WO2019000905A1 - 分诊对话方法、分诊对话设备及系统 - Google Patents

分诊对话方法、分诊对话设备及系统 Download PDF

Info

Publication number
WO2019000905A1
WO2019000905A1 PCT/CN2018/072098 CN2018072098W WO2019000905A1 WO 2019000905 A1 WO2019000905 A1 WO 2019000905A1 CN 2018072098 W CN2018072098 W CN 2018072098W WO 2019000905 A1 WO2019000905 A1 WO 2019000905A1
Authority
WO
WIPO (PCT)
Prior art keywords
triage
task
dialogue
user
intermediate representation
Prior art date
Application number
PCT/CN2018/072098
Other languages
English (en)
French (fr)
Inventor
张振中
Original Assignee
京东方科技集团股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 京东方科技集团股份有限公司 filed Critical 京东方科技集团股份有限公司
Priority to US16/082,031 priority Critical patent/US10872697B2/en
Priority to EP18758796.9A priority patent/EP3660854B1/en
Publication of WO2019000905A1 publication Critical patent/WO2019000905A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/14—Session management
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H40/00—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices
    • G16H40/60—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices for the operation of medical equipment or devices
    • G16H40/67—ICT specially adapted for the management or administration of healthcare resources or facilities; ICT specially adapted for the management or operation of medical equipment or devices for the operation of medical equipment or devices for remote operation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/10—Protocols in which an application is distributed across nodes in the network

