Spoken language named entity recognition method and device, computer device and storage medium

Abstract

The present application relates to the field of natural language processing, and discloses a spoken language named entity recognition method and device, computer equipment and storage medium, the method comprising: processing the voice data set through the first self-training model to obtain the pseudo label and the first uncertainty of each voice data in the voice data set; when the first uncertainty meets the first preset convergence condition, output the first labeled data; processing the first labeled data through the continuous multi-word pseudo label verification model to obtain the second labeled data; processing the second labeled data through the second self-training model to obtain the named entity label and the second uncertainty of each voice data of the second labeled data; when the second uncertainty meets the second preset convergence condition, output the named entity recognition result of the voice data set. The present application greatly reduces the dependence of spoken language named entity recognition on a large number of labeled data and reduces the training cost.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a spoken named entity recognition method, apparatus, computer device, and storage medium. Background Technology

[0002] The financial industry, a service sector, involves numerous customer interactions, such as telephone banking and car insurance claims. Traditionally, these transactions required manual handling by customer service personnel. However, with the development of natural language processing technology, an increasing number of chatbots are replacing human agents, significantly reducing operating costs for various business operations.

[0003] Replacing human agents with chatbots requires the chatbots to accurately understand customers' spoken language. In some cases, chatbots use Spoken Language Understanding (SLU) tasks to convert customers' spoken language into structured information for computer understanding and analysis. Spoken Language Understanding involves mapping speech signals to semantic labels, also known as Spoken Named Entity Recognition. Spoken Named Entity Recognition requires a large amount of labeled data. When dealing with a new task or domain, collecting large amounts of labeled data is difficult and costly, while unlabeled data is readily available and inexpensive. Summary of the Invention

[0004] Therefore, it is necessary to provide a spoken name entity recognition method, apparatus, computer equipment, and storage medium to address the aforementioned technical problems, so as to reduce the dependence of spoken name entity recognition on a large amount of labeled data and reduce training costs.

[0005] A spoken named entity recognition method includes:

[0006] The speech dataset is processed by a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels.

[0007] When the first uncertainty satisfies the first preset convergence condition, the first labeled data is output; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels;

[0008] The first labeled data is processed by a continuous multi-term pseudo-label verification model to obtain the second labeled data;

[0009] The second labeled data is processed by a second self-trained model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels;

[0010] When the second uncertainty satisfies the second preset convergence condition, the named entity recognition result of the speech dataset is output.

[0011] A spoken named entity recognition device, comprising:

[0012] The first training module is used to process the speech dataset through a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels.

[0013] The first output module is used to output first labeled data when the first uncertainty satisfies the first preset convergence condition; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels;

[0014] The correction module is used to process the first labeled data through a continuous multi-term pseudo-label verification model to obtain the second labeled data;

[0015] The second training module is used to process the second labeled data through a second self-training model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels.

[0016] The second output module is used to output the named entity recognition result of the speech dataset when the second uncertainty satisfies the second preset convergence condition.

[0017] A computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor implements the above-described spoken named entity recognition method when executing the computer-readable instructions.

[0018] One or more readable storage media storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the spoken named entity recognition method described above.

[0019] The aforementioned spoken name entity recognition method, apparatus, computer device, and storage medium process a speech dataset using a first self-trained model based on a Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data point in the speech dataset, and a first uncertainty corresponding to the pseudo-labels. When the first uncertainty satisfies a first preset convergence condition, first labeled data is output. The first labeled data includes pseudo-labels and the uncertainties corresponding to the pseudo-labels. Here, generating pseudo-labels through the first self-trained model significantly reduces the time and cost of labeling. The first labeled data is processed by a continuous multi-word pseudo-label verification model to obtain second labeled data, which is then used to evaluate and correct the first labeled data, improving the accuracy of spoken name entity recognition. The second labeled data is processed by a second self-trained model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data point in the second labeled data, and a second uncertainty corresponding to the named entity labels. When the second uncertainty satisfies a second preset convergence condition, the named entity recognition result of the speech dataset is output. Here, a highly accurate named entity recognition result can be obtained. This invention achieves accurate recognition of spoken named entities through two training sessions and one intermediate verification, greatly reducing the dependence of spoken named entity recognition on a large amount of labeled data and lowering training costs. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of an application environment for a spoken named entity recognition method according to an embodiment of the present invention;

[0022] Figure 2 This is a flowchart illustrating a spoken named entity recognition method according to an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of a spoken language named entity recognition device according to an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The spoken named entity recognition method provided in this embodiment can be applied to, for example, Figure 1 In this application environment, the client communicates with the server. Clients include, but are not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0027] In one embodiment, such as Figure 2 As shown, a spoken named entity recognition method is provided, which is then applied to... Figure 1 Taking the server-side as an example, the explanation includes the following steps:

[0028] S10. Process the speech dataset using a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels.

