Model updating method and device, equipment and storage medium

Abstract

Embodiments of the present disclosure disclose a model updating method and device, equipment and a storage medium, wherein the method comprises: in response to receiving audio information sent by a user terminal, identifying the audio information by using a platform model of a platform end to obtain a first identification result; sending the first identification result to the user terminal; receiving an updated end-side model uploaded by the user terminal; wherein the updated end-side model is obtained by updating an initial end-side model based on the first identification result and a second identification result, and the second identification result is obtained by identifying the audio information by using the end-side model; and fusing the updated end-side model and the platform model to obtain an updated platform model. The embodiments of the present disclosure can improve the efficiency and automation of model updating.

Description

Technical Field

[0001] This disclosure relates to, but is not limited to, the field of computer technology, and in particular to a model updating method, apparatus, device, and storage medium. Background Technology

[0002] With the rapid development of artificial intelligence technology, technologies such as intelligent voice are being applied more widely in various industries, effectively improving interaction efficiency. In related technologies, intelligent voice training platforms typically use large-scale corpus data and incremental training of the entire corpus to update the model during training. However, this can affect the overall recognition performance of the model, leading to a decrease in the accuracy of the recognition results. Summary of the Invention

[0003] In view of the above, embodiments of this disclosure provide at least one model update method, apparatus, device, and storage medium.

[0004] The technical solution of this disclosure embodiment is implemented as follows:

[0005] On one hand, this disclosure provides a model update method applied to a platform, comprising: responding to receiving audio information sent by a user terminal, recognizing the audio information using a platform model on the platform to obtain a first recognition result; sending the first recognition result to the user terminal; receiving an updated terminal-side model uploaded by the user terminal; the updated terminal-side model is obtained by the user terminal updating an initial terminal-side model based on the first recognition result and a second recognition result, wherein the second recognition result is obtained by the terminal-side model recognizing the audio information; and fusing the updated terminal-side model and the platform model to obtain an updated platform model.

[0006] On the other hand, this disclosure provides a model update apparatus applied to a platform, comprising: a first recognition module, configured to, in response to receiving audio information sent by a user terminal, recognize the audio information using a platform model on the platform to obtain a first recognition result; a sending module, configured to send the first recognition result to the user terminal; a receiving module, configured to receive an updated terminal-side model uploaded by the user terminal; the updated terminal-side model is obtained by the user terminal updating an initial terminal-side model based on the first recognition result and a second recognition result, wherein the second recognition result is obtained by the terminal-side model recognizing the audio information; and a fusion module, configured to fuse the updated terminal-side model and the platform model to obtain an updated platform model.

[0007] In another aspect, embodiments of this disclosure provide a computer device including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the program to implement some or all of the steps in the above-described method.

[0008] In another aspect, embodiments of this disclosure provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements some or all of the steps in the above-described method.

[0009] In another aspect, embodiments of this disclosure provide a computer program including computer-readable code, which, when executed in a computer device, causes a processor in the computer device to perform some or all of the steps in the above-described method.

[0010] In another aspect, embodiments of this disclosure provide a computer program product, the computer program product including a non-transitory computer-readable storage medium storing a computer program, wherein when the computer program is read and executed by a computer, it implements some or all of the steps in the above method.

[0011] In related technologies, a technical solution involves directly retraining the edge-side few-shot model in the user terminal based on newly added audio information to obtain a new edge-side few-shot model. In this embodiment, firstly, the platform responds to the audio information sent by the user terminal by using its platform model to recognize the audio information, obtaining a first recognition result. This allows the platform to perform one recognition of the audio information. Secondly, the platform can send the first recognition result to the user terminal; wherein the first and second recognition results are used by the user terminal to update its edge-side model, and the second recognition result is obtained by the edge-side model recognizing the audio information. Thus, the user terminal can more accurately update its edge-side model using the platform's first recognition result and its own second recognition result. Then, the platform receives the updated edge-side model uploaded by the user terminal, fuses the updated edge-side model and the platform model, and quickly and accurately obtains the updated platform model. This effectively reduces the impact of model updates on model recognition capabilities across multiple terminals, including the user terminal and the platform, and improves the efficiency and automation of model updates.

[0012] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure. Attached Figure Description

[0013] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0014] Figure 1 A schematic diagram illustrating the implementation flow of the first model update method provided in this embodiment of the disclosure;

[0015] Figure 2 A schematic diagram illustrating the implementation flow of the second model update method provided in this embodiment of the disclosure;

[0016] Figure 3 A schematic diagram illustrating the implementation process of the third model update method provided in this embodiment of the disclosure;

[0017] Figure 4 A schematic diagram illustrating the implementation flow of the fourth model update method provided in this embodiment of the disclosure;

[0018] Figure 5 This is a schematic diagram of the structure of a multi-service voice interaction system provided in an embodiment of the present disclosure;

[0019] Figure 6 A schematic diagram illustrating the implementation process of the fifth model update method provided in this embodiment of the disclosure;

[0020] Figure 7 A schematic diagram illustrating the implementation process of a corpus expansion method provided in this embodiment of the disclosure;

[0021] Figure 8 This is a schematic diagram of the composition structure of a model update device provided in an embodiment of the present disclosure;

[0022] Figure 9 This is a schematic diagram of the hardware entity of a computer device provided in an embodiment of this disclosure. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this disclosure clearer, the technical solutions of this disclosure are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on this disclosure. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0024] In the following description, references to "some embodiments" describe a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "first / second / third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. The terminology used herein is for descriptive purposes only and is not intended to limit the scope of this disclosure.

[0026] This disclosure provides a model update method, which can be executed by a processor on a platform. The platform can refer to a device or platform that stores and runs platform models, such as a cloud server. The platform model can be a model with a large amount of data, a complex structure, and is relatively general. The platform can generate a smaller, user-specific model based on the platform model and send the user-specific model to the user terminal, enabling the user terminal to perform operations such as voice interaction with the user. The user terminal can refer to a server, laptop, tablet, desktop computer, mobile device (e.g., mobile phone, portable gaming device), or other device with model update capabilities. Figure 1 This is a schematic diagram illustrating the implementation flow of a model update method provided in an embodiment of this disclosure, as shown below. Figure 1 As shown, the method includes the following steps S101 to S104:

[0027] Step S101: In response to receiving audio information sent by the user terminal, the audio information is identified using the platform model on the platform side to obtain a first identification result.

