WO2022227203A1

WO2022227203A1 - Triage method, apparatus and device based on dialogue representation, and storage medium

Info

Publication number: WO2022227203A1
Application number: PCT/CN2021/097183
Authority: WO
Inventors: 孙行智; 胡岗; 朱昭苇; 刘卓; 唐蕊; 姚海申
Original assignee: 平安科技（深圳）有限公司
Priority date: 2021-04-30
Filing date: 2021-05-31
Publication date: 2022-11-03
Also published as: CN113223735A

Abstract

A triage method, apparatus and device based on dialogue representation, and a storage medium, which relate to the field of big data. The method comprises: acquiring a plurality of rounds of dialogue generated when an object to be subjected to triage is subjected to medical treatment, and extracting diagnostic inquiry data in the plurality of rounds of dialogue (101); performing data cleaning on the diagnostic inquiry data, so as to obtain target diagnostic inquiry data (102); performing intention recognition on the target diagnostic inquiry data, so as to obtain a sentence pair included in the target diagnostic inquiry data (103); calling a preset target BERT network model to perform feature extraction on the sentence pair and chief complaint information, so as to obtain a sentence pair vector of the sentence pair and a chief complaint vector of the chief complaint information (104); calculating a Euclidean distance between the chief complaint vector and the sentence pair vector, and on the basis of the Euclidean distance, determining a dialogue representation vector corresponding to each round of dialogue from among the plurality of rounds of dialogue (105); and inputting the dialogue representation vector of each round of dialogue into a preset triage model for recognition, so as to obtain triage information (106). By means of representing chief complaint information and a plurality of rounds of dialogue, the technical problem of low triage accuracy is solved.

Description

Diagnosis method, device, device and storage medium based on dialogue representation

This application claims the priority of the Chinese patent application filed on April 30, 2021 with the application number 202110489044.5 and the invention titled "diagnosis method, device, equipment and storage medium based on dialogue representation", the entire content of which is Incorporated in the application by reference.

technical field

The present application relates to the field of big data, and in particular, to a method, device, device and storage medium for triage based on dialogue representation.

Background technique

Due to the huge population base in my country, the number of people seeking medical treatment ranks among the top in the world. However, due to limited medical resources, there are still many problems in the national medical treatment process, one of which occurs during the initial triage of medical treatment. Triage is to assign patients to different departments for medical treatment according to their condition, which is of great significance for improving the efficiency of medical treatment.

Triage is the initial link for patients to seek medical treatment, and the selection of appropriate departments is directly related to the treatment effect or whether effective treatment can be obtained. However, with the deepening of medical research, the inventors realized that the division of labor in medicine was gradually refined, and different departments specialized in the diagnosis and treatment of certain diseases. Most patients do not have profound medical knowledge, so it is difficult to identify their own conditions and choose the most appropriate department according to their own conditions. The number of triage desk staff in hospitals is limited, and the volume of hospital admissions is huge, resulting in a heavy workload for triage desk staff. In addition, the triage desk faces the general practice environment, which further increases the possibility of misdiagnosis, which will lead to more secondary referrals and affect the efficiency of medical treatment. Therefore, how to improve the accuracy of triage has become a technical problem that those skilled in the art need to face.

SUMMARY OF THE INVENTION

The main purpose of this application is to improve the accuracy of triage based on dialogue representation, and to solve the technical problem of low accuracy of triage based on dialogue representation.

In order to achieve the above object, the present application provides a method for triage based on dialogue representation, and the method for triage based on dialogue representation comprises the following steps:

Acquire multiple rounds of dialogues generated by the subject to be triaged when they visit a doctor, and extract the consultation data in the multiple rounds of dialogues;

Data cleaning is performed on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

performing intention identification on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance; input the dialogue representation vector of each round of dialogue into a preset triage model

No, get triage information.

Further, in order to achieve the above object, the present application also provides a triage device based on dialogue representation, and the triage device based on dialogue representation includes the following modules:

a data cleaning module, configured to perform data cleaning on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

an intention recognition module, configured to perform intention recognition on the target consultation data to obtain sentence pairs included in the target consultation data, wherein the target consultation data includes at least one sentence pair;

The feature extraction module is used to call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

a first calculation module, used for calculating the Euclidean distance between the main complaint vector and the sentence pair vector, and determining the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogues based on the Euclidean distance;

The identification module is used to input the dialogue representation vector of each round of dialogue into the preset triage model for identification, and obtain triage information.

Further, in order to achieve the above object, the present application also provides a triage device based on dialogue representation, the triage device based on dialogue representation includes: a memory and at least one processor, wherein instructions are stored in the memory, and the a memory and the at least one processor are interconnected by wires;

The at least one processor invokes the instructions in the memory, so that when the dialogue-characterization-based triage device is executed, the steps of the above-mentioned dialogue-characterization-based triage method include:

Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance; input the dialogue representation vector of each round of dialogue into a preset The triage model is used to identify and obtain triage information.

Further, in order to achieve the above-mentioned purpose, the present application also provides a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, and when it is run on a computer, the above-mentioned dialogue-based dialogue is realized when the computer is executed. The steps of the triage method of characterization include:

In the technical solution provided by the present application, the target consultation data is obtained by acquiring the consultation data entered by the object to be triaged in each round of dialogues and cleaning the data; Sentence pairs contained in the interrogation data; call the preset target BERT network model to perform feature extraction on the sentence pairs and main complaint information to obtain sentence pair vectors and main complaint vectors; calculate the Euclidean distance between the main complaint vectors and the sentence pair vectors respectively, and The dialogue representation vector corresponding to each round of dialogue is determined based on the Euclidean distance; the dialogue representation vector of each round of dialogue is input into the preset triage model for identification, and triage information is obtained. This solution solves the technical problem of low triage accuracy by characterizing the chief complaint information and multiple rounds of dialogue.

Description of drawings

1 is a schematic diagram of the first embodiment of the triage method based on dialogue representation of the present application;

2 is a schematic diagram of a second embodiment of the triage method based on dialogue representation of the present application;

3 is a schematic diagram of a third embodiment of the triage method based on dialogue representation of the present application;

FIG. 4 is a schematic diagram of the fourth embodiment of the triage method based on dialogue representation of the present application;

FIG. 5 is a schematic diagram of the fifth embodiment of the triage method based on dialogue representation of the present application;

FIG. 6 is a schematic diagram of the first embodiment of the triage device based on dialogue representation of the present application;

FIG. 7 is a schematic diagram of a second embodiment of the triage device based on dialogue representation of the present application;

FIG. 8 is a schematic diagram of an embodiment of the triage device based on dialogue representation in the present application.

Detailed ways

The embodiments of the present application provide a method, device, device, and storage medium for triage based on dialogue representation. In the technical solution of the present application, firstly, the consultation data entered by the object to be triaged in each round of dialogue is obtained and the consultation is performed. The data is cleaned to obtain the target consultation data; the intent recognition is performed on the target consultation data to obtain the sentence pairs contained in the target consultation data; the preset target BERT network model is called to perform feature extraction on the sentence pairs and the main complaint information, and the sentence pairs are obtained. pair vector and main complaint vector; calculate the Euclidean distance between the main complaint vector and sentence pair vector respectively, and determine the dialogue representation vector corresponding to each round of dialogue based on the Euclidean distance; input the dialogue representation vector of each round of dialogue into the preset triage model Identify and obtain triage information. This solution solves the technical problem of low triage accuracy by characterizing the chief complaint information and multiple rounds of dialogue.

