[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

WO2020034642A1 - 医疗自动问答方法及装置、存储介质、电子设备 - Google Patents

医疗自动问答方法及装置、存储介质、电子设备 Download PDF

Info

Publication number
WO2020034642A1
WO2020034642A1 PCT/CN2019/080298 CN2019080298W WO2020034642A1 WO 2020034642 A1 WO2020034642 A1 WO 2020034642A1 CN 2019080298 W CN2019080298 W CN 2019080298W WO 2020034642 A1 WO2020034642 A1 WO 2020034642A1
Authority
WO
WIPO (PCT)
Prior art keywords
question
answer
training
medical
construct
Prior art date
Application number
PCT/CN2019/080298
Other languages
English (en)
French (fr)
Inventor
鹿文鹏
张玉腾
张若雨
张旭
岳书通
成金勇
Original Assignee
齐鲁工业大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 齐鲁工业大学 filed Critical 齐鲁工业大学
Priority to AU2019322953A priority Critical patent/AU2019322953B2/en
Publication of WO2020034642A1 publication Critical patent/WO2020034642A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H80/00—ICT specially adapted for facilitating communication between medical practitioners or patients, e.g. for collaborative diagnosis, therapy or health monitoring
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/20—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for electronic clinical trials or questionnaires
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H70/00—ICT specially adapted for the handling or processing of medical references