[0028] Here, audio information can refer to information carrying user speech data, such as user-interacting dialogue with a user terminal. Audio information can be either audio representing the user's speech data or text representing the user's speech data. The user terminal can refer to a device used for voice interaction with the user, such as a smart TV. The smart TV uses its built-in microphone and other audio acquisition components to collect the user's audio information, which is then used to request corresponding television programs. The smart TV can then use a pre-defined communication protocol to send this audio information to the associated platform.

[0029] The platform side can refer to the device used to store and run platform models. Based on the different types of services processed by the user terminal, the platform side can generate different platform models capable of handling different services, such as a platform model for music search, a platform model for video switching, and a platform model for home appliance control. The platform side then generates a small-sample model with a smaller data volume and simpler structure corresponding to the platform model, and synchronizes this small-sample model to the user terminal (i.e., the user terminal's edge model). During voice interaction with the user, the user terminal can update the edge model based on the collected audio information, and the platform side can also update the platform model based on the audio information and the updated edge model, helping to more accurately and promptly identify the user's semantic content. The first recognition result can refer to the result of recognizing the audio information, such as the interaction topic. For example, audio information can be input into a trained platform model, and the first recognition result output by the platform model is "control the air conditioner to turn on," etc.

[0030] Step S102: Send the first identification result to the user terminal.

[0031] Here, the platform can use a preset communication protocol to feed back the first recognition result to the user terminal. First, after the user terminal collects the audio information, it can also use a preset edge-side model to recognize the audio information, obtaining a second recognition result. Then, the user terminal can update the edge-side model based on the first and second recognition results, obtaining an updated edge-side model. For example, the user terminal can determine the prediction probabilities corresponding to the first and second recognition results, and determine the recognition result with the higher prediction probability as the final recognition result. The prediction probability characterizes the likelihood of the recognition result. For example, if the prediction probability corresponding to the first recognition result is 0.8 and the prediction probability corresponding to the second recognition result is 0.9, the user terminal can determine the second recognition result as the final recognition result. Simultaneously, based on the difference between the prediction probabilities of the first and second recognition results, the current loss is determined, and the parameters of the edge-side model are adjusted based on the current loss to obtain the updated edge-side model.

[0032] Step S103: Receive the updated end-side model uploaded by the user terminal.

[0033] Here, after the user terminal obtains the updated parameters of the terminal-side model, it can use a preset communication protocol to send the updated parameters of the terminal-side model to the platform, so that the platform can synchronize the terminal-side model of the user terminal in a timely manner.

[0034] Step S104: The updated end-side model and the platform model are fused to obtain the updated platform model.

[0035] Here, the platform can use the updated parameters of the edge model as the parameters of the corresponding part of the platform model to obtain the updated platform model. For example, by identifying the part of the platform model that has common recognition with the edge model, the parameters of that part can be adjusted based on the updated parameters of the edge model to obtain the updated platform model.

[0036] In related technologies, a technical solution involves directly retraining the edge-side few-shot model in the user terminal based on newly added audio information to obtain a new edge-side few-shot model. In this embodiment, firstly, the platform responds to the audio information sent by the user terminal by using its platform model to recognize the audio information, obtaining a first recognition result. This allows the platform to perform one recognition of the audio information. Secondly, the platform can send the first recognition result to the user terminal; wherein the first and second recognition results are used by the user terminal to update its edge-side model, and the second recognition result is obtained by the edge-side model recognizing the audio information. Thus, the user terminal can more accurately update its edge-side model using the platform's first recognition result and its own second recognition result. Then, the platform receives the updated edge-side model uploaded by the user terminal, fuses the updated edge-side model and the platform model, and quickly and accurately obtains the updated platform model. This effectively reduces the impact of model updates on model recognition capabilities across multiple terminals, including the user terminal and the platform, and improves the efficiency and automation of model updates.

[0037] In some embodiments, the audio information is used for full training of the platform model and for incremental training of the edge model.

[0038] Here, full training can refer to adding new audio information to the historical sample set used to train the platform model, obtaining a combined historical sample set, and then using the combined historical sample set to update the platform model; alternatively, it can refer to the historical sample set used to train the edge model, and then using the combined historical sample set to update the edge model. Incremental training can refer to using newly added audio information to train the platform model or the edge model.

[0039] Since the hardware resources of the platform are superior to those of the user terminal, for the platform, a full training approach can be used to update the platform model based on audio information, which helps improve the accuracy of the platform model. For the user terminal, an incremental training approach can be used to update the terminal model based on audio information, which helps save computing power and improve update efficiency. Hardware resources can include computing resources and / or storage resources, which can be determined by reading the current performance indicators of the hardware. Performance indicators can include graphics processor throughput, video memory size, video memory bus width, video memory bandwidth, and memory storage capacity. The specific methods for full training and incremental training are not limited here.

[0040] In this embodiment of the disclosure, by using full training to update the platform model and incremental training to update the edge model, the accuracy of the platform model and the update efficiency of the edge model can be effectively improved.

[0041] This disclosure provides a model update method, wherein the platform model is a topic recognition model, and the topic recognition model includes at least two sub-linguistic models. For example... Figure 2 As shown, the method includes the following steps S201 to S205:

[0042] Steps S203 to S205 correspond to the aforementioned steps S102 to S104, respectively. When implementing these steps, the specific implementation methods of the aforementioned steps S102 to S104 can be referred to.

[0043] Step S201: Use each of the sub-linguistic models to identify the audio information and obtain the sub-recognition result corresponding to each sub-linguistic model.

[0044] Here, the topic recognition model can refer to a model used to identify the topic of audio information, such as determining whether the current audio information is music, video, sports, or technology, etc. (or domains). Sub-linguistic models for different topics can be pre-set. These sub-linguistic models can be used to identify audio information of different topics, yielding sub-recognition results, which can refer to the results identified by the sub-linguistic models. For example, the first sub-linguistic model is for audio information in the sports domain, and the second sub-linguistic model is for audio information in the technology domain, etc., which helps to accurately identify audio information of different topics. For example, inputting audio information into the first and second sub-linguistic models respectively, the first sub-recognition result indicates a prediction probability of 0.55 for the topic of sports and 0.45 for the topic of technology; the second sub-recognition result indicates a prediction probability of 0.3 for the topic of sports and 0.7 for the topic of technology, etc.