[0029] Understandably, there are two main types of spoken named entity recognition (NER) methods: end-to-end (E2E) methods and pipelined methods. The pipelined method involves performing speech recognition to obtain the corresponding text, and then using the obtained text for named entity recognition (NER). The spoken named entity recognition method provided in this embodiment is an improvement on the pipelined method.

[0030] Specifically, the first self-trained model is a semi-supervised learning model built based on the Bayesian Inconsistent Active Learning (BALD) algorithm. The BALD algorithm selects the samples most likely to improve the classification performance of the first self-trained model for labeling. The speech dataset contains multiple speech data points. By processing the speech data through the first self-trained model, corresponding prediction results, i.e., pseudo-labels for the speech data, can be obtained. Simultaneously, the variance of the estimated class probability based on these pseudo-labels can be calculated, which is the first uncertainty.

[0031] S20. When the first uncertainty satisfies the first preset convergence condition, output the first annotation data; the first annotation data includes pseudo-labels and the uncertainty corresponding to the pseudo-labels.

[0032] Understandably, the first preset convergence condition can be set according to actual needs. In some examples, the first preset convergence condition refers to the first uncertainty approaching a certain convergence value, that is, the difference between the first uncertainty and the convergence value is less than a preset convergence difference. For example, if the convergence value is 1 and the preset convergence difference is 0.05, and the difference between the first uncertainty and the convergence value 1 is less than the preset convergence difference of 0.05 for multiple consecutive times (e.g., it can be set to 10 times), then it can be determined that the first uncertainty meets the first preset convergence condition. After determining that the first uncertainty meets the first preset convergence condition, the pseudo-label of the speech data and the first uncertainty, that is, the first labeled data, can be output.

[0033] S30. Process the first labeled data using a continuous multi-term pseudo-label verification model to obtain the second labeled data.

[0034] Understandably, a continuous multi-term pseudo-labeling verification model is used to check whether the first labeled data is correct. During the processing of the first labeled data, the continuous multi-term pseudo-labeling verification model can also correct incorrect first labeled data to obtain second labeled data. In some examples, the continuous multi-term pseudo-labeling verification model can be a 4-gram model. A 4-gram model can segment the text into a sequence of four consecutive words and then verify the correctness of the sequence. After discovering incorrect first labeled data, a correction procedure can be used to correct it. Here, the correction procedure can involve manual intervention or can be an automatic correction model built based on historical data.

[0035] S40. The second labeled data is processed by a second self-trained model based on the Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels.

[0036] Understandably, the second self-trained model can either reuse the first self-trained model or be a semi-supervised learning model trained on an untrained speech dataset. Similar to the first self-trained model, the second self-trained model is a semi-supervised learning model built based on the Bayesian inconsistent active learning algorithm. The training data used by the second self-trained model is the second labeled data. By processing the second labeled data through the second self-trained model, the corresponding prediction results, i.e., named entity labels for the speech data, can be obtained. Simultaneously, the variance of the estimated class probabilities of these named entity labels can be calculated, which is the second uncertainty.

[0037] S50. When the second uncertainty satisfies the second preset convergence condition, output the named entity recognition result of the speech dataset.

[0038] Understandably, the second preset convergence condition can be set according to actual needs. In some examples, the second preset convergence condition refers to the second uncertainty approaching a certain convergence value, that is, the difference between the second uncertainty and the convergence value is less than a preset convergence difference. For example, if the convergence value is 1 and the preset convergence difference is 0.03, and the difference between the second uncertainty and the convergence value 1 is less than the preset convergence difference of 0.03 for several consecutive times (e.g., it can be set to 8 times), then it can be determined that the second uncertainty meets the second preset convergence condition. After determining that the second uncertainty meets the second preset convergence condition, the named entity recognition result of the speech data can be output.