Definitions

  • the invention relates to the technical field of natural language processing and the field of computer artificial intelligence, in particular to a medical automatic question answering method and device, a storage medium, and an electronic device.
  • Patent document CN107516110A discloses a medical question answering semantic clustering method based on integrated convolutional coding.
  • the method includes the following steps: Step 1: Obtain medical question answering data from a medical platform, and pre-treat the medical question answering data set. Process and get the input matrix; Step 2: Use a convolutional coding network to select different convolution kernels for different input matrices for kernel clustering, calculate the clustering quality and diversity after the kernel clustering, and calculate the clustering quality based on the clustering quality.
  • Step 3 The convolution kernels selected in Step 2 are trained by a convolutional neural network respectively;
  • Step 4 Fusion features of different convolution kernels Represent the results;
  • Step 5 Enter the fused feature representation results into the autoencoder and perform input reconstruction training to obtain the best feature representation;
  • Step 6 Cluster the best feature representations obtained by encoding to obtain the final medical text semantic clustering Class results.
  • this technical solution is only a classification of medical texts, and it is impossible to accurately obtain accurate answers to medical questions.
  • the technical task of the present invention is to provide a medical automatic question answering method and device, storage medium, and electronic equipment to solve how to use natural language processing technology to select a matching answer for a patient from an existing question answering result and recommend it to a user. Questions that reduce the workload of the doctor and increase the speed of answering questions.
  • a medical automatic question answering method includes the following steps:
  • Constructing a medical question-answering knowledge base crawling question-answering pairs from the Internet medical question-answering community, and pre-processing the question-answering pairs to construct a medical question-answering knowledge base;
  • step S2 Construct a training data set for the question answering model; for each medical question, there is at least one correct answer in the medical question answering knowledge base in step S1, and construct a positive training example from the correct answer and question, and answer and question other than the correct answer Construct negative training examples; the number of negative examples is set according to the size of the medical question answering knowledge base;
  • step S4 Training a question answering model and answer selection:
  • the question answering model training data set obtained in step S2 is used to train the question model constructed in step S3.
  • step S1 the specific steps of constructing the medical question answering knowledge base in step S1 are as follows:
  • step S102 Preprocessing the question and answer pairs crawled from the Internet medical question and answer community in step S101, mainly including splitting a question with multiple answers and performing hyphenation processing on the questions and answers;
  • Hyphenation processing of questions and answers refers to hyphenation processing of the contents of questions and answers.
  • each question and answer in the medical question and answer community is based on Chinese characters, and each word Separate with spaces and keep numbers and punctuation.
  • the specific steps of constructing the training data set of the question answering model in the step S2 are as follows:
  • a negative training example select the question in step S201 in the medical question answering knowledge base, and then randomly select a question other than step S201 and the answer corresponding to the question from the medical question answering knowledge base, and randomly combine the question in step S201 with a random answer.
  • a combination of answers corresponding to the selected question other than step S201 is used to construct a negative training example.
  • the form of the negative training example is (question, bad-answer, 0), where question refers to the question and bad-answer refers to a randomly selected division. Answers for questions other than step S201, where 0 represents a negative example;
  • S203 Construct a training data set: combine the samples of all the positive training examples and the negative training examples obtained in steps S201 and S202, and randomly disrupt the samples of the positive training examples and the negative training examples to construct the final training data set;
  • Both positive training examples and negative training examples include three dimensions, namely question, answer or bad-answer, 1 or 0.
  • the first two dimensions construct a two-dimensional tensor representation based on the words actually contained, and the third dimension constructs a vector. Means.
  • step S3 the specific steps of constructing the question answering model in step S3 are as follows:
  • step S301 Construct a character mapping conversion table: according to the medical question answering knowledge base constructed after preprocessing in step S1, construct a character table, and map characters in the character table to unique numeric representations. For example, the characters can be added from the character table in the order 1 starts to increase in order to build a character mapping conversion table;
  • the input layer includes three input contents.
  • the question and answer are obtained from the training positive examples corresponding to the question, and the bad-answer is obtained from the negative training examples.
  • the form of the input layer is (question, answer, bad-answer );
  • the character mapping conversion table obtained in step S301 the characters in the input data are converted into digital representations;
  • a character vector mapping layer specify the dimension of the character vector as embedding_dim, the size of the character table obtained in step S301 is vocab_num, and use a uniform distribution definition of [-1,1] (vocab_num + 1) ⁇ embedding_dim as a two-dimensional matrix And use it as a weight matrix of the character vector mapping layer of the model, where each row corresponds to a vector representation of a single character in the medical question answering knowledge base, where the 0th row represents characters that do not appear in the knowledge base;
  • the multi-scale convolutional layer uses a convolution kernel of different sizes and the data processed in step S303 to perform a convolution operation to capture the character-level n in the question or answer. -gram feature as output; stacking convolution layers is to stack convolution kernels of different sizes, and to perform convolution operations with each convolution kernel from the bottom up to the data processed in step S303 to capture sentences in questions or answers Semantic features at levels are used as outputs; then the output of the multi-scale convolutional layer is connected to the output of the stacked convolutional layers as the feature vector of the question or answer;
  • ⁇ ⁇ ⁇ represents the L2 norm
  • Q and A refer to the vector representation of the question and correct answer or wrong answer after being processed by the Q & A model
  • Q represents the vector of vector_question
  • A represents the vector of vector_answer or vector_bad-answer.
  • step S304 the specific steps of calculating the multi-scale convolutional layer in step S304 are as follows:
  • step S304 The specific steps of calculating the stacked convolution layers in step S304 are as follows:
  • step S4 the specific steps of training the question answering model and answer selection in step S4 are as follows:
  • margin is a hyper-parameter, which is adjusted according to the actual data;
  • max (a, b) represents the maximum value between a and b;
  • betas (0.9, 0.999), betas is a coefficient used to calculate the running average of the gradient and its square;
  • eps 1e-8, eps is added to the denominator to increase numerical stability;
  • weight_decay 0
  • weight_decay is used to control weight attenuation.
  • Answer selection a new question and all candidate answers in the medical question answering knowledge base are used as input samples; and then sent to the question and answer model trained in step S3 to calculate the similarity between the question and the candidate answer; finally The group with the highest similarity value is selected from all the results as the most likely answer to the new question and the answer is pushed to the user.
  • a medical automatic question answering device includes:
  • Medical Q & A knowledge base building unit used to crawl the Q & A pairs from the medical Q & A community on the Internet, and pre-process the questions and answers of the medical Q & A knowledge base.
  • the pre-processing includes splitting multiple questions And perform hyphenation on the content of questions and answers to build a medical question answering knowledge base;
  • Question and answer model training data set generation unit which is used to construct training positive and negative training examples based on the medical question answering knowledge base, and then construct a training data set;
  • Question-and-answer model building unit which is used to construct a character mapping conversion table, an input layer, a character vector mapping layer, a multi-scale convolutional layer, a stacked convolutional layer, and a similarity calculation layer, thereby constructing a question-and-answer model;
  • Question and answer model training and answer selection unit are used to construct loss function and optimization function to complete answer selection.
  • the medical question answering knowledge base construction unit includes:
  • the reptile unit is used to grab content from the medical Q & A community from the Internet and save the results in the form of Q & A pairs;
  • the crawler data processing unit is used to split the question-and-answer pairs obtained by the crawler into a plurality of separate corresponding question-and-answer pairs, and perform hyphenation processing on the questions and answer sentences to construct a question-and-answer knowledge base;
  • the question and answer model training data set generating unit includes:
  • Training positive example construction unit which is used to combine the question with its corresponding answer to construct a positive training example
  • a training negative construction unit is used to select a question q1 from the question and answer knowledge base, and then randomly select a question q2 different from question q1 from the medical question and answer knowledge base, and combine q1 and q2's answer a2 to construct a negative training example;
  • the training data set construction unit is used to combine the training positive sample and training negative sample of all the questions and answers obtained from the training positive construction unit and the training negative construction unit, and randomly shuffle the order to form the final training data set;
  • the question answering model building unit includes:
  • Character mapping conversion table construction unit which is used to construct all characters in the medical question answering knowledge base into a character table, and maps each character in the character table to a unique numeric representation. It can start from 1 in the order of characters added to the character table. Increment one by one to build a character map conversion table.
  • the input layer construction unit is used to obtain input data from the training positive examples and training negative examples corresponding to the problem, and respectively obtain the question question and answer answer in the training positive example and the answer in the training negative example, and formalize them as: ( question, answer, bad-answer);
  • Character vector mapping layer construction unit which is used to construct and store the vector representation of each character in the character mapping conversion table
  • Multi-scale convolutional layer and stacked convolutional layer construction unit which is used to generate multi-scale convolutional layer and stacked convolutional layer.
  • Multi-scale convolutional layer is used to capture and output character-level n-gram features in questions or answers, and stack them.
  • the convolution layer is used to capture and output sentence-level semantic features in the question or answer.
  • the output of the multi-scale convolution layer is connected with the output of the stacked convolution layer as the feature vector of the question or answer;
  • Similarity evaluation construction unit which is used to connect the vector representations of the questions and answers of the output of the multi-scale convolutional layer construction unit and the stacked convolutional layer construction unit to obtain the final representation of the question and answer, and calculate the cosine of the two Similarity
  • the question and answer model training and answer selection unit includes:
  • Loss function building unit used to calculate the similarity between the question and the correct answer and the error between the similarity between the question and the wrong answer;
  • An optimization function building unit is used to adjust parameters in the model so that the similarity between the model and the correct answer to the model is greater than the similarity between the question and the incorrect answer;
  • An answer selection unit is used to select a correct answer from a medical question answering knowledge base using a trained model when processing a new question.
  • a storage medium stores a plurality of instructions, and the instructions are loaded by a processor to execute the steps of the medical automatic question answering method described above.
  • An electronic device includes:
  • a processor configured to execute instructions in the storage medium.
  • the medical automatic question answering method and device, storage medium, and electronic device of the present invention have the following advantages:
  • the present invention can complete the collection and preprocessing of the medical question-answering knowledge base through the crawler technology, generate a medical question-answering knowledge base, and provide data support for the subsequent establishment of a medical question-answering model;
  • the present invention can effectively model the existing question-answer pairs to construct a question-answer model, and extract multiple layers of semantic associations between the question-answer pairs;
  • the present invention can define and implement a complete set of end-to-end automatic question answering models, and directly recommend the most similar answers to user questions;
  • the present invention can use the semantic relationship in the medical question to obtain a medical question answering model, thereby solving the question of answer selection in the medical question answering;
  • the present invention uses character-level vector representations for questions and answers, and combines the strong local abstraction capabilities of convolutional neural networks, which can reduce the interference of typos to a certain extent;
  • the present invention combines the multi-scale convolutional layer and the stacked convolutional layer to represent the n-gram features of the characters and sentences in the medical question and answer database, which can more effectively extract the features of the medical question, and then obtain more accurate answers;
  • the present invention uses natural language processing technology to select matching answers for patients from the existing question and answer results and recommends them to users, reducing the workload of doctors and increasing the speed of answering questions, which is an important breakthrough in computer artificial intelligence.
  • FIG. 1 is a flowchart of a medical automatic question answering method
  • FIG. 2 is a flow block diagram of step S1 for constructing a medical question answering knowledge base
  • FIG. 3 is a flow block diagram of constructing a training data set of a question answering model in step S2;
  • FIG. 4 is a flow block diagram of step S3 to construct a question answering model
  • FIG. 5 is a flow block diagram of step S4 for constructing a training question answering model and answer selection
  • FIG. 6 is a flowchart of the calculation of the multi-scale convolutional layer and the stacked convolutional layer in step S304;
  • FIG. 7 is a structural block diagram of a medical automatic question answering device.
  • the medical automatic question answering method of the present invention includes the following steps:
  • Constructing a medical question-answering knowledge base crawling question-answering pairs from the Internet medical question-answering community, and pre-processing the question-answering pairs to construct a medical question-answering knowledge base;
  • step S2 Construct a training data set for the question answering model. For each medical question, there are two correct answers in the medical question answering knowledge base in step S1. Construct a positive training example from the correct answer and the question, and answer and question other than the correct answer. Construct negative training examples; the number of negative examples is set according to the size of the medical question answering knowledge base;
  • step S4 Training a question answering model and answer selection:
  • the question answering model training data set obtained in step S2 is used to train the question model constructed in step S3.
  • step S1 the specific steps of constructing the medical question answering knowledge base in step S1 are as follows:
  • step S102 Preprocessing the question and answer pairs crawled from the Internet medical question and answer community in step S101, mainly including splitting a question with multiple answers and performing hyphenation processing on the questions and answers;
  • Hyphenation processing of questions and answers refers to hyphenation processing of the contents of questions and answers.
  • each question and answer in the medical question and answer community is based on Chinese characters, and each word Separate with spaces and keep numbers and punctuation.
  • step S2 the specific steps for constructing the training data set of the question answering model in step S2 are as follows:
  • a negative training example select the question in step S201 in the medical question answering knowledge base, and then randomly select a question other than step S201 and the answer corresponding to the question from the medical question answering knowledge base, and randomly match the question in step S201 with The combination of answers selected for the questions other than step S201 is used to construct a negative training example.
  • the negative training example has the form (question, bad-answer, 0), where question refers to the question and bad-answer refers to a randomly selected division. Answers for questions other than step S201, where 0 represents a negative example;
  • Antiviral drugs such as acyclovir ribavirin (triazine Ribavirin).
  • Chinese medicine includes Daqingye, Prunella vulgaris, or Daochisan addition and subtraction.
  • Shuanghuanglian oral solution can also be used. In severe cases, it needs to be treated at the hospital after treatment. ", 0).
  • S203 Construct a training data set: combine the samples of all the positive training examples and the negative training examples obtained in steps S201 and S202, and randomly disrupt the samples of the positive training examples and the negative training examples to construct the final training data set;
  • Both positive training examples and negative training examples include three dimensions, namely question, answer or bad-answer, 1 or 0.
  • the first two dimensions construct a two-dimensional tensor representation based on the words actually contained, and the third dimension constructs a vector. Means.
  • vector_question use vector_question, vector_answer, vector_posneg to represent the vector representation of the question, the vector representation of the answer, and the vector representation of the positive and negative examples of the training data set.
  • vector_num refers to the number of characters in question, answer, or bad-answer.
  • step S3 the specific steps for constructing a question answering model in step S3 are as follows:
  • step S301 Construct a character mapping conversion table: according to the medical question answering knowledge base constructed after preprocessing in step S1, construct a character table, and map characters in the character table to unique numeric representations. For example, the characters can be added from the character table in the order 1 starts to increase in order to build a character mapping conversion table;
  • the character table and character mapping conversion table are constructed as follows:
  • the input layer includes three input contents.
  • the question and answer are obtained from the training positive examples corresponding to the question, and the bad-answer is obtained from the negative training examples.
  • the form of the input layer is (question, answer, bad-answer );
  • the character mapping conversion table obtained in step S301 the characters in the input data are converted into digital representations;
  • a character vector mapping layer specify the dimension of the character vector as embedding_dim, the size of the character table obtained in step S301 is vocab_num, and use a uniform distribution definition of [-1,1] (vocab_num + 1) ⁇ embedding_dim as a two-dimensional matrix And use it as a weight matrix of the character vector mapping layer of the model, where each row corresponds to a vector representation of a single character in the medical question answering knowledge base, where the 0th row represents characters that do not appear in the knowledge base;
  • embedding_layer torch.nn.Embedding ((word_num + 1), embedding_dim).
  • This layer is a general layer of the question answering model, which stores the character vectors corresponding to all characters; the processing operations for quest, answer, and bad-answer are all the same.
  • the multi-scale convolutional layer uses a convolution kernel of different sizes and the data processed in step S303 to perform a convolution operation to capture the character-level n in the question or answer. -gram feature as output; stacking convolution layers is to stack convolution kernels of different sizes, and to perform convolution operations with each convolution kernel from the bottom up to the data processed in step S303 to capture sentences in questions or answers Semantic features at levels are used as outputs; then the output of the multi-scale convolutional layer is connected to the output of the stacked convolutional layers as the feature vector of the question or answer;
  • ⁇ ⁇ ⁇ represents the L2 norm
  • Q and A refer to the vector representation of the question and correct answer or wrong answer after being processed by the Q & A model
  • Q represents the vector of vector_question
  • A represents the vector of vector_answer or vector_bad-answer.
  • step S304 the specific steps of calculating the multi-scale convolutional layer in step S304 are as follows:
  • step S304 The specific steps of calculating the stacked convolutional layers in step S304 are as follows:
  • Convolution code: conv torch.nn.Conv1d (embedding_dim, output_channel, c_i), where output_channel is the size of the output channel, which can be customized by the user.
  • step S4 the specific steps of training the question answering model and answer selection in step S4 are as follows:
  • margin is a hyper-parameter, which is adjusted according to the actual data;
  • max (a, b) represents the maximum value between a and b;
  • the Adam optimization function can be defined by the following code:
  • Answer selection a new question and all candidate answers in the medical question answering knowledge base are used as input samples; and then sent to the question and answer model trained in step S3 to calculate the similarity between the question and the candidate answer; finally The group with the highest similarity value is selected from all the results as the most likely answer to the new question and the answer is pushed to the user.
  • the device includes:
  • Medical Q & A knowledge base building unit used to crawl the Q & A pairs from the medical Q & A community on the Internet, and pre-process the questions and answers of the medical Q & A knowledge base.
  • the pre-processing includes splitting multiple questions And perform hyphenation on the content of questions and answers to build a medical question answering knowledge base;
  • Question and answer model training data set generating unit which is used to construct training positive examples and negative training examples based on the medical question answering knowledge base, and then construct a training data set;
  • Question-and-answer model building unit which is used to construct a character mapping conversion table, an input layer, a character vector mapping layer, a multi-scale convolutional layer, a stacked convolutional layer, and a similarity calculation layer, thereby constructing a question-and-answer model;
  • Question and answer model training and answer selection unit are used to construct loss function and optimization function to complete answer selection.
  • the medical Q & A knowledge base construction unit includes:
  • the reptile unit is used to grab content from the medical Q & A community from the Internet and save the results in the form of Q & A pairs;
  • the crawler data processing unit is used to split the question-and-answer pairs obtained by the crawler into a plurality of separate corresponding question-and-answer pairs, and perform hyphenation processing on the questions and answer sentences to construct a question-and-answer knowledge base;
  • the questionnaire model training data set generation unit includes:
  • Training positive example construction unit which is used to combine the question with its corresponding answer to construct a positive training example
  • a training negative construction unit is used to select a question q1 from the question and answer knowledge base, and then randomly select a question q2 different from question q1 from the medical question and answer knowledge base, and combine q1 and q2's answer a2 to construct a negative training example;
  • the training data set construction unit is used to combine the training positive sample and training negative sample of all the questions and answers obtained from the training positive construction unit and the training negative construction unit, and randomly shuffle the order to form the final training data set;
  • Q & A model building units include:
  • Character mapping conversion table construction unit which is used to construct all characters in the medical question answering knowledge base into a character table, and maps each character in the character table to a unique numeric representation. It can start from 1 in the order of characters added to the character table. Increment one by one to build a character map conversion table.
  • the input layer construction unit is used to obtain input data from the training positive examples and training negative examples corresponding to the problem, and respectively obtain the question question and answer answer in the training positive example and the answer in the training negative example, and formalize them as: ( question, answer, bad-answer);
  • Character vector mapping layer construction unit which is used to construct and store the vector representation of each character in the character mapping conversion table
  • Multi-scale convolutional layer and stacked convolutional layer construction unit which is used to generate multi-scale convolutional layer and stacked convolutional layer.
  • Multi-scale convolutional layer is used to capture and output character-level n-gram features in questions or answers, and stack them.
  • the convolution layer is used to capture and output sentence-level semantic features in the question or answer.
  • the output of the multi-scale convolution layer is connected with the output of the stacked convolution layer as the feature vector of the question or answer;
  • Similarity evaluation construction unit which is used to connect the vector representations of the questions and answers of the output of the multi-scale convolutional layer construction unit and the stacked convolutional layer construction unit to obtain the final representation of the question and answer, and calculate the cosine of the two Similarity
  • Question and answer model training and answer selection units include:
  • Loss function building unit used to calculate the similarity between the question and the correct answer and the error between the similarity between the question and the wrong answer;
  • An optimization function building unit is used to adjust parameters in the model so that the similarity between the model and the correct answer to the model is greater than the similarity between the question and the incorrect answer;
  • An answer selection unit is used to select a correct answer from a medical question answering knowledge base using a trained model when processing a new question.
  • a storage medium based on Embodiment 1 stores a plurality of instructions, and the instructions are loaded by a processor to execute the steps of the medical automatic question answering method of Embodiment 1.
  • An electronic device based on Embodiment 3 includes: a storage medium of Embodiment 3; and a processor, configured to execute instructions in the storage medium.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • Public Health (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Epidemiology (AREA)
  • Molecular Biology (AREA)
  • Primary Health Care (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本发明公开了医疗自动问答方法及装置、存储介质、电子设备，属于自然语言处理技术领域及计算机人工智能领域，本发明要解决的技术问题为如何使用自然语言处理技术为患者在已有的问答结果中选择匹配的答案并将其推荐给用户，减少医生的工作量，提高问题解答的速度，采用的技术方案为：①一种医疗自动问答方法，该方法包括如下步骤：S1、构建医疗问答知识库；S2、构建问答模型训练数据集；S3、构建问答模型；S4、训练问答模型和答案选择。②一种医疗自动问答装置，该装置包括：医疗问答知识库构建单元、问答模型训练数据集生成单元、问答模型构建单元以及问答模型训练和答案选择单元。