[0045] Step S202: Based on all the sub-identification results, determine the first identification result.

[0046] Here, taking the first sub-identification result as an example, which indicates that the predicted probability of the current audio information's theme is sports (0.55) and the predicted probability of the current audio information's theme is technology (0.45); and the second sub-identification result as an example, which indicates that the predicted probability of the current audio information's theme is sports (0.3) and the predicted probability of the current audio information's theme is technology (0.7), and combining the two sub-identification results, we obtain a total predicted probability of 0.85 for the current audio information's theme being sports and 1.15 for the current audio information's theme being technology. Therefore, we can determine that the first identification result indicates that the theme of the current audio information is technology.

[0047] When the platform generates at least two sub-linguistic models based on a historical sample set, it can generate a few-shot model based on these two sub-linguistic models and send this few-shot model to the user terminal so that the user terminal receives the initial edge model. For example, the platform can perform at least one of the following processes on the at least two sub-linguistic models: pruning, quantization, encoding, knowledge distillation, etc., to obtain the few-shot model.

[0048] During voice interaction with users, the user terminal can record users' request statements and feedback actions in different scenarios. This data is used for scenario-based business recognition and self-learning training of edge and platform models. After the user terminal uploads the collected audio information to the platform, in addition to recognition, the platform also caches the intent and corpus templates identified from the audio information. Each user's corpus template is cached in the corpus training library. During data updates and automated lexicon updates, the platform synchronizes the lexicon to the corpus training library, forming a basic user training set. The basic user training set mainly stores small sample user corpora with high sentence frequencies, which can be effective in real time and hot-deployed to edge models. The corpus training library can store all types of audio information, the basic user training set can store basic types (such as high-frequency audio information), and the small sample user corpora can refer to customized functional corpus sets used to store specific audio information.

[0049] After the user terminal reports audio information to the platform, the platform updates the small sample user corpus. Based on the entity extraction model and triples, the platform can then fill in the vocabulary and perform topic recognition to generate a platform model, which is then synchronized to the user terminal, enabling the user terminal to generate its own edge model. The topic recognition model can be hot-deployed; it is only updated and deployed when the user speaks audio information containing newly appearing sentence structures, low-frequency words, or sentence structures.

[0050] During topic recognition on the platform, a topic recognition model is generated from a small sample training set of users. This model aims to reduce the impact of topic differences on the linguistic model (i.e., the topic recognition model itself), and it can be divided into n sub-linguistic models M1, ..., Mn. Topic extraction for the entire topic recognition model can be achieved through entity classification. Specifically, topic extraction involves dividing the platform model (also known as the large model) into smaller models (i.e., sub-linguistic models) for sub-domains such as music, video, and home furnishings. Subsequent iterative training can replace these smaller models with smaller ones. After acquiring audio information, domain grouping is performed first, such as for home furnishings, before parsing the smaller models.

[0051] For example: First, it can be based on Sample representation matrix of the selected entity A non-linear transformation is performed using an activation function (e.g., the Squash function); then, the transformed sample representations are summed using topic-weighted methods; where s is the topic weighting parameter and b represents topic polarization. To improve the performance of the topic recognition model, the class vector can be represented by averaging the k samples in each entity class. in, N i Let be the total number of entities in category i. Label the fused topic features as... Then we have: Where w1 represents the weight value (i.e., prediction probability) of the topic feature score in a certain scene; w2 represents the weight value of the topic feature score in a certain scene. Through the above training method, topic labels can be assigned to each sub-language model of the user's small sample corpus. The small sample corpus can be used to update the sub-linguistic models. The large model can be divided into smaller models for sub-domains such as music, video, and home furnishing. Subsequent iterations can train only the smaller models for replacement. After obtaining the audio information, domain grouping can be performed first, such as home furnishing, and then the specific intent and slots of the smaller models can be analyzed.

[0052] The merged topic vector is then compressed using an activation function (such as the Squash function) to non-linearly map it to the interval [0,1], resulting in a new class vector c. i :in, This represents the fused topic feature labels, and 'e' represents a preset constant. The topic vectors corresponding to the sub-language models are obtained through multiple iterations. After the topic vectors are generated, the small sample model on the platform is generated and quickly synchronized to the user terminal in real time for hot loading, generating the initial client-side model, etc. For example, deployment can be completed by automatically replacing the corresponding software (JAR) package.

[0053] In this embodiment of the disclosure, audio information is identified through at least two sub-linguistic models to obtain sub-identification results, and then all sub-identification results are integrated to obtain a more accurate final first identification result.

[0054] In some embodiments, the first recognition result is the current topic of the audio information, and after the above step S202 is implemented, the following steps S211 to S212 are further included:

[0055] Step S211: If the current topic does not belong to the preset topic set, generate a few-sample model based on at least two of the sub-linguistic models.

[0056] Here, the platform can pre-set a topic set to record historical topics used for interaction. After determining the current topic corresponding to the audio information, the platform can match it with each topic in the topic set. If there is no match, it is determined that the current topic does not belong to the preset topic set, and the user terminal's edge model can be updated; if there is a match, it is determined that the current topic belongs to the preset topic set, and the user terminal's edge model does not need to be updated. A few-shot model can refer to a topic recognition model with a small amount of data and a simple structure. The few-shot model is related to the edge model; for example, the few-shot model can be sent to the user terminal as the user terminal's edge model. For example, a few-shot model can be obtained by quantizing and pruning at least two sub-linguistic models.

[0057] Step S212: Send the software package corresponding to the small sample model to the user terminal.

[0058] Here, a software package (JAR) carrying the parameters of the few-shot model can be generated. This package is used for synchronizing the client-side model. For example, after receiving the package containing the parameters of the few-shot model, the user terminal uses the parameters of the few-shot model as the parameters of the client-side model, thereby updating the client-side model.