The terms "first", "second", "third", "fourth", etc. (if any) in the description and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It is to be understood that data so used may be interchanged under appropriate circumstances so that the embodiments described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those expressly listed steps or units, but may include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

For ease of understanding, the following describes the specific process of the embodiment of the present application, referring to FIG. 1 , the first embodiment of the triage method based on the dialogue representation in the embodiment of the present application includes:

101. Acquire multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extract the consultation data in the multiple rounds of dialogues;

In this embodiment, the multiple rounds of dialogues generated by the object to be triaged when visiting a doctor are acquired, and the consultation data in the multiple rounds of dialogues are extracted. The object to be triaged may issue a triage request instruction to the dialogue robot (doctor) in the form of voice. When the dialogue robot receives the voice of the object to be triaged, it can determine whether it includes preset keywords, and the keywords include but are not limited to words such as "triage", "diagnosis guide", "department" and so on. , if the voice of the subject to be triaged includes the keyword, it can be determined that the voice is a triage request instruction issued by the subject to be triaged.

The object to be triaged can also issue a triage request instruction to the dialogue robot through the physical button or virtual button in the designated human-computer interaction interface. When the object to be triaged needs to issue a triage request instruction to the dialogue robot, the specific button displayed therein can be clicked. After receiving the triage request instruction issued by the object to be triaged, the dialogue robot can carry out N rounds of dialogue with the object to be triaged according to the preset dialogue flow, and obtains information about the number of rounds of the object to be triaged in each round. Conversational statements within a conversation. Among them, N is a positive integer, and its specific value can be set according to the actual situation, for example, it can be set to 2, 3, 5 or other values. That is, the dialogue robot can conduct N rounds of dialogue with the object to be triaged. For example, in the first round of dialogue, the dialogue robot can ask: "Where are you uncomfortable?", and then obtain the dialogue sentence answered by the subject to be triaged in response to the question. In the second round of dialogue, the dialogue The robot can ask: "What main symptoms do you have", and then obtain the dialogue sentence answered by the subject to be triaged for this question, and in the Nth round of dialogue, the dialogue robot can ask: "What accompanying symptoms do you have?" , and then obtain the dialogue sentence answered by the subject to be triaged for the question.

102. Perform data cleaning on the consultation data to obtain the target consultation data;

In this embodiment, the user's consultation data is obtained through multiple rounds of interactions between the object to be triaged and the robot (doctor), wherein the consultation data includes the user's basic information and main complaint content.

The execution body of the embodiment of the present application is a server end with a remote diagnosis function. It should be noted that the user terminal may be a smart terminal such as a PC terminal, a mobile phone, a tablet computer, a smart watch, a smart bracelet, etc. used by the patient, and the user enters the consultation data through the user terminal, wherein the consultation data The information can include: the patient's basic information and main complaints. For example, the content of the main complaint may include: time of illness (onset time, duration or duration of each attack), patient symptoms, patient identity information, patient underlying conditions (for example: what chronic disease the patient has suffered in addition to the current symptoms) )Wait. Data cleaning is the process of re-examining and verifying data with the purpose of removing duplicate information, correcting existing errors, and providing data consistency.

In this embodiment, data cleaning includes data desensitization, data verification, and data conversion. Among them, data desensitization is used to encrypt sensitive data in source business data. For example, the data includes an individual's ID card number, etc., and the ID card number can be encrypted. Data verification is used to query whether there is dirty data in the source business data, and delete the dirty data to eliminate the impact of dirty data on actuarial results. The server sets a dirty data determination method for each type of data, and detects whether it is dirty data according to a preset determination method. For example, you can set the character length range or numerical size range of each type of data, etc. When the character length of a certain type of data is not within the preset range, or the size of the numerical value is not within the preset size range , the data is determined to be dirty data. Data verification is the process of uniformly converting data with multiple different representations into the same preset representation.

Data cleaning, as the name suggests, is to "wash out" the "dirty", which refers to the last process of finding and correcting identifiable errors in data files, including checking data consistency, dealing with invalid and missing values, etc. Because the data in the data warehouse is a collection of data oriented to a certain topic, these data are extracted from multiple business systems and contain historical data, so it is inevitable that some data are wrong data, and some data are interrelated with each other. Conflicts, these erroneous or conflicting data that we obviously don't want, are called "dirty data". We have to "wash out" the "dirty data" according to certain rules, which is data cleaning. The task of data cleaning is to filter those data that do not meet the requirements, and hand over the filtered results to the business department to confirm whether it is filtered out or corrected by the business unit before extracting. The data that does not meet the requirements are mainly divided into three categories: incomplete data, wrong data, and duplicate data.

103. Perform intent recognition on the target consultation data to obtain sentence pairs included in the target consultation data;

In this embodiment, intent recognition is performed on the target consultation data to obtain sentence pairs included in the target consultation data. Train and classify the target consultation data, construct a convolutional neural network, take the target consultation data after replacing the entity as input, map the question after word segmentation, each word to word embedding, and the spliced entity embedding and entity location Embedding together, as the input of the neural network, the annotated intent is used as the classification label to train the classification problem.

Secondly, according to the combination of intent and entity, the subsequent multi-round dialogue process is configured. After identifying the intent and entity, the subsequent process configuration is performed according to the obtained value. Here, the processing method of functionalizing the intent and parameterizing the entity is adopted, that is, the subsequent operation for a certain intent is regarded as a function, and the entity to be extracted is regarded as the parameter of this function, which is adjusted according to different parameters. The return result of the function.

Perform intent recognition and entity extraction on new sentences, and pass the results to the process configuration engine for subsequent multiple rounds of dialog configuration. For the identified intent, its corresponding entity parameters are detected, and the extracted entities are filled into the corresponding parameter list. Based on the identified intent and the filled parameters, the subsequent dialogue flow is triggered. If the required entity parameters are satisfied, the corresponding action can be directly performed. If the required entity is satisfied, multiple rounds of guided question and answer are triggered to guide the user to complete the information.

104. Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

In this embodiment, the preset target BERT network model is invoked to perform feature extraction on the sentence pair and the main complaint information, and the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information are obtained. The full name of the BERT network model in this embodiment is Pre-training of DeepBidirectional Transformers for Language Understanding. Pre-training means that BERT is a pre-training model that learns a large amount of prior language, syntax, word meaning and other information for downstream tasks through unsupervised training of a large number of corpora in the early stage. Bidirectional shows that BERT adopts a two-way language model, which can better integrate the knowledge of the context. In short, BERT is a deep bidirectional pretrained language understanding model using Transformers as feature extractors. During the pre-training process, BERT has learned a wealth of linguistic information. Symptom recognition is named entity recognition, and its essence is a serialized labeling task. The above process of semantic encoding is the process of vectorizing disease information, specifically: the pre-training model BERT vectorizes each character in the disease information, obtains the character vector of each character, and gives each character Mark the position vector to obtain a character position vector, combine each character vector and its corresponding character mark position vector to obtain the encoding vector of each character, and combine the encoding vectors of each character to obtain the disease code.

In this embodiment, a graph neural network is used to perform representation learning on the target consultation data, so as to obtain sentence pair vectors for obtaining all sentence pairs and a main complaint vector for main complaint information contained in the consultation data. Among them, representation learning is a collection of techniques that use computers to learn a feature, and is to convert data into a form of learning that can be learned and developed by machines.

105. Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance;

In this embodiment, the Euclidean distance between the main complaint vector and the sentence pair vector is calculated, and based on the Euclidean distance, a dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue is determined. Among them, a sentence pair is also called a question-and-answer pair, which means that the dialogue between the doctor and the patient is organized into several sentence pairs, that is, a question-and-answer format. If a character speaks several times in a row, the consecutive speeches are spliced together as a whole speech. For example, in each sentence pair, the doctor's speech is the first sentence, and the patient's speech is the next sentence, such as: <CLS>: [How long has the headache been accompanied by dizziness? <SEP> It has been more than a month without dizziness. ] <SEP> The former means the doctor's speech, and the latter means the patient's speech. This sentence takes the vector representation of <CLS> for the final representation. Measure the Euclidean distance between the main complaint vector and each sentence pair vector.

