A dialogue state determination method, device, equipment and storage medium

Abstract

Embodiments of the present application provide a dialogue state determination method, comprising: obtaining input data; wherein the input data is dialogue data input by a user in a dialogue process of any field; processing the input data through a trained neural network to obtain output data; wherein the trained neural network is obtained by meta testing network parameters of a preliminarily adjusted neural network based on first data in sample data; the preliminarily adjusted neural network is obtained by meta training a sub neural network of each field based on second data in the sample data; the first data carries labeled information; the second data does not carry labeled information; the sample data comprises dialogue data of at least one field; and determining a dialogue state of the dialogue based on the output data. Embodiments of the present application also provide a dialogue state determination apparatus, device and computer readable storage medium.

Description

Technical Field

[0001] This application relates to the field of information technology, and in particular to a method, apparatus, device, and computer-readable storage medium for determining dialogue state. Background Technology

[0002] Multi-domain generative dialogue systems can facilitate dialogue interaction with any user within any domain. In related technologies, multi-domain generative dialogue systems often employ neural networks for dialogue processing. The neural networks for these systems are trained using semi-supervised training methods. In this semi-supervised training process, both labeled and unlabeled sample data are typically used as input data for the neural network. However, in practical applications, the amount of labeled sample data is relatively small, thus limiting its impact on improving the network parameters. Therefore, in related technologies, the neural networks of multi-domain generative dialogue systems exhibit poor performance in processing user-input dialogue data. Summary of the Invention

[0003] This application provides a method for determining dialogue state, a device for determining dialogue state, an apparatus for determining dialogue state, and a computer-readable storage medium.

[0004] The dialogue state determination method provided in this application can efficiently and accurately identify dialogue data in any domain through a trained neural network, thereby quickly and accurately determining the state of any dialogue in any domain.

[0005] The dialogue state determination method provided in this application embodiment is implemented as follows:

[0006] A method for determining a dialogue state, the method comprising:

[0007] Acquire input data; wherein, the input data is dialogue data entered by the user during a dialogue in any domain;

[0008] The input data is processed by the trained neural network to obtain output data; wherein, the trained neural network is obtained by meta-testing the network parameters of the neural network with preliminary parameter adjustments based on the first data in the sample data; the neural network with preliminary parameter adjustments is obtained by meta-training the neural network based on the second data in the sample data; the first data carries annotation information; the second data does not carry annotation information; the sample data includes dialogue data from at least one domain.

[0009] Based on the output data, the dialogue state of the dialogue is determined.

[0010] In some implementations, the second data includes dialogue data from at least one domain; the neural network with initially adjusted parameters is obtained by meta-training the neural network based on the second data in the sample data, including:

[0011] Based on the second data, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments.

[0012] Based on the network parameters of the sub-neural network of each domain with preliminary parameter adjustment and the second data, the network parameters of the neural network are adjusted to obtain the neural network with preliminary parameter adjustment; wherein, the neural network after the domain parameter adjustment is completed includes the sub-neural network of each domain with preliminary parameter adjustment.

[0013] In some implementations, adjusting the network parameters of the sub-neural network in each domain based on the second data to obtain a preliminarily adjusted sub-neural network in each domain includes:

[0014] Obtain domain training data from the second data; wherein the domain training data is used to train the sub-neural network of each domain;

[0015] Based on the dialogue data of each domain in the training data of the domain, the network parameters of the sub-neural network of each domain are adjusted to obtain the sub-neural network of each domain with the parameters initially adjusted.

[0016] In some implementations, adjusting the network parameters of the sub-neural network for each domain based on dialogue data from the domain training data to obtain the sub-neural network for each domain with preliminarily adjusted parameters includes:

[0017] Based on the sub-neural network of each domain, meta-learning processing is performed on the dialogue data of each domain in the training data of the domain to obtain the training error of each domain.

[0018] Based on the training error of each domain, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with the parameters initially adjusted.

[0019] In some implementations, the step of performing meta-learning processing on the dialogue data of each domain in the domain training data based on the sub-neural network of each domain to obtain the training error of each domain includes:

[0020] The t-th round of dialogue data from the d-th domain of the training data is input into the sub-neural network of the d-th domain to obtain the t-th round of output data; where d is an integer greater than or equal to 0; and t is an integer greater than or equal to 1.

[0021] Based on the output data of the t-th round, the training error of the d-th domain is obtained.

[0022] In some implementations, the network parameters of the neural network are adjusted based on the network parameters of each sub-neural network in each domain, which are initially adjusted according to the parameters, and the second data, to obtain the neural network with initially adjusted parameters, including:

[0023] Obtain domain detection data from the second data; wherein, the domain detection data is used to train the neural network;

[0024] Based on the network parameters of the sub-neural network in each domain, which have been initially adjusted according to the parameters, and the domain detection data, the network parameters of the neural network are adjusted to obtain the neural network with initially adjusted parameters.

[0025] In some implementations, the network parameters of the sub-neural network in each domain, based on the preliminary parameter adjustment, and the domain detection data, are used to adjust the network parameters of the neural network to obtain the neural network with preliminary parameter adjustment, including:

[0026] Based on the network parameters of the sub-neural network of each domain initially adjusted according to the parameters, the dialogue data of each domain in the domain detection data is processed to obtain the detection error of each domain.

[0027] Based on the detection error in each domain, the network training error is obtained;

[0028] Based on the network training error, the network parameters of the neural network are adjusted to obtain the neural network after preliminary parameter adjustment.

[0029] In some implementations, adjusting the network parameters of the neural network based on the network training error to obtain the neural network with initially adjusted parameters includes:

[0030] If the network parameter adjustment results of the neural network converge, the neural network after the initial parameter adjustment is obtained based on the network parameter adjustment results.

[0031] If the network parameter adjustment result of the neural network fails to converge, repeat the following steps until the network parameter adjustment result of the neural network after the neighborhood parameter adjustment is completed converges:

[0032] The network parameters of the neural network are used as the network parameters of the sub-neural network in each domain.

[0033] Based on the second data, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments.

[0034] Based on the network parameters of the sub-neural network in each domain, which have been initially adjusted, and the second data, the neural network with initially adjusted parameters is obtained.

[0035] In some implementations, the first data includes dialogue data from at least one domain; the trained neural network is obtained through the following methods:

[0036] Each dialogue data point in the first data is input into the neural network whose parameters have been initially adjusted to obtain the dialogue error corresponding to each dialogue data point.

[0037] The network adjustment error is obtained by summing up each of the aforementioned dialogue errors.

[0038] Based on the network adjustment error, the network parameters of the neural network that has been initially adjusted are further adjusted to obtain the trained neural network.

[0039] In some implementations, adjusting the network parameters of the neural network that has undergone preliminary parameter adjustment based on the network adjustment error to obtain the trained neural network includes:

[0040] If the network parameter adjustment results of the neural network with the initial parameters adjusted converge, the trained neural network is obtained based on the network parameter adjustment results.

[0041] If the network parameter adjustment result of the initially adjusted neural network fails to converge, repeat the following steps until the network parameter adjustment result of the initially adjusted neural network converges:

[0042] Based on the first data in the sample data, the network parameters of the neural network, after initial parameter adjustments, are subjected to meta-testing.

[0043] This application embodiment also provides a dialogue state determination device, the device comprising: an acquisition module, a processing module, and a determination module, wherein:

[0044] The acquisition module is used to acquire input data; wherein, the input data is dialogue data entered by the user during a dialogue in any domain;

[0045] The processing module is used to process the input data through a trained neural network to obtain output data; wherein, the trained neural network is obtained by performing meta-testing on the network parameters of the neural network with preliminary parameter adjustments based on the first data in the sample data; the neural network with preliminary parameter adjustments is obtained by performing meta-training on the neural network based on the second data in the sample data; the first data carries annotation information; the second data does not carry annotation information; the sample data includes dialogue data from at least one domain.

[0046] The determining module is used to determine the dialogue state of the dialogue based on the output data.

[0047] This application embodiment also provides a dialogue state determination device, which includes a processor, a memory, and a communication bus, wherein the communication bus is used to establish a communication connection between the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the dialogue state determination method as described above.

[0048] This application also provides a computer-readable storage medium that can be executed by a processor to implement the dialogue state determination method as described above.