[0039] In steps S10-S50, the speech dataset is processed by a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset and a first uncertainty corresponding to the pseudo-labels. When the first uncertainty satisfies a first preset convergence condition, first labeled data is output. The first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels. Here, generating pseudo-labels through the first self-trained model greatly reduces the time and cost of labeling. The first labeled data is processed by a continuous multi-word pseudo-label verification model to obtain second labeled data, which is used to evaluate and correct the first labeled data, improving the accuracy of spoken named entity recognition. The second labeled data is processed by a second self-trained model based on the Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data and a second uncertainty corresponding to the named entity labels. When the second uncertainty satisfies a second preset convergence condition, the named entity recognition result of the speech dataset is output. Here, a highly accurate named entity recognition result can be obtained. This embodiment achieves accurate recognition of spoken named entities through two training sessions and one intermediate verification, greatly reducing the dependence of spoken named entity recognition on a large amount of labeled data and lowering training costs.

[0040] Optionally, before step S10, i.e. before processing the speech dataset using the first self-trained model based on the Bayesian inconsistency active learning algorithm, the following steps are included:

[0041] S11. Obtain the unlabeled speech dataset;

[0042] S12. Label a portion of the speech data in the unlabeled speech dataset to obtain initial labeled data; wherein, the unlabeled speech data in the unlabeled speech dataset is the initial unlabeled data; the number of initial labeled data is less than the number of initial unlabeled data;

[0043] S13. Construct the speech dataset based on the initial labeled data and the initial unlabeled data.

[0044] Understandably, the speech data in the unlabeled speech dataset is unlabeled. A small portion of the speech data in the unlabeled speech dataset can be labeled to obtain initial labeled data. The majority of the speech data remains unlabeled; this unlabeled speech data constitutes the initial unlabeled data. A speech dataset can be constructed using the initial labeled data and the initial unlabeled data.

[0045] In this embodiment, only a small amount of voice data can be labeled, which can greatly reduce the workload and cost of labeling.

[0046] Optionally, step S10, namely, processing the speech dataset using a first self-trained model based on a Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels, includes:

[0047] S101. The first self-trained model containing a random deactivation layer is used to process the speech data multiple times to obtain multiple prediction results;

[0048] S102. Determine the pseudo-label of the speech data based on the multiple prediction results;

[0049] S103. Calculate the probability distribution of the multiple prediction results, and determine the first uncertainty of the speech data based on the probability distribution.

[0050] In essence, a randomized dropout layer refers to a computational layer based on the Monte Carlo dropout algorithm. During the training of the first self-trained model, the randomized dropout layer randomly discards some neurons, forcing the first self-trained model to learn multiple independent sub-models. These sub-models can complement each other, reducing the risk of overfitting and making the first self-trained model more robust to small perturbations in the input. The randomized dropout layer can reduce overfitting in the first self-trained model, enhance its generalization ability, and accelerate its training speed. Each time the first self-trained model processes speech data, it obtains one prediction result; processing multiple times yields multiple prediction results.

[0051] Pseudo-labels for speech data can be determined based on multiple prediction results. The pseudo-label can be the average of the multiple prediction results. The probability distribution of the multiple prediction results can be calculated, and the first uncertainty of the speech data can be determined based on this probability distribution. Here, the first uncertainty can be an entropy function, variance, etc.

[0052] This embodiment calculates the pseudo-labels and the first uncertainty of the speech data through the first self-trained model, which greatly reduces the time and money costs of annotation.

[0053] Optionally, step S30, namely, processing the first labeled data through the continuous multi-term pseudo-label verification model to obtain the second labeled data, includes:

[0054] S301. The first labeled data is verified by the continuous multi-term pseudo-label verification model to find the erroneous labeled data contained in the first labeled data.

[0055] S302. Receive a correction instruction and correct the erroneous labeled data according to the correction instruction;

[0056] S303. Generate the second annotation data based on the corrected first annotation data.

[0057] Understandably, a continuous multi-term pseudo-label validation model can be used to verify the correctness of the first labeled data, in order to find erroneous labeled data contained in the first labeled data. In some examples, the continuous multi-term pseudo-label validation model can be an n-gram model, such as a 4-gram model. The correction instructions can be manually entered instructions or instructions automatically generated by the correction program. Through the correction instructions, erroneous labeled data can be corrected. Finally, based on the corrected first labeled data, second labeled data is generated.

[0058] In this embodiment, the first labeled data is corrected and the second labeled data is generated by using a continuous multi-term pseudo-label verification model, which can improve the accuracy of spoken name entity recognition.