In this embodiment, it is often necessary to estimate the similarity measurement (Similarity Measurement) between different samples during classification, and the method usually adopted in this case is to calculate the "distance" (Distance) between the samples. The smaller the "distance", the more similar the samples are. The most common representation of the distance between two or more points, also known as the Euclidean metric. It is defined in Euclidean space, such as the distance between x=(x ₁ , x ₂ ,..., x _n ) and y=(y ₁ , y ₂ ,..., y _n ) as:

where, the Euclidean distance between two n-dimensional vectors a(x ₁₁ , x ₁₂ ,...,x _1n ) and b(x ₂₁ , x ₂₂ ,...,x _2n ):

106. Input the dialogue representation vector of each round of dialogue into a preset triage model for identification, and obtain triage information.

In this embodiment, the dialogue representation vector of each round of dialogue is input into a preset triage model for identification, and triage information is obtained. The triage model is a model for automatically determining the corresponding department for the user according to the user's symptoms. Different softmax classifiers are used in the fully connected layer of the prediction model, and the output result of the fully connected layer is input into the softmax classifier to obtain the prediction result of the claim settlement data set.

In the embodiment of the present application, the target consultation data is obtained by acquiring the consultation data entered by the object to be triaged in each round of dialogues and cleaning the consultation data; Sentence pairs contained in the data; call the preset target BERT network model to perform feature extraction on sentence pairs and main complaint information, and obtain sentence pair vectors and main complaint vectors; The distance is used to determine the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue; the dialogue representation vector of each round of dialogue is input into the preset triage model for identification, and the triage information is obtained. This solution solves the technical problem of low triage accuracy by characterizing the chief complaint information and multiple rounds of dialogue.

Referring to FIG. 2 , the second embodiment of the triage method based on dialogue representation in the embodiment of the present application includes:

201. Acquire multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extract the consultation data in the multiple rounds of dialogues;

202. Obtaining the cleaning needs of consultation data;

In this embodiment, the cleaning requirements of the consultation data are obtained. Understandably, real-world data is often multi-dimensional, incomplete, noisy, and inconsistent. The purpose of data cleaning is to fill in missing values, smooth noise and identify outliers, and correct inconsistencies in the data. Wait.

In this embodiment, after acquiring the data to be cleaned that needs to be cleaned, the electronic device further acquires the cleaning requirement of the data to be cleaned. In layman's terms, the cleaning requirement describes the cleaning effect that the data to be cleaned needs to be cleaned. For example, the original data to be cleaned contains data of multiple dimensions, and these dimensions are often not independent, that is to say, maybe one of them There is an association between several dimensions, and the cleaning requirement of the data to be executed may be to reduce the dimension of the data to be cleaned to a specified dimension.

203. Determine the target cleaning rules for data cleaning of the consultation data according to the consultation data and cleaning requirements;

In this embodiment, a target cleaning rule for performing data cleaning on the consultation data is determined according to the consultation data and cleaning requirements. All possible cleaning rules can be integrated in advance, and the sample data to be cleaned corresponding to each cleaning rule and its cleaning effect can be collected at the same time; then, the cleaning rule features that can characterize the cleaning rule, and the sample data that can characterize the cleaning effect and its cleaning effect are obtained. Then, take each joint feature as training input, take the cleaning rule feature corresponding to each joint feature as target output, and carry out model training according to the preset training algorithm, so as to obtain which cleaning rule to be used for training. Cleaning rule classification model for cleaning data.

Therefore, after the electronic device obtains the data to be cleaned that needs to be cleaned, and obtains the cleaning requirement of the data to be cleaned, the electronic device can input the data to be cleaned and the cleaning requirement into the cleaning rule classification model, so that the cleaning rule classification model outputs A cleaning rule that can perform data cleaning on the data to be cleaned and the cleaning effect meets the cleaning requirements, and the cleaning rule is used as the target cleaning rule for data cleaning of the data to be cleaned.

204. Perform data cleaning on the consultation data according to the target cleaning rule to obtain the target consultation data;

In this embodiment, data cleaning is performed on the consultation data according to the target cleaning rule to obtain the target consultation data. After determining the target cleaning rule for cleaning the data to be cleaned, the data to be cleaned can be cleaned according to the target cleaning rule, so that the cleaning effect of the data to be cleaned meets the aforementioned cleaning requirements, and finally the required data is obtained.

In this embodiment, the data to be cleaned that needs to be cleaned, and the cleaning requirements of the data to be cleaned are obtained first, and then according to the obtained data to be cleaned, the cleaning requirements and the pre-trained cleaning rule classification model, a classification model for the data to be cleaned is determined. The target cleaning rules for data cleaning are performed on the cleaned data, and finally the data to be cleaned is cleaned according to the determined target cleaning rules, and the target consultation data that meets the requirements are obtained.

205. Acquire all question and answer sentences in the target consultation data and object identifiers of objects corresponding to the question and answer sentences according to a preset representation rule;

In this embodiment, all question and answer sentences in the target consultation data and object identifiers of objects corresponding to the question and answer sentences are acquired according to preset characterization rules. The target consultation data refers to the online consultation record between the robot (doctor) and the patient, which may include multiple question and answer sentences between the doctor and the patient, that is, may include the question or answer sentences corresponding to the doctor or the patient. A question and answer sentence is a sentence corresponding to each conversation between the doctor or the patient in the dialogue between the doctor and the patient.

In this embodiment, the object identifier is an identifier used to distinguish a doctor and a patient in the target consultation data, and may include a doctor identifier and a patient identifier. The doctor identification may refer to the unified doctor identification indication applicable to all target consultation data, for example, the patient identification refers to the unified patient identification indication applicable to all target consultation data.

206. Based on each object identifier, perform intention identification on each question and answer sentence in the target medical consultation data, to obtain an intention identification sequence corresponding to the medical consultation data;

In this embodiment, based on each object identifier, intention identification is performed on each question and answer sentence in the target consultation data, respectively, to obtain an intention identification sequence corresponding to the consultation data. The intent identifier refers to an identifier that identifies the intent of each question and answer sentence, for example, a medication inquiry, a registered inquiry, a medication response, a registered response, and the like. In this embodiment, the intent identifiers of the patient's question may be represented by P ₁ -P _n , the intent identifier of the patient's answer is represented by P _A , the intent identifier of the doctor's question is denoted as D ₁ -D _n , and the intent identifier of the doctor's answer is denoted by P 1 -P n . Denoted as D _A1 ~ D _An .

The intent identifier sequence refers to a sequence composed of multiple intent identifiers, for example, P ₁ , D _A1 , D _A1 , P ₂ , P ₂ , D _A2 and so on. In this embodiment, the server may perform intention identification on each question and answer sentence according to the object identifier corresponding to each question and answer sentence, and generate each corresponding intent identifier, and may obtain the corresponding target consultation data according to a plurality of target consultation data of each intent identifier sequence.

207. Determine, according to the intent identifier sequence, the intent identifier of the target answer sentence corresponding to the question;

In this embodiment, the intent identifier of the target answer sentence corresponding to the question is determined according to the intent identifier sequence. Each question sentence intent identifier corresponds to the number of answer sentence intent identifiers of each initial answer sentence, and based on each quantity, a target answer sentence intent identifier corresponding to each question sentence intent identifier is determined. Further, after determining each question sentence intent identifier and the corresponding target answer sentence intent identifier, the server may generate a corresponding target sentence pair according to the question sentence intent identifier and the corresponding target answer sentence intent identifier. For example, if the server determines that the intent identifier of the corresponding question sentence P ₁ corresponds to the intent identifier of the target answer sentence as D _A1 , the server may generate a pair of target sentences as P ₁ -D _A1 .