Description

医疗自动问答方法及装置、存储介质、电子设备 技术领域
本发明涉及自然语言处理技术领域及计算机人工智能领域，具体地说是医疗自动问答方法及装置、存储介质、电子设备。
背景技术
随着当前互联网的快速发展，传统的医疗行业开始拥抱互联网，涌现了一大批方便患者的问答的网址，其中的在线医疗问答社区发展较为迅速，因为它极大地减少了患者向医生求助的环节，患者在互联网社区发布问题，由相关专业的医生解疑答惑，避免了传统就医环节中，前往医院、挂号、排队的流程。然而，随着在线提问的用户数量的增加，问题的增加速度远远大于解答的速度，主要包括两方面的原因：一方面是患者的数量远远高于专业医生的数量，并且在社区中解答问题的医生数量更少；另一方面是由于患者缺乏专业的医学知识，无法专业地描述问题，因此会对同样的问题提出不同的描述，这就造成了问答社区中存在着大量的重复问题。医生数量稀少的问题暂无解决办法，但是可以从问题重复现象出发进行处理，如何使用自然语言处理技术为患者在已有的问答结果中选择匹配的答案并将其推荐给用户，减少医生的工作量，提高问题解答的速度是目前现有技术中急需解决的技术问题。
专利号为CN107516110A的专利文献公开了一种基于集成卷积编码的医疗问答语义聚类方法，所述方法包括以下步骤：步骤1：从医疗平台上获取医疗问答数据，对医疗问答数据集进行预处理，并得到输入矩阵；步骤2：用卷积编码网络对不同的输入矩阵选取不同的卷积核进行核聚类，对核聚类后的聚类质量和多样性进行计算，根据聚类质量和多样性挑选出表示文本特征最好的n个卷积核；步骤3：将步骤2中挑选的卷积核分别通过卷积神经网络来进行训练操作；步骤4：融合不同卷积核的特征表示结果；步骤5：将融合后的特征表示结果输入自编码机，进行输入重构训练得到最佳特征表示；步骤6：将编码得到的最佳特征表示进行聚类，得到最终医疗文本语义聚类结果。但是该技术方案仅是对医疗文本的分类，无法精准获取医疗问题的精准答案。
发明内容
本发明的技术任务是提供一种医疗自动问答方法及装置、存储介质、电子 设备，来解决如何使用自然语言处理技术为患者在已有的问答结果中选择匹配的答案并将其推荐给用户，减少医生的工作量，提高问题解答的速度的问题。
本发明的技术任务是按以下方式实现的，一种医疗自动问答方法，该方法包括如下步骤：
S1、构建医疗问答知识库：从互联网的医疗问答社区爬取问答对，并对问答对进行预处理构建成医疗问答知识库；
S2、构建问答模型训练数据集；对于每个医疗问题，在步骤S1中的医疗问答知识库中至少有一个正确答案，将正确答案与问题构建训练正例，将除正确答案以外的答案与问题构建训练负例；其中，负例的数量根据医疗问答知识库的大小设定；
S3、构建问答模型：利用卷积神经网络构建问答模型；
S4、训练问答模型和答案选择：在步骤S2所得到问答模型训练数据集中对步骤S3构建的问题模型进行训练。
作为优选，所述步骤S1中构建医疗问答知识库的具体步骤如下：
S101、利用爬虫技术，从互联网的医疗问答社区爬取问答对：在互联网上有着大量的医疗问答社区，其内容是开放可获取的；医疗问答社区均采用Web形式进行展示，利用Scrapy工具获取问题和所对应的答案，构建问答知识库；
S102、对步骤S101中从互联网的医疗问答社区爬取的问答对进行预处理，主要包括拆分有多个答案的问题以及对问题和答案进行断字处理；
S10201、拆分有多个答案的问题就是将问题与答案拆分成一一对应的形式，具体来说：医疗问答社区中同一个问题至少有两个正确答案，将问题与答案拆分成多个问答对；
S10202、对问题和答案进行断字处理就是对问题和答案的内容进行断字处理，具体来说：将医疗问答社区中每个问题和答案，以汉字字符为基本单位，将每个字之间用空格分开且保留数字和标点符号。
作为优选，所述步骤S2中构建问答模型训练数据集的具体步骤如下：
S201、构建训练正例：将问题与其所对应的正确答案进行组合，构建训练正例，训练正例的形式为：(question,answer,1)，其中，question指问题，answer指该问题所对应的正确答案，1表示正例；
S202、构建训练负例：在医疗问答知识库中选中步骤S201的问题，再从医疗问答知识库中随机选择除步骤S201以外的问题及其该问题所对应的答案，将步骤S201的问题与随机选择的除步骤S201以外的问题所对应的答案组合， 构建训练负例，训练负例的形式为：(question,bad-answer,0)，其中，question指问题，bad-answer指随机选择的除步骤S201以外的问题所对应的答案，0表示负例；
S203、构建训练数据集：将步骤S201和步骤S202所的的全部训练正例和训练负例的样本组合，并随机打乱训练正例和训练负例的样本构建最终的训练数据集；其中，训练正例和训练负例均包含三个维度，即question、answer或bad-answer、1或0，前两个维度根据实际包含的词语构建二维的张量表示，第三个维度构建一个向量表示。
作为优选，所述步骤S3中构建问答模型的具体步骤如下：
S301、构建字符映射转换表：根据步骤S1中预处理以后构建的医疗问答知识库，构建字符表，并将字符表中的字符映射为唯一的数字表示，例如可按照字符加入字符表的顺序从1开始依次递增，构建字符映射转换表；
S302、构建输入层：输入层包括三个输入内容，从问题对应的训练正例中获取question和answer，从训练负例中获取bad-answer，输入层的形式为(question,answer,bad-answer)；再根据步骤S301中得到的字符映射转换表将输入数据中的字符转化为数字表示；
S303、构建字符向量映射层：指定字符向量的维度大小为embedding_dim，步骤S301中获取的字符表大小为vocab_num，使用[-1,1]的均匀分布定义(vocab_num+1)×embedding_dim的二维矩阵并将其作为模型的字符向量映射层的权重矩阵，其中，每一行对应医疗问答知识库中单个字符的向量表示，其中，第0行表示在知识库中没有出现的字符；
S304、构建多尺度卷积层和堆叠卷积层：多尺度卷积层是利用不同尺寸的卷积核与经过步骤S303处理后的数据做卷积操作，捕获问题或答案里的字符级别的n-gram特征作为输出；堆叠卷积层是将不同尺寸的卷积核堆叠起来，将经过步骤S303处理后的数据自下而上地与各个卷积核做卷积操作，捕获问题或答案里句子级别的语义特征作为输出；再将多尺度卷积层的输出与堆叠卷积层的输出联接起来，作为问题或答案的特征向量；
S305、构建相似度计算层：使用余弦相似度公式计算得到问题和正确答案的相似度sim_pos以及问题与错误答案的相似度sim_neg；余弦相似度的计算公式如下所示：
Figure PCTCN2019080298-appb-000001
其中，‖·‖表示L2范数，Q和A分别指问题和正确答案或错误答案经过问答模型处理后获得的向量表示；Q表示vector_question的向量，A表示vector_answer或vector_bad-answer的向量。
作为优选，所述步骤S304中多尺度卷积层计算的具体步骤如下：
(1)、定义卷积核的数量为n，卷积核的集合C＝{c 1,c 2,…，c i,…，c n}，其中，c i表示卷积核的尺寸；
(2)、根据步骤S303处理后的数据与卷积核的集合C内的卷积核分别做卷积运算得到结果O＝{o 1,o 2,…，o i,…，o n}；
(3)、对每个o i依次使用ReLU激活函数，做基于字符向量维度的取最大值操作，得到最终的输出O multi。ReLU激活函数的公式如下为：
Figure PCTCN2019080298-appb-000002
所述步骤S304中堆叠卷积层计算的具体步骤如下：
(1)、定义卷积核的数量为m，卷积核的集合C＝{c 1,c 2,…，c i,…，c m}，其中，c i表示卷积核的尺寸；
(2)、根据步骤S303处理后的数据与卷积核的集合C内的第1个卷积核c 1做卷积操作，再经过ReLU激活得到o 1；
(3)、将o 1与卷积核的集合C内的第2个卷积核c 2做卷积操作，再经过ReLU激活得到o 2；
(4)、如此重复操作直至得到o m，对其做基于字符向量维度的取最大值操作，得到最终的输出O stack。
作为优选，所述步骤S4中训练问答模型和答案选择的具体步骤如下：
S401、构建损失函数：根据步骤S305得到的问题和正确答案的相似度sim_pos以及问题与错误答案的相似度sim_neg使用Max-margin损失函数计算模型的损失值，用损失值来优化模型的结果；Max-margin损失函数的公式为：
L＝max(0，margin-sim_pos+sim_neg)；
其中，margin为超参数，根据实际数据调整；max(a,b)的形式表示取a和b之间最大值；
S402、构建优化函数：使用Adam算法作为模型的优化函数；其中，学习率参数设置为0.01，其他超参数使用PyTorch中的默认值即可，其他超参数包 括如下：
betas＝(0.9,0.999)，betas用于计算梯度及其平方的运行平均值的系数；
eps＝1e-8，eps用于添加到分母中以增加数值稳定性；
weight_decay＝0，weight_decay用于控制权重衰减。
S403、答案选择：将一个新的问题与医疗问答知识库中的所有候选答案组成输入样本；再分别送入利用步骤S3训练完成的问答模型，计算问题与候选答案之间相似度的数值；最后从所有结果中选取相似度值最高的一组作为新问题最有可能的答案并该答案推送给用户。
一种医疗自动问答装置，该装置包括：
医疗问答知识库构建单元，用于利用爬虫技术，从互联网上的医疗问答社区爬取问答对，并对医疗问答知识库的问题和答案进行预处理，预处理包括拆分有多个答案的问题以及对问题和答案的内容进行断字处理，构建医疗问答知识库；
问答模型训练数据集生成单元，用于根据医疗问答知识库构建训练正例和训练负例，进而构建训练数据集；
问答模型构建单元，用于构建字符映射转换表、输入层、字符向量映射层、多尺度卷积层和堆叠卷积层以及相似度计算层，进而构建问答模型；
问答模型训练和答案选择单元，用于构建损失函数和优化函数，完成答案选择。
作为优选，所述医疗问答知识库构建单元包括：
爬虫单元，用于从互联网上抓取医疗问答社区中的内容，将结果保存为问答对的形式；
爬虫数据处理单元，用于将爬虫获取的问答对拆分为多个单独对应的问答对，并将问题和答案句子进行断字处理，构建问答知识库；
所述问答模型训练数据集生成单元包括：
训练正例构建单元，用于将问题与其所对应的答案进行组合，构建为训练正例；
训练负例构建单元，用于从问答知识库选中一个问题q1，再从医疗问答知识库中随机选择一个与问题q1不同的问题q2，将q1与q2的答案a2组合，构建训练负例；
训练数据集构建单元，用于从训练正例构建单元和训练负例构建单元所得 的全部问答的训练正例样本、训练负例样本组合在一起，并随机打乱顺序构成最终的训练数据集；
所述问答模型构建单元包括：
字符映射转换表构建单元，用于将医疗问答知识库中的所有字符构建为字符表，并将字符表中的每个字符映射为唯一的数字表示，可按照字符加入字符表的顺序从1开始依次递增，构建字符映射转换表。