Definitions

  • Embodiments of the present disclosure relate to a triage dialogue method, a triage dialogue device, and a system.
  • Intelligent triage refers to the diagnosis of possible diseases based on the patient's main symptoms and signs, the priority of the disease and its affiliated specialists, and the recommendation of effective treatment routes. Compared with traditional outpatient triage, intelligent triage can judge the disease more quickly and accurately and give reasonable advice. Therefore, under the current situation of social medical resources, intelligent triage has been widely concerned by the medical community and has broad application prospects.
  • the main task of the triage dialogue system is to quickly and accurately determine the subject's specialty according to the patient's symptoms and signs, and to provide an effective treatment path.
  • the intelligent triage system needs to interact with the patient to collect the patient's symptoms and signs.
  • the more interactions with the patient the more comprehensive the collected symptom and sign information, and the more accurate the judgment is made.
  • the more interactions with the patient the more time is required, which can delay the patient's visit time and waste medical resources. Therefore, the key to the successful completion of the triage task by the intelligent triage system is to accurately judge whether the task is successfully completed in the process of human-computer interaction.
  • Traditional triage systems are often based on manually written rule bases or machine learning based methods (such as the Airdoc triage system). However, neither of these methods explicitly assesses whether the smart triage task was successfully completed.
  • An embodiment of the present disclosure provides a method for a triage dialogue, including: receiving a triage conversation content; dividing the triage conversation content into a series of conversation features, and generating a conversation feature vector that corresponds one-to-one with the conversation feature; Inputting the dialog feature vector to an encoder, and generating an intermediate representation vector by the encoder; determining, according to the intermediate representation vector, whether the triage task is successfully completed; and adjusting the triage according to the judgment result of the triage task Dialogue strategy.
  • the encoder employs a two-way gated loop unit network; based on the forward hidden sequence and the backward hidden sequence of the two-way gated loop unit network, the intermediate representation vector h is:
  • determining whether the triage task is successfully completed according to the intermediate representation vector includes: calculating a success probability of completion of the triage task based on the previous intermediate representation vector and the intermediate representation vector.
  • the determining whether the triage task is successfully completed is characterized as a Gaussian process, and the success probability is:
  • H represents the previous intermediate representation vector
  • h represents the intermediate representation vector, which is the current intermediate representation
  • ⁇ ( ⁇ ) a distribution function that is a standard normal distribution
  • ⁇ * and They are the posterior mean and posterior variance of f(h), respectively
  • f( ⁇ ) represents the latent function, which is characterized as a Gaussian process.
  • the determination result is that the triage task is determined to be failed; when the success probability falls within the uncertainty interval, the determination result is that the triage task is uncertain. Whether the successful completion is successful; or, when the success probability falls within the success interval, the determination result is that the diagnosis of the triage task is completed.
  • adjusting the triage dialogue policy according to the judgment result of the triage task includes: when the judgment result is determined to be that the triage task fails, continuing to interact with the user; when the judgment result is uncertain When the triage task is successfully completed, prompting the user to provide feedback information, and determining whether to continue interacting with the user according to the feedback information; or, when the determination result is determining that the triage task is completed , terminating the conversation with the user.
  • determining whether to continue interacting with the user according to the feedback information includes: when the feedback information indicates that the triage task is completed Terminating the conversation with the user; or, when the feedback information indicates that the triage task has not been completed, the information with the highest information gain is selected to interact with the user.
  • selecting information with the greatest information gain to interact with the user includes calculating an information gain for different symptoms and determining a symptom associated with the maximum information gain; and selecting a symptom associated with the maximum information gain The user interacts.
  • An embodiment of the present disclosure provides a triage dialogue device, including a processor and a memory, where the memory is used to store an instruction, wherein when the instruction is executed by the processor, the following operations are performed: receiving a triage conversation content; The content of the triage dialogue is divided into a series of dialog features, and a dialog feature vector corresponding to the dialog feature is generated; the dialog feature vector is input to an encoder, and an intermediate representation vector is generated by the encoder; The intermediate representation vector determines whether the triage task is successfully completed; and adjusts the triage dialogue strategy according to the judgment result of the triage task.
  • the encoder employs a two-way gated loop unit network; based on the forward hidden sequence and the backward hidden sequence of the two-way gated loop unit network, the intermediate representation vector h is:
  • the determining, according to the intermediate representation vector, whether the triage task is successfully completed comprises: calculating a success probability of completion of the triage task based on a previous intermediate representation vector and the intermediate representation vector.
  • the determining whether the triage task is successfully completed is characterized as a Gaussian process, and the success probability is:
  • H represents the previous intermediate representation vector
  • h represents the intermediate representation vector, which is the current intermediate representation
  • ⁇ ( ⁇ ) a distribution function that is a standard normal distribution
  • ⁇ * and They are the posterior mean and posterior variance of f(h), respectively
  • f( ⁇ ) represents the latent function, which is characterized as a Gaussian process.
  • the determination result is that the triage task is determined to be failed; when the success probability falls within the uncertainty interval, the determination result is that the triage task is uncertain. Whether the successful completion is successful; or, when the success probability falls within the success interval, the determination result is that the diagnosis of the triage task is completed.
  • the adjusting the operation of the triage dialogue policy according to the judgment result of the triage task includes: when the judgment result is determined to be that the triage task fails, continuing to interact with the user; when the judgment result is In order to determine whether the triage task is successfully completed, prompting the user to provide feedback information, and determining whether to continue interacting with the user according to the feedback information; or, when the determination result is determining the triage When the task has completed, the conversation with the user is terminated.
  • determining whether to continue the operation of interacting with the user according to the feedback information comprises: when the feedback information indicates the score The diagnosis task has been completed, and the conversation with the user is terminated; or, when the feedback information indicates that the triage task has not been completed, the information with the largest information gain is selected to interact with the user.
  • the operation of interacting with the user with the information having the largest gain of selection information includes: calculating an information gain of different symptoms, and determining a symptom associated with the maximum information gain; and selecting a correlation with the maximum information gain The symptoms of the association interact with the user.
  • the triage dialog device further includes an interaction interface, wherein the interaction interface is configured to effect interaction with a user.
  • the embodiment of the present disclosure further provides a computer readable storage medium, where computer instructions are stored, and when the computer instructions are executed by the processor, the following operations are performed: receiving a triage conversation content; and dividing the triage conversation content into a series of dialog features, and generating a dialog feature vector corresponding to the dialog feature; inputting the dialog feature vector to an encoder, and generating an intermediate representation vector by the encoder; determining a score according to the intermediate representation vector Whether the diagnosis task is successfully completed; and adjusting the triage dialogue strategy according to the judgment result of the triage task.
  • FIG. 1 is a schematic block diagram of a triage dialogue system according to an embodiment of the present disclosure
  • 2A is a flowchart of a method for a triage dialogue according to an embodiment of the present disclosure
  • 2B is a second flowchart of a method for a triage dialogue according to an embodiment of the present disclosure
  • FIG. 3 is a schematic block diagram of a triage dialogue device according to an embodiment of the present disclosure.
  • FIG. 4 is a schematic block diagram of a triage dialogue device according to an embodiment of the present disclosure.
  • FIG. 5 is a schematic diagram of an encoder and a decoder according to an embodiment of the present disclosure.
  • the embodiment of the present disclosure proposes a cross-diagnosis dialogue device, a triage dialogue method, and a triage dialogue system based on active reward learning, which can quickly and accurately determine a subordinate subject according to symptoms and symptoms of a user (for example, a patient). Give valid medical advice.
  • active reward learning which can quickly and accurately determine a subordinate subject according to symptoms and symptoms of a user (for example, a patient).
  • the dialogue strategy is adjusted according to the judgment result.
  • the reward model is modeled by an active learning method using a Gaussian process, and a triage dialogue policy is adjusted based on the enhanced signal output by the reward model to maximize the expected reward value (ie, to quickly and accurately complete the score) Diagnosis task).
  • the triage dialogue device, the triage dialogue method, and the triage dialogue system have, but are not limited to, the following advantages: (1) determining whether user feedback is required by using an active learning method, thereby reducing user burden; (2) helping to alleviate The impact of noise data on dialogue strategy learning, improve performance; (3) through the user's explicit feedback, learning dialogue strategy, help to complete the triage task quickly and accurately.
  • the triage dialogue device learned that the injury was a snake injury by interacting with the user, but at this time, the triage dialogue device could not determine whether it was a poisonous snake wound or a non-toxic snake wound (the two snake wounds were treated differently). Therefore, the triage dialogue device continues to interact with the user, and it is inferred to be a poisonous snake wound by symptoms, but it is still impossible to determine which kind of snake wound (such as hemolytic toxin, neurotoxin or mixed toxin, different toxins correspond to different serums). Through further interaction with the user, it is judged by the shape characteristics of the snake that the user may be bitten by the python. Finally, it is recommended that the user register the emergency department and recommend treatment suggestions, and the triage task is successfully completed.
  • a cross-diagnosis dialogue device, a triage dialogue method, and a triage dialogue system provided by the embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.
  • FIG. 1 is a schematic block diagram of a triage dialogue system 100 according to an embodiment of the present disclosure.
  • the triage dialogue system 100 can include a triage dialogue device 101, a user device 106, a server 150, and the like.
  • Devices and/or servers in system 100 can be connected via network 160.
  • Each device and/or server of system 100 can communicate with one another directly or indirectly, for example, devices and/or servers of system 100 can transmit and receive data and/or signals to each other over network 160.
  • Network 160 may include wireless networks, wired networks, and/or any combination of wireless networks and wired networks.
  • network 160 may include a local area network, the Internet, a telecommunications network, an Internet of Things based Internet and/or telecommunications network, and/or any combination of the above, and the like.
  • network 160 can be a medical network.
  • the type and function of the network 160 in the present disclosure are not limited herein.
  • Server 150 can be a computing device that includes a processor and a memory.
  • server 150 can be a server or a cloud server in a local area network.
  • the triage dialogue device 101 can be a device for completing a triage task.
  • the triage dialogue device 101 will be described in detail below in conjunction with Figures 2A-5.
  • User device 106 can be a computing device that includes a processor and a memory.
  • user device 106 can be a television, smart home device, desktop computer, laptop, smart phone, tablet, game controller, music player (eg, mp3 player, etc.), and other terminals including a processor and memory (eg, , mobile terminal, intelligent terminal).
  • user device 106 can include a processor, memory, and other components such as input devices and output devices.
  • the user may interact with the triage dialogue device 101 via the user device 106.
  • a user may use an application (app) in the user device 106 to complete a conversation with the triage dialogue device 101.
  • each device and/or server of system 100 may also include display devices (eg, LCD, OLED, etc.), input devices (eg, touch devices, keyboards, microphones, mice, etc.), speakers, etc., as desired.
  • display devices eg, LCD, OLED, etc.
  • input devices eg, touch devices, keyboards, microphones, mice, etc.
  • speakers etc.
  • FIG. 2A is a flowchart of a method for a triage dialogue based on active reward learning provided by an embodiment of the present disclosure.
  • the triage dialogue method 200 includes: step S202, receiving the triage conversation content; step S204, generating an intermediate representation of the triage conversation content; step S206, determining, according to the intermediate representation, whether the triage task is successfully completed; S208: Adjust a triage dialogue strategy according to the judgment result of the triage task.
  • FIG. 2B shows a second flowchart of a method for a triage dialogue based on active reward learning provided by an embodiment of the present disclosure.
  • the triage dialogue method 250 includes: step S202, receiving a triage conversation content; step S2041, dividing the triage conversation content into a series of conversation features, and generating a conversation feature vector corresponding to the conversation feature one-to-one; Step S2042, inputting the dialog feature vector to the encoder, and generating an intermediate representation vector by the encoder; step S206', determining whether the triage task is successfully completed according to the intermediate representation vector; and step S208, according to the step S208 The judgment result of the triage task is adjusted to adjust the triage dialogue strategy.
  • the method 200 of FIG. 2A is similar to the method 250 of FIG. 2B.
  • the differences include: step S204 in FIG. 2A is decomposed into steps S2041 and S2042 of FIG. 2B; step S206 in FIG. 2A is replaced with step S206' of FIG. 2B.
  • step S206 in Fig. 2A is “intermediate representation vector”
  • step S206 in Fig. 2A is the same as step S206' in Fig. 2B.
  • receiving the triage conversation content includes: receiving the conversation content of the interaction between the triage dialogue device 101 and the user.
  • Table 1 above shows three sets of conversations between the triage dialogue device 101 and the user (Human Machine Dialogue 1, Human Machine Dialogue 2, Human Machine Dialogue 3).
  • the user can interact with the triage dialogue device 101 by voice, a display screen (eg, a touch screen with a touch function), a somatosensory device, a keyboard, a mouse, or an application (app) of the user device.
  • the intermediate representation is an intermediate representation vector h having a fixed dimension dim(h).
  • generating an intermediate representation of the triage conversation content first includes: dividing the triage conversation content into a series of conversation features, and generating a conversation feature corresponding to the conversation feature one-to-one Vector (step 2041 of Figure 2B).
  • the triage conversation cut into a series of dialogue features ⁇ d 1, d 2, ... , d T ⁇ , wherein each feature comprises a set d i interactive, 1 ⁇ i ⁇ T.
  • the dialogue feature d 1 ⁇ the snake bites the right hand; can you describe the symptoms? ⁇
  • the dialogue feature d 2 ⁇ right hand swelling, pain, palpitations, chest tightness; you can describe the type or color of the snake Shape?
  • dialogue characteristics d 3 ⁇ body short and thick, tail is particularly short, dark brown; it is recommended to register emergency department, timely injection of python anti-venom serum ⁇ , and so on.
  • the dialog features can be digitized (ie, the dialog features ⁇ d 1 , d 2 ,..., d T ⁇ can be converted into conversations Feature vector ).
  • a word representation ie, using a real vector of a specified length to represent a word
  • the word can be trained through neural network or deep learning to output a vector of a specified dimension, which can be used as an expression of the input word (eg, word2vec).
  • the present disclosure does not limit the method of digitizing the dialog features.
  • step S204 of FIG. 2A generating an intermediate representation of the triage conversation content, further comprising: constructing an encoder and a decoder; and using the conversation feature vector
  • the encoder is input to the encoder, and the intermediate representation vector h is generated by the encoder and output (step S2042 in Fig. 2B).
  • An example of the encoder and decoder is shown in FIG.
  • the encoder 502 can employ a bi-directional gated recurrent unit network (BGRU).
  • the gated loop unit network is a Recurrent Neural Network (RNN) that can alleviate the gradient dispersion problem.
  • RNN Recurrent Neural Network
  • BGRU encoder 502 with dialog feature vector As input, information for different directions of the feature sequence (from front to back and from back to front) is calculated. For example, the ith element in the forward hidden sequence And the ith element in the backward hidden sequence They are:
  • GRU ( ⁇ ) represents the activation function of the gated loop unit network.
  • the intermediate representation vector h is:
  • decoder 504 can be implemented with a forward-gated loop unit network that uses the intermediate representation vector h output by encoder 502 as input to generate a series of dialog feature vectors.
  • the target function used may be the output dialog feature vector.
  • input dialog feature vector Mean Square Error (MSE) calculated as follows:
  • N the number of dialog feature vectors in the training data
  • 2 the L2 paradigm. Since the functions used by both the encoder and the decoder are both steerable, the Stochastic Gradient Decent (SGD) can be used to train the encoder and decoder.