[0059] In some embodiments, the platform can summarize the newly added sentence pattern rules reported by the user terminal, load the corresponding vocabulary, and perform thematic segmentation and summarization of audio information. The platform has a complete set of thematic sentence patterns. When a user adds a new theme, the platform dynamically loads the corresponding sentence patterns and vocabulary, and uses these corpora to build a specific sub-linguistic model. Based on the user's usage history and different theme usage patterns, the platform can use the sub-linguistic model to extract and merge thematic corpora, and train a user-personalized end-to-end cloud linguistic model (such as a few-shot model). After the overall model (platform model) is trained on the platform, a software package can be generated and the user terminal can be synchronized for replacement.

[0060] In this embodiment of the disclosure, by generating a few-sample model based on at least two sub-linguistic models, determining the software package corresponding to the few-sample model, and sending the software package to the user terminal, it is helpful to quickly and accurately achieve synchronization of the terminal-side model.

[0061] In some embodiments, the first recognition result is the current topic of the audio information, and the above step S211 includes the following steps S2111 to S2112:

[0062] Step S2111: Determine the current sentence structure of the audio information.

[0063] Here, sentence structure can be a subclass name derived from the local characteristics of a sentence, such as declarative sentence structure, interrogative sentence structure, etc. The platform can use a sentence structure recognition model to identify the current sentence structure of the audio information. A sentence structure recognition model can be understood as a pre-set machine learning model, such as a trained neural network model, that can be used to identify the sentence structure of audio information.

[0064] Step S2112: If the current sentence structure meets the preset conditions, generate the small sample model based on the audio information and at least two sub-linguistic models according to the preset time interval.

[0065] Here, the platform can pre-set a set of sentence patterns, including sentence patterns and historical sentence patterns used for storing voice interactions with users. The platform matches the current sentence pattern with historical sentence patterns in the set. If a match is successful, it indicates that the current sentence pattern meets preset conditions. The platform can then update the current small sample model at preset time intervals and simultaneously update the on-device model on the user terminal. For example, based on audio information, at least two sub-linguistic models can be pruned, quantized, and processed to obtain a sample model.

[0066] In some embodiments, step S212 above includes the following step S2113:

[0067] Step S2113: The software package is sent to the user terminal using a hot deployment method.

[0068] Here, hot deployment refers to upgrading the client-side model while the user terminal is running, without needing to restart the client-side model. The platform can use hot deployment to update the client-side model of the user terminal.

[0069] In some embodiments, the platform can load and train only on newly added user interaction topics. Due to its short training time, it can be trained daily, with each training session lasting less than 2 minutes, and the client-side model can be synchronized in real time. During training, the platform summarizes topics based on the templates involved in the user corpus model. Different topics have different frequencies of occurrence, so the same vocabulary phrases will have a bias in hit rate. This is a characteristic of the platform model and an advantage after strengthening user feature training. The user corpus model can be used to extract intent and slots from the corpus to achieve semantic understanding. The client-side model and platform model used for voice interaction can involve multiple business functions and adjust training phrases according to user behavior, with rapid response and hot deployment to user terminals. During the semantic understanding process, because the platform is a unified model (i.e., the platform model), it has strong universality when sending semantic understanding to the user terminal. After receiving the semantic understanding, the user terminal will use the client-side model to perform secondary recognition of the user's sentences and compare the hit scores of the platform's recognition, taking the one with the higher score.

[0070] In this embodiment of the disclosure, the platform generates a small sample model based on audio information and at least two sub-linguistic models at preset time intervals when the current sentence structure meets preset conditions, which helps to quickly and accurately update the terminal model.

[0071] This disclosure provides a model update method, wherein the platform model includes at least a semantic recognition model. For example... Figure 3 As shown, the method includes the following steps S301 to S304:

[0072] Steps S302 to S304 correspond to the aforementioned steps S102 to S104 respectively. When implementing these steps, you can refer to the specific implementation methods of the aforementioned steps S102 to S104.

[0073] Step S301: In response to receiving audio information sent by the user terminal, if the predicted probability of the topic recognition model is less than a preset probability threshold, the semantic recognition model is used to recognize the audio information to obtain the first recognition result.

[0074] Here, the topic recognition model, after determining the current topic of the audio information, can also obtain the predicted probability corresponding to the current topic. The predicted probability is used to characterize the likelihood of the current topic. The semantic recognition model can perform knowledge-based semantic understanding. The topic recognition model is based on language rules, sentence structure, and grammar recognition, while the semantic recognition model is based on knowledge recognition and serves as a supplement.

[0075] In some embodiments, the platform acquires audio information and its popularity information from a specified storage address. The popularity information includes the quantity of audio information, the earliest publication time, and associated user behavior data. The platform can then segment the crawled and cleaned news content into multiple sentences, resulting in multiple sentence sets corresponding to each news content. Each sentence in each set is then segmented into words, and sentence vectors and corresponding popularity values ​​are obtained based on the word segmentation. The popularity value has a detection mechanism; when the popularity value corresponding to a sentence vector exceeds a popularity threshold, the corresponding sentence vector set is proportionally added to the training set for incremental training. After the popularity determination model for the popularity value is trained, it outputs the popularity value corresponding to each major intent category. The popularity value (N) can be calculated in the following ways: Where, N i Let Nj be the current popularity of the intent for category i, and M be the popularity value of the crawled information content for this category. k Let k be the number of sentence patterns in the k-th major category. This indicates the frequency of voice usage by users in the corresponding major category.

[0076] Once the popularity value output by the popularity determination model matches the popularity value found in the news search, the model is confirmed to be consistent with knowledge popularity, and the platform model can be deployed. The platform model is trained on popularity knowledge, and after the platform model identifies the current topic, the client-side model will also perform a secondary correction. Because the correction is based on the conclusions of the client-side user habit model, it has a more personalized training effect, so the weight of the client-side model will be higher. The client-side typically receives general recognition results and personalized recognition results, and will execute corresponding actions based on the client-side's capability clusters and hit scores.

[0077] In this embodiment of the disclosure, when the predicted probability of the topic recognition model is less than a preset probability threshold, the platform can also use the semantic recognition model to recognize the audio information and quickly and accurately obtain the first recognition result.