输入层构建单元，用于从问题对应的训练正例和训练负例中获取输入数据，分别获取训练正例中的问题question和答案answer以及训练负例中的答案，将其形式化为：(question,answer,bad-answer)；
字符向量映射层构建单元，用于构建并存储字符映射转换表中每个字符的向量表示；
多尺度卷积层和堆叠卷积层构建单元，用于生成多尺度卷积层和堆叠卷积层，多尺度卷积层用于捕获问题或答案里字符级别的n-gram特征并输出，堆叠卷积层用于捕获问题或答案里句子级别的语义特征并输出；最后将多尺度卷积层的输出与堆叠卷积层的输出联接起来，作为问题或答案的特征向量；
相似度评价构建单元，用于将多尺度卷积层构建单元和堆叠卷积层构建单元的输出的问题与答案的向量表示，进行联接，获得问题与答案的最终表示，并计算两者的余弦相似度；
所述问答模型训练和答案选择单元包括：
损失函数构建单元，用于计算问题和正确答案的相似度以及问题和错误答案的相似度之间的误差；
优化函数构建单元，用于调整模型中的参数，使模型对问题与正确答案的相似度大于问题与错误答案的相似度；
答案选择单元，用于处理新的问题时，从医疗问答知识库中使用训练好的模型选择正确的答案。
一种存储介质，其中存储有多条指令，所述指令由处理器加载，执行上述的医疗自动问答方法的步骤。
一种电子设备，所述电子设备包括：
上述的存储介质；以及
处理器，用于执行所述存储介质中的指令。
本发明的医疗自动问答方法及装置、存储介质、电子设备具有以下优点：
(1)、本发明通过爬虫技术能够完成医疗问答知识库的采集和预处理，生 成医疗问答知识库，为后续建立医疗问答模型提供数据支持；
(2)、本发明能够有效地对已有问答对进行建模构建问答模型，并提取问答对之间的多层语义关联关系；
(3)、本发明能够定义并实现一套完整的端到端的自动问答模型，直接为用户的问题推荐相似度最高的答案；
(4)、本发明能够利用医疗问题中的语义关系得到医疗问答模型，从而解决医疗问答中的答案选择问题；
(5)、本发明对问题和答案采用字符级别的向量表征，并结合了卷积神经网络的强大的局部抽象能力，能够在一定程度上减小错别字的干扰；
(6)、本发明将多尺度卷积层与堆叠卷积层组合共同表示医疗问答数据库中字符及句子的n-gram特征，能够更有效的提取医疗问题的特征，进而获取更加准确的答案；
(7)、本发明使用自然语言处理技术为患者在已有的问答结果中选择匹配的答案并将其推荐给用户，减少医生的工作量，提高问题解答的速度，是计算机人工智能重要突破。
附图说明
下面结合附图对本发明进一步说明。
附图1为医疗自动问答方法流程框图；
附图2为步骤S1构建医疗问答知识库的流程框图；
附图3为步骤S2构建问答模型训练数据集的流程框图；
附图4为步骤S3构建问答模型的流程框图；
附图5为步骤S4构建训练问答模型和答案选择的流程框图；
附图6为步骤S304中多尺度卷积层和堆叠卷积层计算的流程框图；
附图7为医疗自动问答装置的结构框图。
具体实施方式
参照说明书附图和具体实施例对本发明的医疗自动问答方法及装置、存储介质、电子设备作以下详细地说明。
实施例1：
如附图1所示，本发明的医疗自动问答方法,该方法包括如下步骤：
S1、构建医疗问答知识库：从互联网的医疗问答社区爬取问答对，并对问答对进行预处理构建成医疗问答知识库；
S2、构建问答模型训练数据集；对于每个医疗问题，在步骤S1中的医疗 问答知识库中有两个正确答案，将正确答案与问题构建训练正例，将除正确答案以外的答案与问题构建训练负例；其中，负例的数量根据医疗问答知识库的大小设定；
S3、构建问答模型：利用卷积神经网络构建问答模型；
S4、训练问答模型和答案选择：在步骤S2所得到问答模型训练数据集中对步骤S3构建的问题模型进行训练。
如附图2所示，步骤S1中构建医疗问答知识库的具体步骤如下：
S101、利用爬虫技术，从互联网的医疗问答社区爬取问答对：在互联网上有着大量的医疗问答社区，其内容是开放可获取的；医疗问答社区均采用Web形式进行展示，利用Scrapy工具获取问题和所对应的答案，构建问答知识库；
医疗问答社区中的问答对示例，如下表所示(将该问题作为示例问题，下面表述中统一称为“示例问题)”：
Figure PCTCN2019080298-appb-000003
Figure PCTCN2019080298-appb-000004
S102、对步骤S101中从互联网的医疗问答社区爬取的问答对进行预处理，主要包括拆分有多个答案的问题以及对问题和答案进行断字处理；
S10201、拆分有多个答案的问题就是将问题与答案拆分成一一对应的形式，具体来说：医疗问答社区中同一个问题有两个正确答案，将问题与答案拆分成两个问答对；
如上述示例问题有两个正确答案，将示例问题与其对应的正确答案拆分为两个问答对，如下表所示：
Figure PCTCN2019080298-appb-000005
Figure PCTCN2019080298-appb-000006
S10202、对问题和答案进行断字处理就是对问题和答案的内容进行断字处理，具体来说：将医疗问答社区中每个问题和答案，以汉字字符为基本单位，将每个字之间用空格分开且保留数字和标点符号。
对示例问题进行断字处理，得到“前几天孩子得了病毒性脑炎在医院住了十来天出院了，过了一周去医院复查，血小板计数509偏高，怎么办？孩子现在11个月了。现在吃的药只有双嘧达莫片。请问医生孩子这是怎么了？血小板怎么才能降下来？谢谢！”。
如附图3所示，步骤S2中构建问答模型训练数据集的具体步骤如下：
S201、构建训练正例：将问题与其所对应的正确答案进行组合，构建训练正例，训练正例的形式为：(question,answer,1)，其中，question指问题，answer指该问题所对应的正确答案，1表示正例；
对上述经常断字处理后的问题及其对应正确答案组合构建训练正例：
(“前几天孩子得了病毒性脑炎在医院住了十来天出院了，过了一周去医院复查，血小板计数509偏高，怎么办？孩子现在11个月了。现在吃的药只有双嘧达莫片。请问医生孩子这是怎么了？血小板怎么才能降下来？谢谢！”,“你好孩子得了病毒性脑炎在医院住了十几天复查发现血小板计数偏高指导意见：建议你积极接受医院的治疗病毒性脑炎血小板偏高是普遍的现象，除了采取吃药最好进行输液治疗”,1)。
S202、构建训练负例：在医疗问答知识库中选中步骤S201的问题，再从医疗问答知识库中随机选择除步骤S201以外的问题及其该问题所对应的答案， 将步骤S201的问题与随机选择的除步骤S201以外的问题所对应的答案组合，构建训练负例，训练负例的形式为：(question,bad-answer,0)，其中，question指问题，bad-answer指随机选择的除步骤S201以外的问题所对应的答案，0表示负例；
如：示例问题q 1，再从医疗问答知识库中随机选择一个与示例问题q 1不同的问题q 2，将q 1与q 2的答案a 2组合，可构建训练负例，可形式化为：
(“前几天孩子得了病毒性脑炎在医院住了十来天出院了，过了一周去医院复查，血小板计数509偏高，怎么办？孩子现在11个月了。现在吃的药只有双嘧达莫片。请问医生孩子这是怎么了？血小板怎么才能降下来？谢谢！”,“你好,手足口病是一种肠道病毒病,具有肠道病毒感染的共同特征.主要侵犯手,足,口,臀四个部位,口腔溃疡疼痛,患儿流涎拒食.本病经1周后可自愈,一般对症治疗,可应用抗病毒药如阿昔洛韦利巴韦林(三氮唑核苷).中药有大青叶,夏枯草,或导赤散加减.还可以用双黄连口服液，严重的话，是需要到医院检查后治疗的。”,0)。
S203、构建训练数据集：将步骤S201和步骤S202所的的全部训练正例和训练负例的样本组合，并随机打乱训练正例和训练负例的样本构建最终的训练数据集；其中，训练正例和训练负例均包含三个维度，即question、answer或bad-answer、1或0，前两个维度根据实际包含的词语构建二维的张量表示，第三个维度构建一个向量表示。
如：使用vector_question、vector_answer、vector_posneg分别表示训练数据集的问题的向量表示、答案的向量表示、正负例标志的向量表示。假定每个字符用一个embedding_dim维的向量表示，则vector_question和vector_answer的张量维度为word_num×embedding_dim，其中，word_num指question、answer或bad-answer中字符的个数。
如附图4所示，步骤S3中构建问答模型的具体步骤如下：
S301、构建字符映射转换表：根据步骤S1中预处理以后构建的医疗问答知识库，构建字符表，并将字符表中的字符映射为唯一的数字表示，例如可按照字符加入字符表的顺序从1开始依次递增，构建字符映射转换表；
以进行断字处理后的示例问题为例，构建字符表及字符映射转换表如下：
字符 映射 字符 映射 字符 映射 字符 映射 字符 映射
前 1 医 14 小 27 个 40 请 53
几 2 院 15 板 28 月 41 问 54
天 3 住 16 计 29 。 42 生 55
孩 4 十 17 数 30 吃 43 这 56
子 5 来 18 509 31 的 44 是 57
得 6 ， 19 偏 32 药 45 才 58
了 7 过 20 高 33 只 46 能 59
病 8 一 21 怎 34 有 47 降 60
毒 9 周 22 么 35 双 48 下 61
性 10 去 23 办 36 嘧 49 谢 62
脑 11 复 24 ？ 37 达 50 ！ 63
炎 12 查 25 现 38 莫 51    
在 13 血 26 11 39 片 52    
S302、构建输入层：输入层包括三个输入内容，从问题对应的训练正例中获取question和answer，从训练负例中获取bad-answer，输入层的形式为(question,answer,bad-answer)；再根据步骤S301中得到的字符映射转换表将输入数据中的字符转化为数字表示；
示例问题的训练正例和训练负例构建一条输入数据，结果如下：
(“前几天孩子得了病毒性脑炎在医院住了十来天出院了，过了一周去医院复查，血小板计数509偏高，怎么办？孩子现在11个月了。现在吃的药只有双嘧达莫片。请问医生孩子这是怎么了？血小板怎么才能降下 来？谢谢！”,“你好孩子得了病毒性脑炎在医院住了十几天复查发现血小板计数偏高指导意见：建议你积极接受医院的治疗病毒性脑炎血小板偏高是普遍的现象，除了采取吃药最好进行输液治疗”,“你好,手足口病是一种肠道病毒病,具有肠道病毒感染的共同特征.主要侵犯手,足,口,臀四个部位,口腔溃疡疼痛,患儿流涎拒食.本病经1周后可自愈,一般对症治疗,可应用抗病毒药如阿昔洛韦利巴韦林(三氮唑核苷).中药有大青叶,夏枯草,或导赤散加减.还可以用双黄连口服液，严重的话，是需要到医院检查后治疗的。”)
根据词表中的映射将上述的输入数据转换为数值表示(由于实际情况中词表数量大约有5000个，暂不在此全部展示)，结果如下：