In this embodiment, the target sentence pair is not limited to a binary combination relationship composed of a question sentence intent identifier and a target answer sentence intent identifier, but may also be a ternary combination relationship.

208. Based on the intent identification of the target answer sentence, obtain sentence pairs included in the target consultation data;

In this embodiment, by acquiring the target consultation data, the target consultation data includes a plurality of question and answer sentences and object identifiers corresponding to the question and answer sentences, and based on the object identifiers, intention recognition is performed on each question and answer sentence in the target consultation data. , obtain the intent identification sequence corresponding to the target inquiry data, and then determine the answer sentences of multiple initial answer sentences corresponding to the question intent identification of each question in the intent identification sequence according to the intent identification sequences corresponding to the multiple target inquiry data Identifying the identifier, further, determining the target answer sentence intent identifier corresponding to each question sentence from the answer sentence identifier identifiers of the multiple initial answer sentences, and generating each target sentence based on each question sentence intent identifier and the corresponding target answer sentence intent identifier right. Therefore, the target sentence pair can be determined after the intention recognition processing of the question and answer sentences of multiple target medical consultation data, so that the generation of the target sentence pair can cover a plurality of different forms of question and answer sentences, and the target sentence pair can be improved. coverage and accuracy.

In one embodiment, based on the identifiers of each object, the intent recognition is performed on each question and answer sentence in the target consultation data, respectively, to obtain a sequence of intent identifiers corresponding to the target consultation data, which may include: determining the target consultation according to the object identifiers Each question and answer sentence corresponding to each object in the data; each question and answer sentence of each object is identified separately, and each question and answer sentence is determined to be a question sentence of the corresponding object or the answer sentence of the corresponding object; the question sentence of each object or the answer sentence of each object Perform intention identification respectively to obtain the corresponding intention identification; according to the obtained intention identification, obtain the intention identification sequence corresponding to the target consultation data. As mentioned above, the target consultation data may include multiple question and answer sentences between doctors and patients, and each question and answer sentence includes a corresponding object identifier, such as a doctor identifier or a patient identifier.

209. Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

210. Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance;

211. Input the dialogue representation vector of each round of dialogue into a preset triage model for identification, and obtain triage information.

Steps 201, 209-211 in this embodiment are similar to steps 101, 104-106 in the first embodiment, and are not repeated here.

Referring to FIG. 3, the third embodiment of the triage method based on dialogue representation in the embodiment of the present application includes:

301. Acquire multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extract the consultation data in the multiple rounds of dialogues;

302. Perform data cleaning on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

303. Perform intention identification on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

304. Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

305. Establish a pre-training task, and call a preset vectorization model to process historical user consultation data into several word vectors;

In this embodiment, a pre-training task is established, and a preset vectorization model is invoked to process historical user consultation data into several word vectors. Wherein, the pre-training task is mainly used to perform the loop iterative calculation of steps S305-S306. Before the loop iterative calculation, it is necessary to convert the symptom samples into word vectors through a vectorized model, such as the word2vec model. Among them, the word2vec model is a model for generating word vectors. In an example, the symptom data is {cough: three days; bloodshot}, and word vectors emb1 (cough), emb2 (three days), and emb3 (bloodshot) can be obtained after transformation by the word2vec model. Here, emb1 is the first word vector, and emb2 and emb3 are the second word vectors. Symptom samples refer to the training data used to train the initial BERT network model, generally historical symptom data in a certain area.

306. Input several word vectors into the initial BERT network model, and obtain the training representation vector output by the initial BERT network model;

In this embodiment, several word vectors are input into the initial BERT network model, and the training representation vector output by the initial BERT network model is obtained. After obtaining the word vector, each word vector can be input into the initial BERT network model as input data, and the training representation vector can be generated, and the corresponding loss value can be calculated. Specifically, the loss value can be obtained by calculating the loss function. The loss function is defined as:

Among them, L(Vs, sym ⁽ⁿ⁾ ) is the loss value of the nth symptom; sym ⁽ⁿ⁾ represents the nth symptom in the symptom list; Vs represents the overall characterization vector; it is the nth symptom in the characterization vector The loss term of , is the loss term of other symptoms in the representation vector. Through the loss function, it can be known that the symptoms appearing in the representation vector should be as small as possible, otherwise the loss should be as large as possible.

307. Calculate the loss value of the initial BERT network model according to the training representation vector;

In this embodiment, the loss value of the initial BERT network model is calculated according to the training characterization vector, and the loss value of the initial BERT network model is calculated according to the training characterization vector, and the obtained loss value can be used to adjust model parameters and determine whether the model is convergence. If the loss value is outside the preset range, the model parameters of the initial BERT network model are adjusted, and the training representation vector of the symptom sample is recalculated, so as to perform iterative calculation when the model does not converge.

308. According to the loss value of the initial BERT network model, adjust the model parameters of the initial BERT network model to obtain the target BERT network model;

In this embodiment, the several word vectors are input into the initial BERT network model, and the training representation vector output by the initial BERT network model is obtained, so as to perform the training step of the initial BERT network model. The loss value of the initial BERT network model is calculated according to the training characterization vector, and the obtained loss value can be used to adjust model parameters and judge whether the model converges. If the loss value is outside the preset range, the model parameters of the initial BERT network model are adjusted, and the training representation vector of the symptom sample is recalculated, so as to perform iterative calculation when the model does not converge. If the loss value is within the preset range, the pre-training task is trained, and the initial BERT network model after the training is completed is the target BERT network model.

309. Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance;

310. Input the dialogue representation vector of each round of dialogue into a preset triage model for identification, and obtain triage information.

Steps 301 - 304 and 309 - 310 in this embodiment are similar to steps 101 - 104 and 105 - 106 in the first embodiment, and are not repeated here.

Referring to FIG. 4 , the fourth embodiment of the triage method based on dialogue representation in the embodiment of the present application includes:

401. Acquire multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extract the consultation data in the multiple rounds of dialogues;

402. Perform data cleaning on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

403. Perform intent recognition on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

404. Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

405. Calculate the Euclidean distance between the main complaint vector and each sentence pair vector in each round of dialogue;

In this embodiment, the Euclidean distance between the main complaint vector and each sentence pair vector in each round of dialogue is calculated. The Euclidean distances between all sentence pairs in the consultation data and the main complaint vector are calculated respectively.

In this embodiment, the Euclidean distance means that the Euclidean distance is an effective method for calculating the closest distance between a sample and a "centre of gravity" of a sample set, or for effectively calculating the similarity between two unknown sample sets. It takes into account the connection between various properties, can exclude the interference of the correlation between variables, and the Euclidean distance is scale-independent, that is, independent of the measurement scale. When ∑ is the identity matrix, the Euclidean distance is the Euclidean distance. To sum up, the Euclidean distance can easily measure the distance between the observed sample and the known sample set, so it is very suitable for fault diagnosis.

406. Determine the weight of each sentence pair in the corresponding round of dialogue based on the Euclidean distance;

In this embodiment, the weight of each sentence pair in a corresponding round of dialogue is determined based on the Euclidean distance. The representation of the main complaint is completed based on the pre-trained word vector (vector representation), and the correlation between the vector of each sentence pair and the main complaint vector is calculated as the basis for evaluating the importance of the sentence pair. The specific calculation method is:

Measure the Euclidean distance between the main complaint vector and each sentence pair vector, that is, calculate the weight ai of each sentence pair.