[0049] The trained neural network used in the dialogue state determination method provided in this application is obtained through meta-training. Thus, even if the amount of sample data carrying labeled information is small, the neural network can still fully acquire feature information from a large amount of unlabeled sample data through meta-training, thereby improving the neural network's processing performance for dialogue data in any domain. Furthermore, after the meta-training of the neural network based on unlabeled sample data is completed, the parameters of the meta-trained neural network can be adjusted based on a small amount of labeled sample data, further improving the trained neural network's processing performance for dialogue data. This enables the trained neural network to have efficient and accurate recognition capabilities for dialogue data in any domain, thus quickly and accurately determining the state of any dialogue. Attached Figure Description

[0050] Figure 1 A schematic diagram of the neural network model structure for a semi-supervised training multi-domain generative dialogue system in related technologies;

[0051] Figure 2 A flowchart illustrating the first dialogue state determination method provided in this application embodiment;

[0052] Figure 3 This is a schematic diagram of the neural network structure provided in an embodiment of this application;

[0053] Figure 4 A schematic diagram of the training process of the neural network provided in the embodiments of this application;

[0054] Figure 5 This is a schematic diagram of the meta-training process of a neural network provided in an embodiment of this application;

[0055] Figure 6 The flowchart illustrating the process of obtaining a trained neural network based on first data, as provided in the embodiments of this application;

[0056] Figure 7 This is a schematic diagram of the structure of the dialogue state determination device provided in the embodiments of this application;

[0057] Figure 8 This is a schematic diagram of the structure of the dialogue state determination device provided in an embodiment of this application. Detailed Implementation

[0058] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0059] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0060] This application relates to the field of information technology, and in particular to a method, apparatus, device for determining dialogue state, and computer-readable storage medium.

[0061] Multi-domain generative dialogue systems can achieve dialogue interaction with users in any domain. They typically use neural networks as the system model to process dialogue data. During the training of the neural network, the error of the neural network in processing the dialogue data is usually calculated based on a loss function. The network parameters are then updated using an error backpropagation algorithm until the neural network converges.

[0062] In practical applications, accurate determination and tracking of dialogue state are crucial for understanding user intent and generating high-quality system responses. Dialogue state can be determined based on the input and output data of a multi-domain generative dialogue system. Dialogue state represents the user's intent at the current moment; therefore, determining dialogue state is a vital function of multi-domain generative dialogue systems. In related technologies, to improve the performance of multi-domain generative dialogue systems, explicit dialogue state tracking is commonly employed.

[0063] Explicit dialogue state tracking relies on a large amount of dialogue data with annotation information. However, adding annotation information to such large amounts of dialogue data requires significant manpower and time. Therefore, although explicit dialogue state tracking can theoretically improve the performance of multi-domain generative dialogue systems, its practical application is diminished due to the low feasibility of annotating large amounts of dialogue state data.

[0064] To improve the performance of neural networks in multi-domain generative dialogue systems, a semi-supervised method for training neural networks has been proposed in related technologies.

[0065] A semi-supervised training method for multi-domain generative dialogue systems uses both labeled and unlabeled sample data as input data, enabling them to participate in the training process of the neural network model simultaneously.

[0066] Figure 1 This is a schematic diagram of the neural network model structure for a semi-supervised training multi-domain generative dialogue system in related technologies.

[0067] exist Figure 1 In a semi-supervised training, the neural network model of a multi-domain generative dialogue system can include three parts: an input text encoder, a dialogue state decoder, and a system response decoder. These three parts work together to generate system responses based on the explicit dialogue state.

[0068] The input text encoder-dialogue state decoder can include two sets of network parameters: prior network parameters and posterior network parameters. These two sets of network parameters correspond to the same network structure, but process different data streams.

[0069] The data stream processed by the prior network parameters is Figure 1 The dashed line shown represents the system response R at time t-1. t-1 And the statement U entered by the user at time t t The input is fed into the input text encoder, which processes it to obtain the hidden vector. Then, using the CopyFlow mechanism, combined with the attention mechanism in the dialogue state decoder, the dialogue state S at time t is decoded. t Then S t The input is fed into the system response decoder to obtain the system response R at time t. t .

[0070] The data stream processed by the posterior network parameters is as follows Figure 1 The solid line in the figure represents the system response R at time t-1.t-1 The statement U entered by the user at time t t and the system response R at time t t The input is fed into the input text encoder, and then processed sequentially by the text encoder and the dialogue state decoder to obtain S. t .

[0071] In semi-supervised training of neural networks, both the prior and posterior network parameters are updated simultaneously. Furthermore, two loss functions are defined to calculate the error loss during unsupervised training on unlabeled sample data and supervised training on labeled sample data: the unsupervised training loss function and the supervised training loss function. The supervised training loss function calculates the loss terms including the dialogue state decoder loss of the prior network, the dialogue state decoder loss of the posterior network, and the system response decoder loss. The unsupervised training loss function calculates the loss terms including the system response decoder loss and the KL (Kullback-Leibler) divergence of the probability distributions generated by the prior and posterior networks. The KL divergence is an asymmetric measure of the difference between two probability distributions. Here, the KL divergence approximates the prior probability distribution by using the probability generated by the posterior network as the standard distribution.

[0072] The sum of the errors calculated using the two error functions above can be used as the error of the neural network. During the training process, the network parameters can be updated based on this error. After the neural network training is complete, it can be tested. During the testing process, the final system response can be obtained by processing the input dialogue data using the prior network parameters of the neural network and the system response decoder.

[0073] In the above training of the neural network, although different loss functions were defined for the two types of sample data, the training process involved simultaneously inputting unlabeled sample data and labeled sample data into the neural network. Under the condition that the amount of unlabeled sample data was much larger than that of labeled sample data, the neural network could only obtain very little effective information from the labeled sample data through the above training process. Therefore, the multi-domain generative dialogue system trained in a semi-supervised manner had poor performance in actual dialogue processing.

[0074] To address the above issues, this application provides a method for determining dialogue state. Figure 2 This is a flowchart illustrating the first dialogue state determination method provided in an embodiment of this application.

[0075] It should be noted that the dialogue state determination method provided in this application embodiment can be implemented by a processor in the dialogue state determination device. The processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor.

[0076] like Figure 2 As shown, the dialogue state determination method provided in this application embodiment may include the following steps:

[0077] Step 201: Obtain input data.

[0078] The input data refers to the dialogue data entered by the user during the dialogue process in any domain.

[0079] In one implementation, the dialogue process can be active at the current moment.

[0080] In one implementation, the dialogue process can be a historical dialogue process, that is, a dialogue process that has ended or is in an inactive state.

[0081] In one implementation, the input data can be input by any user for the current conversation.

[0082] In one implementation, the input data can be entered by any user during the history of conversations.

[0083] In one implementation, any field may include any of the following: education, finance, shopping, current affairs, health and wellness, science and technology Q&A, city navigation, travel guidance, weather, itinerary planning, etc.

[0084] In one implementation, acquiring input data can be an operation performed in real time.

[0085] In one implementation, acquiring input data can be an operation performed after receiving an instruction to acquire input data. For example, the input data can be historical dialogue data acquired from the historical dialogue process, or historical dialogue data acquired from the currently active dialogue process, after receiving the instruction to acquire input data.

[0086] Step 202: Process the input data using the trained neural network to obtain the output data.

[0087] The trained neural network is obtained by meta-testing the network parameters of the neural network with preliminary parameter adjustments based on the first data in the sample data; the neural network with preliminary parameter adjustments is obtained by meta-training the sub-neural network of each domain based on the second data in the sample data; the first data carries annotation information; the second data does not carry annotation information; the sample data includes dialogue data from at least one domain.

[0088] In one implementation, the amount of dialogue data from at least one domain included in the sample data can be greater than or equal to a sample data volume threshold. This sample data volume threshold can be a large value, such as 100,000.

[0089] In one implementation, the amount of dialogue data in each of the at least one domain included in the sample data may be different. In this case, the minimum amount of dialogue data in each domain can be taken as the training data for the neural network.

[0090] In one implementation, the labeling information may include predefined slots and slot values.

[0091] In one implementation, the first data and the second data may include all dialogue data contained in a complete dialogue process, such as user input data and system response data.

[0092] In one implementation, the first data and the second data may include a portion of the dialogue data from all the dialogue data corresponding to a complete dialogue process.

[0093] In one implementation, the ratio of the amount of the first data to the amount of the second data in the sample data can be less than or equal to a first preset ratio. For example, the first preset ratio can be 2%. Thus, in the training process of the neural network in this embodiment, the amount of sample data carrying labeled information is small, while the amount of sample data without labeled information is large.

[0094] In this embodiment, meta-training of the neural network employs Model-Agnostic Meta-Learning (MAML) technology, which is independent of the neural network model. This technology is applicable to all models that satisfy gradient descent and aims to learn the intrinsic representations and features of different tasks and data. Therefore, using meta-training technology can enable the trained neural network to adapt to new learning tasks more quickly.

[0095] Accordingly, in the embodiments of this application, the meta-test also uses MAML technology to adjust the network parameters of the neural network after the initial parameter adjustment.

[0096] In the embodiments of this application, during the training process of the neural network, by adopting a semi-supervised method based on MAML, even when the amount of data in the second data is much larger than the amount of data in the first data, it is still possible to obtain a large amount of effective information carried in the second data, thereby improving the performance of the system.