(“112,332,70,167,73,274,24,19,302,27,213,77,33,30,86,707,24,489,98,70,55,86,24,2,61,24,10,163,122,30,86,120,29,2,32,110,750,883,408,0,584,168,2,142,52,318,68,167,73,54,33,0,48,66,24,4,54,33,57,3,28,323,6,488,1869,599,905,237,4,244,131,30,22,167,73,40,5,142,52,24,68,32,110,750,142,52,340,23,528,65,98,68,356,356,201”,“17,9,167,73,274,24,19,302,27,213,77,33,30,86,707,24,489,332,70,120,29,31,54,32,110,750,883,408,584,168,111,84,45,103,51,42,44,17,347,449,437,267,30,86,3,12,14,19,302,27,213, 77,549,1815,763,584,168,5,991,1832,3,54,491,2,249,24,416,484,57,28,132,9,106,118,602,208,12,14”,“17,9,15,91,385,146,19,5,10,108,270,176,19,302,19,15,445,6,270,176,19,302,83,189,3,945,196,407,780,26,181,16,1284,1253,91,15,385,15,146,15,1050,359,48,88,292,15,146,431,864,933,155,67,15,199,153,156,2437,2171,39,26,469,19,37,87,163,36,8,247,434,15,10,127,99,35,12,14,15,8,101,20,259,19,302,28,53,703,1603,1061,1344,497,724,1344,770,828,275,1994,729,586,1616,821,26,100,28,6,56,776,809,15,1196,1977,826,15,80,84,1770,659,188,303,26,85,8,13,20,488,326,666,146,105,208,2,304,141,3,262,2,5,126,16,81,30,86,38,29,36,12,14,3,4”)。
S303、构建字符向量映射层：指定字符向量的维度大小为embedding_dim，步骤S301中获取的字符表大小为vocab_num，使用[-1,1]的均匀分布定义(vocab_num+1)×embedding_dim的二维矩阵并将其作为模型的字符向量映射层的权重矩阵，其中，每一行对应医疗问答知识库中单个字符的向量表示，其中，第0行表示在知识库中没有出现的字符；
在PyTorch中，示例的代码实现如下：
matrix＝numpy.random.uniform(-1,1,((word_num+1),embedding_di m))
matrix[0,:]＝0
embedding_layer＝torch.nn.Embedding((word_num+1),embedding_di m)。
该层是问答模型的通用层，存储了全部字符对应的字符向量；对于quest ion、answer、bad-answer的处理操作均完全相同。
S304、构建多尺度卷积层和堆叠卷积层：多尺度卷积层是利用不同尺寸的卷积核与经过步骤S303处理后的数据做卷积操作，捕获问题或答案里的字符 级别的n-gram特征作为输出；堆叠卷积层是将不同尺寸的卷积核堆叠起来，将经过步骤S303处理后的数据自下而上地与各个卷积核做卷积操作，捕获问题或答案里句子级别的语义特征作为输出；再将多尺度卷积层的输出与堆叠卷积层的输出联接起来，作为问题或答案的特征向量；
S305、构建相似度计算层：使用余弦相似度公式计算得到问题和正确答案的相似度sim_pos以及问题与错误答案的相似度sim_neg；余弦相似度的计算公式如下所示：
Figure PCTCN2019080298-appb-000007
其中，‖·‖表示L2范数，Q和A分别指问题和正确答案或错误答案经过问答模型处理后获得的向量表示；Q表示vector_question的向量，A表示vector_answer或vector_bad-answer的向量。
如附图6所示，步骤S304中多尺度卷积层计算的具体步骤如下：
(1)、定义卷积核的数量为n，卷积核的集合C＝{c 1,c 2,…，c i,…，c n}，其中，c i表示卷积核的尺寸；
(2)、根据步骤S303处理后的数据与卷积核的集合C内的卷积核分别做卷积运算得到结果O＝{o 1,o 2,…，o i,…，o n}；
(3)、对每个o i依次使用ReLU激活函数，做基于字符向量维度的取最大值操作，得到最终的输出O multi。ReLU激活函数的公式如下为：
Figure PCTCN2019080298-appb-000008
步骤S304中堆叠卷积层计算的具体步骤如下：
(1)、定义卷积核的数量为m，卷积核的集合C＝{c 1,c 2,…，c i,…，c m}，其中，c i表示卷积核的尺寸；
(2)、根据步骤S303处理后的数据与卷积核的集合C内的第1个卷积核c 1做卷积操作，再经过ReLU激活得到o 1；
(3)、将o 1与卷积核的集合C内的第2个卷积核c 2做卷积操作，再经过ReLU激活得到o 2；
(4)、如此重复操作直至得到o m，对其做基于字符向量维度的取最大值操作，得到最终的输出O stack。
在PyTorch中定义卷积、ReLU激活函数、基于字符向量维度的最大值操 作的代码如下所示：
卷积代码：conv＝torch.nn.Conv1d(embedding_dim,output_channel,c_i)，其中output_channel为输出通道大小，用户可自定义。
ReLU激活函数代码：relu＝torch.nn.ReLU()。
基于字符向量维度的最大值操作代码：max＝torch.nn.MaxPool1d()。
如附图5所示，所述步骤S4中训练问答模型和答案选择的具体步骤如下：
S401、构建损失函数：根据步骤S305得到的问题和正确答案的相似度sim_pos以及问题与错误答案的相似度sim_neg使用Max-margin损失函数计算模型的损失值，用损失值来优化模型的结果；Max-margin损失函数的公式为：
L＝max(0，margin-sim_pos+sim_neg)；
其中，margin为超参数，根据实际数据调整；max(a,b)的形式表示取a和b之间最大值；
在PyTorch中，定义max-margin函数可用如下代码实现：
loss＝torch.nn.MarginRankingLoss(margin)。
S402、构建优化函数：使用Adam算法作为模型的优化函数；其中，学习率参数设置为0.01，其他超参数使用PyTorch中的默认值即可；其中，超参数是指在开始训练过程之前，需要人工设置值的参数；该参数不能通过训练而自动优化；根据实际数据集的不同，该参数需要用户自行人工设置。
在PyTorch中，定义Adam优化函数可用如下代码实现：
optim＝torch.optim.Adam(lr＝0.001)。
S403、答案选择：将一个新的问题与医疗问答知识库中的所有候选答案组成输入样本；再分别送入利用步骤S3训练完成的问答模型，计算问题与候选答案之间相似度的数值；最后从所有结果中选取相似度值最高的一组作为新问题最有可能的答案并该答案推送给用户。
实施例2：
如附图7所示，基于实施例1的医疗自动问答装置，该装置包括：
医疗问答知识库构建单元，用于利用爬虫技术，从互联网上的医疗问答社区爬取问答对，并对医疗问答知识库的问题和答案进行预处理，预处理包括拆分有多个答案的问题以及对问题和答案的内容进行断字处理，构建医疗问答知识库；
问答模型训练数据集生成单元，用于根据医疗问答知识库构建训练正例和 训练负例，进而构建训练数据集；
问答模型构建单元，用于构建字符映射转换表、输入层、字符向量映射层、多尺度卷积层和堆叠卷积层以及相似度计算层，进而构建问答模型；
问答模型训练和答案选择单元，用于构建损失函数和优化函数，完成答案选择。
其中，医疗问答知识库构建单元包括：
爬虫单元，用于从互联网上抓取医疗问答社区中的内容，将结果保存为问答对的形式；
爬虫数据处理单元，用于将爬虫获取的问答对拆分为多个单独对应的问答对，并将问题和答案句子进行断字处理，构建问答知识库；
问答模型训练数据集生成单元包括：
训练正例构建单元，用于将问题与其所对应的答案进行组合，构建为训练正例；
训练负例构建单元，用于从问答知识库选中一个问题q1，再从医疗问答知识库中随机选择一个与问题q1不同的问题q2，将q1与q2的答案a2组合，构建训练负例；
训练数据集构建单元，用于从训练正例构建单元和训练负例构建单元所得的全部问答的训练正例样本、训练负例样本组合在一起，并随机打乱顺序构成最终的训练数据集；
问答模型构建单元包括：
字符映射转换表构建单元，用于将医疗问答知识库中的所有字符构建为字符表，并将字符表中的每个字符映射为唯一的数字表示，可按照字符加入字符表的顺序从1开始依次递增，构建字符映射转换表。
输入层构建单元，用于从问题对应的训练正例和训练负例中获取输入数据，分别获取训练正例中的问题question和答案answer以及训练负例中的答案，将其形式化为：(question,answer,bad-answer)；
字符向量映射层构建单元，用于构建并存储字符映射转换表中每个字符的向量表示；
多尺度卷积层和堆叠卷积层构建单元，用于生成多尺度卷积层和堆叠卷积层，多尺度卷积层用于捕获问题或答案里字符级别的n-gram特征并输出，堆叠卷积层用于捕获问题或答案里句子级别的语义特征并输出；最后将多尺度卷积层的输出与堆叠卷积层的输出联接起来，作为问题或答案的特征向量；
相似度评价构建单元，用于将多尺度卷积层构建单元和堆叠卷积层构建单元的输出的问题与答案的向量表示，进行联接，获得问题与答案的最终表示，并计算两者的余弦相似度；
问答模型训练和答案选择单元包括：
损失函数构建单元，用于计算问题和正确答案的相似度以及问题和错误答案的相似度之间的误差；
优化函数构建单元，用于调整模型中的参数，使模型对问题与正确答案的相似度大于问题与错误答案的相似度；
答案选择单元，用于处理新的问题时，从医疗问答知识库中使用训练好的模型选择正确的答案。
实施例3：
基于实施例1的一种存储介质，其中存储有多条指令，指令由处理器加载，执行实施例1的医疗自动问答方法的步骤。
实施例4：
基于实施例3的一种电子设备，电子设备包括：实施例3的存储介质；以及处理器，用于执行所述存储介质中的指令。
最后应说明的是：以上各实施例仅用以说明本发明的技术方案，而非对其限制；尽快参照前述各实施例对本发明进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分或者全部技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本发明各实施例技术方案的范围。