407. Based on the weights, perform a weighted average of the weights corresponding to each sentence pair vector to obtain a dialogue representation vector corresponding to each round of dialogues in multiple rounds of dialogues;

In this embodiment, based on the weights, the weights corresponding to each sentence pair vector are weighted and averaged to obtain a dialogue representation vector corresponding to each round of dialogues in multiple rounds of dialogues. For example, in a round of dialogue, if the answer sentence in the sentence pair (question and answer sentence and answer sentence) is the answer to the question sentence, the answer sentence usually contains the keywords in the question sentence; the common key between the question sentence and the answer sentence The more and longer the number of words, the more effective information it contains, and the more important the corresponding round of dialogue is. For example, the answer sentence in the sentence pair is more likely to be the answer to the question sentence; since the target question-answer pair expected to be obtained is a complete question and answer, the longer the question sentence can better describe a complete question , and the longer the answer sentence can better describe a complete answer; if the answer sentence in the sentence pair is the answer to the question sentence, the themes between the answer sentence and the question sentence are usually consistent; if the answer sentence in the sentence pair is the answer to the question sentence is the answer to the question sentence, then there is usually a certain syntactic connection between the answer sentence and the question sentence.

In this embodiment, considering that not every sentence spoken by the patient during the dialogue contains valid information, a weighted importance measurement module (hereinafter referred to as the weighting module) is introduced to evaluate the importance of each round of dialogue, and weighted representation of all dialogues Then make a judgment.

408. Determine the target disease information in the target consultation data based on the dialogue representation vector corresponding to each round of dialogue;

In this embodiment, the target disease information in the target consultation data is determined based on the dialogue representation vector corresponding to each round of dialogue. The target disease information refers to text information or voice information input by the person to be triaged. If the person to be triaged inputs voice information, the system will now convert the voice information into text information.

409. Encode the target disease information and the chief complaint information corresponding to each round of dialogue to obtain the disease entity vector of the target consultation data;

In this embodiment, the target disease information and the main complaint information corresponding to each round of dialogue are encoded to obtain the disease entity vector of the target consultation data. Among them, coding is the process of converting information from one form or format to another, also known as coding in computer programming languages for short. Digitize characters, numbers or other objects with a predetermined method, or convert information and data into specified electrical pulse signals. Coding is widely used in electronic computers, television, remote control and communication. Encoding is the process of converting information from one form or format to another. Decoding is the reverse process of encoding.

410. Input the disease entity vector into the preset triage model for prediction, and obtain the triage probability of different departments;

In this embodiment, the disease entity vector is input into the preset triage model for prediction, and the triage probability of different departments is obtained.

In this embodiment, the convolution operation is performed on the splicing vector at the convolution layer to obtain the convolution correlation vector, and the convolution correlation vector is input to the output layer to obtain the prediction output result. Based on the prediction output results and department labels, the prediction error loss is calculated, and the parameters of the neural network model are updated according to the prediction error loss. When the model converges, the triage model is obtained. The disease entity vector is input into the preset triage model for prediction, and the triage probability of different departments is obtained.

411. Sort the triage probabilities to obtain triage information, where the triage information includes a preset number of recommended departments with the highest ranking and the triage probabilities corresponding to the recommended departments.

In this embodiment, the triage probability is sorted to obtain triage information. According to the fighting probability of each different department, the triage probability of each department is sorted, and the recommended department with the highest set number and the corresponding triage probability are returned to the patient. Preferably, the triage model includes: multiple convolution layers, multiple pooling layers, fully connected layers and Softmax layers.

Steps 401-404 in this embodiment are similar to steps 101-104 in the first embodiment, and are not repeated here.

Referring to FIG. 5 , the fifth embodiment of the triage method based on dialogue representation in the embodiment of the present application includes:

501. Acquire multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extract the consultation data in the multiple rounds of dialogues;

502. Perform data cleaning on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

503. Perform intention identification on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

504. Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

505. Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance;

506. Obtain historical consultation data from a preset database;

In this embodiment, historical consultation data is obtained from a preset database. The N rounds of dialogue with the historical patient according to the preset dialogue flow can be acquired from a preset database, and the dialogue sentences of the historical patient in each round of dialogue can be acquired. Among them, N is a positive integer, and its specific value can be set according to the actual situation, for example, it can be set to 2, 3, 5 or other values. That is, the dialogue robot can conduct N rounds of dialogue with the historical patient. For example, in the first round of dialogue, the dialogue robot can ask: "Where are you uncomfortable", and then obtain the dialogue sentences answered by the historical patient for this question, and in the second round of dialogue, the dialogue robot can Ask: "What main symptoms do you have", and then obtain the dialogue sentences answered by the historical patient for this question, in the Nth round of dialogue, the dialogue robot can ask: "What accompanying symptoms do you have", and then obtain all the Dialogue sentences answered by the patient in response to the question.

507. Perform representation learning on the historical consultation data through the target BERT network model, and obtain a second dialogue representation vector of the historical consultation data;

In this embodiment, the target BERT network model is used to perform representation learning on the historical consultation data to obtain the second dialogue representation vector of the historical consultation data. The target BERT network model is obtained after being trained on the pre-training task. The pre-training task is a custom task, and the pre-training task is defined as inferring the symptom name and symptom attribute contained in the representation vector according to the current representation vector. The pre-training task can ensure that the target BERT network model can learn the information contained in the output representation vector, that is, the relationship between the representation vector and the symptom feature data is determined through the pre-training task. It should be noted that here, the relationship is embodied in the model parameters of the target BERT network model. In this way, symptom names and symptom attributes can be accurately converted into an overall vector, that is, a representation vector. The number of generated representation vectors is equal to the number of symptoms in the symptom data. That is, how many symptoms there are in the historical consultation data, the same number and corresponding representation vectors are generated.

508. Obtain a preset training symptom and a department label corresponding to the training symptom, screen the preset node set association vector based on the training symptom, and obtain a target vector corresponding to the training symptom;

In this embodiment, a preset training symptom and a department label corresponding to the training symptom are obtained, and the preset node set association vector is screened based on the training symptom, and a target vector corresponding to the training symptom is obtained. Among them, the training symptom refers to the symptom used to train the BERT network model. The department label is the department corresponding to the training symptom, and the department label is the training label. For example, if the training symptom is a skin problem, the corresponding department label is dermatology.

The target vector refers to the vector corresponding to the training symptom. For example, if the symptom corresponding to the training symptom is cough, the target vector is filtered from the association vector of the node set according to the symptom. In this embodiment, the node set association vectors are screened according to the training symptoms to match the target vectors corresponding to the training symptoms, so as to ensure that the model training samples have a corresponding relationship and ensure the feasibility of model training.

509. Perform a mapping operation on the second dialogue representation vector and the target vector to obtain the dialogue embedding vector and the target embedding vector;

In this embodiment, a mapping operation is performed on the second dialogue representation vector and the target vector to obtain the dialogue embedding vector and the target embedding vector. A mapping operation is performed on the graph symptom vector in the second dialogue representation vector to obtain the dialogue embedding vector. The first embedding layer is a layer for performing dimension reduction processing on the second dialogue representation vector, so that the second dialogue representation vector is mapped to a dialogue embedding vector with a lower dimension.

In this embodiment, a mapping operation is performed on the target vector in the second embedding layer to obtain the target embedding vector. Among them, the second embedding layer is a layer used to reduce the dimension of the target vector, so that the target vector is mapped to a target embedding vector with a lower dimension. Specifically, in the second embedding layer, a preset mapping table is used to process the target vector to obtain the target embedding vector, thereby reducing the dimension of the target embedding vector and reducing the difficulty of subsequent operations.

510. Based on the training symptom, splicing the dialogue embedding vector and the target embedding vector to obtain the splicing vector;

In this embodiment, the dialogue embedding vector and the target embedding vector are spliced based on the training symptom to obtain the splicing vector. Among them, the splicing vector is a vector obtained from the dialogue embedding vector and the target embedding vector to form a vector with deep meaning.

In this embodiment, the splicing vector with more comprehensive information is fully utilized in the process of triage model training, so that the acquired triage model has strong generalization ability and high robustness. In this embodiment, the tf.concat() function of TensorFlow is used to splicing the dialogue embedding vector and the target embedding vector to quickly obtain the splicing vector.