[0097] Figure 3 This is a schematic diagram of the neural network structure provided in an embodiment of this application. (In conjunction with...) Figure 1 As can be seen, the neural network structure trained in this application embodiment, compared with the neural network structure in related technologies, adds an auto-encoder in the posterior network parameter data flow stage. Figure 3 The autoencoder in the code is used to decode the input data for reconstructing the posterior network. Specifically, it decodes the predicted dialogue state S′. t The system response R′ at time t-1 is obtained by inputting the data into the autoencoder and outputting the posterior network. t-1 The statement U′ entered by the user at time t t and the system response R′ at time t t .

[0098] It should be noted that, Figure 3 The schematic diagram of the neural network shown can represent the structure of the neural network described in the embodiments of this application, and can also represent the structure of the sub-neural network in each domain. That is, in the embodiments of this application, the structure of the neural network and the structure of the sub-neural network in each domain may be the same, but the network parameters of the neural network and the network parameters of the sub-neural network in each domain may be different.

[0099] exist Figure 3 In this context, the prior network can include a text encoder, a dialogue state decoder, and a system response decoder; the posterior network can include an input text encoder, a dialogue state decoder, and an autoencoder.

[0100] In one implementation, the neural network with initially adjusted parameters can be obtained through the following method:

[0101] Based on the second data and the meta-learning algorithm, the network parameters of the sub-neural network in each domain are adjusted. Then, based on the processing effect of the network parameters of the sub-neural network in each domain on the specified sample data, the neural network with initially adjusted parameters is obtained.

[0102] In this embodiment of the application, the neural network with initially adjusted parameters can be obtained through the following methods:

[0103] Based on the second data, Figure 3 The parameters of each module corresponding to the unsupervised training data flow of the sub-neural network in each domain are adjusted. That is, the data flow corresponding to the prior network parameters in the meta-training includes the network structure of input text encoder - dialogue state decoder - system response decoder; while the data flow corresponding to the posterior network parameters can include input text encoder - dialogue state decoder - autoencoder. The autoencoder is used to reconstruct the input of the input text encoder.

[0104] In one implementation, the network parameters of the neural network, which have been initially adjusted based on the first data, are subjected to meta-testing to obtain the trained neural network. This can be achieved in the following way:

[0105] Based on the initial data and the meta-learning algorithm, the network parameters of the neural network with initially adjusted parameters are tested. Based on the test results, an adjustment strategy for the neural network parameters is determined. Then, the network parameters are adjusted according to this strategy, ultimately resulting in a trained neural network. For example, during the adjustment of the network parameters of the initially adjusted neural network, the annotation information carried by the initial data can be fully utilized to correct any deviations in the initially adjusted network parameters, thereby further improving the performance of the trained neural network.

[0106] In this embodiment, the training of each sub-neural network in each domain, and the training of the neural network itself, can include two stages: unsupervised training and supervised training. Unsupervised training can include the process of performing meta-training on the neural network based on second data to obtain a neural network with initially adjusted parameters; supervised training can include the process of adjusting the network parameters of the initially adjusted neural network based on first data to obtain a fully trained neural network. In other words, in this embodiment, the training of the neural network is achieved through a semi-supervised training method.

[0107] In the embodiments of this application, based on the different sample data used in the above two stages and the different roles played by the sample data in each training stage, the unsupervised training stage of neural network training can also be called the meta-training stage, while the supervised training stage can be called the meta-testing stage.

[0108] The neural network training process provided in this application uses unlabeled sample data and labeled sample data as input data for two different training stages, respectively, thereby reducing mutual interference between the two types of sample data during training. This training method means that when a small amount of newly added labeled sample data is added, after the neural network has completed training based on the current sample data, it is not necessary to execute all the training operations of the two different training stages; instead, a supervised training process can be performed only based on the newly added sample data. Thus, compared with related technologies where adding new sample data requires retraining the neural network, the neural network training method in this application reduces the consumption of time and computing resources, thereby accelerating the optimization speed of the neural network and reducing computational overhead.

[0109] Furthermore, since the neural network training method provided in this application uses MAML technology, when the dialogue is transferred to a new domain, good performance can be obtained by training the neural network with only a small amount of new domain sample data. In contrast, in related technologies, when the new domain dataset contains only a small amount of sample data with labeled information, the positive impact of this small amount of labeled new domain sample data on the trained neural network is small, resulting in poor performance of the trained neural network in the new domain.

[0110] In one implementation, the dialogue data in each domain includes sample data carrying annotation information and sample data without annotation information.

[0111] In one implementation, the output data can be the response data obtained by processing the input data in a trained neural network.

[0112] For example, the response data may be at least one of the following: text data, voice data, image data, etc., based on the user's input dialogue data.

[0113] In one implementation, the response data can be option data obtained from dialogue data input by the user. For example, for relatively ambiguous dialogue data input by the user, the trained neural network can provide several options for the user to choose from, so that the trained neural network can further determine the user's intention.

[0114] Step 203: Determine the dialogue state based on the output data.

[0115] In one implementation, the dialogue state can be determined by a trained neural network.

[0116] In one implementation, the dialogue state can be determined by analyzing the output data.

[0117] In one implementation, the user's intent regarding the current conversation can also be obtained through analysis of the output data.

[0118] In one implementation, by analyzing the output data, information about the user who initiated the local conversation can also be obtained, such as the user's identity information, age information, occupation information, personal hobby information, and health status information.

[0119] As described above, the dialogue state determination method provided in this application, after acquiring user input data during a dialogue in any domain, processes the user input data using a trained neural network to obtain output data, and determines the dialogue state based on the output data. Furthermore, the trained neural network is obtained by performing meta-training on second data (without labeled information) to obtain a neural network with initially adjusted parameters, and then performing meta-testing on the neural network with initially adjusted parameters based on first data (with labeled information). Therefore, the neural network obtained in this way can more deeply acquire the feature information carried in the sample data, thereby improving the processing performance of the trained neural network on input data, and thus enabling more accurate and efficient determination of the dialogue state.

[0120] Based on the foregoing embodiments, this application also provides the training process of the trained neural network in the dialogue state determination method.

[0121] In the embodiments of this application, the sub-neural networks of each domain can be independent of each other.

[0122] In one implementation, the sub-neural networks of each domain can be interconnected. For example, sub-neural networks of related domains can be interconnected. For instance, sub-neural networks of the health and wellness domain and sub-neural networks of the health consultation domain can be interconnected.

[0123] In this application embodiment, the second data includes dialogue data from at least one domain.

[0124] In one implementation, the amount of dialogue data in each of the at least one domain included in the second data may be similar to that in the others.

[0125] In one implementation, the amount of dialogue data in each of the at least one domain included in the second data can vary considerably. For example, for popular domains, there are many sources of sample data, and correspondingly, there are also many sample data in that domain that do not carry annotation information; while for some less common domains, there are fewer sources of sample data, and correspondingly, there are also fewer sample data in that domain that do not carry annotation information.

[0126] In the embodiments of this application, meta-training includes two steps: adjusting the network parameters of the sub-neural network for each domain, and adjusting the network parameters of the neural network.

[0127] In one implementation, meta-training may include adjusting the network parameters of sub-neural networks for each domain to obtain the network parameter adjustment results of sub-neural networks for each domain, and then adjusting the network parameters of the neural network based on sample data and the network parameter adjustment results of sub-neural networks for each domain, thereby obtaining a neural network with initially adjusted parameters.

[0128] Figure 4 This is a schematic diagram of the training process of a neural network provided in an embodiment of this application.

[0129] In this embodiment, the neural network with initially adjusted parameters is obtained through meta-training of the neural network based on the second data in the sample data. Figure 4 Steps 401 to 402 are implemented as follows:

[0130] Step 401: Based on the second data, adjust the network parameters of the sub-neural network in each domain to obtain the sub-neural network in each domain with preliminary parameter adjustments.

[0131] In one implementation, the sub-neural network of each domain, whose parameters are initially adjusted, can be implemented in the following way:

[0132] Based on all sample data in the second dataset, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments.

[0133] In one implementation, the sub-neural network of each domain, whose parameters are initially adjusted, can be implemented in the following way:

[0134] Based on a portion of the sample data in the second dataset, the network parameters of the sub-neural network for each domain are adjusted to obtain the sub-neural network for each domain with initially adjusted parameters.

[0135] For example, step 401 can be achieved through steps A1-A2:

[0136] Step A1: Obtain domain training data from the second dataset.

[0137] Among them, the domain training data is used to train the sub-neural networks for each domain.

[0138] In this embodiment of the application, the domain training data of the d-th domain can be denoted as: Where d is an integer greater than or equal to 0, used to represent the index of different domains.

[0139] In one implementation, the domain training data may be obtained from the second data based on a specified proportion.

[0140] In one implementation, the specified ratio can be higher than or equal to a second preset ratio. For example, the second preset ratio can be 80% or 90%.

[0141] In one implementation, the domain training data may include sample data for each domain.