Claims (10)

  1. 一种医疗自动问答方法，其特征在于，该方法包括如下步骤：
    S1、构建医疗问答知识库：从互联网的医疗问答社区爬取问答对，并对问答对进行预处理构建成医疗问答知识库；
    S2、构建问答模型训练数据集；对于每个医疗问题，在步骤S1中的医疗问答知识库中至少有一个正确答案，将正确答案与问题构建训练正例，将除正确答案以外的答案与问题构建训练负例；
    S3、构建问答模型：利用卷积神经网络构建问答模型；
    S4、训练问答模型和答案选择：在步骤S2所得到问答模型训练数据集中对步骤S3构建的问题模型进行训练。
  2. 根据权利要求1所述的医疗自动问答方法，其特征在于，所述步骤S1中构建医疗问答知识库的具体步骤如下：
    S101、利用爬虫技术，从互联网的医疗问答社区爬取问答对：医疗问答社区均采用Web形式进行展示，利用Scrapy工具获取问题和所对应的答案，构建问答知识库；
    S102、对步骤S101中从互联网的医疗问答社区爬取的问答对进行预处理，主要包括拆分有多个答案的问题以及对问题和答案进行断字处理；
    S10201、拆分有多个答案的问题就是将问题与答案拆分成一一对应的形式；
    S10202、对问题和答案进行断字处理就是对问题和答案的内容进行断字处理，具体来说：将医疗问答社区中每个问题和答案，以汉字字符为基本单位，将每个字之间用空格分开且保留数字和标点符号。
  3. 根据权利要求1所述的医疗自动问答方法，其特征在于，所述步骤S2中构建问答模型训练数据集的具体步骤如下：
    S201、构建训练正例：将问题与其所对应的正确答案进行组合，构建训练正例，训练正例的形式为：(question,answer,1)，其中，question指问题，answer指该问题所对应的正确答案，1表示正例；
    S202、构建训练负例：在医疗问答知识库中选中步骤S201的问题，再从医疗问答知识库中随机选择除步骤S201以外的问题及其该问题所对应的答案，将步骤S201的问题与随机选择的除步骤S201以外的问题所对应的答案组合，构建训练负例，训练负例的形式为：(question,bad-answer,0)，其中， question指问题，bad-answer指随机选择的除步骤S201以外的问题所对应的答案，0表示负例；
    S203、构建训练数据集：将步骤S201和步骤S202所的的全部训练正例和训练负例的样本组合，并随机打乱训练正例和训练负例的样本构建最终的训练数据集。
  4. 根据权利要求1所述的医疗自动问答方法，其特征在于，所述步骤S3中构建问答模型的具体步骤如下：
    S301、构建字符映射转换表：根据步骤S1中预处理以后构建的医疗问答知识库，构建字符表，并将字符表中的字符映射为唯一的数字表示；
    S302、构建输入层：输入层包括三个输入内容，从问题对应的训练正例中获取question和answer，从训练负例中获取bad-answer，输入层的形式为(question,answer,bad-answer)；再根据步骤S301中得到的字符映射转换表将输入数据中的字符转化为数字表示；
    S303、构建字符向量映射层：指定字符向量的维度大小为embedding_dim，步骤S301中获取的字符表大小为vocab_num，使用[-1,1]的均匀分布定义(vocab_num+1)×embedding_dim的二维矩阵并将其作为模型的字符向量映射层的权重矩阵，其中，每一行对应医疗问答知识库中单个字符的向量表示；
    S304、构建多尺度卷积层和堆叠卷积层：多尺度卷积层是利用不同尺寸的卷积核与经过步骤S303处理后的数据做卷积操作，捕获问题或答案里的字符级别的n-gram特征作为输出；堆叠卷积层是将不同尺寸的卷积核堆叠起来，将经过步骤S303处理后的数据自下而上地与各个卷积核做卷积操作，捕获问题或答案里句子级别的语义特征作为输出；再将多尺度卷积层的输出与堆叠卷积层的输出联接起来，作为问题或答案的特征向量；
    S305、构建相似度计算层：使用余弦相似度公式计算得到问题和正确答案的相似度sim_pos以及问题与错误答案的相似度sim_neg；余弦相似度的计算公式如下所示：
    Figure PCTCN2019080298-appb-100001
    其中，‖·‖表示L2范数，Q和A分别指问题和正确答案或错误答案经过问答模型处理后获得的向量表示；Q表示vector_question的向量，A表示vector_answer或vector_bad-answer的向量。
  5. 根据权利要求4所述的医疗自动问答方法，其特征在于，所述步骤 S304中多尺度卷积层计算的具体步骤如下：
    (1)、定义卷积核的数量为n，卷积核的集合C＝{c 1,c 2,…，c i,…，c n}，其中，c i表示卷积核的尺寸；
    (2)、根据步骤S303处理后的数据与卷积核的集合C内的卷积核分别做卷积运算得到结果O＝{o 1,o 2,…，o i,…，o n}；
    (3)、对每个o i依次使用ReLU激活函数，做基于字符向量维度的取最大值操作，得到最终的输出O multi。ReLU激活函数的公式如下为：
    Figure PCTCN2019080298-appb-100002
    所述步骤S304中堆叠卷积层计算的具体步骤如下：
    (1)、定义卷积核的数量为m，卷积核的集合C＝{c 1,c 2,…，c i,…，c m}，其中，c i表示卷积核的尺寸；
    (2)、根据步骤S303处理后的数据与卷积核的集合C内的第1个卷积核c 1做卷积操作，再经过ReLU激活得到o 1；
    (3)、将o 1与卷积核的集合C内的第2个卷积核c 2做卷积操作，再经过ReLU激活得到o 2；
    (4)、如此重复操作直至得到o m，对其做基于字符向量维度的取最大值操作，得到最终的输出O stack。
  6. 根据权利要求1所述的医疗自动问答方法，其特征在于，所述步骤S4中训练问答模型和答案选择的具体步骤如下：
    S401、构建损失函数：根据步骤S305得到的问题和正确答案的相似度sim_pos以及问题与错误答案的相似度sim_neg使用Max-margin损失函数计算模型的损失值，用损失值来优化模型的结果；Max-margin损失函数的公式为：
    L＝max(0，margin-sim_pos+sim_neg)；
    其中，margin为超参数，根据实际数据调整；max(a,b)的形式表示取a和b之间最大值；
    S402、构建优化函数：使用Adam算法作为模型的优化函数；
    S403、答案选择：将一个新的问题与医疗问答知识库中的所有候选答案组成输入样本；再分别送入利用步骤S3训练完成的问答模型，计算问题与候选答案之间相似度的数值；最后从所有结果中选取相似度值最高的一组作为新问 题最有可能的答案并该答案推送给用户。
  7. 一种医疗自动问答装置，其特征在于，该装置包括：
    医疗问答知识库构建单元，用于利用爬虫技术，从互联网上的医疗问答社区爬取问答对，并对医疗问答知识库的问题和答案进行预处理，预处理包括拆分有多个答案的问题以及对问题和答案的内容进行断字处理，构建医疗问答知识库；
    问答模型训练数据集生成单元，用于根据医疗问答知识库构建训练正例和训练负例，进而构建训练数据集；
    问答模型构建单元，用于构建字符映射转换表、输入层、字符向量映射层、多尺度卷积层和堆叠卷积层以及相似度计算层，进而构建问答模型；
    问答模型训练和答案选择单元，用于构建损失函数和优化函数，完成答案选择。
  8. 根据权利要求7所述的医疗自动问答装置，其特征在于，所述医疗问答知识库构建单元包括：
    爬虫单元，用于从互联网上抓取医疗问答社区中的内容，将结果保存为问答对的形式；
    爬虫数据处理单元，用于将爬虫获取的问答对拆分为多个单独对应的问答对，并将问题和答案句子进行断字处理，构建问答知识库；
    所述问答模型训练数据集生成单元包括：
    训练正例构建单元，用于将问题与其所对应的答案进行组合，构建为训练正例；
    训练负例构建单元，用于从问答知识库选中一个问题q 1，再从医疗问答知识库中随机选择一个与问题q 1不同的问题q 2，将q 1与q 2的答案a 2组合，构建训练负例；
    训练数据集构建单元，用于从训练正例构建单元和训练负例构建单元所得的全部问答的训练正例样本、训练负例样本组合在一起，并随机打乱顺序构成最终的训练数据集；
    所述问答模型构建单元包括：
    字符映射转换表构建单元，用于将医疗问答知识库中的所有字符构建为字符表，并将字符表中的每个字符映射为唯一的数字表示，构建字符映射转换表。
    输入层构建单元，用于从问题对应的训练正例和训练负例中获取输入数据，分别获取训练正例中的问题question和答案answer以及训练负例中的答 案，将其形式化为：(question,answer,bad-answer)；
    字符向量映射层构建单元，用于构建并存储字符映射转换表中每个字符的向量表示；
    多尺度卷积层和堆叠卷积层构建单元，用于生成多尺度卷积层和堆叠卷积层，多尺度卷积层用于捕获问题或答案里字符级别的n-gram特征并输出，堆叠卷积层用于捕获问题或答案里句子级别的语义特征并输出；最后将多尺度卷积层的输出与堆叠卷积层的输出联接起来，作为问题或答案的特征向量；
    相似度评价构建单元，用于将多尺度卷积层构建单元和堆叠卷积层构建单元的输出的问题与答案的向量表示，进行联接，获得问题与答案的最终表示，并计算两者的余弦相似度；
    所述问答模型训练和答案选择单元包括：
    损失函数构建单元，用于计算问题和正确答案的相似度以及问题和错误答案的相似度之间的误差；
    优化函数构建单元，用于调整模型中的参数，使模型对问题与正确答案的相似度大于问题与错误答案的相似度；
    答案选择单元，用于处理新的问题时，从医疗问答知识库中使用训练好的模型选择正确的答案。
  9. 一种存储介质，其中存储有多条指令，其特征在于，所述指令由处理器加载，执行权利要求1-6中所述的医疗自动问答方法的步骤。
  10. 一种电子设备，其特征在于，所述电子设备包括：
    权利要求9所述的存储介质；以及
    处理器，用于执行所述存储介质中的指令。
PCT/CN2019/080298 2018-08-17 2019-03-29 医疗自动问答方法及装置、存储介质、电子设备 WO2020034642A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2019322953A AU2019322953B2 (en) 2018-08-17 2019-03-29 Method, system, storage medium and electric device of medical automatic question answering