[0078] This disclosure provides a model update method, such as... Figure 4 As shown, the method includes the following steps S401 to S406:

[0079] Steps S404 to S406 correspond to the aforementioned steps S102 to S104, respectively. When implementing these steps, the specific implementation methods of the aforementioned steps S102 to S104 can be referred to.

[0080] Step S401: In response to receiving audio information sent by the user terminal, determine the sentence similarity between the current sentence structure of the audio information and the preset sentence structure template.

[0081] Here, the preset sentence template can refer to a set of historical sentence patterns used to store voice interactions with users. The platform can use a preset sentence pattern recognition model to determine the current sentence pattern of the audio information and determine the sentence pattern similarity between the current sentence pattern and each historical sentence pattern in the preset sentence pattern template. For example, the feature distance between the vector features used to represent the current sentence pattern and the vector features used to represent historical sentence patterns can be determined, and this feature distance can be defined as the sentence pattern similarity.

[0082] Step S402: If the sentence similarity is less than the similarity threshold, slot replacement processing is performed on the text corresponding to the audio information to obtain the replaced audio information.

[0083] Here, a slot can refer to a keyword or word in the text corresponding to the audio information, such as the slot for time "today" or the slot for weather type "rainy". Slot replacement refers to replacing the slot in the text corresponding to the audio information, such as replacing the time "today" with "yesterday". For example, identify words with the same attributes as the current slot (such as synonyms), and randomly replace the current slot with words with the same attributes to obtain the replaced audio information.

[0084] Step S403: Use the platform model on the platform to identify the replaced audio information and obtain the first identification result corresponding to the replaced audio information.

[0085] Here, the replaced audio information can be input into the platform model to obtain the first recognition result corresponding to the replaced audio information, such as the topic corresponding to the replaced audio information.

[0086] In some embodiments, the platform can use a preset semantic parsing engine (for semantic understanding, etc.) to synchronize unrecognized corpora to the annotation module. The annotation module (for annotating key information in the corpora, etc.) performs similarity matching in the sentence pattern template pool. For example, if the sentence pattern similarity is c... n , Where, x k a n-k Let x represent the sentence similarity between the k-th template and the current sentence, where x represents the preset similarity weight, a represents the number of similar slots between the current sentence and the template, M represents the average number of slots between the current sentence and the template, and N represents the sentence similarity between the template and the template. K This can indicate the number of key slots in the template. If the sentence similarity is determined to be greater than or equal to 70%, the current template can be considered to match the current sentence pattern, and no manual annotation is required. If the sentence similarity is less than 70%, manual sentence pattern template extraction will be performed, such as annotating the domain, intent, and key slots.

[0087] The system retrieves key slots from the database used for corpus collection and replaces the slot values ​​with existing slot data. The semantic parsing engine determines if the data can be recognized. If it can, it indicates the issue stems from the absence of a vocabulary. The semantic annotation engine then extracts entities according to the sentence template and adds the extracted entities to the slots that replace the corresponding slots. If the semantic parsing engine cannot recognize the data, it indicates the sentence structure is the cause of the failure. Adding the current sentence structure to the sentence structure set corresponding to the intended meaning allows for sentence structure expansion, etc.

[0088] In this embodiment of the disclosure, by performing slot replacement processing on the text corresponding to the audio information when the sentence similarity is less than the similarity threshold, the replaced audio information can be obtained. This can achieve semi-automatic word list filtering and sentence pattern filtering, thereby realizing word list expansion and sentence pattern expansion of the existing platform model.

[0089] The following describes the application of the model update provided in this disclosure in a real-world scenario, using a multi-service voice interaction system scenario composed of user terminals and platform devices as an example. Figure 5 As shown, the multi-service voice interaction system may include a user terminal 501 and a platform terminal 502.

[0090] The user terminal 501 may include an on-device corpus collector 5011 and an on-device model 5012. The on-device corpus collector 5011 can be used to collect the user's audio information, and the on-device model 5012 can be used to identify the second recognition result of the audio information. The platform terminal 502 may include a corpus collection 5021, an annotation module 5022, a platform model 5023, and a semantic recognition model (also known as a knowledge training model) 5024. The corpus collection 5021 is used to store the corpus training library, the basic user training set, and the small sample user corpus, etc. The annotation module 5022 is used for user annotation and other processing of audio information. The platform model 5023 is used for recognition and other processing of audio information, and the semantic recognition model 5024 can also be used for recognition and other processing of audio information. The platform model 5023 may include a general language model 511, a personalized language model 512, and a small sample model 513, etc. The general language model 511 can be used for common grammar recognition, the personalized language model 512 can be used for user speech feature recognition, and the few-sample model 513 can be fine-tuned based on the audio information reported by the user terminal 501. When the predicted probability obtained by the platform model 5023 is less than the preset probability, the semantic recognition model 5024 can be used to perform reasoning retrieval based on objective knowledge, without being limited by grammatical rules.

[0091] In some embodiments, user terminals upload activity times and user tags weekly. Leveraging user habits across multiple terminals, the platform monitors the information flow on the device side. After recognizing semantics, the platform generates a personal corpus for each user, enabling dynamic modeling of the device-side model (also known as the device-side few-shot model). If the topic identified by the platform is a popular topic, the corpus is proportionally synchronized to the few-shot corpus. After each sentence is uploaded to the platform, the platform compares the first recognition result of the platform model with the second recognition result of the device-side model, selecting the higher-scoring result for dynamic feedback. After a period of positive feedback, the device-side model actively merges with the platform model, completing the platform model update. The platform periodically performs knowledge training based on popularity (i.e., training the semantic recognition model), and after fusing the device-side model and the popularity model, the final cloud model update is completed.

[0092] This disclosure provides a model update method, such as... Figure 6 As shown, the method may include the following steps S601 to S610:

[0093] Step S601: Record the user's request action and obtain audio information.

[0094] Here, the user terminal can use the user's audio information.

[0095] Step S602: Recognize the audio information.

[0096] Here, the user terminal can first use the on-device model to recognize the audio information and obtain a second recognition result. At the same time, the user terminal can also report the audio information to the platform.