[0142] In one implementation, the amount of sample data for each domain in the domain training data can be comparable.

[0143] Step A2: Based on the dialogue data of each domain in the domain training data, adjust the network parameters of the sub-neural network of each domain to obtain the sub-neural network of each domain with preliminary parameter adjustments.

[0144] In one implementation, the sub-neural network of each domain, with its parameters initially adjusted, can be obtained by:

[0145] The dialogue data within a specified time period is obtained from the dialogue data of each domain in the domain training data and recorded as the specified time period training data. Based on the specified time period training data, the network parameters of the sub-neural network of each domain are adjusted to obtain the sub-neural network of each domain with preliminary parameter adjustments.

[0146] In one implementation, the sub-neural network of each domain, with its parameters initially adjusted, can be obtained by:

[0147] The dialogue data of a specific group in each domain of the domain training data is obtained and recorded as the specific group training data. Based on the specific group training data, the network parameters of the sub-neural network of each domain are adjusted to obtain the sub-neural network of each domain with preliminary parameter adjustments.

[0148] For example, step A2 can be achieved through steps B1-B2:

[0149] Step B1: Based on the sub-neural network of each domain, perform meta-learning processing on the dialogue data of each domain in the domain training data to obtain the training error of each domain.

[0150] The meta-learning processing in the embodiments of this application can be performed using the MAML technology described above.

[0151] In the neural network training method provided in this application embodiment, the training error corresponding to the unsupervised training of the sub-neural network of each domain based on the second data may include the decoding error of the system response decoder, the decoding error of the sub-encoder, and the KL divergence of the probability distribution generated by the prior and posterior networks.

[0152] In one implementation, the training error for each domain can be obtained by processing each sample data in the domain training data based on the sub-neural network of that domain, obtaining the processing error of each sample data, and then taking a weighted average of these errors.

[0153] In one implementation, before training the sub-neural network for each domain, the initial parameters of the neural network can be set to M, and the initial parameters M can be copied to the network parameters of the sub-neural network for each domain. For example, the network parameters of the sub-neural network for the d-th domain can be denoted as M′. d .

[0154] For example, step B1 can be achieved through steps C1-C3:

[0155] Step C1: Input the t-th round of dialogue data from the d-th domain of the domain training data into the sub-neural network of the d-th domain to obtain the t-th round of output data.

[0156] Where d is an integer greater than or equal to 0; t is an integer greater than or equal to 1.

[0157] In one implementation, the d-th domain can be any of the multiple domains.

[0158] In one implementation, the t-th round of dialogue data can be any sample data from the d-th domain dialogue data belonging to the training domain data.

[0159] In this embodiment, the t-th round of dialogue data may include user input data for the t-th round of dialogue, wherein the user's input data may include at least one of text data, image data, audio data, and emoticon data. For example, the user's input data in the t-th round of dialogue can be denoted as U. t .

[0160] In this embodiment, the output data of round t can represent the response data output by the sub-neural network in the d-th domain, and the output data of round t can be denoted as R. t .

[0161] For example, the output data of any round of the sub-neural network in the d-th domain can be obtained in the manner described above.

[0162] For example, the output data of the (t-1)th round can be obtained by inputting the dialogue data of the (t-1)th round into the sub-neural network of the d-th domain, and then processing the dialogue data of the (t-1)th round by the sub-neural network of the d-th domain.

[0163] Step C3: Based on the output data of round t, obtain the training error of the domain d.

[0164] For example, after the above processing, as the value of t changes, the output data of the sub-neural network in the d-th domain for each round of dialogue data can be obtained.

[0165] In one implementation, the training error of the d-th domain can be obtained by processing the dialogue state data of the t-th round of dialogue data, the output data of the t-th round of dialogue data, and the dialogue state data of the t-th round of dialogue data.

[0166] In one implementation, the training error in the d-th domain can be calculated by taking into account the error generated in each processing step of the t-th round of dialogue data.

[0167] In one implementation, the training error in the d-th domain can be obtained by the probability distribution of the output data of the system response decoder corresponding to the t-th round of dialogue data, the probability distribution of the final output data of the posterior network, and the KL divergence of the output probability distribution of the antecedent and posterior networks.

[0168] In one implementation, the training error of the d-th domain can be obtained by the error function defined by equation (1):

[0169]

[0170] In equation (1), M′ represents the network parameters of the sub-neural network based on the d-th domain. d Using domain training data from the d-th domain The training error in the d-th domain is calculated.

[0171] In equation (1), P(R) t |R t-1 U t ,S t ) is used to indicate that when R is known t-1 U t and St Under the condition of R t The probability distribution of P is the probability distribution of the output data of the system response decoder in the vocabulary space.

[0172] In equation (1), Q Φ For dialogue state S t The posterior probability distribution of . Where, This indicates the dialogue state output by the posterior network.

[0173] In equation (1), p i Let q be the probability distribution of the i-th word output by the prior network dialogue state decoder in the vocabulary space; i Let be the probability distribution of the i-th word output by the posterior network dialogue decoder in the vocabulary space.

[0174] In equation (1), KL(q) i ||p i ) is based on q i p as a standard distribution i With q i The KL divergence.

[0175] In equation (1), N is the length of the dialogue state data. When the dialogue state data is a sequence, N can be the length of the dialogue state sequence; U is the domain training data set; and α is the weight of the KL divergence.

[0176] Step B2: Based on the training error of each domain, adjust the network parameters of the sub-neural network of each domain to obtain the sub-neural network of each domain with preliminary parameter adjustments.

[0177] In one implementation, adjusting the network parameters of the sub-neural network for each domain can be achieved in the following way:

[0178] A first error threshold is obtained, and the network parameters of the sub-neural network in each domain are adjusted based on the relationship between the training error in each domain and the first error threshold. Here, the first error threshold represents the expected training error of the sub-neural network in each domain.

[0179] For example, when the training error of each domain is greater than or equal to a first error threshold, the network parameters of the sub-neural network of each domain are adjusted based on the training error of each domain; correspondingly, when the training error of each domain is less than the first error threshold, the above adjustment operation can be stopped.

[0180] In one implementation, adjusting the network parameters of the sub-neural network for each domain can be achieved in the following way:

[0181] Based on the training error of each domain and the network parameter state of the sub-neural network in each domain, the network parameters of the sub-neural network in each domain are adjusted.

[0182] For example, the network parameter state of the sub-neural network in each domain can be evaluated by whether the network parameters of the sub-neural network in each domain have converged.

[0183] In one implementation, adjusting the network parameters of the sub-neural network for each domain can be achieved in the following way:

[0184] Based on whether the network parameters of the sub-neural network in each domain converge, and combined with the training error in each domain, the network parameters of the sub-neural network in each domain are adjusted through a certain network parameter adjustment algorithm.

[0185] For example, the network parameters M′ of the sub-neural network in the d-th domain d The adjustment can be based on the training error in the d-th domain, according to the gradient descent strategy. The specific implementation process is shown in equation (2):

[0186]

[0187] In equation (2), δ is the learning rate.

[0188] Through the above process, the neural network training method of this application embodiment can adjust the network parameters of each domain sub-neural network based on the domain training data in the second data, thereby obtaining the sub-neural network of each domain with initially adjusted parameters.

[0189] Step 402: Based on the network parameters of the sub-neural networks in each domain that have undergone preliminary parameter adjustment, and the second data, adjust the network parameters of the neural network to obtain a neural network with preliminary parameter adjustment.

[0190] In one implementation, the neural network with initially adjusted parameters can be obtained through the following method:

[0191] Take any sample data from each domain, excluding the domain training data, as input data and input it into the sub-neural network of each domain whose parameters have been initially adjusted to obtain the corresponding output data. Then, determine the adjustment method of the neural network parameters based on the output data and adjust the neural network parameters according to the determined adjustment method.

[0192] In one implementation, the neural network with initially adjusted parameters can be obtained through the following method:

[0193] The sample data corresponding to several domains in the second data, excluding the domain training data, are used as input data and input into the corresponding domain sub-neural network whose parameters have been initially adjusted to obtain the corresponding output data. Based on the above input data and output data, the adjustment method of the network parameters of the neural network is determined, and then the network parameters of the neural network are adjusted according to the determined adjustment method.

[0194] For example, step 402 can be implemented through steps D1-D2:

[0195] Step D1: Obtain the domain detection data from the second data.

[0196] Among them, the domain detection data is used to train the neural network.

[0197] In one implementation, the domain detection data can be obtained from second data based on a specified ratio. In this embodiment, the domain detection data can be denoted as...

[0198] In one implementation, the amount of domain detection data can be much smaller than the domain training data. Thus, during the training of each domain's sub-neural network, each sub-neural network can acquire sufficient dialogue data feature information from a large amount of domain training data, while the detection results are obtained based only on a small amount of domain detection data and after the domain parameters of each sub-neural network have been adjusted.