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810939302.3A CN108932349B (zh) 2018-08-17 2018-08-17 医疗自动问答方法及装置、存储介质、电子设备
CN201810939302.3 2018-08-17

Publications (1)

Publication Number Publication Date
WO2020034642A1 true WO2020034642A1 (zh) 2020-02-20

Family

ID=64446026

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/080298 WO2020034642A1 (zh) 2018-08-17 2019-03-29 医疗自动问答方法及装置、存储介质、电子设备

Country Status (4)

Country Link
CN (1) CN108932349B (zh)
AU (1) AU2019322953B2 (zh)
LU (1) LU101290B1 (zh)
WO (1) WO2020034642A1 (zh)

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200151542A1 (en) * 2018-11-12 2020-05-14 Alibaba Group Holding Limited Question and answer matching method, system and storage medium
CN111339277A (zh) * 2020-02-28 2020-06-26 中国工商银行股份有限公司 基于机器学习的问答交互方法及装置
CN111939529A (zh) * 2020-08-25 2020-11-17 山东大学齐鲁医院 一种基于肌力测量装置的主动式康复训练方法、系统
CN112084312A (zh) * 2020-07-23 2020-12-15 江苏海洋大学 一种基于知识图构建的智能客服系统
CN112131393A (zh) * 2020-08-11 2020-12-25 淮阴工学院 一种基于bert和相似度算法的医疗知识图谱问答系统构建方法
CN112200216A (zh) * 2020-09-03 2021-01-08 上海眼控科技股份有限公司 汉字识别方法、装置、计算机设备和存储介质
CN112380843A (zh) * 2020-11-18 2021-02-19 神思电子技术股份有限公司 一种基于随机扰动网络的开放性答案生成方法
CN112863627A (zh) * 2021-03-12 2021-05-28 云知声智能科技股份有限公司 医疗质控信息检测方法、系统以及存储介质
CN113469049A (zh) * 2021-06-30 2021-10-01 平安科技（深圳）有限公司 一种疾病信息识别方法、系统、装置及存储介质
CN114821245A (zh) * 2022-05-30 2022-07-29 大连大学 一种基于全局视觉信息干预的医学视觉问答方法
WO2023283739A1 (en) * 2021-07-15 2023-01-19 Pottie Kevin Health information based communities and knowledge incentive systems and methods
CN115964471A (zh) * 2023-03-16 2023-04-14 成都安哲斯生物医药科技有限公司 医疗数据近似查询方法
CN116955576A (zh) * 2023-09-21 2023-10-27 神州医疗科技股份有限公司 基于人类反馈与强化学习的问答回复方法、系统及设备
CN117194602A (zh) * 2023-09-06 2023-12-08 书音(上海)文化科技有限公司 基于大语言模型和bert模型的本地知识库更新方法及系统
CN117312512A (zh) * 2023-09-25 2023-12-29 星环信息科技(上海)股份有限公司 一种基于大模型的问答方法、装置、电子设备及存储介质
CN117592567A (zh) * 2023-11-21 2024-02-23 广州方舟信息科技有限公司 药品问答模型训练方法、装置、电子设备和存储介质