  • SGD Stochastic Gradient Decent
  • determining whether the triage task is successfully completed according to the intermediate representation includes: calculating a success probability of completion of the triage task based on the past intermediate representation and the intermediate representation. For example, based on a portion or all of the previous intermediate representations and the current intermediate representation, the probability of success of the completion of the triage task is calculated.
  • the completion of the triage task can be to infer the patient's condition, determine the department to which the condition belongs, and recommend treatment recommendations.
  • determining whether the triage task is successfully completed according to the intermediate representation vector includes: calculating the success of the completion of the triage task based on the previous intermediate representation vector and the intermediate representation vector. Probability. For example, based on a portion or all of the previous intermediate representations and the current intermediate representation, the probability of success of the completion of the triage task is calculated.
  • the determination of whether the triage task is successfully completed may be characterized as a Gaussian Process (GP), ie, a probability of success p(y
  • GP Gaussian Process
  • H indicates intermediate representation vector of the current conversation
  • H is the intermediate representation vector of the previous conversation (For example, the middle representation vector of some or all of the previous conversations, also known as the previous intermediate representation vector).
  • ⁇ ( ⁇ ) is the distribution function of the standard normal distribution
  • H) is a potential function
  • H) maps the vector of the dim(h) dimension to a real number (ie, R dim(h) ⁇ R).
  • Embodiments of the present disclosure characterize the latent function f( ⁇ ) as a Gaussian process, ie, f(h) ⁇ GP(m(h),k(h,h')), where m( ⁇ ) represents the mean function, k( ⁇ , ⁇ ) is a kernel function, the formula is as follows:
  • exp( ⁇ ) is an exponential function based on the natural constant e
  • ⁇ n is used to model noise.
  • the parameters p, l and ⁇ n can be parameterized by a gradient-based method.
  • the posterior mean ⁇ * and the posterior variance of f(h) can be calculated.
  • a posteriori mean ⁇ * and posterior variance for the calculation method see Y Engel, S Mannor, and R Meir, 2005. Reinforcement learning with Gaussian processes. In Proceedings of ICML).
  • the probability that the vector representation h of the current conversation can successfully complete the triage task is:
  • the triage process it is necessary to determine whether the current triage task is successfully completed. If the triage task has been successfully completed, there is no need to continue interacting with the user to prevent wasting user visit time. Of course, it is up to the user to feedback whether the triage task is successful. For example, the user can select the triage task to complete or not complete each time a tribune dialogue device answers a sentence. If the user selects the triage task to complete, the triage dialogue device terminates the conversation. However, this will increase the burden on the user and bring a bad experience to the user. Accordingly, embodiments of the present disclosure assess whether a current triage task requires user feedback through an active learning approach.
  • the uncertainty interval [ ⁇ , 1- ⁇ ] (0 ⁇ ⁇ ⁇ 0.5), the failure interval [0, ⁇ ), and the success interval ( ⁇ , 1] can be set.
  • the success probability p falls within the failure interval [ 0, ⁇ )
  • the judgment result of step S206 and/or step S206' is to determine that the triage task fails, indicating that there is a great possibility that the triage task is not completed, and at this time, it is necessary to continue to interact with the user. Make sure the task is completed or adjust the conversation strategy.
  • the success probability p falls within the success interval ( ⁇ , 1)
  • the result of the determination in step S206 and/or step S206' is to determine that the triage task has been completed, indicating that there is a high probability of completing the task.
  • the success probability p falls within the uncertainty interval [ ⁇ , 1- ⁇ ]
  • the result of the determination in step S206 and/or step S206′ is to determine whether the triage task is successfully completed, and the dialogue strategy needs to be adjusted ( As shown in step S208
  • step S208 according to the judgment result of the triage task, the triage dialogue strategy is adjusted, including:
  • Step S2080 when the determination result is that the triage task fails to be determined, continue to interact with the user and/or adjust the dialogue policy
  • Step S2082 When the determination result is that the diagnosis of the triage task is successful, the user is prompted to provide feedback information, and according to the feedback information, whether to continue to interact with the user; or
  • Step S2084 when the result of the determination is to determine that the triage task has been completed, the conversation with the user is terminated.
  • step S2082 when the determination result is uncertain whether the triage task is successfully completed, determining whether to continue interacting with the user according to the feedback information, including: when the feedback information indicates the The triage task has been completed, terminating the conversation with the user; or, when the feedback information indicates that the triage task has not been completed, the information with the largest information gain (IG) is selected to interact with the user.
  • IG information gain
  • information that maximizes information gain can be selected to interact with the user, including: calculating information gains for different symptoms, and determining symptoms associated with the maximum information gain; and selecting and The symptoms associated with the maximum information gain interact with the user.
  • Trigger Headache dizziness, nausea, urine Migraine Trigger Headache, dizziness, nausea, vomiting Neurasthenia Trigger Headache, dizziness, insomnia, anxiety, irritability
  • the triage dialogue device cannot determine which of the diseases in Table 2 the patient is suffering from.
  • the triage dialogue device needs to ask the patient for more questions to gather information to more accurately analyze the patient's condition. For example, a triage dialogue device can ask "Do you have any recent nausea symptoms?" or "Have you recently had more urine?" Specifically, it is first necessary to determine a symptom and then ask the patient based on the symptom. For example, the triage dialogue device can ask the patient "Do you have more urine recently?", and if the patient answers yes, the triage dialogue device can conclude that the patient is most likely to have diabetes. Therefore, how to choose a symptom to ask a patient is a problem that needs to be solved. For example, a triage dialogue device needs to select one of the symptoms such as "nausea", “more urine”, and "vomiting" to ask the patient. An exemplary solution is to select the symptom with the greatest information gain for interrogation.
  • the information gain is calculated as follows:
  • symptom means symptoms
  • diseases means disease
  • H( ⁇ ) means entropy.
  • IG (symptom) represents the information gain of the symptom of the symptom
  • H (diseases) represents the entropy of the disease
  • symptom) represents the entropy of the disease given the symptom of the symptom.
  • hypertension "nausea” and "more urine”
  • IG information gain
  • IG information gain
  • the entropy of the disease is:
  • an embodiment of the present disclosure further provides a triage dialogue setting 101 based on active reward learning, including: a processor 302 and a memory 304, wherein the memory 304 is configured to store an instruction, wherein the instruction is processed by the
  • the device 302 performs the following operations: receiving the triage conversation content; dividing the triage conversation content into a series of conversation features, and generating a conversation feature vector corresponding to the conversation feature; inputting the conversation feature vector Go to the encoder, and generate an intermediate representation vector by the encoder; determine whether the triage task is successfully completed according to the intermediate representation vector; and adjust the triage dialogue strategy according to the judgment result of the triage task.
  • Processor 302 can process data signals and can include various computing structures, such as a Complex Instruction Set Computer (CISC) architecture, a Structured Reduced Instruction Set Computer (RISC) architecture, or a structure that implements a combination of multiple instruction sets.
  • processor 302 can also be a microprocessor, such as an X86 processor or an ARM processor, or can be a digital processor (DSP) or the like.
  • the processor 302 can control other components in the triage dialogue device 101 to perform the desired functions.
  • Memory 304 can hold instructions and/or data executed by processor 302.
  • memory 304 can include one or more computer program products, which can include various forms of computer readable storage media, such as volatile memory and/or nonvolatile memory.
  • the volatile memory may include, for example, a random access memory (RAM) and/or a cache or the like.
  • the nonvolatile memory may include, for example, a read only memory (ROM), a hard disk, a flash memory, or the like.
  • One or more computer program instructions may be stored on the computer readable storage medium, and the processor 302 may execute the program instructions to implement the triage dialog function and/or other desired functions provided by embodiments of the present disclosure.
  • Various applications and various data may also be stored in the computer readable storage medium, such as various data used and/or generated by the application, and the like.
  • the encoder employs a two-way gated loop unit network; based on the forward hidden sequence and the backward hidden sequence of the two-way gated loop unit network, the intermediate representation vector h is:
  • the operation of determining, according to the intermediate representation vector, whether the triage task is successfully completed according to the intermediate representation vector, when the instruction is executed by the processor 302, includes: calculating based on a previous intermediate representation vector and the intermediate representation vector The probability of success of the completion of the triage task.
  • the determining whether the triage task is successfully completed is characterized as a Gaussian process, and the success probability is:
  • H represents a part or all of the intermediate representation vectors
  • h represents the intermediate representation vector, which is the current intermediate representation vector
  • ⁇ ( ⁇ ) is the distribution function of the standard normal distribution
  • ⁇ * and They are the posterior mean and posterior variance of f(h), respectively
  • f( ⁇ ) represents the latent function, which is characterized as a Gaussian process.
  • the determination result is that the triage task is determined to be failed; when the success probability falls within the uncertainty interval, the determination result is that the triage task is uncertain. Whether the successful completion is successful; or when the success probability falls within the success interval, the determination result is to determine that the triage task has been completed.
  • the operation of adjusting the triage dialogue policy according to the judgment result of the triage task implemented by the processor 302 when the instruction is executed includes: when the judgment result is determined to determine that the triage task fails And continuing to interact with the user; when the determination result is uncertain whether the triage task is successfully completed, prompting the user to provide feedback information, and determining whether to continue interacting with the user according to the feedback information; Alternatively, when the result of the determination is to determine that the triage task has been completed, the conversation with the user is terminated.
  • the instruction when the instruction is executed by the processor 302, "when the determination result is uncertain whether the triage task is successfully completed, whether to continue to interact with the user according to the feedback information is determined according to the feedback information.
  • the operation includes: when the feedback information indicates that the triage task has been completed, terminating the conversation with the user; or, when the feedback information indicates that the triage task has not been completed, selecting the information with the greatest information gain Interact with the user.
  • the operation of "the information that maximizes the information gain gain interacts with the user" implemented when the instruction is executed by the processor 302 includes: calculating an information gain of different symptoms, and determining that the maximum information gain is related to Symptoms; and selecting a symptom associated with the maximum information gain to interact with the user.
  • the triage dialogue device 101 also includes an interaction interface 306, wherein the interaction interface 306 is configured to effect interaction with the user 308.
  • the interactive interface 306 can include a microphone, a speaker, a camera, a button, a keyboard, a mouse, a display screen, and/or a somatosensory device.
  • the embodiment of the present disclosure further provides a computer readable storage medium, where computer instructions are stored, and when the computer instructions are executed by the processor, the following operations are performed: receiving a triage conversation content; and dividing the triage conversation content into a series of dialog features, and generating a dialog feature vector corresponding to the dialog feature; inputting the dialog feature vector to an encoder, and generating an intermediate representation vector by the encoder; determining a score according to the intermediate representation vector Whether the diagnosis task is successfully completed; and adjusting the triage dialogue strategy according to the judgment result of the triage task.
  • the computer readable storage medium can include, for example, volatile memory and/or nonvolatile memory.
  • the volatile memory may include, for example, a random access memory (RAM) and/or a cache or the like.
  • the nonvolatile memory may include, for example, a read only memory (ROM), a hard disk, a flash memory, or the like.
  • One or more computer program instructions can be stored on the computer readable storage medium, and the processor can execute the program instructions to implement the triage dialog function provided by embodiments of the present disclosure.
  • Various applications and various data may also be stored in the computer readable storage medium, such as various data used and/or generated by the application, and the like.
  • an embodiment of the present disclosure further provides a triage dialogue device 400.
  • the triage dialogue device 400 includes an interaction device 402, a parsing device 404, a judging device 406, and a policy adjustment device 408.
  • the triage dialogue device 400 can be implemented using software, hardware, or a combination of software and hardware.
  • the interaction device 402, the parsing device 404, the judging device 406, and the policy adjustment device 408 include code and programs stored in a memory; the processor can execute the code and program to implement some of the methods provided by the embodiments of the present disclosure or All features.
  • the interaction device 402, the parsing device 404, the judging device 406, and the policy adjustment device 408 can be dedicated hardware devices for implementing some or all of the functions provided by embodiments of the present disclosure.
  • the interaction device 402, the parsing device 404, the judging device 406, and the policy adjustment device 408 can be a single circuit board or a combination of multiple circuit boards.
  • the one or a combination of the plurality of boards may include: (1) one or more processors; (2) one or more non-transitory computer readable memories coupled to the processor; and (3) A firmware executable by the processor that is stored in the memory.
  • the interaction device 402 is configured to receive triage conversation content from the user 410.
  • the interaction device 402 can be the interaction interface 306 shown in FIG.
  • the parsing device 404 is configured to generate an intermediate representation of the triage conversation content.
  • the determining means 406 is configured to determine whether the triage task is successfully completed according to the intermediate representation.
  • the policy adjustment device 408 is configured to adjust the triage dialogue policy according to the judgment result of the triage task.
  • the intermediate representation is an intermediate representation vector
  • the parsing device 404 includes a decoder
  • the parsing device 404 is further configured to: split the triage conversation content into a series of conversation features, and generate and a dialog feature vector corresponding to the dialog feature; and using the dialog feature vector as an input to the encoder, the intermediate representation vector is generated by the encoder and output.
  • the encoder employs a two-way gated loop unit network.
  • the determining means 406 is further configured to calculate a success probability of completion of the triage task based on the past intermediate representation and the intermediate representation. For example, the determining means 406 is further configured to calculate a success probability of completion of the triage task based on the past intermediate representation vector and the intermediate representation vector.
  • the determining device 406 uses a Gaussian process to determine whether the triage task is successfully completed. The success probability is:
  • the determination result is that the triage task fails to be determined; when the success probability falls within the uncertainty interval, the determination result is uncertain whether the triage task is successful. Completion; or, when the success probability falls within the success interval, the determination result is to determine that the triage task has been completed.
  • the policy adjustment device 408 is further configured to continue to interact with the patient when the determination result is determined to be that the triage task fails; when the determination result is that the triage task is successfully completed, The patient is prompted to provide feedback information, and determines whether to continue interacting with the patient according to the feedback information; or, when the determination result is to determine that the triage task has been completed, the conversation with the patient is terminated.
  • the policy adjustment device 408 is further configured to: terminate the conversation with the patient when the feedback information indicates that the triage task has been completed Or, when the feedback information indicates that the triage task has not been completed, the information with the highest information gain is selected to interact with the patient.
  • the policy adjustment device 408 is further configured to: calculate an information gain for different symptoms, and determine a symptom associated with the maximum information gain; and select a symptom associated with the maximum information gain to interact with the patient.
  • the triage dialogue device, the triage dialogue method, the triage dialogue system, and the triage dialogue device provided by the embodiments of the present disclosure can determine whether the user feedback is needed by using the active learning method, thereby reducing the burden on the user and being able to quickly Complete the triage task accurately.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Biomedical Technology (AREA)
  • Public Health (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Machine Translation (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