[0059] Optionally, step S301, namely, verifying the first labeled data using the continuous multi-term pseudo-label verification model to find erroneous labeled data contained in the first labeled data, includes:

[0060] S3011. Obtain the text to be evaluated from the first labeled data;

[0061] S3012. Calculate the probability of occurrence of the text to be evaluated;

[0062] S3013. If the occurrence probability is lower than the preset occurrence probability threshold, the text to be evaluated is marked as the error annotation data.

[0063] Understandably, the text to be evaluated can be the result of processing speech data through a pre-trained self-trained model. When using a continuous multi-term pseudo-label validation model to verify the text, the probability of its occurrence can be calculated first. A higher probability indicates a lower error rate, and vice versa. A preset probability threshold can be set; if the probability is lower than the threshold, the text is marked as incorrectly labeled; if the probability is higher than or equal to the threshold, it is marked as correctly labeled. The preset probability threshold can be set according to actual needs.

[0064] This embodiment improves the accuracy of the annotated text by verifying each text to be evaluated in the first annotated data.

[0065] Optionally, step S3012, namely calculating the occurrence probability of the text to be evaluated, includes:

[0066] S30121. Obtain the predicted occurrence probability of each word in the text to be evaluated;

[0067] S30122. Multiply all the predicted occurrence probabilities to obtain the occurrence probability of the text to be evaluated.

[0068] Understandably, the predicted occurrence probability of each word in the text to be evaluated can be obtained. Here, the predicted occurrence probability of each word can be queried from a database associated with the continuous multi-word pseudo-label validation model. The database used here can be an industry database in the same domain as the speech dataset. For example, if the speech dataset comes from customer dialogue records of a specific insurance type, the database could be a database built based on historical customers of that insurance type (or all insurance types). After obtaining the predicted occurrence probabilities, all predicted occurrence probabilities can be multiplied to obtain the occurrence probability of the text to be evaluated.

[0069] Taking the 4-gram model as an example, for a text to be evaluated consisting of m words, the 4-gram model is represented as follows:

[0070]

[0071] Where w1 is the first word of the text to be evaluated; w2 is the second word of the text w to be evaluated; w m Let m be the m-th word in the text to be evaluated;

[0072] P(w1,w2,…,w m ) represents the predicted probability of the text to be evaluated;

[0073] P(w i |w i-3 wi-2 w i-1 () represents the predicted probability of the i-th word in the text to be evaluated. The probability of the i-th word's occurrence depends only on the probability of the three words w preceding it. i-3 w i-2 w i-1 Relevant. Here, the predicted occurrence probability P(w1) of the first word in the text to be evaluated, the predicted occurrence probability P(w2|w1), the predicted occurrence probability P(w3|w1w2), and the predicted occurrence probability P(w4|w1w2w3) of the second word, the predicted occurrence probability P(w4|w1w2w3) of the third word, are given.

[0074] This embodiment obtains the occurrence probability of the text to be evaluated, which greatly improves the efficiency of finding incorrectly labeled data and has a high accuracy.

[0075] Optionally, step S50, namely, outputting the named entity recognition result of the speech dataset when the second uncertainty satisfies the second preset convergence condition, includes:

[0076] S501. When the second uncertainty satisfies the second preset convergence condition, it is determined that the second self-training model has been trained and the trained second self-training model is determined as a spoken name entity recognition model.

[0077] S502. Process the speech dataset using the spoken name entity recognition model to obtain the name entity recognition result.

[0078] Understandably, the second preset convergence condition can be set according to actual needs. When the second uncertainty satisfies the second preset convergence condition, the second self-trained model can be considered trained successfully. At this point, the trained second self-trained model can be identified as the spoken name entity recognition model. The spoken name entity recognition model can be used to process speech samples and obtain corresponding recognition results. Using the spoken name entity recognition model to process speech datasets, named entity recognition results for the speech datasets can be obtained.

[0079] This embodiment obtains a spoken name entity recognition model for processing speech samples. This spoken name entity recognition model has a high recognition accuracy and can output accurate name entity recognition results.

[0080] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0081] In one embodiment, a spoken name entity recognition device is provided, which corresponds one-to-one with the spoken name entity recognition method described in the above embodiments. For example... Figure 3As shown, the spoken name entity recognition device includes a first training module 10, a first output module 20, a correction module 30, a second training module 40, and a second output module 50. Detailed descriptions of each functional module are as follows:

[0082] The first training module 10 is used to process the speech dataset through a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels.