[0097] Step S603: Determine whether the platform model is being recognized.

[0098] Here, the platform can use the platform model to identify audio information and obtain a first identification result; at the same time, it can determine the prediction probability corresponding to the first identification result; the platform can compare the prediction probability corresponding to the first identification result with a preset probability threshold to determine whether the prediction probability corresponding to the first identification result is less than the preset probability threshold; if it is less than the preset probability threshold, it is determined that the first identification result obtained by using the platform model to identify audio information is inaccurate, and then proceeds to step S604; if it is greater than or equal to the preset probability threshold, it is determined that the first identification result obtained by using the platform model to identify audio information is accurate, and then proceeds to step S606.

[0099] Step S604: Synchronize the audio information to the semantic recognition model.

[0100] Here, the platform can use a semantic recognition model trained based on popularity to identify audio information and obtain the first recognition result. Then, the platform can send the first recognition result to the user terminal.

[0101] Step S605: Extract knowledge from the audio information to obtain a semantic corpus.

[0102] Here, the platform can also extract knowledge from the audio information to obtain the knowledge entities corresponding to the audio information, and synchronize the knowledge entities to the semantic corpus. In some embodiments, the platform can use the synchronized semantic corpus to update the semantic recognition model to obtain an updated semantic recognition model.

[0103] Step S606: Extract audio information using a model to obtain a small sample corpus.

[0104] Here, the platform can utilize pre-defined natural language processing modules and other functional modules to extract data from audio information based on attributes such as syntax and sentence structure, obtaining corpus information such as entities corresponding to the audio information, and defining all corpus information such as entities corresponding to the audio information as a small sample corpus. In some embodiments, the platform can use the small sample corpus to perform quantization, pruning, and other processing on general language models and personalized language models to obtain small sample models.

[0105] Step S607: Determine whether the current topic is a popular topic.

[0106] Here, the platform can further match the current topic to which the audio information belongs with historical topics. If the match fails, it is determined that the current topic belongs to a less popular topic, and the process can proceed to step S608; if the match succeeds, it is determined that the current topic belongs to a popular topic, and the process can proceed to step S609.

[0107] Step S608: Obtain the first corpus corresponding to the unpopular topic, and update the semantic recognition model based on the first corpus to obtain the updated semantic recognition model.

[0108] Here, when the current topic is a niche topic, based on the number of corpora in the semantic corpus and a preset first ratio, the corpus of the niche topic corresponding to the first ratio is obtained; using the corpus of the niche topic corresponding to the first ratio, the current semantic recognition model is updated to obtain the updated semantic recognition model.

[0109] Step S609: Obtain the second corpus corresponding to the popular topics, and update the small sample model based on the second corpus to obtain the updated small sample model.

[0110] Here, when the current topic is a popular topic, based on the number of corpora in the small sample corpus and the preset second ratio, the corpus of popular topics corresponding to the second ratio is obtained; using the corpus of popular topics corresponding to the second ratio, the current small sample model is updated to obtain the updated small sample model.

[0111] Step S610: Synchronize the updated small sample model software package to the user terminal.

[0112] Here, the platform can generate a software package containing the parameters of the updated small sample model and hot-deploy the software package to the user terminal, so that the user terminal can update the terminal model in a timely manner.

[0113] In some embodiments, the platform may further include a voice information processing module, a natural language processing module, a few-shot model management module, a popularity determination model, and a natural language processing (NLP) engine module. The voice information processing module converts user voice files into text information, prioritizing matching based on user-uploaded hot words from various domains. The natural language processing module filters sensitive words based on the received user-spoken text information, prioritizes question-answer pairs, and finally performs NLP parsing, transmitting the parsed skill domains and results to the user terminal. The few-shot model management module stores various user-level few-shot corpora and adjusts the corpus proportions based on topic popularity. The popularity determination model periodically acquires target data and its popularity information from a specified address and trains the platform model based on the popularity. The NLP engine module distributes NLP parsing results to skill domains. The platform can perform corresponding logical processing based on various skill domains. When voice interaction information is incomplete, multi-turn interactions are triggered on the platform, and this module performs layered parsing of skill domains, etc.

[0114] This disclosure provides a corpus expansion method, such as... Figure 7 As shown, the method may include the following steps S701 to S713:

[0115] Step S701: Obtain the full corpus corresponding to the audio information.

[0116] Here, if the audio information is in audio format, the platform can convert the audio information into text information; the platform can then use a pre-defined entity extraction model to extract the full corpus corresponding to the text information.

[0117] Step S702: Perform an initial screening on the entire corpus to obtain the initial screening of the entire corpus.

[0118] Here, the platform can use a preset filtering method to perform the first filtering on the full corpus, and obtain the full corpus after the first filtering, such as the filtering program (Kenlm).

[0119] Step S703: Determine whether the full corpus after the initial screening is recognizable.

[0120] Here, the platform can match the full corpus after the initial screening with the pre-set historical corpus. If the match is successful, it is determined that the full corpus after the initial screening can be recognized, and then proceeds to step S704; if the match fails, it is determined that the full corpus after the initial screening cannot be recognized, and then proceeds to step S705.

[0121] Step S704: Determine the full corpus after the initial screening as identifiable corpus.

[0122] Here, if the full corpus after the initial screening is identifiable, the platform model on the platform side does not need to be updated.

[0123] Step S705: Manually annotate the full corpus after the initial screening to obtain the annotated full corpus.

[0124] Here, manual labeling can be used to annotate the intent, domain, and slot of the full corpus after the initial screening, resulting in an annotated full corpus.

[0125] Step S706: Replace the slots in the annotated full corpus to obtain the replaced full corpus.

[0126] Here, the target slot can be randomly selected from multiple slots in the annotated full corpus, and the content of the target operation can be processed by synonym replacement and other methods to obtain the replaced full corpus.

[0127] Step S707: Perform a second screening on the replaced full corpus to obtain the second-screened full corpus.

[0128] Here, the platform can match the full corpus with a preset filtered corpus, and then filter out the corpus that does not exist in the filtered corpus.

[0129] Step S708: Determine whether the full corpus after the second screening is recognizable.