一种分诊对话方法、分诊对话设备以及分诊对话系统。所述分诊对话方法包括：接收分诊对话内容(S202)；将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量(S2041)；输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量(S2042)；根据所述中间表示向量，判断分诊任务是否成功完成(S206')；以及根据所述分诊任务的判断结果，调整分诊对话策略(S208)。

Description

分诊对话方法、分诊对话设备及系统
本公开要求于2017年6月28日递交的中国专利申请第201710507287.0号的优先权，在此全文引用上述中国专利申请公开的内容以作为本公开的一部分。
技术领域
本公开的实施例涉及一种分诊对话方法、分诊对话设备及系统。
背景技术
智能分诊是指根据患者的主要症状及体征，诊断出可能的疾病，判断病情的轻重缓急及其隶属专科，并推荐有效的就诊路径等。相比传统的门诊分诊，智能分诊能够更快速更准确地判断疾病并给出合理建议。因此，在目前社会医疗资源紧张的情况下，智能分诊受到医疗界的广泛关注，具有广阔的应用前景。
具体来说，智能分诊系统(triage dialogue system)的主要任务是依据患者的症状体征快速准确地判断其隶属专科，并给出有效的就诊路径。为了完成这一任务，智能分诊系统需要同患者进行人机交互，收集患者的症状体征信息。通常来说，同患者交互的次数越多，收集的症状体征信息就越全面，从而做出的判断就越准确。但是，同患者交互的次数越多，也就意味着需要的时间越多，而这会耽误患者的就诊时间以及浪费医疗资源。因此，智能分诊系统能顺利完成分诊任务的关键在于在人机交互过程中能够准确地判断任务是否成功完成。传统的分诊系统通常基于人工编写的规则库或者基于机器学习的方法(例如Airdoc分诊系统)。但是这两种方式都没有显式地评估智能分诊任务是否成功完成。
发明内容
本公开实施例提供一种分诊对话方法，包括：接收分诊对话内容；将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的 对话特征向量；输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；根据所述中间表示向量，判断分诊任务是否成功完成；以及根据所述分诊任务的判断结果，调整分诊对话策略。
例如，所述编码器采用双向门控循环单元网络；基于所述双向门控循环单元网络的前向隐藏序列和后向隐藏序列，所述中间表示向量h为：
Figure PCTCN2018072098-appb-000001
其中，
Figure PCTCN2018072098-appb-000002
表征所述前向隐藏序列
Figure PCTCN2018072098-appb-000003
中的第i个元素
Figure PCTCN2018072098-appb-000004
和所述后向隐藏序列
Figure PCTCN2018072098-appb-000005
中第i个元素
Figure PCTCN2018072098-appb-000006
的连接，以及T表征所述对话特征的数目。
例如，根据所述中间表示向量，判断分诊任务是否成功完成，包括：基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。
例如，所述判断分诊任务是否成功完成被刻画为高斯过程，所述成功概率为：
Figure PCTCN2018072098-appb-000007
其中，H表示以往的中间表示向量；h表示所述中间表示向量，其为当前的中间表示；y∈{-1,1}，y＝1表示所述分诊任务成功完成；φ(·)为标准正态分布的分布函数；μ *和
Figure PCTCN2018072098-appb-000008
分别为f(h)的后验均值和后验方差，f(·)表示潜在函数，被刻画为高斯过程。
例如，当所述成功概率落入失败区间，则所述判断结果为确定所述分诊任务失败；当所述成功概率落入不确定区间，则所述判断结果为不确定所述分诊任务是否成功完成；或者，当所述成功概率落入成功区间，则所述判断结果为确定所述分诊任务已完成。
例如，根据所述分诊任务的判断结果，调整分诊对话策略，包括：当所述判断结果为确定所述分诊任务失败时，继续与用户进行交互；当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述用户提供反馈信息，并根据所述反馈信息确定是否继续与所述用户进行交互；或者，当所述判断 结果为确定所述分诊任务已完成时，终止与所述用户的对话。
例如，当所述判断结果为不确定所述分诊任务是否成功完成时，根据所述反馈信息确定是否继续与所述用户进行交互，包括：当所述反馈信息表示所述分诊任务已完成，终止与所述用户的对话；或者，当所述反馈信息表示所述分诊任务尚未完成，选择信息增益最大的信息与所述用户进行交互。
例如，选择信息增益最大的信息与所述用户进行交互，包括：计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及选择与所述最大信息增益相关联的症状与所述用户进行交互。
本公开实施例提供一种分诊对话设备，包括处理器和存储器，所述存储器用于存储指令，其中，所述指令被所述处理器执行时实现以下操作：接收分诊对话内容；将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；根据所述中间表示向量，判断分诊任务是否成功完成；以及根据所述分诊任务的判断结果，调整分诊对话策略。
例如，所述编码器采用双向门控循环单元网络；基于所述双向门控循环单元网络的前向隐藏序列和后向隐藏序列，所述中间表示向量h为：
Figure PCTCN2018072098-appb-000009
其中，
Figure PCTCN2018072098-appb-000010
表征所述前向隐藏序列
Figure PCTCN2018072098-appb-000011
中的第i个元素
Figure PCTCN2018072098-appb-000012
和所述后向隐藏序列
Figure PCTCN2018072098-appb-000013
中第i个元素
Figure PCTCN2018072098-appb-000014
的连接，T表征所述对话特征的数目。
例如，所述根据所述中间表示向量，判断分诊任务是否成功完成的操作，包括：基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。
例如，所述判断分诊任务是否成功完成被刻画为高斯过程，所述成功概率为：
Figure PCTCN2018072098-appb-000015
其中，H表示以往的中间表示向量；h表示所述中间表示向量，其为当 前的中间表示；y∈{-1,1}，y＝1表示所述分诊任务成功完成；φ(·)为标准正态分布的分布函数；μ *和
Figure PCTCN2018072098-appb-000016
分别为f(h)的后验均值和后验方差，f(·)表示潜在函数，被刻画为高斯过程。
例如，当所述成功概率落入失败区间，则所述判断结果为确定所述分诊任务失败；当所述成功概率落入不确定区间，则所述判断结果为不确定所述分诊任务是否成功完成；或者，当所述成功概率落入成功区间，则所述判断结果为确定所述分诊任务已完成。
例如，所述根据所述分诊任务的判断结果，调整分诊对话策略的操作，包括：当所述判断结果为确定所述分诊任务失败时，继续与用户进行交互；当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述用户提供反馈信息，并根据所述反馈信息确定是否继续与所述用户进行交互；或者，当所述判断结果为确定所述分诊任务已完成时，终止与所述用户的对话。
例如，所述当所述判断结果为不确定所述分诊任务是否成功完成时，根据所述反馈信息确定是否继续与所述用户进行交互的操作，包括：当所述反馈信息表示所述分诊任务已完成，终止与所述用户的对话；或者，当所述反馈信息表示所述分诊任务尚未完成，选择信息增益最大的信息与所述用户进行交互。
例如，所述选择信息增益最大的信息与所述用户进行交互的操作，包括：计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及选择与所述最大信息增益相关联的症状与所述用户进行交互。
例如，所述分诊对话设备还包括交互接口，其中，所述交互接口被配置为实现与用户之间的交互。
本公开实施例还提供一种计算机可读存储介质，其上存储有计算机指令，所述计算机指令被处理器执行时实现以下操作：接收分诊对话内容；将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；根据所述中间表示向量，判断分诊任务是否成功完成；以及根据所述分诊任务的判断结果，调整分诊对话策略。
附图说明
为了更清楚地说明本公开实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本公开的一些实施例，而非对本公开的限制，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本公开实施例提供的一种分诊对话系统的示意性框图；
图2A为本公开实施例提供的一种分诊对话方法的流程图之一；
图2B为本公开实施例提供的一种分诊对话方法的流程图之二；
图3为本公开实施例提供的一种分诊对话设备的示意性框图；
图4为本公开实施例提供的一种分诊对话装置的示意性框图；
图5为本公开实施例提供的一种编码器和解码器的示意图。
具体实施方式
为使本公开的目的、技术方案和优点更加清楚，以下举实施例对本公开作进一步详细说明。显然，所描述的实施例仅仅是本公开一部分实施例，而不是全部的实施例。基于本公开中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例，都属于本公开保护的范围。
本公开实施例提出一种基于主动奖励学习(active reward learning)的分诊对话设备、分诊对话方法以及分诊对话系统，能够依据用户(例如，患者)的症状体征快速准确地判断隶属专科并给出有效的就诊建议。在每轮分诊对话过程中，显式地判断分诊任务是否可以成功完成，并依据该判断结果调整对话策略。例如，通过使用高斯过程的主动学习方法建模奖励模型，并依据奖励模型输出的强化信号调整分诊对话策略(triage dialogue policy)，以便使期望奖励值达到最大(即，以便快速准确地完成分诊任务)。所述分诊对话设备、分诊对话方法以及分诊对话系统具有，但不限于，如下优点：(1)通过使用主动学习方法判断是否需要用户反馈，减轻用户负担；(2)有助于减轻噪音数据对对话策略学习的影响，提高性能；(3)通过用户的显式反馈，学习对话策略，有助于快速准确地完成分诊任务。
例如，当用户(例如，患者)与分诊对话系统进行交互时，本公开实施例提供的分诊对话设备、分诊对话方法以及分诊对话系统可以与用户一起完 成如下表格1所示的分诊对话：
表格1：
Figure PCTCN2018072098-appb-000017
在上述示例中，分诊对话设备通过和用户交互了解到伤情是蛇伤，但此时分诊对话设备还无法确定是有毒蛇伤还是无毒蛇伤(这两种蛇伤的处理方法不同)。因此，分诊对话设备继续和用户交互，通过症状推断是有毒蛇伤，但还无法确定是哪种蛇伤(例如溶血性毒素、神经性毒素还是混合毒素，不同的毒素对应不同的血清)。通过和用户的进一步交互，由蛇的外形特征判断用户可能被蝮蛇咬伤，最后建议用户挂号急诊科并推荐治疗建议，至此分诊任务成功完成。
下面将结合附图对本公开实施例提供的一种分诊对话设备、分诊对话方法以及分诊对话系统进行详细的说明。
图1为本公开实施例提供的一种分诊对话系统100的示意性框图。如图1所示，分诊对话系统100可以包括分诊对话设备101、用户设备106和服务器150等。系统100中的各设备和/或服务器可以通过网络160连接。系统100的各设备和/或服务器之间可以直接或间接地互相通信，例如，系统100的各设备和/或服务器可以通过网络160互相发送和接收数据和/或信号。
网络160可以包括无线网络、有线网络、和/或无线网络和有线网络的任意组合。例如，网络160可以包括局域网、互联网、电信网、基于互联网和/或电信网的物联网(Internet of Things)、和/或以上网络的任意组合等。例如，网络160可以为医疗网络。本公开对网络160的类型和功能在此不作限制。
服务器150可以为一种包括处理器和存储器的计算设备。例如，服务器150可以为局域网中的服务器或云端服务器。
分诊对话设备101可以为一种用于完成分诊任务的设备。分诊对话设备 101将在下面结合图2A-5进行详细的说明。