[0083] The first output module 20 is used to output first labeled data when the first uncertainty satisfies the first preset convergence condition; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels;

[0084] The correction module 30 is used to process the first labeled data through a continuous multi-term pseudo-label verification model to obtain the second labeled data;

[0085] The second training module 40 is used to process the second labeled data through a second self-training model based on the Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels.

[0086] The second output module 50 is used to output the named entity recognition result of the speech dataset when the second uncertainty satisfies the second preset convergence condition.

[0087] Optionally, the spoken named entity recognition device further includes a dataset construction module, which includes:

[0088] Obtain initial dataset units to acquire unlabeled speech datasets;

[0089] An annotation unit is used to annotate a portion of the speech data in the unannotated speech dataset to obtain initial annotated data; wherein, the unannotated speech data in the unannotated speech dataset is the initial unannotated data; the number of initial annotated data is less than the number of initial unannotated data;

[0090] A dataset construction unit is used to construct the speech dataset based on the initial labeled data and the initial unlabeled data.

[0091] Optionally, the first training module 10 includes:

[0092] The prediction unit is used to process the speech data multiple times using the first self-trained model containing a random deactivation layer to obtain multiple prediction results;

[0093] A pseudo-label determination unit is used to determine pseudo-labels for the speech data based on the multiple prediction results;

[0094] A first uncertainty unit is determined, which is used to calculate the probability distribution of the plurality of prediction results, and the first uncertainty of the speech data is determined based on the probability distribution.

[0095] Optionally, the calibration module 30 includes:

[0096] The error data finding unit is used to verify the first labeled data through the continuous multi-term pseudo-label verification model in order to find the erroneous labeled data contained in the first labeled data;

[0097] A correction unit is configured to receive a correction instruction and correct the erroneous labeled data according to the correction instruction;

[0098] A second annotation data unit is generated, which is used to generate the second annotation data based on the corrected first annotation data.

[0099] Optionally, the error data lookup unit includes:

[0100] Obtain the text unit to be evaluated, used to obtain the text to be evaluated from the first labeled data;

[0101] Obtain an occurrence probability unit to calculate the occurrence probability of the text to be evaluated;

[0102] The error data unit is used to mark the text to be evaluated as the error annotation data if the occurrence probability is lower than a preset occurrence probability threshold.

[0103] Optionally, calculating the occurrence probability of the text to be evaluated includes:

[0104] A predictive occurrence probability unit is used to obtain the predicted occurrence probability of each word in the text to be evaluated;

[0105] An occurrence probability unit is obtained, which is used to multiply all the predicted occurrence probabilities to obtain the occurrence probability of the text to be evaluated.

[0106] Optionally, the second output module 50 includes:

[0107] The identification model unit is used to determine that the second self-training model has been trained when the second uncertainty meets the second preset convergence condition, and to identify the trained second self-training model as a spoken name entity recognition model.

[0108] The recognition result obtaining unit is used to process the speech dataset through the spoken name entity recognition model to obtain the name entity recognition result.

[0109] Specific limitations regarding the spoken name entity recognition device can be found in the limitations of the spoken name entity recognition method described above, and will not be repeated here. Each module in the aforementioned spoken name entity recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0110] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes a readable storage medium and internal memory. The readable storage medium stores an operating system, computer-readable instructions, and a database. The internal memory provides an environment for the operation of the operating system and computer-readable instructions in the readable storage medium. The database stores data related to the spoken name entity recognition method. The network interface communicates with external terminals via a network connection. When the computer-readable instructions are executed by the processor, they implement a spoken name entity recognition method. The readable storage medium provided in this embodiment includes both non-volatile and volatile readable storage media.

[0111] In one embodiment, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor performs the following steps when executing the computer-readable instructions:

[0112] The speech dataset is processed by a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels.

[0113] When the first uncertainty satisfies the first preset convergence condition, the first labeled data is output; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels;

[0114] The first labeled data is processed by a continuous multi-term pseudo-label verification model to obtain the second labeled data;

[0115] The second labeled data is processed by a second self-trained model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels;

[0116] When the second uncertainty satisfies the second preset convergence condition, the named entity recognition result of the speech dataset is output.

[0117] In one embodiment, one or more computer-readable storage media storing computer-readable instructions are provided. The readable storage media provided in this embodiment include non-volatile readable storage media and volatile readable storage media. The readable storage media stores computer-readable instructions, which, when executed by one or more processors, perform the following steps:

[0118] The speech dataset is processed by a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels.