[0199] In one implementation, the domain detection data may include sample data for each domain.

[0200] In one implementation, the domain detection data may include sample data from several domains.

[0201] In one implementation, the sample data in each domain of the domain detection data can be of similar size or can differ significantly from each other.

[0202] In one implementation, when there is new sample data, this new sample data can also be divided into two parts: domain detection data and domain training data.

[0203] Step D2: Based on the network parameters of the sub-neural network in each domain and the domain detection data, adjust the network parameters of the neural network to obtain the neural network with preliminary parameter adjustment.

[0204] In one implementation, the neural network with initially adjusted parameters can be obtained through the following method:

[0205] The dialogue data within a specified time period is obtained from the dialogue data of each domain in the domain detection data and recorded as the specified time period detection data. Based on the processing results of the specified time period detection data by the sub-neural network of each domain with initially adjusted parameters, the network parameters of the neural network are adjusted to obtain the neural network with initially adjusted parameters.

[0206] In one implementation, the neural network with initially adjusted parameters can be obtained through the following method:

[0207] The dialogue data of a specific group in each domain of the domain detection data is obtained and denoted as the specific group detection data. Based on the processing results of the specific group detection data by the sub-neural network of each domain with initially adjusted parameters, the network parameters of the neural network are adjusted to obtain the neural network with initially adjusted parameters.

[0208] For example, step D2 can be achieved through steps E1-E3:

[0209] Step E1: Based on the preliminary adjustment of the network parameters of the sub-neural network for each domain, process the dialogue data of each domain in the domain detection data to obtain the detection error of each domain.

[0210] For example, the domain detection data of each domain can be input into the sub-neural network of each domain, which has undergone preliminary parameter adjustment after the domain parameter adjustment of the neural network has been completed, so as to obtain the detection error of each domain.

[0211] For example, the detection error for each domain can be achieved using equation (3):

[0212]

[0213] In equation (3), M′ represents the network parameters of the sub-neural network based on the d-th domain. d Using domain training data from the d-th domain The training error in the d-th domain is calculated.

[0214] In equation (3), P(R) t |R t-1 U t ,S t ) is used to indicate that when R is known t-1 U t and S t Under the condition of R t The probability distribution of P is the probability distribution of the output data of the system response decoder in the vocabulary space.

[0215] In equation (3), and Q ΦFor dialogue state S t The posterior probability distribution of . Where, This indicates the dialogue state output by the posterior network.

[0216] In equation (3), p i Let q be the probability distribution of the i-th word output by the prior network dialogue state decoder in the vocabulary space; i Let be the probability distribution of the i-th word output by the posterior network dialogue decoder in the vocabulary space.

[0217] In equation (3), KL(q) i ||p i ) is based on q i p as a standard distribution i With q i The KL divergence.

[0218] In equation (3), N is the length of the dialogue state sequence; U is the domain training data set; and α is the weight of the KL divergence.

[0219] It should be noted that, in the embodiments of this application, since both the domain training data and the domain detection data are sample data without labeling information, the intermediate variables for processing the sample data and the sample data are represented by the same symbols in equations (1) and (3).

[0220] Step E2: Based on the detection error in each domain, obtain the network training error.

[0221] In one implementation, the network training error can be calculated by weighting the detection error of each domain based on the proportion of sample data in the total sample data.

[0222] In one implementation, the network training error can be calculated by superimposing the detection errors for each domain.

[0223] In the embodiments of this application, network training error can be used to represent the overall error of the entire neural network model trained based on the second data.

[0224] Step E3: Based on the network training error, adjust the network parameters of the neural network to obtain the neural network after preliminary parameter adjustment.

[0225] For example, before adjusting the network parameters of a neural network, the initial network parameters of the neural network can be set first, such as M.

[0226] In one implementation, the network parameters of the neural network can be adjusted in the following way:

[0227] A second error threshold is obtained; based on the relationship between the network training error and the second error threshold, the network parameters of the neural network are adjusted. The second error threshold represents the expected error of adjusting the network parameters of the neural network based on the network detection data.

[0228] For example, when the network training error is greater than or equal to the second error threshold, the network parameters of the neural network are adjusted based on the network training error; correspondingly, when the network training error is less than the second error threshold, the above adjustment operation can be stopped.

[0229] In one implementation, the network parameters of the neural network can be adjusted in the following way:

[0230] Based on the network training error and the state of the neural network parameters, adjust the network parameters of the neural network.

[0231] For example, the network parameter state of a neural network can be evaluated by whether the network parameter state of the neural network converges.

[0232] In one implementation, adjusting the network parameters of the neural network can be achieved in the following way:

[0233] Based on whether the network parameters of the neural network converge, and taking into account the network training error, the network parameters of the neural network are adjusted using a certain network parameter adjustment algorithm.

[0234] For example, step E3 can be implemented in the following way:

[0235] If the network parameter adjustment results converge, the neural network with initially adjusted parameters is obtained based on these results. If the network parameter adjustment results do not converge, the following steps are repeated until the network parameter adjustment results converge:

[0236] The network parameters of the neural network are used as the network parameters of the sub-neural network in each domain. Based on the second data, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments. Based on the network parameters of the sub-neural network in each domain with preliminary parameter adjustments and the second data, the neural network with preliminary parameter adjustments is obtained.

[0237] In this embodiment of the application, after adjusting the network parameters of the sub-neural network of each domain and the domain detection data based on the domain training data in the second data, if the network parameters of the neural network do not converge, the network parameters of the neural network can be copied to the network parameters of the sub-neural network of each domain. On the one hand, the results of the previous meta-training can be retained, and on the other hand, the deficiencies of the current meta-training can be made up for in the next meta-training process based on the results of the current meta-training.

[0238] In this embodiment, the network parameters of the sub-neural network in each domain are adjusted based on the second data to obtain the sub-neural network in each domain with initially adjusted parameters. The neural network with initially adjusted parameters is obtained based on the network parameters of the sub-neural network in each domain with initially adjusted parameters and the second data. This can be achieved through the meta-training process described in the foregoing embodiments, which will not be repeated here.

[0239] In one implementation, when the network parameter adjustment results of the neural network converge, the network parameter adjustment results can be saved, so that a neural network with initially adjusted parameters can be obtained based on the network parameter adjustment results.

[0240] In one implementation, the network parameters M of the neural network can be adjusted based on the network training error and according to the gradient descent strategy, as shown in equation (4):

[0241]

[0242] In equation (4), η is the learning rate; This represents the network training error obtained by summing the detection errors for each domain.

[0243] Through the above steps, the neural network training method of this application embodiment can adjust the network parameters of the neural network based on the domain training data in the second data and the network parameters of each domain sub-neural network with preliminary parameter adjustment, thereby obtaining a neural network with preliminary parameter adjustment.

[0244] The neural network training method provided in this application embodiment may further include steps 403-405, namely, based on the first data carrying labeled information, the network parameters of the neural network are fine-tuned by performing meta-tests on the network parameters of the neural network with initially adjusted parameters, thereby improving the convergence of the network parameters of the neural network with initially adjusted parameters.

[0245] Step 403: Input each dialogue data in the first data into the neural network with initially adjusted parameters to obtain the dialogue error corresponding to each dialogue data.

[0246] In this embodiment of the application, the first data can be denoted as D. test .

[0247] For example, the method of obtaining the dialogue error corresponding to each dialogue data is similar to the process of obtaining the training error of each domain from the domain training data described in the previous embodiments, and will not be repeated here.

[0248] For example, the dialogue error corresponding to each dialogue data can be achieved through equation (5):

[0249]

[0250] In equation (5), L′ represents the dialogue error corresponding to each dialogue data.

[0251] In equation (5), R′ t 、R′ t-1 These are used to represent the neural network after initial parameter adjustment, the output data of the t-th round of dialogue in the first data, and the (t-1)-th round of dialogue, respectively.

[0252] In equation (5), U′ t This is used to represent the input data corresponding to the t-th round of dialogue in the first data.

[0253] In equation (5), S′ t , These are used to represent the dialogue states output by the prior network and the posterior network, respectively.

[0254] In equation (5), S′ t-1 This is used to represent the dialogue state of the prior network in the neural network after initial parameter adjustment, based on the dialogue data of the (t-1)th round.

[0255] In equation (5), p′ i This represents the probability distribution of the i-th word in the dialogue state decoder of the prior network in the neural network after initial parameter adjustment; q′ i This represents the probability distribution of the i-th word in the dialogue state decoder of the neural network and the posterior network after initial parameter adjustments.

[0256] As can be seen from Equation (5), the dialogue error corresponding to each dialogue data retains all the error loss defined in the unsupervised meta-training process, and also adds the decoding error of the posterior network dialogue state decoder.

[0257] In equation (5), A is used to represent the dataset of the second data.

[0258] Step 404: Accumulate the errors of each dialogue to obtain the network adjustment error.