用户设备106可以为一种包括处理器和存储器的计算设备。例如，用户设备106可以为电视、智能家电设备、台式电脑、笔记本电脑、智能手机、平板电脑、游戏控制器、音乐播放器(例如mp3播放器等)以及其他包括处理器和存储器的终端(例如，移动终端，智能终端)。在一些实施例中，用户设备106可以包括处理器、存储器以及诸如输入设备和输出设备等其他部件。在一些例子中，用户可以通过用户设备106来实现与分诊对话设备101的交互。例如，用户可以使用用户设备106中的应用(app)来完成与分诊对话设备101的对话。
在一些实施例中，系统100的各设备和/或服务器根据需要还可以包括显示装置(例如LCD、OLED等)、输入装置(例如触控装置、键盘、麦克风、鼠标等)、扬声器等。本公开在此不作限定。
图2A示出了本公开实施例提供的一种基于主动奖励学习的分诊对话方法200的流程图之一。该分诊对话方法200包括：步骤S202，接收分诊对话内容；步骤S204，生成所述分诊对话内容的中间表示；步骤S206，根据所述中间表示，判断分诊任务是否成功完成；以及步骤S208，根据所述分诊任务的判断结果，调整分诊对话策略。
图2B示出了本公开实施例提供的一种基于主动奖励学习的分诊对话方法200的流程图之二。该分诊对话方法250包括：步骤S202，接收分诊对话内容；步骤S2041，将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；步骤S2042，输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；步骤S206’，根据所述中间表示向量，判断分诊任务是否成功完成；以及步骤S208，根据所述分诊任务的判断结果，调整分诊对话策略。
图2A的方法200和图2B的方法250相类似，不同点包括：图2A中的步骤S204被分解为图2B的步骤S2041和S2042；图2A中的步骤S206被替换为图2B的步骤S206’。当图2A的步骤S206的“中间表示”为“中间表示向量”时，图2A中的步骤S206和图2B的步骤S206’相同。
在步骤S202中，接收分诊对话内容，包括：接收分诊对话设备101与用户之间交互的对话内容。例如，上述表格1示出了分诊对话设备101与用 户的三组对话(人机对话1、人机对话2、人机对话3)。用户可以通过语音、显示屏(例如，带触摸功能的触摸屏)、体感装置、键盘、鼠标或用户设备的应用(app)等，来实现与分诊对话设备101的交互。
在一些实施例中，所述中间表示为中间表示向量h，具有固定维数dim(h)。在图2A的步骤S204中，生成所述分诊对话内容的中间表示，首先包括：将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量(图2B的步骤2041)。
具体地，将分诊对话内容切分成一系列的对话特征{d 1,d 2,…,d T}，其中，每一个特征d i包含了一组人机对话，1≤i≤T。例如，结合上述表格1，对话特征d 1＝{被蛇咬了右手；您能描述一下症状吗}，对话特征d 2＝{右手红肿疼痛、心悸、胸闷；您能描述一下蛇的种类或者颜色外形吗}，对话特征d 3＝{身体短粗，尾巴特别短，暗褐色；建议挂号急诊科，及时注射蝮蛇抗毒血清}，依此类推。
由于使用计算机对自然语言进行处理需要将自然语言转换为机器能够识别的符号，因此，可以将对话特征进行数值化(即，将对话特征{d 1,d 2,…,d T}转化为对话特征向量
Figure PCTCN2018072098-appb-000018
)。例如，可以利用词向量(word representation)(即，使用一个指定长度的实数向量来表示一个词)实现对话特征的数值化。又例如，可以通过神经网络或者深度学习对词进行训练，输出指定维度的向量，该向量可以作为输入词的表达(例如，word2vec)。本公开对对话特征的数值化的方法不作限定。
在图2A的步骤S204中，生成所述分诊对话内容的中间表示，还包括：构建编码器和解码器；以及将所述对话特征向量
Figure PCTCN2018072098-appb-000019
输入至所述编码器，并通过所述编码器生成所述中间表示向量h并输出(图2B中的步骤S2042)。所述编码器和解码器的一个例子如图5所示。
例如，参照图5，编码器502可以采用双向门控循环单元网络(bi-directional gated recurrent unit network,BGRU)。门控循环单元网络是一种能够缓解梯度弥散问题的循环神经网络(Recurrent Neural Network,RNN)。BGRU编码器502以对话特征向量
Figure PCTCN2018072098-appb-000020
作为输入，计算特征序列不同方向(从前向后和从后向前)的信息。例如，前向隐藏序列中的第i个元素
Figure PCTCN2018072098-appb-000021
和后向隐藏序列中的第i个元素
Figure PCTCN2018072098-appb-000022
分别为：
Figure PCTCN2018072098-appb-000023
Figure PCTCN2018072098-appb-000024
其中，GRU(·)表示门控循环单元网络的激活函数。基于所述双向门控循环单元网络的前向隐藏序列和后向隐藏序列，所述中间表示向量h为：
Figure PCTCN2018072098-appb-000025
其中，
Figure PCTCN2018072098-appb-000026
用以表征所述前向隐藏序列
Figure PCTCN2018072098-appb-000027
中的第i个元素
Figure PCTCN2018072098-appb-000028
和所述后向隐藏序列
Figure PCTCN2018072098-appb-000029
中第i个元素
Figure PCTCN2018072098-appb-000030
的连接(concatenation)；以及T表征所述对话特征的数目。
继续参照图5，解码器504可以采用前向门控循环单元网络来实现，其以编码器502输出的中间表示向量h作为输入，产生一系列对话特征向量
Figure PCTCN2018072098-appb-000031
Figure PCTCN2018072098-appb-000032
编码器502和解码器504进行训练时，采用的目标函数可以为输出的对话特征向量
Figure PCTCN2018072098-appb-000033
和输入的对话特征向量
Figure PCTCN2018072098-appb-000034
之间的均方误差(Mean Square Error,MSE)，计算方式如下：
Figure PCTCN2018072098-appb-000035
其中，N表示训练数据中对话特征向量的个数，||·|| 2表示L2范式。因为编码器和解码器所用的函数都是可导的，可以采用随机梯度下降方法(Stochastic Gradient Decent,SGD)去训练编码器和解码器。
继续参照图2A，在步骤S206中，根据所述中间表示，判断分诊任务是否成功完成，包括：基于以往的中间表示以及所述中间表示，计算所述分诊任务完成的成功概率。例如，基于之前的一部分或所有的中间表示以及目前的中间表示，计算所述分诊任务完成的成功概率。分诊任务完成可以为推断 出患者的患病情况、判断病症隶属的科室并推荐治疗建议等。
例如，在图2B的步骤S206’中，根据所述中间表示向量，判断分诊任务是否成功完成，包括：基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。例如，基于之前的一部分或所有的中间表示以及目前的中间表示，计算所述分诊任务完成的成功概率。例如，在图2A的步骤S206和/或图2B的步骤S206’中，所述判断分诊任务是否成功完成可以被刻画为高斯过程(Gaussian Process,GP)，即计算成功概率p(y|h,H)，其中：y∈{-1,1}，y＝-1表示分诊失败，y＝1表示分诊成功；h表示当前对话的中间表示向量；H是以前的对话的中间表示向量(例如，以前的一部分或所有的对话的中间表示向量，又称作以往的中间表示向量)。
可以定义p(y＝1|h,H)＝φ(f(h|H))，其中，φ(·)是标准正态分布的分布函数，以及f(h|H)是一个潜在函数，f(h|H)将dim(h)维的向量映射为实数(即R dim(h)→R)。本公开实施例将潜在函数f(·)刻画为高斯过程，即f(h)～GP(m(h),k(h,h'))，其中，m(·)表示均值函数，k(·,·)是核函数，其计算公式如下：
Figure PCTCN2018072098-appb-000036
其中，exp(·)是以自然常数e为底的指数函数，ε n用来建模噪音。参数p、l以及ε n可以通过基于梯度的方法进行参数学习。给定当前对话的中间表示向量h和之前对话的中间表示向量H，可以计算f(h)的后验均值μ *和后验方差
Figure PCTCN2018072098-appb-000037
(后验均值μ *和后验方差
Figure PCTCN2018072098-appb-000038
的计算方法可以参见Y Engel,S Mannor,and R Meir,2005.Reinforcement learning with Gaussian processes.In Proceedings of ICML)。当前对话的中间表示向量h能够成功完成分诊任务的概率(即，成功概率)为：
Figure PCTCN2018072098-appb-000039
在分诊过程中，需要判断当前分诊任务是否成功完成。如果已经成功完成分诊任务，则不需要继续和用户进行交互，以防止浪费用户的就诊时间。当然，可以由用户反馈分诊任务是否成功。例如，分诊对话设备每回答一句话，用户都可以选择分诊任务完成或者没完成，如果用户选择分诊任务完成， 则分诊对话设备终止此次对话。然而，这样会增加用户的负担，给用户带来不好的体验。因此，本公开的实施例通过主动学习方法评估当前分诊任务是否需要用户反馈。
例如，可以设定不确定区间[λ,1-λ](0<λ<0.5)，失败区间[0,λ)以及成功区间(λ,1]。当所述成功概率p落入失败区间[0,λ)，则步骤S206和/或步骤S206’的所述判断结果为确定所述分诊任务失败，表示有很大的可能性没有完成分诊任务，此时，需要继续和用户交互以确保任务完成或者调整对话策略。当所述成功概率p落入成功区间(λ,1]，则步骤S206和/或步骤S206’的所述判断结果为确定所述分诊任务已完成，表示有很大的可能性完成任务。当所述成功概率p落入不确定区间[λ,1-λ]，则步骤S206和/或步骤S206’的所述判断结果为不确定所述分诊任务是否成功完成，需要调整对话策略(如步骤S208所示)。
在步骤S208中，根据所述分诊任务的判断结果，调整分诊对话策略，包括：
步骤S2080，当所述判断结果为确定所述分诊任务失败时，继续与用户交互和/或调整对话策略；
步骤S2082，当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述用户提供反馈信息，并根据所述反馈信息确定是否继续与所述用户进行交互；或者，
步骤S2084，当所述判断结果为确定所述分诊任务已完成时，终止与所述用户的对话。
例如，在步骤S2082中，当所述判断结果为不确定所述分诊任务是否成功完成时，根据所述反馈信息确定是否继续与所述用户进行交互，包括：当所述反馈信息表示所述分诊任务已完成，终止与所述用户的对话；或者，当所述反馈信息表示所述分诊任务尚未完成，选择信息增益(information gain,IG)最大的信息与所述用户进行交互。
例如，在调整对话策略时，可以选择信息增益最大的信息与所述用户进行交互，其包括：计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及选择与所述最大信息增益相关联的症状与所述用户进行交互。
下面以表格2中所示的疾病和症状为例，对选择信息增益最大的信息与用户进行交互的操作进行说明。为了简单清楚地描述示例，下面的说明只局限在表格2所列的疾病和症状，然而，其应用并不限于表格2中的3种疾病以及8种症状，其可以应用于任何疾病以及任何症状，本公开在此不作限定。
表格2：
疾病 关系 症状或体征
高血压 引发 头疼、眩晕、恶心、尿多
偏头痛 引发 头疼、眩晕、恶心、呕吐
神经衰弱 引发 头疼、眩晕、失眠、焦虑、烦躁
假设有患者出现“头疼”和“眩晕”的症状，此时分诊对话设备无法判断该患者是患有表格2中的哪一种疾病。分诊对话设备需要向患者询问更多的问题来收集信息以便更准确的分析患者的情况。例如，分诊对话设备可以问“请问您最近出现恶心的症状吗？”或者“您最近尿多吗？”等。具体来说，首先要确定一种症状，然后依据该症状询问患者。例如，分诊对话设备可以询问患者“您最近尿多吗？”，如果患者回答是，则分诊对话设备可以断定该患者有很大的可能是患有糖尿病。因此，如何选择症状来询问患者是一个需要解决的问题。例如，分诊对话设备需要从“恶心”、“尿多”、“呕吐”等症状中选择一个来询问患者。一种示例性的解决方案是选择信息增益最大的症状来进行询问。
