based on hierarchical maximal marginal relevance(MMR)was proposed. Firstly, the Khmer multi-document text was input into the trained deep learning model to extract all the single-document summaries. Then, all single document summaries were iteratively merged according to a similar hierarchical waterfall method, and the improved MMR algorithm was used to reasonably select summary sentences to obtain the final multi-document summary. The experimental results show that the R1, R2, R3, RL values of the Khmer multi-document summary obtained by using the deep learning method combined with the hierarchical MMR algorithm increases by 4.31%, 5.33%, 645% and 4.26% respectively compared with other methods. The Khmer multi-document extractive summarization method based on hierarchical MMR can effectively improve the quality of Khmer multi-document summary while ensuring the diversity and difference of the summary sentences.

Aiming at the problem of high cost and low expansibility of traditional user intention recognition, which mainly uses template matching or artificial feature set, a hybrid neural network intention recognition model based on BERT word embedding and BiGRU-Attention was proposed. First, the word embedding pre-trained by BERT was used as the input, and the features of the interrogative sentences were extracted by BiGRU. Then, the attention mechanism was introduced to extract the information of words that have important influence on the meaning of sentences and allocate the corresponding weights, so as to obtain the sentence embedding that integrates the word-level weights and input it into the softmax classifier to realize intention classification. According to the experiment on the crawling corpus, it shows that the performance of BERT-BiGRU-Attention method is better than that of traditional template matching, SVM and lately popular CNN-LSTM deep learning combined model. The proposed method can effectively improve the performance of intention recognition model and the quality of online health information service, which provide technical support for the online health community question answering system.