[0119] When the first uncertainty satisfies the first preset convergence condition, the first labeled data is output; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels;

[0120] The first labeled data is processed by a continuous multi-term pseudo-label verification model to obtain the second labeled data;

[0121] The second labeled data is processed by a second self-trained model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels;

[0122] When the second uncertainty satisfies the second preset convergence condition, the named entity recognition result of the speech dataset is output.

[0123] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When executed, these computer-readable instructions can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0124] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0125] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A spoken named entity recognition method, characterized in that, include: The speech dataset is processed by a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels. When the first uncertainty satisfies the first preset convergence condition, the first labeled data is output; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels; The first labeled data is processed by a continuous multi-term pseudo-label verification model to obtain the second labeled data; The second labeled data is processed by a second self-trained model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels; When the second uncertainty satisfies the second preset convergence condition, the named entity recognition result of the speech dataset is output; The step of processing the first labeled data through a continuous multi-term pseudo-label verification model to obtain the second labeled data includes: The first labeled data is verified using the continuous multi-term pseudo-label verification model to find erroneous labeled data contained in the first labeled data. Receive a correction instruction and correct the erroneous labeled data according to the correction instruction; The second annotation data is generated based on the corrected first annotation data; The step of verifying the first labeled data using the continuous multi-term pseudo-label verification model to find erroneous labeled data contained in the first labeled data includes: Obtain the text to be evaluated from the first labeled data; Calculate the probability of occurrence of the text to be evaluated; If the occurrence probability is lower than a preset occurrence probability threshold, the text to be evaluated is marked as the erroneous annotation data; The calculation of the occurrence probability of the text to be evaluated includes: Obtain the predicted occurrence probability of each word in the text to be evaluated; Multiply all the predicted occurrence probabilities to obtain the occurrence probability of the text to be evaluated.

2. The spoken named entity recognition method as described in claim 1, characterized in that, Before processing the speech dataset using the first self-trained model based on the Bayesian inconsistency active learning algorithm, the process includes: Obtain the unlabeled speech dataset; A portion of the speech data in the unlabeled speech dataset is labeled to obtain initial labeled data; wherein, the unlabeled speech data in the unlabeled speech dataset is the initial unlabeled data; the number of initial labeled data is less than the number of initial unlabeled data; The speech dataset is constructed based on the initial labeled data and the initial unlabeled data.

3. The spoken named entity recognition method as described in claim 1, characterized in that, The step of processing the speech dataset using a first self-trained model based on a Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data point in the speech dataset, and a first uncertainty corresponding to the pseudo-labels, includes: The first self-trained model, which includes a random deactivation layer, is used to process the speech data multiple times to obtain multiple prediction results. Based on the multiple prediction results, pseudo-labels for the speech data are determined; Calculate the probability distribution of the multiple prediction results, and determine the first uncertainty of the speech data based on the probability distribution.

4. The spoken named entity recognition method as described in claim 1, characterized in that, When the second uncertainty satisfies the second preset convergence condition, the named entity recognition result of the speech dataset is output, including: When the second uncertainty satisfies the second preset convergence condition, it is determined that the second self-trained model has been trained and the trained second self-trained model is identified as a spoken name entity recognition model. The speech dataset is processed by the spoken name entity recognition model to obtain the name entity recognition result.

5. A spoken named entity recognition device, characterized in that, The apparatus is used to perform the spoken named entity recognition method according to claim 1, including: The first training module is used to process the speech dataset through a first self-trained model based on the Bayesian inconsistency active learning algorithm to obtain pseudo-labels for each speech data in the speech dataset, and a first uncertainty corresponding to the pseudo-labels. The first output module is used to output first labeled data when the first uncertainty satisfies the first preset convergence condition; the first labeled data includes pseudo-labels and uncertainties corresponding to the pseudo-labels; The correction module is used to process the first labeled data through a continuous multi-term pseudo-label verification model to obtain the second labeled data; The second training module is used to process the second labeled data through a second self-training model based on a Bayesian inconsistency active learning algorithm to obtain named entity labels for each speech data in the second labeled data, and a second uncertainty corresponding to the named entity labels. The second output module is used to output the named entity recognition result of the speech dataset when the second uncertainty satisfies the second preset convergence condition.

6. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the computer-readable instructions, it implements the spoken named entity recognition method as described in any one of claims 1 to 4.

7. One or more readable storage media storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the spoken named entity recognition method as described in any one of claims 1 to 4.

Images