511. Perform a convolution operation on the splicing vector at the convolution layer to obtain the convolution correlation vector, input the convolution correlation vector into the output layer, and obtain the predicted output result;

In this embodiment, the convolution operation is performed on the splicing vector at the convolution layer to obtain the convolution correlation vector, and the convolution correlation vector is input to the output layer to obtain the prediction output result. Among them, the predicted output result is the predicted department result corresponding to the training symptom.

In this embodiment, the output layer is used to calculate the probability of possible departments corresponding to the training symptoms, and the department with the highest probability is used as the prediction output result, so as to obtain the corresponding prediction output result according to the model training samples.

512. Calculate the prediction error loss based on the prediction output result and the department label, and update the parameters of the target BERT network model according to the prediction error loss until the target BERT network model converges, and obtain the triage model based on the dialogue representation;

In this embodiment, the prediction error loss is calculated based on the prediction output result and the department label, and the parameters of the target BERT network model are updated according to the prediction error loss until the target BERT network model converges to obtain a triage model based on dialogue representation. Specifically, the partial derivative of the prediction error loss is obtained to obtain the gradient value, and the parameters of the target BERT network model are updated according to the gradient value to realize the optimization of the target BERT network model. When the prediction error loss is less than the preset threshold, the target BERT network model Convergence, the target BERT network model is determined as the triage model.

513. Input the dialogue representation vector of each round of dialogue into a preset triage model for identification, and obtain triage information.

Steps 501 - 505 and 513 in this embodiment are similar to steps 101 - 105 and 106 in the first embodiment, and are not repeated here.

In the embodiment of the present application, the target consultation data is obtained by acquiring the consultation data entered by the object to be triaged in each round of dialogues and cleaning the consultation data to obtain the target consultation data; The sentence pairs contained in the diagnosis data; call the preset target BERT network model to extract the features of the sentence pairs and the main complaint information, and obtain the sentence pair vector and the main complaint vector; respectively calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and based on the The Euclidean distance determines the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue; the dialogue representation vector of each round of dialogue is input into the preset triage model for identification, and triage information is obtained. This solution solves the technical problem of low triage accuracy by characterizing the chief complaint information and multiple rounds of dialogue.

The triage method based on the dialogue representation in the embodiment of the present application is described above, and the triage device based on the dialogue representation in the embodiment of the present application is described below. Please refer to FIG. 6 , the triage device based on the dialogue representation in the embodiment of the present application is described The first embodiment includes:

The data extraction module 601 is used for acquiring multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extracting the consultation data in the multiple rounds of dialogues;

A data cleaning module 602, configured to perform data cleaning on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

Intention recognition module 603, configured to perform intention recognition on the target consultation data to obtain sentence pairs included in the target consultation data, wherein the target consultation data includes at least one sentence pair;

The feature extraction module 604 is used to call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

The first calculation module 605 is used to calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue based on the Euclidean distance;

The identification module 606 is configured to input the dialogue representation vector of each round of dialogue into a preset triage model for identification, and obtain triage information.

Please refer to FIG. 7 , the second embodiment of the triage device based on dialogue representation in the embodiment of the present application, the triage device based on dialogue representation specifically includes:

In this embodiment, the intent recognition module 603 includes:

Obtaining unit 6031, configured to obtain all question and answer sentences in the target consultation data and object identifiers of objects corresponding to the question and answer sentences according to preset characterization rules;

an intent identification unit 6032, configured to perform intent identification on each question and answer sentence in the target medical consultation data based on each of the object identifiers, to obtain an intent identification sequence corresponding to the medical consultation data;

The determining unit 6033 is configured to determine, according to the intent identifier sequence, a target answer sentence intent identifier corresponding to the question; and based on the target answer sentence intent identifier, obtain sentence pairs included in the target questioning data.

In this embodiment, the triage device based on dialogue representation further includes:

The vectorization module 607 is used for establishing a pre-training task, calling a preset vectorization model to process historical user consultation data into several word vectors, wherein the word vectors include word vectors based on the historical user consultation data;

The first acquisition module 608 is used to input the several word vectors into the initial BERT network model, and obtain the training representation vector output by the initial BERT network model;

The second calculation module 609 is configured to calculate the loss value of the initial BERT network model according to the training characterization vector; according to the loss value of the initial BERT network model, adjust the model parameters of the initial BERT network model to obtain the target BERT network model.

In this embodiment, the first calculation module 605 is specifically used for:

Calculate the Euclidean distance between the main complaint vector and each sentence pair vector in each round of dialogue;

Based on the Euclidean distance, determine the weight of each sentence pair in the corresponding round of dialogue;

Based on the weights, the weights corresponding to each sentence pair vector are weighted and averaged to obtain a dialogue representation vector corresponding to each round of dialogues in multiple rounds of dialogues.

In this embodiment, the identifying module 606 is specifically used for:

Determine the target disease information in the target consultation data based on the dialogue representation vector corresponding to each round of dialogue;

Encoding the target disease information and the main complaint information corresponding to each round of dialogue to obtain the disease entity vector of the target consultation data;

Inputting the disease entity vector into a preset triage model for prediction, and obtaining the triage probability of different departments;

Sorting the triage probabilities to obtain triage information, wherein the triage information includes a preset number of recommended departments with the highest ranking and triage probabilities corresponding to the recommended departments.

The second obtaining module 610 is configured to obtain historical consultation data from a preset database;

A representation learning module 611, configured to perform representation learning on the historical consultation data through the target BERT network model, and obtain a second dialogue representation vector of the historical consultation data;

A screening module 612, configured to obtain a preset training symptom and a department label corresponding to the training symptom, screen a preset node set association vector based on the training symptom, and obtain a target vector corresponding to the training symptom;

A mapping module 613, configured to perform a mapping operation on the second dialogue representation vector and the target vector to obtain a dialogue embedding vector and a target embedding vector;

A splicing module 614, configured to splicing the dialogue embedding vector and the target embedding vector based on the training symptom to obtain a splicing vector;

The convolution processing module 615 is configured to perform a convolution operation on the splicing vector at the convolution layer, obtain a convolution association vector, input the convolution association vector into the output layer, and obtain a prediction output result;

The updating module 616 is configured to calculate the prediction error loss based on the prediction output result and the department label, and update the parameters of the target BERT network model according to the prediction error loss, until the target BERT network model converges, obtain Diagnosis model based on dialogue representation.

In this embodiment, the data cleaning module 602 is specifically used for:

Obtain the cleaning requirements of the consultation data;

According to the consultation data and the cleaning requirement, determine a target cleaning rule for performing data cleaning on the consultation data;

Data cleaning is performed on the consultation data according to the target cleaning rule to obtain target consultation data.

6 and 7 above describe in detail the triage device based on dialogue representation in the embodiment of the present application from the perspective of modular functional entities, and the following describes the triage device based on dialogue representation in the embodiment of the present application from the perspective of hardware processing in detail. describe.

FIG. 8 is a schematic structural diagram of a dialogue representation-based triage device provided by an embodiment of the present application. The dialogue representation-based triage device 800 may vary greatly due to different configurations or performances, and may include one or more than one Central processing units (CPU) 810 (eg, one or more processors) and memory 820, one or more storage media 830 (eg, one or more mass storage devices) that store application programs 833 or data 832. Among them, the memory 820 and the storage medium 830 may be short-term storage or persistent storage. The program stored in the storage medium 830 may include one or more modules (not shown), and each module may include a series of instructions to operate on the dialogue characterization-based triage device 800 . Further, the processor 810 may be configured to communicate with the storage medium 830, and execute a series of instruction operations in the storage medium 830 on the dialogue-characterization-based triage device 800, so as to realize the dialogue-characterization-based diagnosis provided by the above method embodiments. Steps of the triage method.