例如，信息增益的计算方式如下：
IG(symptom)＝H(diseases)-H(diseases|sympton),
其中，symptom表示症状，diseases表示疾病，H(·)表示熵。例如，IG(symptom)表示该症状symptom的信息增益，H(diseases)表示患病的熵，H(diseases|symptom)表示给定该症状symptom时患病的熵。
下面以高血压的两个相关症状“恶心”和“尿多”为例，说明信息增益的计算过程以及症状的选择过程。对于这两个症状，可以分别计算它们的信息增益(IG)，并选择信息增益最大的症状向患者询问。假设患者患上“高血压”，“偏头痛”和“神经衰弱”的概率均服从均匀分布，例如，概率p(高血压)＝p(偏头痛)＝p(神经衰弱)＝1/3，其中，p(·)表示概率。在这种情况下，患病的熵为：
Figure PCTCN2018072098-appb-000040
当症状为“恶心”的时候，概率p(高血压)＝1/2，p(偏头痛)＝1/2，p(神经衰弱)＝0，这是因为在本例中出现恶心的疾病只有高血压和偏头痛。此时，给定该恶心症状时患病的熵H(diseases|恶心)、以及恶心症状的信息增益IG(恶心)为：
Figure PCTCN2018072098-appb-000041
则
Figure PCTCN2018072098-appb-000042
当症状为“尿多”的时候，概率p(高血压)＝1，p(偏头痛)＝0，p(神经衰弱)＝0，这是因为在本例中出现尿多的疾病只有高血压。此时，给定该尿多症状时患病的熵H(diseases|尿多)、以及尿多症状的信息增益IG(尿多)为
Figure PCTCN2018072098-appb-000043
Figure PCTCN2018072098-appb-000044
因为尿多的信息增益大于恶心的信息增益(即，IG(尿多)>IG(恶心))，所以可以选择症状“尿多”向患者提问(例如，提问“您最近出现尿多情况吗？”)。如果患者回答“是”，则确定疾病为高血压。如果回答“否”，则排除高血压，并将剩下的疾病(对于本例来说，“偏头痛”、“神经衰弱”)作为候选的疾病来重复上述步骤，直到确定疾病或者患者终止该过程。
参照图3，本公开实施例还提供一种基于主动奖励学习的分诊对话设101，包括：处理器302和存储器304，所述存储器304用于存储指令，其中，所述指令被所述处理器302执行时实现以下操作：接收分诊对话内容；将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；根据所述中间表示向量，判断分诊任务是否成功完成；以及根据所述分诊任务的判断结果，调整分诊对话策略。
处理器302可以处理数据信号，可以包括各种计算结构，例如复杂指令集计算机(CISC)结构、结构精简指令集计算机(RISC)结构或者一种实行多种指令集组合的结构。在一些实施例中，处理器302也可以是微处理器，例如X86处理器或ARM处理器，或者可以是数字处理器(DSP)等。处理器302可以控制分诊对话设备101中的其它部件以执行期望的功能。
存储器304可以保存处理器302执行的指令和/或数据。例如，存储器 304可以包括一个或多个计算机程序产品，所述计算机程序产品可以包括各种形式的计算机可读存储介质，例如易失性存储器和/或非易失性存储器。所述易失性存储器例如可以包括随机存取存储器(RAM)和/或高速缓冲存储器(cache)等。所述非易失性存储器例如可以包括只读存储器(ROM)、硬盘、闪存等。在所述计算机可读存储介质上可以存储一个或多个计算机程序指令，处理器302可以运行所述程序指令，以实现本公开实施例提供的分诊对话功能以及/或者其它期望的功能。在所述计算机可读存储介质中还可以存储各种应用程序和各种数据，例如所述应用程序使用和/或产生的各种数据等。
例如，所述编码器采用双向门控循环单元网络；基于所述双向门控循环单元网络的前向隐藏序列和后向隐藏序列，所述中间表示向量h为：
Figure PCTCN2018072098-appb-000045
其中，
Figure PCTCN2018072098-appb-000046
用以表征所述前向隐藏序列
Figure PCTCN2018072098-appb-000047
中的第i个元素
Figure PCTCN2018072098-appb-000048
和所述后向隐藏序列
Figure PCTCN2018072098-appb-000049
中第i个元素
Figure PCTCN2018072098-appb-000050
的连接，T表征所述对话特征的数目。
例如，所述指令被所述处理器302执行时实现的“根据所述中间表示向量，判断分诊任务是否成功完成”的操作，包括：基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。
例如，所述判断分诊任务是否成功完成被刻画为高斯过程，所述成功概率为：
Figure PCTCN2018072098-appb-000051
其中，H表示以往的一部分或全部中间表示向量；h表示所述中间表示向量，其为当前的中间表示向量；y∈{-1,1}，y＝1表示所述分诊任务成功完成；φ(·)为标准正态分布的分布函数；μ *和
Figure PCTCN2018072098-appb-000052
分别为f(h)的后验均值和后验方差，f(·)表示潜在函数，被刻画为高斯过程。
例如，当所述成功概率落入失败区间，则所述判断结果为确定所述分诊任务失败；当所述成功概率落入不确定区间，则所述判断结果为不确定所述 分诊任务是否成功完成；或者当所述成功概率落入成功区间，则所述判断结果为确定所述分诊任务已完成。
例如，所述指令被所述处理器302执行时实现的“根据所述分诊任务的判断结果，调整分诊对话策略”的操作，包括：当所述判断结果为确定所述分诊任务失败时，继续与用户进行交互；当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述用户提供反馈信息，并根据所述反馈信息确定是否继续与所述用户进行交互；或者，当所述判断结果为确定所述分诊任务已完成时，终止与所述用户的对话。
例如，所述指令被所述处理器302执行时实现的“所述当所述判断结果为不确定所述分诊任务是否成功完成时，根据所述反馈信息确定是否继续与所述用户进行交互”的操作，包括：当所述反馈信息表示所述分诊任务已完成，终止与所述用户的对话；或者，当所述反馈信息表示所述分诊任务尚未完成，选择信息增益最大的信息与所述用户进行交互。
例如，所述指令被所述处理器302执行时实现的“选择信息增益最大的信息与所述用户进行交互”的操作，包括：计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及选择与所述最大信息增益相关联的症状与所述用户进行交互。
如图3所示，所述分诊对话设备101还包括交互接口306，其中，所述交互接口306被配置为实现与用户308之间的交互。所述交互接口306可以包括麦克风、扬声器、摄像头、按钮、键盘、鼠标、显示屏和/或体感装置等。
本公开实施例还提供一种计算机可读存储介质，其上存储有计算机指令，所述计算机指令被处理器执行时实现以下操作：接收分诊对话内容；将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；根据所述中间表示向量，判断分诊任务是否成功完成；以及根据所述分诊任务的判断结果，调整分诊对话策略。
计算机可读存储介质可以包括，例如易失性存储器和/或非易失性存储器。所述易失性存储器例如可以包括随机存取存储器(RAM)和/或高速缓冲存储器(cache)等。所述非易失性存储器例如可以包括只读存储器(ROM)、硬盘、闪存等。在所述计算机可读存储介质上可以存储一个或多个计算机程 序指令，处理器可以运行所述程序指令，以实现本公开实施例提供的分诊对话功能。在所述计算机可读存储介质中还可以存储各种应用程序和各种数据，例如所述应用程序使用和/或产生的各种数据等。
对图3所示的所述分诊对话设备101以及所述计算机可读存储介质的说明，可以参考上述图2A中的分诊对话方法200和图2B的分诊对话方法250的描述，本公开在此不再赘述。
如图4所示，本公开实施例还提供了一种分诊对话装置400。所述分诊对话装置400包括交互装置402、解析装置404、判断装置406以及策略调整装置408。
所述分诊对话装置400可以使用软件、硬件或软硬件结合的方式来实现。在一些实施例中，交互装置402、解析装置404、判断装置406以及策略调整装置408包括存储在存储器中的代码和程序；处理器可以执行该代码和程序以实现本公开实施例提供的一些或全部功能。
在一些实施例中，交互装置402、解析装置404、判断装置406以及策略调整装置408可以是专用硬件器件，用来实现本公开实施例提供的一些或全部功能。例如，交互装置402、解析装置404、判断装置406以及策略调整装置408可以是一个电路板或多个电路板的组合。该一个电路板或多个电路板的组合可以包括：(1)一个或多个处理器；(2)与处理器相连接的一个或多个非暂时的计算机可读的存储器；以及(3)处理器可执行的存储在存储器中的固件。
例如，所述交互装置402被配置为接收来自用户410的分诊对话内容。交互装置402可以为图3所示的交互接口306。所述解析装置404被配置为生成所述分诊对话内容的中间表示。所述判断装置406被配置为根据所述中间表示，判断分诊任务是否成功完成。所述策略调整装置408被配置为根据所述分诊任务的判断结果，调整分诊对话策略。
例如，所述中间表示为中间表示向量，所述解析装置404包括解码器，所述解析装置404还被配置为：将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；以及将所述对话特征向量作为所述编码器的输入，通过所述编码器生成所述中间表示向量并输出。所述编码器采用双向门控循环单元网络。
所述判断装置406还被配置为：基于以往的中间表示以及所述中间表示，计算所述分诊任务完成的成功概率。例如，所述判断装置406还被配置为：基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。所述判断装置406使用高斯过程来判断所述分诊任务是否成功完成，所述成功概率为：
Figure PCTCN2018072098-appb-000053
当所述成功概率落入失败区间，则所述判断结果为确定所述分诊任务失败；当所述成功概率落入不确定区间，则所述判断结果为不确定所述分诊任务是否成功完成；或者，当所述成功概率落入成功区间，则所述判断结果为确定所述分诊任务已完成。
所述策略调整装置408还被配置为：当所述判断结果为确定所述分诊任务失败时，继续与患者进行交互；当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述患者提供反馈信息，并根据所述反馈信息确定是否继续与所述患者进行交互；或者，当所述判断结果为确定所述分诊任务已完成时，终止与所述患者的对话。
当所述判断结果为不确定所述分诊任务是否成功完成时，所述策略调整装置408还被配置为：当所述反馈信息表示所述分诊任务已完成，终止与所述患者的对话；或者，当所述反馈信息表示所述分诊任务尚未完成，选择信息增益最大的信息与所述患者进行交互。
所述策略调整装置408还被配置为：计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及选择与所述最大信息增益相关联的症状与所述患者进行交互。
综上所述，本公开实施例提供的一种分诊对话设备、分诊对话方法、分诊对话系统以及分诊对话装置，通过使用主动学习方法判断是否需要用户反馈，减轻用户负担，能够快速准确地完成分诊任务。
在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过 程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
显然，本领域的技术人员可以对本公开进行各种改动和变型而不脱离本公开的精神和范围。这样，倘若本公开的这些修改和变型属于本公开权利要求及其等同技术的范围之内，则本公开也意图包含这些改动和变型在内。
以上所述，仅为公开的具体实施方式，但本公开的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本公开揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应以所述权利要求的保护范围为准。