[0259] For example, step 404 can be similar to the process of obtaining network training error in the foregoing embodiments, and will not be described again here.

[0260] Step 405: Based on the network adjustment error, perform meta-adjustment on the network parameters of the neural network that has been initially adjusted to obtain the trained neural network.

[0261] For example, based on the network adjustment error, the network parameters of the neural network that has been initially adjusted are adjusted to obtain the trained neural network. This process is the same as the process of obtaining the neural network after initial parameter adjustment in the previous embodiment, and will not be repeated here.

[0262] For example, the network parameters of the trained neural network can be determined by equation (6):

[0263]

[0264] For example, step 405 can be achieved through the following steps:

[0265] If the initial parameter adjustment of the neural network converges, the neural network is trained based on the parameter adjustment results. If the initial parameter adjustment of the neural network fails to converge, the following steps are repeated until the initial parameter adjustment of the neural network converges:

[0266] Based on the first data in the sample data, the network parameters of the neural network, which have been initially adjusted, are further adjusted.

[0267] In one implementation, if the network parameter adjustment results of the neural network with initially adjusted parameters converge, the network parameter adjustment results can be saved, so that the trained neural network can be obtained based on the network parameter adjustment results.

[0268] Through the above steps, the neural network training method provided in this application embodiment, by using first data carrying labeled information, completes further fine-tuning of the network parameters of the neural network whose parameters have been initially adjusted, thereby further improving the performance of the trained neural network.

[0269] Figure 5 This is a schematic diagram of the meta-training process of a neural network provided in an embodiment of this application. For example... Figure 5 As shown in the diagram above, the training process can include the following steps:

[0270] Step 501: Divide the data without annotation information into third data and fourth data.

[0271] For example, the data without annotation information can be the second data described in the foregoing embodiments.

[0272] For example, the data in the d-th domain of the third data can be from the data in the foregoing embodiments. The data in the d-th domain of the fourth data can be from the data in the aforementioned embodiments.

[0273] Step 502: Copy the parameters M of the neural network to obtain the network parameters of the sub-neural network in each domain.

[0274] For example, the network parameters of the sub-neural network in the d-th domain can be M′ d .

[0275] Step 503: Calculate the first error generated by each sub-neural network when processing the third data.

[0276] For example, the first error can be the domain training error as described in the foregoing embodiments.

[0277] For example, It can be achieved by... The input is fed into the sub-neural network of the d-th domain to obtain the output data, and according to... The results are obtained by calculating the corresponding output data, which can be achieved through equation (1).

[0278] Step 504: Update the network parameters of the sub-neural network based on the first error.

[0279] For example, step 504 can be performed on M′ using the gradient descent algorithm shown in equation (2). d This is achieved through updates.

[0280] The sub-neural network with updated parameters obtained after step 504 can be the sub-neural network of each domain with the parameters initially adjusted as described in the foregoing embodiments.

[0281] Step 505: Calculate the second error generated by the sub-neural network after updating the network parameters for the fourth data processing.

[0282] For example, the second error can be as described in the foregoing embodiments.

[0283] For example, It can be done through equation (3) It is obtained through processing and calculation.

[0284] Step 506: Accumulate the second error across all domains and update the neural network parameters based on the accumulation result.

[0285] For example, step 506 can be implemented by equation (4).

[0286] Step 507: Determine whether M has converged.

[0287] If M converges, proceed to step 508. If M does not converge, proceed to steps 502-506.

[0288] Step 508: Obtain the neural network with preliminary parameter adjustments.

[0289] By performing meta-training on the neural network using only the second data without labeled information, a neural network with initially adjusted parameters can be obtained. Furthermore, due to the use of meta-training, the neural network with initially adjusted parameters can more accurately and efficiently acquire various feature data carried in the second data, thereby improving the performance of the neural network.

[0290] Figure 6 This document describes the flowchart for obtaining a trained neural network based on first data, as provided in the embodiments of this application. Figure 6 As shown, the process may include the following steps:

[0291] Step 601, D test A neural network with initial input parameter adjustments.

[0292] For example, D test It can be the first data as described in the previous embodiments, i.e., sample data carrying annotation information.

[0293] Step 602: Calculate the network adjustment error.

[0294] For example, the network adjustment error can be calculated by superimposing the dialogue errors corresponding to each dialogue data obtained by equation (5).

[0295] Step 603: Based on the network adjustment error, update the network parameters M of the neural network to make preliminary adjustments.

[0296] For example, step 603 can be implemented by equation (6).

[0297] Step 604: Determine whether M has converged.

[0298] If convergence occurs, proceed to step 605; if convergence does not occur, proceed to steps 601-603.

[0299] Step 605: Obtain the trained neural network.

[0300] As can be seen from the above steps, in the process of training the neural network based on the first data to initially adjust the parameters, the overall structure of the neural network remains unchanged. It is only the network parameters of the unsupervised neural network based on the second data that are finely adjusted in a supervised manner. Therefore, it can retain the training results of the unsupervised neural network based on the second data to the greatest extent and improve the overall performance of the trained neural network.

[0301] Compared to neural network training methods in related technologies, the neural network training method provided in this application utilizes unlabeled sample data in a different way. In related technologies, the difference in probability distributions of unlabeled data in the prior and posterior networks is obtained by calculating the KL divergence, and this difference is used to update the model parameters. This training method cannot fully explore the effective internal information of the unlabeled data.

[0302] The neural network training method provided in this application uses meta-learning technology to perform domain-specific meta-training on unlabeled data, fully exploring the inherent features contained in a large number of unlabeled sample data in each domain, thereby maximizing the sensitivity of the loss function. On this basis, combined with a meta-testing process using a small number of labeled sample data, a significant performance improvement can be achieved.

[0303] Table 1 shows the performance improvement data of the trained neural network provided in the embodiments of this application in terms of dialogue state accuracy and user intent recognition accuracy.

[0304] Table 1 shows a performance comparison of the two neural network types at different specified ratios in terms of dialogue state accuracy and user intent recognition rate.

[0305]

[0306] Table 1

[0307] In Table 1, the specified ratio can be the specified ratio as in the aforementioned embodiments, which is used to represent the ratio between the amount of first data carrying annotation information and the amount of second data not carrying annotation information in the sample data.

[0308] In Table 1, the types of neural networks may include the Semi-supervised Explicit Dialogue State Tracker (SEDST) and the Meta-Learning SEDST (Meta-SEDST) provided in this application.

[0309] In Table 1, the Dialogue State Accuracy is used to represent the accuracy of the dialogue state obtained by processing the dialogue data in the sample data by two different types of neural networks.

[0310] In Table 1, the User Intent Recognition Rate represents the accuracy of user intent recognition after processing dialogue data in the sample data by two different types of neural networks.

[0311] In Table 1, with a specified proportion of 2%, SEDST's dialogue state accuracy is 0.546, while Meta-SEDST's is 0.688. Regarding user intent recognition, Meta-SEDST's is 0.680, while SEDST's is 0.579. In other words, with a sample data proportion carrying labeled information of 2%, Meta-SEDST's dialogue state accuracy after processing the sample data is 14.2% higher than SEDST's, and Meta-SEDST's user intent recognition rate is 10.1% higher than SEDST's.

[0312] In Table 1, with a specified proportion of 4%, SEDST's dialogue state accuracy is 0.587, while Meta-SEDST's is 0.683. Regarding user intent recognition rate, Meta-SEDST's is 0.605, and SEDST's is 0.667. In other words, with a sample data proportion carrying labeled information of 4%, Meta-SEDST's dialogue state accuracy after processing the sample data is 9.6% higher than SEDST's, and Meta-SEDST's user intent recognition rate is 6.2% higher than SEDST's.

[0313] In Table 1, with a specified proportion of 6%, SEDST's dialogue state accuracy is 0.607, while Meta-SEDST's is 0.717. Regarding user intent recognition, Meta-SEDST's is 0.605, and SEDST's is 0.703. In other words, with a sample data proportion carrying labeled information of 6%, Meta-SEDST's dialogue state accuracy after processing the sample data is 11% higher than SEDST's, and Meta-SEDST's user intent recognition rate is 9.8% higher than SEDST's.

[0314] In Table 1, with a specified proportion of 8%, SEDST's dialogue state accuracy is 0.699, while Meta-SEDST's is 0.710. Regarding user intent recognition rate, Meta-SEDST's is 0.658, and SEDST's is 0.717. In other words, with 8% of the sample data carrying labeled information, Meta-SEDST's dialogue state accuracy after processing the sample data is 1.1% higher than SEDST's, and Meta-SEDST's user intent recognition rate is 5.9% higher than SEDST's.

[0315] As can be seen from the data in Table 1, regardless of the specified ratio, Meta-SEDST shows significant improvements over SEDST in both dialogue state accuracy and user intent recognition rate. Furthermore, the improvement of Meta-SEDST over SEDST in these two dimensions is particularly noticeable at smaller specified ratios, such as 2%. This demonstrates that Meta-SEDST can completely overcome the problem of insufficient sample data with labeled information in related technologies, which leads to a decline in neural network performance.