[0130] Here, the platform can match the full corpus after secondary filtering with the pre-set historical corpus. If the match is successful, it is determined that the full corpus after secondary filtering can be recognized, and then proceeds to step S709; if the match fails, it is determined that the full corpus after secondary filtering cannot be recognized, and then proceeds to step S711.

[0131] Step S709: Perform word list filtering on the full corpus after the second filtering to obtain the word list after word list filtering.

[0132] Here, the platform can determine the word list to be filtered corresponding to the full corpus after secondary filtering. By matching the word list to be filtered with the word lists in the preset word list library, the word lists to be filtered that do not match the preset word list library are deleted, and the word list after filtering is obtained.

[0133] Step S710: Expand the vocabulary based on the filtered vocabulary.

[0134] Here, the vocabulary in the vocabulary library is expanded by adding the filtered vocabulary to the preset vocabulary library.

[0135] Step S711: Perform sentence pattern filtering on the full corpus after the second filtering to obtain the sentence patterns after sentence pattern filtering.

[0136] Here, the platform can determine the sentence patterns to be filtered corresponding to the full corpus after secondary filtering. By matching the sentence patterns to be filtered with the sentence patterns in the preset sentence pattern library, the sentence patterns to be filtered that do not match the preset sentence pattern library are deleted, and the sentence patterns after word list filtering are obtained.

[0137] Step S712: Based on the sentence patterns filtered from the vocabulary list, expand the sentence patterns.

[0138] Here, sentence patterns are added to a preset sentence pattern library after being filtered from the vocabulary list, thereby expanding the sentence patterns in the library.

[0139] Step S713: The annotated full corpus is determined as the training corpus set.

[0140] Here, the training corpus can be used for training and updating platform models, semantic recognition models, etc.

[0141] In this embodiment, training corpora from different domains can be dynamically loaded, and the number of training corpora from different domains can be configured according to the user's usage habits and behaviors on multiple terminals. When adding new names or expressions, a combination of full training and incremental training is selected to achieve hot deployment of the model and automated iterative updates, enabling seamless model upgrades during idle time. For voice-based smart hardware products, considering the long training time of large-scale semantic recognition engines, a small-sample augmentation method is used to iteratively upgrade the recognition engine (i.e., the platform model). This method allows users to dynamically record hit corpora during daily use of voice and other smart devices, extract models based on these corpora, and then generalize by loading a corpus vocabulary pool to complete self-learning semantic modeling. For unrecognized online corpora, slot annotation is performed by the operations team. Then, slot replacement is performed automatically, and the problem is automatically determined as a word list issue or a sentence structure issue based on the results, thereby achieving semi-automatic semantic modeling. The topic of semantic recognition can be dynamically adjusted. The hit rate of the knowledge engine is adjusted daily based on the frequency of appearance of popular information and knowledge in general media. The hit rate can be dynamically adjusted for information updates and iterations.

[0142] Based on the foregoing embodiments, this disclosure provides a model update device, which includes the included units and the modules included in each unit, which can be implemented by a processor in a computer device; of course, it can also be implemented by specific logic circuits; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0143] Figure 8 This is a schematic diagram of the composition structure of a model update device provided in an embodiment of the present disclosure, as shown below. Figure 8 As shown, the model update device 800 includes: a first identification module 810, a sending module 820, a receiving module 830, and a fusion module 840, wherein:

[0144] The first recognition module 810 is used to respond to audio information sent by the user terminal, and to recognize the audio information using the platform model on the platform to obtain a first recognition result; the sending module 820 is used to send the first recognition result to the user terminal; the receiving module 830 is used to receive the updated terminal-side model uploaded by the user terminal; the updated terminal-side model is obtained by the user terminal updating the initial terminal-side model based on the first recognition result and the second recognition result, and the second recognition result is obtained by the terminal-side model recognizing the audio information; the fusion module 840 is used to fuse the updated terminal-side model and the platform model to obtain an updated platform model.

[0145] In some embodiments, the platform model is a topic recognition model, which includes at least two sub-linguistic models; the first recognition module is further configured to: recognize the audio information using each of the sub-linguistic models, and determine the sub-recognition result corresponding to each of the sub-linguistic models; and obtain the first recognition result based on all the sub-recognition results; wherein at least two of the sub-linguistic models are used to generate the edge model.

[0146] In some embodiments, the first identification result is the current topic of the audio information; the device further includes: a generation module, configured to generate a few-shot model based on at least two of the sublinguistic models when the current topic does not belong to a preset topic set; and a synchronization module, configured to send the software package corresponding to the few-shot model to the user terminal; wherein the software package is used for the synchronization of the terminal model.

[0147] In some embodiments, the generation module is further configured to: determine the current sentence structure of the audio information; and, if the current sentence structure satisfies a preset condition, generate the few-sample model based on the audio information and at least two sub-linguistic models at a preset time interval; the synchronization module is further configured to: send the software package to the user terminal using a hot deployment method.

[0148] In some embodiments, the platform model includes at least a semantic recognition model; the first recognition module is further configured to: when the predicted probability of the topic recognition model is less than a preset probability threshold, use the semantic recognition model to recognize the audio information to obtain the first recognition result; wherein the semantic recognition model is updated based on the popularity of the audio information.

[0149] In some embodiments, the apparatus further includes: a second determining module, configured to determine the sentence similarity between the current sentence structure of the audio information and a preset sentence structure template; a replacement module, configured to perform slot replacement processing on the text corresponding to the audio information when the sentence similarity is less than a similarity threshold, to obtain the replaced audio information; and a second recognizing module, configured to recognize the replaced audio information using the platform model on the platform side, to obtain a first recognition result corresponding to the replaced audio information.

[0150] In some embodiments, the audio information is used for full training of the platform model and for incremental training of the edge model.

[0151] The descriptions of the apparatus embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or modules included in the apparatus provided in this disclosure can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the apparatus embodiments of this disclosure, please refer to the descriptions of the method embodiments of this disclosure for understanding.

[0152] It should be noted that, in the embodiments of this disclosure, if the above-described model update method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this disclosure, or the part that contributes to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk. Thus, the embodiments of this disclosure are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0153] This disclosure provides a computer device including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements some or all of the steps in the above-described method.