Claims (18)

  1. 一种分诊对话方法，包括：
    接收分诊对话内容；
    将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；
    输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；
    根据所述中间表示向量，判断分诊任务是否成功完成；以及
    根据所述分诊任务的判断结果，调整分诊对话策略。
  2. 如权利要求1所述的分诊对话方法，其中，
    所述编码器采用双向门控循环单元网络；
    基于所述双向门控循环单元网络的前向隐藏序列和后向隐藏序列，所述中间表示向量h为：
    Figure PCTCN2018072098-appb-100001
    其中，
    Figure PCTCN2018072098-appb-100002
    用以表征所述前向隐藏序列
    Figure PCTCN2018072098-appb-100003
    中的第i个元素
    Figure PCTCN2018072098-appb-100004
    和所述后向隐藏序列
    Figure PCTCN2018072098-appb-100005
    中第i个元素
    Figure PCTCN2018072098-appb-100006
    的连接，以及T表征所述对话特征的数目。
  3. 如权利要求1-2任一项所述的分诊对话方法，其中，根据所述中间表示向量，判断分诊任务是否成功完成，包括：
    基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。
  4. 如权利要求3所述的分诊对话方法，其中，
    所述判断分诊任务是否成功完成被刻画为高斯过程，所述成功概率为：
    Figure PCTCN2018072098-appb-100007
    其中，H表示所述以往的中间表示向量；h表示所述中间表示向量，其为当前的中间表示；y∈{-1,1}，y＝1表示所述分诊任务成功完成；φ(·)为标准正态分布的分布函数；μ *和
    Figure PCTCN2018072098-appb-100008
    分别为f(h)的后验均值和后验方差，f(·) 表示潜在函数，被刻画为高斯过程。
  5. 如权利要求3-4任一项所述的分诊对话方法，其中，
    当所述成功概率落入失败区间，则所述判断结果为确定所述分诊任务失败；
    当所述成功概率落入不确定区间，则所述判断结果为不确定所述分诊任务是否成功完成；或者
    当所述成功概率落入成功区间，则所述判断结果为确定所述分诊任务已完成。
  6. 如权利要求1-5任一项所述的分诊对话方法，其中，根据所述分诊任务的判断结果，调整分诊对话策略，包括：
    当所述判断结果为确定所述分诊任务失败时，继续与用户进行交互；
    当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述用户提供反馈信息，并根据所述反馈信息确定是否继续与所述用户进行交互；或者
    当所述判断结果为确定所述分诊任务已完成时，终止与所述用户的对话。
  7. 如权利要求6所述的分诊对话方法，其中，当所述判断结果为不确定所述分诊任务是否成功完成时，根据所述反馈信息确定是否继续与所述用户进行交互，包括：
    当所述反馈信息表示所述分诊任务已完成，终止与所述用户的对话；或者
    当所述反馈信息表示所述分诊任务尚未完成，选择信息增益最大的信息与所述用户进行交互。
  8. 如权利要求7所述的分诊对话方法，其中，选择信息增益最大的信息与所述用户进行交互，包括：
    计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及
    选择与所述最大信息增益相关联的症状与所述用户进行交互。
  9. 一种分诊对话设备，包括处理器和存储器，所述存储器用于存储指令，其中，所述指令被所述处理器执行时实现以下操作：
    接收分诊对话内容；
    将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；
    输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；
    根据所述中间表示向量，判断分诊任务是否成功完成；以及
    根据所述分诊任务的判断结果，调整分诊对话策略。
  10. 如权利要求9所述的分诊对话设备，其中，
    所述编码器采用双向门控循环单元网络；
    基于所述双向门控循环单元网络的前向隐藏序列和后向隐藏序列，所述中间表示向量h为：
    Figure PCTCN2018072098-appb-100009
    其中，
    Figure PCTCN2018072098-appb-100010
    用以表征所述前向隐藏序列
    Figure PCTCN2018072098-appb-100011
    中的第i个元素
    Figure PCTCN2018072098-appb-100012
    和所述后向隐藏序列
    Figure PCTCN2018072098-appb-100013
    中第i个元素
    Figure PCTCN2018072098-appb-100014
    的连接，T表征所述对话特征的数目。
  11. 如权利要求9-10任一项所述的分诊对话设备，其中，所述根据所述中间表示向量，判断分诊任务是否成功完成的操作，包括：
    基于以往的中间表示向量以及所述中间表示向量，计算所述分诊任务完成的成功概率。
  12. 如权利要求11所述的分诊对话设备，其中，所述判断分诊任务是否成功完成被刻画为高斯过程，所述成功概率为：
    Figure PCTCN2018072098-appb-100015
    其中，H表示所述以往的中间表示向量；h表示所述中间表示向量，其为当前的中间表示；y∈{-1,1}，y＝1表示所述分诊任务成功完成；φ(·)为标准正态分布的分布函数；μ *和
    Figure PCTCN2018072098-appb-100016
    分别为f(h)的后验均值和后验方差，f(·)表示潜在函数，被刻画为高斯过程。
  13. 如权利要求11-12任一项所述的分诊对话设备，其中，
    当所述成功概率落入失败区间，则所述判断结果为确定所述分诊任务失败；
    当所述成功概率落入不确定区间，则所述判断结果为不确定所述分诊任务是否成功完成；或者
    当所述成功概率落入成功区间，则所述判断结果为确定所述分诊任务已完成。
  14. 如权利要求9-13任一项所述的分诊对话设备，其中，所述根据所述分诊任务的判断结果，调整分诊对话策略的操作，包括：
    当所述判断结果为确定所述分诊任务失败时，继续与用户进行交互；
    当所述判断结果为不确定所述分诊任务是否成功完成时，提示所述用户提供反馈信息，并根据所述反馈信息确定是否继续与所述用户进行交互；或者
    当所述判断结果为确定所述分诊任务已完成时，终止与所述用户的对话。
  15. 如权利要求14所述的分诊对话设备，其中，所述当所述判断结果为不确定所述分诊任务是否成功完成时，根据所述反馈信息确定是否继续与所述用户进行交互的操作，包括：
    当所述反馈信息表示所述分诊任务已完成，终止与所述用户的对话；或者
    当所述反馈信息表示所述分诊任务尚未完成，选择信息增益最大的信息与所述用户进行交互。
  16. 如权利要求15所述的分诊对话设备，其中，所述选择信息增益最大的信息与所述用户进行交互的操作，包括：
    计算不同症状的信息增益，并确定与所述最大信息增益相关联的症状；以及
    选择与所述最大信息增益相关联的症状与所述用户进行交互。
  17. 如权利要求9-16任一项所述的分诊对话设备，还包括交互接口，其中，所述交互接口被配置为实现与用户之间的交互。
  18. 一种计算机可读存储介质，其上存储有计算机指令，所述计算机指令被处理器执行时实现以下操作：
    接收分诊对话内容；
    将所述分诊对话内容切分为一系列对话特征，并生成与所述对话特征一一对应的对话特征向量；
    输入所述对话特征向量至编码器，并通过所述编码器生成中间表示向量；
    根据所述中间表示向量，判断分诊任务是否成功完成；以及
    根据所述分诊任务的判断结果，调整分诊对话策略。
PCT/CN2018/072098 2017-06-28 2018-01-10 分诊对话方法、分诊对话设备及系统 WO2019000905A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US16/082,031 US10872697B2 (en) 2017-06-28 2018-01-10 Triage dialogue method, device and system
EP18758796.9A EP3660854B1 (en) 2017-06-28 2018-01-10 Triage dialogue method, device, and system