[0316]

[0317] Table 2

[0318] Accordingly, Table 2 shows a comparison of Meta-SEDST and SEDST in three different target domains, with the specified proportion in the target domain remaining constant at 5%, using any two of the three domains (weather, scheduling, and navigation) as source domains and the other as the target domain, and using the source domain data as unlabeled data for unsupervised meta-training of the neural network and the target domain data as labeled data for supervised meta-testing of the neural network. BLEU is a commonly used metric in natural language processing for evaluating the quality of generated language.

[0319] In Table 2, under the specified proportion of 5% and with the source domains being navigation and scheduling, and the target domain being weather, Meta-SEDST achieved a BLEU score of 0.203, a dialogue state accuracy of 0.757, and a user intent recognition rate of 0.675 after processing the weather sample data. In contrast, SEDST achieved a BLEU score of 0.180, a dialogue state accuracy of 0.630, and a user intent recognition rate of 0.540 after processing the sample data. In other words, under the specified proportion of 5%, Meta-SEDST showed improvements of 2.3%, 12.7%, and 13.5% in BLEU score, dialogue state accuracy, and user intent recognition rate, respectively, compared to SEDST in the weather domain.

[0320] In Table 2, under the specified proportion of 5% and with the source domains being navigation and weather, and the target domain being scheduling, Meta-SEDST achieved a BLEU score of 0.199, a dialogue state accuracy of 0.711, and a user intent recognition rate of 0.694 after processing the sample data for the scheduling domain. In contrast, SEDST achieved a BLEU score of 0.183, a dialogue state accuracy of 0.651, and a user intent recognition rate of 0.663 after processing the sample data. In other words, under the specified proportion of 5%, Meta-SEDST showed improvements of 1.6%, 6%, and 3.1% respectively in the scheduling domain compared to SEDST in terms of BLEU score, dialogue state accuracy, and user intent recognition rate.

[0321] In Table 2, under the specified proportion of 5% and with the source domain being both weather and navigation, and the target domain being navigation, Meta-SEDST, after processing the sample data in the navigation domain, achieved a BLEU score of 0.168, a dialogue state accuracy of 0.457, and a user intent recognition rate of 0.462; while SEDST, after processing the sample data, achieved a BLEU score of 0.164, a dialogue state accuracy of 0.350, and a user intent recognition rate of 0.299. In other words, under the specified proportion of 5%, in the navigation domain, Meta-SEDST shows improvements of 0.4%, 10.7%, and 16.3% in BLEU score, dialogue state accuracy, and user intent recognition rate, respectively.

[0322] As can be seen from the data in Table 2, under the condition that the specified ratio remains unchanged, Meta-SEDST shows a more significant improvement in BLEU index, dialogue state accuracy and user intent recognition rate when processing sample data in the target domain compared with SEDST.

[0323] In summary, the trained neural network Meta-SEDST obtained through the above steps not only improves the processing performance of dialogue data when sample data without labeled information is very limited, but also demonstrates good dialogue processing performance in new dialogue domains in terms of domain transfer. Therefore, Meta-SEDST can overcome the problems of decreased neural network performance and insufficient domain transfer ability caused by insufficient sample data with labeled information in related technologies.

[0324] Based on the foregoing embodiments, this application provides a dialogue state determination device 7. Figure 7 This is a schematic diagram of the structure of the dialogue state determination device 7 provided in an embodiment of this application. Figure 7 As shown, the dialogue state determination device 7 may include: an acquisition module 701, a processing module 702, and a determination module 703, wherein:

[0325] The acquisition module 701 is used to acquire input data; wherein, the input data is the dialogue data entered by the user during the dialogue process in any domain.

[0326] The processing module 702 is used to process the input data through the trained neural network to obtain the output data; wherein, the trained neural network is obtained by meta-testing the network parameters of the neural network with preliminary parameter adjustment based on the first data in the sample data; the neural network with preliminary parameter adjustment is obtained by meta-training the neural network based on the second data in the sample data; the first data carries annotation information; the second data does not carry annotation information; the sample data includes dialogue data from at least one domain.

[0327] The determination module 703 is used to determine the dialogue state of the dialogue based on the output data.

[0328] In some implementations, the second data includes dialogue data from at least one domain; meta-training includes adjusting the network parameters of sub-neural networks for each domain, and adjusting the network parameters of the neural network.

[0329] The processing module 702 is used to adjust the network parameters of the sub-neural network in each domain based on the second data, so as to obtain the sub-neural network in each domain with preliminary parameter adjustment.

[0330] The processing module 702 is also used to adjust the network parameters of the neural network based on the network parameters of the sub-neural network of each domain, which are initially adjusted in terms of parameters, and the second data, to obtain a neural network with initially adjusted parameters.

[0331] In some implementations, the acquisition module 701 is used to acquire domain training data from the second data; wherein the domain training data is used to train sub-neural networks for each domain.

[0332] The processing module 702 is used to adjust the network parameters of the sub-neural network of each domain based on the dialogue data of each domain in the domain training data, so as to obtain the sub-neural network of each domain with preliminary parameter adjustment.

[0333] In some implementations, the processing module 702 is used to perform meta-learning processing on the dialogue data of each domain in the domain training data based on the sub-neural network of each domain, so as to obtain the training error of each domain.

[0334] The processing module 702 is also used to adjust the network parameters of the sub-neural network of each domain based on the training error of each domain, so as to obtain the sub-neural network of each domain with preliminary parameter adjustment.

[0335] In some implementations, the processing module 702 is used to input the t-th round dialogue data of the d-th domain dialogue data in the domain training data into the sub-neural network of the d-th domain to obtain the t-th round output data; where d is an integer greater than or equal to 0; and t is an integer greater than or equal to 1.

[0336] The processing module 702 is also used to obtain the training error of the d-th domain based on the output data of the t-th round.

[0337] In some implementations, the acquisition module 701 is used to acquire neighborhood detection data from the second data; wherein the neighborhood detection data is used to train the neural network.

[0338] The processing module 702 is used to adjust the network parameters of the neural network based on the network parameters of the sub-neural network of each domain and the domain detection data, so as to obtain the neural network with preliminary parameter adjustment.

[0339] In some implementations, the processing module 702 is used to process the dialogue data of each domain in the domain detection data based on the network parameters of the sub-neural network of each domain, which are initially adjusted based on the parameters, to obtain the detection error of each domain.

[0340] The processing module 702 is also used to obtain the network training error based on the detection error of each domain; and to adjust the network parameters of the neural network based on the network training error to obtain the neural network with preliminary parameter adjustment.

[0341] In some implementations, the determining module 703 is used to, if the network parameter adjustment result of the neural network converges, obtain the neural network after preliminary parameter adjustment based on the network parameter adjustment result; if the network parameter adjustment result of the neural network does not converge, repeat the following steps until the network parameter adjustment result of the neural network after the neighborhood parameter adjustment is completed converges:

[0342] The network parameters of the neural network are used as the network parameters of the sub-neural network in each domain. Based on the second data, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments. Based on the network parameters of the sub-neural network in each domain with preliminary parameter adjustments and the second data, the neural network with preliminary parameter adjustments is obtained.

[0343] In some implementations, the processing module 702 is used to input each dialogue data in the first data into a neural network with preliminarily adjusted parameters to obtain the dialogue error corresponding to each dialogue data.

[0344] The processing module 702 is also used to accumulate each dialogue error to obtain the network adjustment error; based on the network adjustment error, the network parameters of the neural network with initially adjusted parameters are adjusted to obtain the trained neural network.

[0345] In some implementations, the determining module 703 is used to, if the network parameter adjustment result of the initially adjusted neural network converges, save the network parameters of the initially adjusted neural network and the trained neural network; if the network parameter adjustment result of the initially adjusted neural network does not converge, repeat the following steps until the network parameter adjustment result of the initially adjusted neural network converges:

[0346] Based on the first data in the sample data, the network parameters of the neural network, which have been initially adjusted, are further adjusted.

[0347] It should be noted that in practical applications, the acquisition module 701, the processing module 702, and the determination module 703 can be implemented using a processor in an electronic device. The processor can be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor.

[0348] As described above, the dialogue state determination device 7 provided in this application, after acquiring user input data during a dialogue in any domain, processes the user input data through a trained neural network to obtain output data, and determines the dialogue state based on the input and output data. Furthermore, the trained neural network is obtained by performing meta-training on the neural network with second data (without labeled information) to obtain a neural network with initially adjusted parameters, and then performing meta-testing on the neural network with initially adjusted parameters based on first data (with labeled information). Therefore, the neural network obtained in this way can more deeply acquire the feature information carried in the sample data, thereby improving the processing performance of the trained neural network on input data, and thus enabling more accurate and efficient determination of the dialogue state.