[0154] This disclosure provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the above-described method. The computer-readable storage medium may be transient or non-transient.

[0155] This disclosure provides a computer program including computer-readable code, wherein when the computer-readable code is executed in a computer device, a processor in the computer device performs some or all of the steps in the above-described method.

[0156] This disclosure provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-described method. This computer program product can be implemented specifically through hardware, software, or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium; in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK), etc.

[0157] It should be noted that the descriptions of the various embodiments above tend to emphasize the differences between them, while their similarities or commonalities can be referenced interchangeably. The descriptions of the above embodiments of the device, storage medium, computer program, and computer program product are similar to the descriptions of the above method embodiments and have similar beneficial effects. For technical details not disclosed in the embodiments of the device, storage medium, computer program, and computer program product of this disclosure, please refer to the descriptions of the method embodiments of this disclosure for understanding.

[0158] It should be noted that, Figure 9 This is a schematic diagram of a hardware entity of a computer device in an embodiment of this disclosure, such as... Figure 9 As shown, the hardware entity of the computer device 900 includes: a processor 901, a communication interface 902, and a memory 903, wherein:

[0159] Processor 901 typically controls the overall operation of computer device 900.

[0160] Communication interface 902 enables computer devices to communicate with other terminals or servers over a network.

[0161] The memory 903 is configured to store instructions and applications executable by the processor 901, and can also cache data to be processed or already processed (e.g., image data, audio data, voice communication data, and video communication data) in the processor 901 and various modules in the computer device 900. It can be implemented using flash memory or random access memory (RAM). Data transfer between the processor 901, the communication interface 902, and the memory 903 can be performed via bus 904.

[0162] It should be understood that the phrase "an embodiment" or "one embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this disclosure. Therefore, "in one embodiment" or "one embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this disclosure, the sequence numbers of the above steps / processes do not imply a sequential order of execution; the execution order of each step / process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this disclosure. The sequence numbers of the above embodiments of this disclosure are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0163] 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.

[0164] In the several embodiments provided in this disclosure, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0165] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0166] In addition, each functional unit in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0167] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0168] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, magnetic disks, or optical disks.

[0169] The methods disclosed in the several method embodiments provided in this disclosure can be arbitrarily combined without conflict to obtain new method embodiments.

[0170] The above description is merely an embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A model update method, characterized in that, Applied to the platform side, including: In response to receiving audio information sent by a user terminal, the platform model on the platform side is used to identify the audio information to obtain a first identification result; wherein, the platform model is a topic identification model, and the topic identification model includes at least two sub-linguistic models; the step of using the platform model on the platform side to identify the audio information to obtain the first identification result includes: using each of the sub-linguistic models to identify the audio information to obtain a sub-identification result corresponding to each sub-linguistic model; determining the first identification result based on all the sub-identification results; and using at least two of the sub-linguistic models to generate an edge model; The first identification result is sent to the user terminal; The system receives an updated edge model uploaded by the user terminal; the updated edge model is obtained by the user terminal updating the initial edge model based on the first recognition result and the second recognition result, whereby the second recognition result is obtained by the edge model recognizing the audio information. The updated endpoint model and the platform model are fused to obtain the updated platform model; Wherein, the first recognition result is the current topic of the audio information; the method further includes: when the current topic does not belong to a preset topic set, generating a few-shot model based on at least two of the sub-linguistic models; sending the software package corresponding to the few-shot model to the user terminal; wherein, the software package is used for the synchronization of the terminal model; The step of generating a few-shot model based on at least two of the sub-linguistic models includes: determining the current sentence structure of the audio information; and generating the few-shot model based on the audio information and at least two of the sub-linguistic models at preset time intervals, provided that the current sentence structure meets preset conditions.

2. The method according to claim 1, characterized in that, Sending the software package to the user terminal includes: The software package is sent to the user terminal using a hot deployment method.

3. The method according to claim 1, characterized in that, The platform model includes at least a semantic recognition model; the step of using the platform model on the platform to recognize the audio information and obtain a first recognition result includes: If the predicted probability of the topic recognition model is less than a preset probability threshold, the semantic recognition model is used to recognize the audio information to obtain the first recognition result. The semantic recognition model is updated based on the popularity of the audio information.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Determine the sentence similarity between the current sentence structure of the audio information and the preset sentence structure template; If the sentence similarity is less than the similarity threshold, slot replacement processing is performed on the text corresponding to the audio information to obtain the replaced audio information; The platform model on the platform is used to identify the replaced audio information to obtain the first identification result corresponding to the replaced audio information.

5. The method according to any one of claims 1 to 3, characterized in that, The audio information is used for full training of the platform model and for incremental training of the edge model.

6. A model updating device, characterized in that, Applied to the platform side, including: A first recognition module is configured to, in response to receiving audio information sent by a user terminal, recognize the audio information using a platform model on the platform side to obtain a first recognition result; wherein, the platform model is a topic recognition model, and the topic recognition model includes at least two sub-linguistic models; the first recognition module is further configured to: recognize the audio information using each of the sub-linguistic models to obtain a sub-recognition result corresponding to each sub-linguistic model; determine the first recognition result based on all the sub-recognition results; and the at least two sub-linguistic models are used to generate an edge-side model; A sending module is used to send the first identification result to the user terminal; A receiving module is configured to receive an updated end-side model uploaded by the user terminal; the updated end-side model is obtained by the user terminal updating the initial end-side model based on the first recognition result and the second recognition result, wherein the second recognition result is obtained by the end-side model recognizing the audio information; The fusion module is used to fuse the updated end-side model and the platform model to obtain the updated platform model; Wherein, the first recognition result is the current topic of the audio information; the device further includes: a generation module, used to generate a few-shot model based on at least two of the sub-linguistic models when the current topic does not belong to a preset topic set; and a synchronization module, used to send the software package corresponding to the few-shot model to the user terminal; wherein, the software package is used for the synchronization of the terminal model; The generation module is further configured to: determine the current sentence structure of the audio information; and, if the current sentence structure meets preset conditions, generate the small sample model based on the audio information and at least two sub-linguistic models at preset time intervals.

7. A computer device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 5.

Images