Dialog-based triage device 800 may also include one or more power supplies 840, one or more wired or wireless network interfaces 850, one or more input-output interfaces 860, and/or, one or more operating systems 831, For example Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art can understand that the structure of the triage device based on dialogue representation shown in FIG. 8 does not constitute a limitation on the triage device based on dialogue representation provided by the present application, and may include more or less components than those shown in the figure. Either some components are combined, or different component arrangements.

The present application also provides a computer-readable storage medium. The computer-readable storage medium may be a non-volatile computer-readable storage medium. The computer-readable storage medium may also be a volatile computer-readable storage medium. The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the above-mentioned method for triage based on dialogue representation.

The blockchain referred to in this application is a new application mode of computer technologies such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm. Blockchain, essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information to verify its Validity of information (anti-counterfeiting) and generation of the next block. The blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

The integrated unit, if implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, and the computer software products are stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and other media that can store program codes .

As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand: The technical solutions recorded in the embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions in the embodiments of the present application.

Claims

A method for triage based on dialogue representation, wherein the method for triage based on dialogue representation comprises:

Acquire multiple rounds of dialogues generated by the subject to be triaged when they visit a doctor, and extract the consultation data in the multiple rounds of dialogues;

Data cleaning is performed on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

performing intention identification on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue based on the Euclidean distance;

The dialogue representation vector of each round of dialogue is input into the preset triage model for identification, and the triage information is obtained.
The method for triage based on dialogue representation according to claim 1, wherein the performing intention identification on the target interrogation data to obtain sentence pairs included in the target interrogation data comprises:

Acquiring all question and answer sentences in the target consultation data and object identifiers of objects corresponding to the question and answer sentences according to preset characterization rules;

Based on each of the object identifiers, intention identification is performed on each question and answer sentence in the target consultation data, to obtain an intention identification sequence corresponding to the consultation data;

According to the sequence of intent identifiers, determine the intent identifier of the target answer corresponding to the question;

Based on the intent identification of the target answer sentence, sentence pairs included in the target consultation data are obtained.
The triage method based on dialogue representation according to claim 1, wherein, in the call preset target BERT network model, feature extraction is performed on the sentence pair and the main complaint information, and a sentence pair vector of the sentence pair is obtained. And before the chief complaint vector of the chief complaint information, it also includes:

Establishing a pre-training task, calling a preset vectorization model to process historical user consultation data into several word vectors, wherein the word vectors include word vectors based on the historical user consultation data;

Input the several word vectors into the initial BERT network model, and obtain the training representation vector output by the initial BERT network model;

Calculate the loss value of the initial BERT network model according to the training characterization vector;

According to the loss value of the initial BERT network model, the model parameters of the initial BERT network model are adjusted to obtain the target BERT network model.
The triage method based on dialogue representation according to any one of claims 1-2, wherein the calculation of the Euclidean distance between the main complaint vector and the sentence pair vector is based on the Euclidean distance Determining the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue includes:

Calculate the Euclidean distance between the main complaint vector and each sentence pair vector in each round of dialogue;

Based on the Euclidean distance, determine the weight of each sentence pair in the corresponding round of dialogue;

Based on the weights, the weights corresponding to each sentence pair vector are weighted and averaged to obtain a dialogue representation vector corresponding to each round of dialogues in multiple rounds of dialogues.
The method for triage based on dialogue representation according to claim 1, wherein the dialogue representation vector of each round of dialogue is input into a preset triage model for identification, and the triage information obtained comprises:

Determine the target disease information in the target consultation data based on the dialogue representation vector corresponding to each round of dialogue;

Encoding the target disease information and the main complaint information corresponding to each round of dialogue to obtain the disease entity vector of the target consultation data;

Inputting the disease entity vector into a preset triage model for prediction, and obtaining the triage probability of different departments;

Sorting the triage probabilities to obtain triage information, wherein the triage information includes a preset number of recommended departments with the highest ranking and triage probabilities corresponding to the recommended departments.
The method for triage based on dialogue representation according to claim 5, wherein, before the dialogue representation vector of each round of dialogue is input into a preset triage model for identification and triage information is obtained, further comprising:

Obtain historical consultation data from a preset database;

Perform representation learning on the historical consultation data through the target BERT network model, and obtain a second dialogue representation vector of the historical consultation data;

Obtaining a preset training symptom and a department label corresponding to the training symptom, screening a preset node set association vector based on the training symptom, and obtaining a target vector corresponding to the training symptom;

performing a mapping operation on the second dialogue representation vector and the target vector to obtain a dialogue embedding vector and a target embedding vector;

splicing the dialogue embedding vector and the target embedding vector based on the training symptom to obtain a splicing vector;

Perform a convolution operation on the splicing vector at the convolution layer to obtain a convolution association vector, input the convolution association vector into the output layer, and obtain a predicted output result;

Calculate the prediction error loss based on the prediction output result and the department label, and update the parameters of the target BERT network model according to the prediction error loss, until the target BERT network model converges, and obtain the triage based on the dialogue representation Model.
The method for triage based on dialogue representation according to claim 1, wherein said performing data cleaning on said interrogation data to obtain target interrogation data comprises:

Obtain the cleaning requirements of the consultation data;

According to the consultation data and the cleaning requirement, determine a target cleaning rule for performing data cleaning on the consultation data;

Data cleaning is performed on the consultation data according to the target cleaning rule to obtain target consultation data.
A dialogue characterization-based triage device, wherein the dialogue characterization-based triage device comprises: a memory and at least one processor, wherein the memory stores instructions, and the memory and the at least one processor are connected through a circuit interconnection;

The at least one processor invokes the instructions in the memory to cause the dialogue-representation-based triage device to perform the steps of the dialogue-representation-based triage method as described below:

Acquire multiple rounds of dialogues generated by the subject to be triaged when they visit a doctor, and extract the consultation data in the multiple rounds of dialogues;

Data cleaning is performed on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

performing intention identification on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue based on the Euclidean distance;

The dialogue representation vector of each round of dialogue is input into the preset triage model for identification, and the triage information is obtained.
The triage device based on the dialogue representation according to claim 8, wherein the triage procedure based on the dialogue representation is executed by the processor to realize the intention recognition of the target inquiry data to obtain the target inquiry When the sentence pair step included in the diagnostic data is executed, the following steps are also performed:

Acquiring all question and answer sentences in the target consultation data and object identifiers of objects corresponding to the question and answer sentences according to preset characterization rules;

Based on each of the object identifiers, intention identification is performed on each question and answer sentence in the target consultation data, to obtain an intention identification sequence corresponding to the consultation data;

According to the sequence of intent identifiers, determine the intent identifier of the target answer corresponding to the question;

Based on the intent identification of the target answer sentence, sentence pairs included in the target consultation data are obtained.
The triage device based on dialogue representation according to claim 8, wherein the triage procedure based on dialogue representation is executed by the processor to realize the call preset target BERT network model for the sentence pair and the main complaint Before the step of performing feature extraction on the information to obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information, the following steps are also performed:

Establishing a pre-training task, calling a preset vectorization model to process historical user consultation data into several word vectors, wherein the word vectors include word vectors based on the historical user consultation data;

Input the several word vectors into the initial BERT network model, and obtain the training representation vector output by the initial BERT network model;

Calculate the loss value of the initial BERT network model according to the training characterization vector;

According to the loss value of the initial BERT network model, the model parameters of the initial BERT network model are adjusted to obtain the target BERT network model.
The dialogue-representation-based triage device according to any one of claims 8-9, wherein the dialogue-representation-based triage procedure is executed by the processor to realize the calculation of the main complaint vector and the sentence pair Euclidean distance between vectors, and when determining the dialogue representation vector step corresponding to each round of dialogue in multiple rounds of dialogue based on the Euclidean distance, the following steps are also performed:

Calculate the Euclidean distance between the main complaint vector and each sentence pair vector in each round of dialogue;

Based on the Euclidean distance, determine the weight of each sentence pair in the corresponding round of dialogue;

Based on the weights, the weights corresponding to each sentence pair vector are weighted and averaged to obtain a dialogue representation vector corresponding to each round of dialogues in multiple rounds of dialogues.
The dialogue representation-based triage device according to claim 8, wherein the dialogue representation-based triage procedure is executed by the processor to realize the input of the dialogue representation vector of each round of dialogue into a preset triage model for identification. , when the triage information step is obtained, the following steps are also performed:

Determine the target disease information in the target consultation data based on the dialogue representation vector corresponding to each round of dialogue;

Encoding the target disease information and the main complaint information corresponding to each round of dialogue to obtain the disease entity vector of the target consultation data;

Inputting the disease entity vector into a preset triage model for prediction, and obtaining the triage probability of different departments;

Sorting the triage probabilities to obtain triage information, wherein the triage information includes a preset number of recommended departments with the highest ranking and triage probabilities corresponding to the recommended departments.
The dialogue representation-based triage device according to claim 12, wherein the dialogue representation-based triage procedure is executed by the processor to realize the input of the dialogue representation vector of each round of dialogue into a preset triage model for identification. , before the steps to obtain triage information, also perform the following steps:

Obtain historical consultation data from a preset database;

Perform representation learning on the historical consultation data through the target BERT network model, and obtain a second dialogue representation vector of the historical consultation data;

Obtaining a preset training symptom and a department label corresponding to the training symptom, screening a preset node set association vector based on the training symptom, and obtaining a target vector corresponding to the training symptom;

performing a mapping operation on the second dialogue representation vector and the target vector to obtain a dialogue embedding vector and a target embedding vector;

splicing the dialogue embedding vector and the target embedding vector based on the training symptom to obtain a splicing vector;

Perform a convolution operation on the splicing vector at the convolution layer to obtain a convolution association vector, input the convolution association vector into the output layer, and obtain a predicted output result;

Calculate the prediction error loss based on the prediction output result and the department label, and update the parameters of the target BERT network model according to the prediction error loss, until the target BERT network model converges, and obtain the triage based on the dialogue representation Model.
A computer-readable storage medium on which a computer program is stored, wherein, when the computer program is executed by a processor, the steps of the method for triage based on dialogue characterization as described below are implemented:

Acquire multiple rounds of dialogues generated by the subject to be triaged when they visit a doctor, and extract the consultation data in the multiple rounds of dialogues;

Data cleaning is performed on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

performing intention identification on the target medical consultation data to obtain sentence pairs included in the target medical consultation data, wherein the target medical consultation data includes at least one sentence pair;

Call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

Calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and determine the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogue based on the Euclidean distance;

The dialogue representation vector of each round of dialogue is input into the preset triage model for identification, and the triage information is obtained.
The computer-readable storage medium according to claim 14, wherein the computer program is executed by the processor to perform the step of recognizing the intent of the target interview data to obtain sentence pairs included in the target interview data , also perform the following steps:

Acquiring all question and answer sentences in the target consultation data and object identifiers of objects corresponding to the question and answer sentences according to preset characterization rules;

Based on each of the object identifiers, intention identification is performed on each question and answer sentence in the target consultation data, to obtain an intention identification sequence corresponding to the consultation data;

According to the sequence of intent identifiers, determine the intent identifier of the target answer corresponding to the question;

Based on the intent identification of the target answer sentence, sentence pairs included in the target consultation data are obtained.
The computer-readable storage medium according to claim 14, wherein the computer program is executed by the processor to perform the call preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information to obtain the sentence Before the step of pairing the sentence pair vector and the main complaint vector of the main complaint information, the following steps are also performed:

Establishing a pre-training task, calling a preset vectorization model to process historical user consultation data into several word vectors, wherein the word vectors include word vectors based on the historical user consultation data;

Input the several word vectors into the initial BERT network model, and obtain the training representation vector output by the initial BERT network model;

Calculate the loss value of the initial BERT network model according to the training characterization vector;

According to the loss value of the initial BERT network model, the model parameters of the initial BERT network model are adjusted to obtain the target BERT network model.
The computer-readable storage medium of any one of claims 14-15, wherein the computer program is executed by a processor to calculate the Euclidean distance between the main complaint vector and the sentence pair vector, and When determining the dialogue characterization vector corresponding to each round of dialogue in the multi-round dialogue based on the Euclidean distance, the following steps are also performed:

Calculate the Euclidean distance between the main complaint vector and each sentence pair vector in each round of dialogue;

Based on the Euclidean distance, determine the weight of each sentence pair in the corresponding round of dialogue;

Based on the weights, the weights corresponding to each sentence pair vector are weighted and averaged to obtain a dialogue representation vector corresponding to each round of dialogues in multiple rounds of dialogues.
The computer-readable storage medium according to claim 14, wherein, when the computer program is executed by the processor of the step of inputting the dialogue representation vector of each round of dialogue into a preset triage model for identification and obtaining triage information, Also perform the following steps:

Determine the target disease information in the target consultation data based on the dialogue representation vector corresponding to each round of dialogue;

Encoding the target disease information and the main complaint information corresponding to each round of dialogue to obtain the disease entity vector of the target consultation data;

Inputting the disease entity vector into a preset triage model for prediction, and obtaining the triage probability of different departments;

Sorting the triage probabilities to obtain triage information, wherein the triage information includes a preset number of recommended departments with the highest ranking and triage probabilities corresponding to the recommended departments.
The computer-readable storage medium according to claim 18, wherein before the computer program is executed by the processor, the step of inputting the dialogue representation vector of each round of dialogue into a preset triage model for identification and obtaining triage information, Also perform the following steps:

Obtain historical consultation data from a preset database;

Perform representation learning on the historical consultation data through the target BERT network model, and obtain a second dialogue representation vector of the historical consultation data;

Obtaining a preset training symptom and a department label corresponding to the training symptom, screening a preset node set association vector based on the training symptom, and obtaining a target vector corresponding to the training symptom;

performing a mapping operation on the second dialogue representation vector and the target vector to obtain a dialogue embedding vector and a target embedding vector;

splicing the dialogue embedding vector and the target embedding vector based on the training symptom to obtain a splicing vector;

Perform a convolution operation on the splicing vector at the convolution layer to obtain a convolution association vector, input the convolution association vector into the output layer, and obtain a predicted output result;

Calculate the prediction error loss based on the prediction output result and the department label, and update the parameters of the target BERT network model according to the prediction error loss, until the target BERT network model converges, and obtain the triage based on the dialogue representation Model.
A triage device based on dialogue representation, wherein the triage device based on dialogue representation comprises:

The first acquisition module is used for acquiring multiple rounds of dialogues generated by the subject to be triaged when visiting a doctor, and extracting the consultation data in the multiple rounds of dialogues;

a data cleaning module, configured to perform data cleaning on the consultation data to obtain target consultation data, wherein the target consultation data includes the main complaint information of the user;

an intention recognition module, configured to perform intention recognition on the target consultation data to obtain sentence pairs included in the target consultation data, wherein the target consultation data includes at least one sentence pair;

The feature extraction module is used to call the preset target BERT network model to perform feature extraction on the sentence pair and the main complaint information, and obtain the sentence pair vector of the sentence pair and the main complaint vector of the main complaint information;

a first calculation module, used for calculating the Euclidean distance between the main complaint vector and the sentence pair vector, and determining the dialogue representation vector corresponding to each round of dialogue in multiple rounds of dialogues based on the Euclidean distance;

The identification module is used to input the dialogue representation vector of each round of dialogue into the preset triage model for identification, and obtain triage information.