[0349] Based on the foregoing embodiments, this application also provides a second dialogue state determination device 8. Figure 8 This is a schematic diagram of the structure of the dialogue state determination device 8 provided in an embodiment of this application. Figure 8 As shown, the dialogue state determination device 8 includes a processor 801, a memory 802, and a communication bus, wherein the communication bus is used to establish a communication connection between the processor 801 and the memory 802; the processor 801 is used to execute a computer program stored in the memory 802 to implement the dialogue state determination method as described in any of the previous embodiments.

[0350] The processor 801 described above can be at least one of an application-specific integrated circuit (ASIC), DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, or microprocessor. It is understood that the electronic device used to implement the functions of the processor can also be other types, and this embodiment of the invention does not impose specific limitations.

[0351] The aforementioned memory 802 may be volatile memory, such as RAM; or non-volatile memory, such as ROM, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or a combination of the above types of memory, and provides instructions and data to the processor.

[0352] Based on the foregoing embodiments, this application also provides a computer-readable storage medium that can be executed by a processor to implement the dialogue state determination method as described in any of the preceding embodiments.

[0353] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0354] The methods disclosed in the various method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict.

[0355] The features disclosed in the various product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0356] The features disclosed in the various method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0357] It should be noted that the aforementioned computer-readable storage media can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices including one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0358] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0359] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0360] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0361] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0362] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0363] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0364] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for determining dialogue state, characterized in that, The method includes: Acquire input data; wherein, the input data is dialogue data entered by the user during a dialogue in any domain; The input data is processed by the trained neural network to obtain output data. The trained neural network is obtained by performing meta-testing on the network parameters of the initially adjusted neural network based on the annotation information carried by the first data in the sample data, and then correcting the network parameters. The initially adjusted neural network is obtained by performing meta-training on sub-neural networks of each domain based on the second data in the sample data. The first data carries annotation information; the second data does not carry annotation information. The sample data includes dialogue data from at least one domain. The neural network structure includes a system response decoder and an autoencoder in the posterior network. The system response decoder generates the output data of the antecedent network in the neural network structure. The autoencoder decodes and reconstructs the input data of the posterior network in the neural network structure. The error of the system response decoder, the decoding error of the autoencoder, and the KL divergence value of the probability distribution generated by the antecedent and posterior networks of the neural network structure are used to construct a training error. This training error is used to adjust the network parameters of the sub-neural networks of each domain, resulting in the initially adjusted sub-neural networks of each domain. The initially adjusted sub-neural networks of each domain are used to obtain the initially adjusted neural network. Based on the output data, the dialogue state of the dialogue is determined.

2. The method according to claim 1, characterized in that, The second data includes dialogue data from at least one domain; the neural network with initially adjusted parameters is obtained by meta-training sub-neural networks for each domain based on the second data in the sample data, including: Based on the second data, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments. Based on the network parameters of the sub-neural network in each domain, which have been initially adjusted, and the second data, the network parameters of the neural network are adjusted to obtain the neural network with initially adjusted parameters.

3. The method according to claim 2, characterized in that, The step of adjusting the network parameters of the sub-neural network in each domain based on the second data to obtain the sub-neural network in each domain with initially adjusted parameters includes: Obtain domain training data from the second data; wherein the domain training data is used to train the sub-neural network of each domain; Based on the dialogue data of each domain in the training data of the domain, the network parameters of the sub-neural network of each domain are adjusted to obtain the sub-neural network of each domain with the parameters initially adjusted.

4. The method according to claim 3, characterized in that, The process of adjusting the network parameters of the sub-neural network for each domain based on dialogue data from the domain training data to obtain the sub-neural network for each domain with initially adjusted parameters includes: Based on the sub-neural network of each domain, meta-learning processing is performed on the dialogue data of each domain in the training data of the domain to obtain the training error of each domain. Based on the training error of each domain, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with the parameters initially adjusted.

5. The method according to claim 4, characterized in that, The sub-neural network based on each domain performs meta-learning processing on the dialogue data of each domain in the domain training data to obtain the training error of each domain, including: The t-th round of dialogue data from the d-th domain of the training data is input into the sub-neural network of the d-th domain to obtain the t-th round of output data; where d is an integer greater than or equal to 0; and t is an integer greater than or equal to 1. Based on the output data of the t-th round, the training error of the d-th domain is obtained.

6. The method according to claim 2, characterized in that, The network parameters of the sub-neural network in each domain, initially adjusted based on the parameters, and the second data, are used to adjust the network parameters of the neural network to obtain the neural network with initially adjusted parameters, including: Obtain domain detection data from the second data; wherein, the domain detection data is used to train the neural network; Based on the network parameters of the sub-neural network in each domain, which have been initially adjusted according to the parameters, and the domain detection data, the network parameters of the neural network are adjusted to obtain the neural network with initially adjusted parameters.

7. The method according to claim 6, characterized in that, The network parameters of each domain's sub-neural network, based on the initial parameter adjustment, and the domain detection data, are used to obtain the initially parameter-adjusted neural network, including: Based on the network parameters of the sub-neural network of each domain initially adjusted according to the parameters, the dialogue data of each domain in the domain detection data is processed to obtain the detection error of each domain. Based on the detection error in each domain, the network training error is obtained; Based on the network training error, the network parameters of the neural network are adjusted to obtain the neural network after preliminary parameter adjustment.

8. The method according to claim 7, characterized in that, The step of adjusting the network parameters of the neural network based on the network training error to obtain the neural network after preliminary parameter adjustment includes: If the network parameter adjustment results of the neural network converge, the neural network after the initial parameter adjustment is obtained based on the network parameter adjustment results. If the network parameter adjustment results of the neural network do not converge, repeat the following steps until the network parameters of the neural network converge: The network parameters of the neural network are used as the network parameters of the sub-neural network in each domain. Based on the second data, the network parameters of the sub-neural network in each domain are adjusted to obtain the sub-neural network in each domain with preliminary parameter adjustments; Based on the network parameters of the sub-neural network in each domain, which have been initially adjusted, and the second data, the neural network with initially adjusted parameters is obtained.

9. The method according to claim 1, characterized in that, The first data includes dialogue data from at least one domain; the trained neural network is obtained through the following methods: Each dialogue data point in the first data is input into the neural network whose parameters have been initially adjusted to obtain the dialogue error corresponding to each dialogue data point. The network adjustment error is obtained by summing up each of the aforementioned dialogue errors. Based on the network adjustment error, the network parameters of the neural network that has been initially adjusted are further adjusted to obtain the trained neural network.

10. The method according to claim 9, characterized in that, The step of adjusting the network parameters of the neural network, which has been initially adjusted based on the network adjustment error, to obtain the trained neural network includes: If the network parameter adjustment results of the neural network with the initial parameters adjusted converge, the trained neural network is obtained based on the network parameter adjustment results. If the network parameter adjustment result of the initially adjusted neural network fails to converge, repeat the following steps until the network parameter adjustment result of the initially adjusted neural network converges: Based on the first data in the sample data, the network parameters of the neural network, after initial parameter adjustments, are subjected to meta-testing.

11. A dialogue state determination device, characterized in that, The device includes: an acquisition module, a processing module, and a determination module, wherein: The acquisition module is used to acquire input data; wherein, the input data is dialogue data entered by the user during a dialogue in any domain; The processing module is used to process the input data using a trained neural network to obtain output data. The trained neural network is obtained by performing meta-testing on the network parameters of the initially adjusted neural network based on the annotation information carried by the first data in the sample data, and then correcting the network parameters. The initially adjusted neural network is obtained by performing meta-training on sub-neural networks for each domain based on the second data in the sample data. The first data carries annotation information; the second data does not carry annotation information; the sample data includes dialogue data from at least one domain; the neural network of the neural network... The structure includes a systematic response decoder and an autoencoder in the posterior network; the systematic response decoder is used to generate the output data of the antecedent network in the neural network structure; the autoencoder is used to decode and reconstruct the input data of the posterior network in the neural network structure; the error of the systematic response decoder, the decoding error of the autoencoder, and the KL divergence value of the probability distribution generated by the antecedent and posterior networks of the neural network structure are used to construct a training error, which is used to adjust the network parameters of the sub-neural network in each domain to obtain the sub-neural network in each domain with initially adjusted parameters; the sub-neural network in each domain with initially adjusted parameters is used to obtain the neural network with initially adjusted parameters. The determining module is used to determine the dialogue state of the dialogue based on the output data.

12. A dialogue state determination device, characterized in that, The dialogue state determination device includes a processor, a memory, and a communication bus, wherein the communication bus is used to establish a communication connection between the processor and the memory; the processor is used to execute a computer program stored in the memory to implement the dialogue state determination method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The readable storage medium can be executed by a processor to implement the dialogue state determination method as described in any one of claims 1-10.

Images