Families Citing this family (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108932349B (zh) * 2018-08-17 2019-03-26 齐鲁工业大学 医疗自动问答方法及装置、存储介质、电子设备
CN109783617B (zh) * 2018-12-11 2024-01-26 平安科技（深圳）有限公司 用于答复问题的模型训练方法、装置、设备及存储介质
CN109657047B (zh) * 2018-12-27 2020-09-29 四川新网银行股份有限公司 一种基于爬虫技术和机器学习的语音自动问答方法及系统
CN109783631B (zh) * 2019-02-02 2022-05-17 北京百度网讯科技有限公司 社区问答数据的校验方法、装置、计算机设备和存储介质
CN110083692B (zh) * 2019-04-22 2023-01-24 齐鲁工业大学 一种金融知识问答的文本交互匹配方法及装置
CN110032635B (zh) * 2019-04-22 2023-01-20 齐鲁工业大学 一种基于深度特征融合神经网络的问题对匹配方法和装置
US11763944B2 (en) * 2019-05-10 2023-09-19 Tencent America LLC System and method for clinical decision support system with inquiry based on reinforcement learning
CN110347807B (zh) * 2019-05-20 2023-08-08 平安科技（深圳）有限公司 问题信息处理方法及装置
CN110322959B (zh) * 2019-05-24 2021-09-28 山东大学 一种基于知识的深度医疗问题路由方法及系统
CN111104492B (zh) * 2019-09-19 2023-12-19 中国民航信息网络股份有限公司 一种基于层次化Attention机制的民航领域自动问答方法
CN111125328B (zh) * 2019-12-12 2023-11-07 深圳数联天下智能科技有限公司 文本处理方法及相关设备
CN111078853B (zh) * 2019-12-13 2023-05-02 上海智臻智能网络科技股份有限公司 问答模型的优化方法、装置、计算机设备和存储介质
CN113257372B (zh) * 2020-02-12 2024-05-07 阿里巴巴集团控股有限公司 口腔健康管理相关系统、方法、装置及设备
CN112085158A (zh) * 2020-07-21 2020-12-15 西安工程大学 一种基于堆栈降噪自编码器的图书推荐方法
CN113806474A (zh) * 2020-08-24 2021-12-17 北京沃东天骏信息技术有限公司 数据匹配方法及装置、电子设备、存储介质
CN113761142A (zh) * 2020-09-25 2021-12-07 北京沃东天骏信息技术有限公司 一种生成答案摘要的方法和装置
CN112966518B (zh) * 2020-12-22 2023-12-19 西安交通大学 一种面向大规模在线学习平台的优质答案识别方法
CN112597291B (zh) * 2020-12-26 2024-09-17 中国农业银行股份有限公司 一种智能问答的实现方法、装置及设备
CN113434629A (zh) * 2021-06-23 2021-09-24 平安科技（深圳）有限公司 基于向量匹配的问答对录入方法、装置、设备及介质
CN114490969B (zh) * 2021-12-29 2023-03-07 北京百度网讯科技有限公司 基于表格的问答方法、装置以及电子设备
CN118689967B (zh) * 2024-08-23 2024-11-08 中科厚立信息技术(成都)有限公司 一种基于医疗问答大模型的医疗数据收集方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103577556A (zh) * 2013-10-21 2014-02-12 北京奇虎科技有限公司 一种获取问答对的相关联程度的装置和方法
CN106845061A (zh) * 2016-11-02 2017-06-13 百度在线网络技术（北京）有限公司 智能问诊系统和方法
CN107368547A (zh) * 2017-06-28 2017-11-21 西安交通大学 一种基于深度学习的智能医疗自动问答方法
CN108182262A (zh) * 2018-01-04 2018-06-19 华侨大学 基于深度学习和知识图谱的智能问答系统构建方法和系统
CN108932349A (zh) * 2018-08-17 2018-12-04 齐鲁工业大学 医疗自动问答方法及装置、存储介质、电子设备

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663129A (zh) * 2012-04-25 2012-09-12 中国科学院计算技术研究所 医疗领域深度问答方法及医学检索系统
CN104965992B (zh) * 2015-07-13 2018-01-09 南开大学 一种基于在线医疗问答信息的文本挖掘方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103577556A (zh) * 2013-10-21 2014-02-12 北京奇虎科技有限公司 一种获取问答对的相关联程度的装置和方法
CN106845061A (zh) * 2016-11-02 2017-06-13 百度在线网络技术（北京）有限公司 智能问诊系统和方法
CN107368547A (zh) * 2017-06-28 2017-11-21 西安交通大学 一种基于深度学习的智能医疗自动问答方法
CN108182262A (zh) * 2018-01-04 2018-06-19 华侨大学 基于深度学习和知识图谱的智能问答系统构建方法和系统
CN108932349A (zh) * 2018-08-17 2018-12-04 齐鲁工业大学 医疗自动问答方法及装置、存储介质、电子设备

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11599769B2 (en) * 2018-11-12 2023-03-07 Alibaba Group Holding Limited Question and answer matching method, system and storage medium
US20200151542A1 (en) * 2018-11-12 2020-05-14 Alibaba Group Holding Limited Question and answer matching method, system and storage medium
CN111339277A (zh) * 2020-02-28 2020-06-26 中国工商银行股份有限公司 基于机器学习的问答交互方法及装置
CN112084312B (zh) * 2020-07-23 2023-11-14 江苏海洋大学 一种基于知识图构建的智能客服系统
CN112084312A (zh) * 2020-07-23 2020-12-15 江苏海洋大学 一种基于知识图构建的智能客服系统
CN112131393A (zh) * 2020-08-11 2020-12-25 淮阴工学院 一种基于bert和相似度算法的医疗知识图谱问答系统构建方法
CN112131393B (zh) * 2020-08-11 2023-09-22 淮阴工学院 一种基于bert和相似度算法的医疗知识图谱问答系统构建方法
CN111939529A (zh) * 2020-08-25 2020-11-17 山东大学齐鲁医院 一种基于肌力测量装置的主动式康复训练方法、系统
CN112200216A (zh) * 2020-09-03 2021-01-08 上海眼控科技股份有限公司 汉字识别方法、装置、计算机设备和存储介质
CN112380843A (zh) * 2020-11-18 2021-02-19 神思电子技术股份有限公司 一种基于随机扰动网络的开放性答案生成方法
CN112863627A (zh) * 2021-03-12 2021-05-28 云知声智能科技股份有限公司 医疗质控信息检测方法、系统以及存储介质
CN112863627B (zh) * 2021-03-12 2023-11-03 云知声智能科技股份有限公司 医疗质控信息检测方法、系统以及存储介质
CN113469049A (zh) * 2021-06-30 2021-10-01 平安科技（深圳）有限公司 一种疾病信息识别方法、系统、装置及存储介质
CN113469049B (zh) * 2021-06-30 2024-05-10 平安科技（深圳）有限公司 一种疾病信息识别方法、系统、装置及存储介质
WO2023283739A1 (en) * 2021-07-15 2023-01-19 Pottie Kevin Health information based communities and knowledge incentive systems and methods
CN114821245A (zh) * 2022-05-30 2022-07-29 大连大学 一种基于全局视觉信息干预的医学视觉问答方法
CN114821245B (zh) * 2022-05-30 2024-03-26 大连大学 一种基于全局视觉信息干预的医学视觉问答方法
CN115964471A (zh) * 2023-03-16 2023-04-14 成都安哲斯生物医药科技有限公司 医疗数据近似查询方法
CN117194602A (zh) * 2023-09-06 2023-12-08 书音(上海)文化科技有限公司 基于大语言模型和bert模型的本地知识库更新方法及系统
CN117194602B (zh) * 2023-09-06 2024-04-19 书音(上海)文化科技有限公司 基于大语言模型和bert模型的本地知识库更新方法及系统
CN116955576A (zh) * 2023-09-21 2023-10-27 神州医疗科技股份有限公司 基于人类反馈与强化学习的问答回复方法、系统及设备
CN117312512A (zh) * 2023-09-25 2023-12-29 星环信息科技(上海)股份有限公司 一种基于大模型的问答方法、装置、电子设备及存储介质
CN117592567A (zh) * 2023-11-21 2024-02-23 广州方舟信息科技有限公司 药品问答模型训练方法、装置、电子设备和存储介质
CN117592567B (zh) * 2023-11-21 2024-05-28 广州方舟信息科技有限公司 药品问答模型训练方法、装置、电子设备和存储介质

Also Published As

Publication number Publication date
AU2019322953B2 (en) 2021-08-19
LU101290B1 (en) 2019-11-29
AU2019322953A1 (en) 2020-09-03
CN108932349B (zh) 2019-03-26
CN108932349A (zh) 2018-12-04

Similar Documents

Publication Publication Date Title
WO2020034642A1 (zh) 医疗自动问答方法及装置、存储介质、电子设备
US8972445B2 (en) Systems and methods for storage of declarative knowledge accessible by natural language in a computer capable of appropriately responding
CN108182262B (zh) 基于深度学习和知识图谱的智能问答系统构建方法和系统
May Uses and abuses of mathematics in biology
CN111708874A (zh) 基于复杂意图智能识别的人机交互问答方法与系统
CN110675944A (zh) 分诊方法及装置、计算机设备及介质
CN110838368A (zh) 一种基于中医临床知识图谱的机器人主动问诊方法
CN111048167A (zh) 一种层级式病例结构化方法及系统
Lee et al. Influences of ketogenic diet on body fat percentage, respiratory exchange rate, and total cholesterol in athletes: A systematic review and meta-analysis
Keli ‘iholokai et al. Reclaiming ʻāina health in Waimānalo
Grue et al. Notes on the History of Normality–Reflections on the Work of Quetelet and Galton
Bystranowski et al. Half a century of bioethics and philosophy of medicine: A topic‐modeling study
Ricci et al. Nursing procedures during continuous renal replacement therapies: a national survey
Klugar et al. The personal active aging strategies of older adults in Europe: a systematic review of qualitative evidence
US11355239B1 (en) Cross care matrix based care giving intelligence
US11791025B2 (en) Personalized health system, method and device having a recommendation function
Liao et al. Medical data inquiry using a question answering model
US20200402672A1 (en) Systems and methods to group related medical results derived from a corpus of medical literature
CN109840275A (zh) 一种医疗搜索语句的处理方法、装置和设备
Wang et al. A residual LSTM and Seq2Seq neural network based on GPT for Chinese rice-related question and answer system
Murphy et al. Exploring progressive mental model representation of core physiology concepts in physician assistant students through word frequency and association analyses
Weber Data representation and algorithms for biomedical informatics applications
Wang A multi-modal knowledge graph platform based on medical data lake
Gillon Natural language semantics: Formation and valuation
CN103870712A (zh) 中医知识数据智能响应方法与设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19850195

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2019322953

Country of ref document: AU

Date of ref document: 20190329

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19850195

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 09/04/2021)