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710507287.0A CN109147930A (zh) 2017-06-28 2017-06-28 分诊对话方法、分诊对话设备及系统
CN201710507287.0 2017-06-28

Publications (1)

Publication Number Publication Date
WO2019000905A1 true WO2019000905A1 (zh) 2019-01-03

Family

ID=64740971

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/072098 WO2019000905A1 (zh) 2017-06-28 2018-01-10 分诊对话方法、分诊对话设备及系统

Country Status (4)

Country Link
US (1) US10872697B2 (zh)
EP (1) EP3660854B1 (zh)
CN (1) CN109147930A (zh)
WO (1) WO2019000905A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115905502A (zh) * 2022-07-19 2023-04-04 北京中关村科金技术有限公司 一种话术挖掘与推荐方法、装置及存储介质

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10387463B2 (en) * 2017-07-06 2019-08-20 International Business Machines Corporation Dialog agent for conducting task-oriented computer-based communications
CN110838359B (zh) * 2019-10-16 2023-07-07 平安科技（深圳）有限公司 基于对话机器人的分诊方法、装置、存储介质及机器人
CN111326245A (zh) * 2020-02-11 2020-06-23 北京字节跳动网络技术有限公司 数据信息的处理方法、装置、电子设备及计算机存储介质
TWI823067B (zh) * 2020-03-11 2023-11-21 宏達國際電子股份有限公司 控制方法以及醫療系統
CN111951959A (zh) * 2020-08-23 2020-11-17 云知声智能科技股份有限公司 基于强化学习的对话式导诊方法、装置及存储介质
CN111708866B (zh) * 2020-08-24 2020-12-11 北京世纪好未来教育科技有限公司 会话切分方法、装置、电子设备和存储介质
CN112035674B (zh) * 2020-08-28 2023-05-23 康键信息技术（深圳）有限公司 导诊数据获取方法、装置、计算机设备和存储介质
CN112017742A (zh) * 2020-09-08 2020-12-01 平安科技（深圳）有限公司 分诊数据处理方法、装置、计算机设备及存储介质
CN112185567B (zh) * 2020-11-30 2021-06-04 成都信息工程大学 中医临床辅助辨证模型的建立方法及系统
CN113223735A (zh) * 2021-04-30 2021-08-06 平安科技（深圳）有限公司 基于对话表征的分诊方法、装置、设备及存储介质
CN115482926B (zh) * 2022-09-20 2024-04-09 浙江大学 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140006013A1 (en) * 2012-05-24 2014-01-02 International Business Machines Corporation Text mining for large medical text datasets and corresponding medical text classification using informative feature selection
CN104462024A (zh) * 2014-10-29 2015-03-25 百度在线网络技术（北京）有限公司 生成对话动作策略模型的方法和装置
EP3032472A1 (en) * 2014-12-12 2016-06-15 Xerox Corporation Spectral diagnostic engine for customer support call center
CN105786798A (zh) * 2016-02-25 2016-07-20 上海交通大学 一种人机交互中自然语言意图理解方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
AU762361B2 (en) * 1997-03-13 2003-06-26 First Opinion Corporation Disease management system
US9081879B2 (en) * 2004-10-22 2015-07-14 Clinical Decision Support, Llc Matrix interface for medical diagnostic and treatment advice system and method
US20140019162A1 (en) * 2012-07-12 2014-01-16 Keona Health, Inc. Methods, systems, and devices for online triage
GB201506824D0 (en) * 2015-04-22 2015-06-03 Trailreach Ltd TrailReach Multitrial
US10455088B2 (en) * 2015-10-21 2019-10-22 Genesys Telecommunications Laboratories, Inc. Dialogue flow optimization and personalization

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140006013A1 (en) * 2012-05-24 2014-01-02 International Business Machines Corporation Text mining for large medical text datasets and corresponding medical text classification using informative feature selection
CN104462024A (zh) * 2014-10-29 2015-03-25 百度在线网络技术（北京）有限公司 生成对话动作策略模型的方法和装置
EP3032472A1 (en) * 2014-12-12 2016-06-15 Xerox Corporation Spectral diagnostic engine for customer support call center
CN105786798A (zh) * 2016-02-25 2016-07-20 上海交通大学 一种人机交互中自然语言意图理解方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115905502A (zh) * 2022-07-19 2023-04-04 北京中关村科金技术有限公司 一种话术挖掘与推荐方法、装置及存储介质
CN115905502B (zh) * 2022-07-19 2024-01-05 北京中关村科金技术有限公司 一种话术挖掘与推荐方法、装置及存储介质

Also Published As

Publication number Publication date
EP3660854B1 (en) 2023-08-16
EP3660854A4 (en) 2021-04-21
US20190385739A1 (en) 2019-12-19
CN109147930A (zh) 2019-01-04
EP3660854A1 (en) 2020-06-03
US10872697B2 (en) 2020-12-22

Similar Documents

Publication Publication Date Title
WO2019000905A1 (zh) 分诊对话方法、分诊对话设备及系统
CN110516161B (zh) 一种推荐方法及装置
WO2021159775A1 (zh) 音频分离网络的训练方法、音频分离方法、装置及介质
TWI709102B (zh) 電腦輔助醫學方法、非暫態電腦可讀取媒體以及醫學系統
WO2018133761A1 (zh) 一种人机对话的方法和装置
US10204097B2 (en) Efficient dialogue policy learning
RU2708941C1 (ru) Способ и устройство распознавания сегментированных предложений для человеко-машинной интеллектуальной вопросно-ответной системы
TWI745693B (zh) 控制方法以及醫學系統
US20180226067A1 (en) Modifying a language conversation model
CN110018735A (zh) 智能个人助理接口系统
US20220083868A1 (en) Neural network training method and apparatus, and electronic device
Mittal et al. Web-based chatbot for frequently asked queries (FAQ) in hospitals
JP6309539B2 (ja) 音声入力を実現する方法および装置
US11776269B2 (en) Action classification in video clips using attention-based neural networks
US20220092441A1 (en) Training method and apparatus, dialogue processing method and system, and medium
CN112418059B (zh) 一种情绪识别的方法、装置、计算机设备及存储介质
WO2021121020A1 (zh) 问答方法、装置及设备
CN112580359B (zh) 计算机实施的方法、训练系统和计算机程序产品
CN113707299A (zh) 基于问诊会话的辅助诊断方法、装置及计算机设备
US20210192365A1 (en) Computer device, system, readable storage medium and medical data analysis method
WO2021159755A1 (zh) 智能诊疗数据处理方法、设备、装置及存储介质
CN111160049B (zh) 文本翻译方法、装置、机器翻译系统和存储介质
WO2021077834A1 (zh) 一种基于对话系统对用户问句提出反问的方法和装置
CN117575008A (zh) 训练样本生成方法、模型训练方法、知识问答方法及设备
US20230062127A1 (en) Method for collaborative knowledge base development

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: PCT/CN2018/0720

Country of ref document: EP

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2018758796

Country of ref document: EP

Effective date: 20200128