Error entity detection device, method, and training method of error entity detection model

Abstract

The embodiment of the application provides an error entity detection device, method and training method of an error entity detection model, and relates to the technical field of error entity detection. The error entity detection device comprises an error entity detection model, wherein the error entity detection model comprises: an error entity probability detection network configured to detect an error entity probability corresponding to each word segmentation in to-be-detected text; and a labeling network configured to label an error entity and a classification field to which the error entity belongs according to the error entity probability, to obtain an error entity labeling result and a classification field labeling result. The embodiment of the application is used for improving the exposure deviation problem and improving the generalization detection capability for error entity detection.

Description

TECHNICAL FIELD

[0001] The embodiments of the present application relate to the technical field of error entity detection. More particularly, the embodiments of the present application relate to an error entity detection device, an error entity detection method and a training method of an error entity detection model. BACKGROUND

[0002] Error entity detection can identify error entities in text so as to correct the error entities subsequently, and is widely applied in the field of voice interaction and the like. Voice recognition is the first step after the voice interaction system is woken up and is the basis for subsequent processing. However, in the voice recognition process, due to the accent of the speaker, environmental noise, homonym polysemy and other factors, voice recognition errors can be caused. In a serial process voice interaction system, the voice recognition errors can also be amplified step by step and accumulated to the subsequent link. Therefore, it is a basic premise to ensure the accuracy of the text corresponding to the voice. For this purpose, it is necessary to identify error entities in the text corresponding to the voice so as to correct the error entities.

[0003] At present, error entity detection is mainly integrated in an end-to-end error correction model. The structure of the end-to-end error correction model is mainly a segmented structure of error entity detection, candidate recall and error correction. When the text is input into the end-to-end error correction model, the error-corrected text can be obtained. However, the end-to-end error correction model has a large exposure bias problem, which leads to poor generalization detection capability for error entity detection. SUMMARY

[0004] The exemplary embodiments of the present application provide an error entity detection device, an error entity detection method and a training method of an error entity detection model, which are used to improve the exposure bias problem and improve the generalization detection capability for error entity detection.

[0005] The technical solutions of the embodiments of the present application are as follows:

[0006] In a first aspect, the embodiments of the present application provide an error entity detection device, comprising an error entity detection model, wherein the error entity detection model comprises:

[0007] an error entity probability detection network configured to detect an error entity probability corresponding to each word segmentation in a to-be-detected text;

[0008] a labeling network configured to label an error entity and label a classification field to which the error entity belongs according to the error entity probability, to obtain an error entity labeling result and a classification field labeling result.

[0009] In a second aspect, the embodiments of the present application provide an error entity detection method applied to an error entity detection device, comprising:

[0010] detect an error entity probability corresponding to each wordpiece in the to-be-detected text;

[0011] According to the error entity probability, an error entity is labeled and a classification field to which the error entity belongs is labeled.

[0012] In a third aspect, an embodiment of the present application provides a training method of an error entity detection model, applied to a model training device, and the method comprises the following steps.

[0013] Obtaining a training sample, wherein the training sample comprises a sample text, a real error entity probability corresponding to each wordpiece in the sample text, and a real error entity labeling result and a real classification field labeling result corresponding to the sample text;

[0014] Inputting the training sample into a to-be-trained error entity detection model to obtain an error entity probability corresponding to each wordpiece in the sample text, an error entity labeling result and a classification field labeling result output by the to-be-trained error entity detection model;

[0015] Determining an error entity probability loss function according to the real error entity probability and the error entity probability output by the to-be-trained error entity detection model, determining an error entity labeling loss function according to the real error entity labeling result and the error entity labeling result output by the to-be-trained error entity detection model, and determining a classification field labeling loss function according to the real classification field labeling result and the classification field labeling result output by the to-be-trained error entity detection model;

[0016] Determining a model loss function according to the error entity probability loss function, the error entity labeling loss function and the classification field labeling loss function;

[0017] Adjusting parameters in the to-be-trained error entity detection model according to the model loss function until the model loss function meets a preset training end condition.

[0018] In a fourth aspect, an embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory is configured to store a computer program, and the processor is configured to execute the computer program, so that the electronic device implements the error entity detection method in the second aspect or any one of the embodiments of the second aspect, or the training method of the error entity detection model in the third aspect or any one of the embodiments of the third aspect.

[0019] In a fifth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program. When the computer program is executed by a computing device, the computing device implements the error entity detection method in the second aspect or any of the embodiments of the second aspect, or the training method of the error entity detection model in the third aspect or any of the embodiments of the third aspect.

[0020] In a sixth aspect, an embodiment of the present application provides a computer program product, which, when running on a computer, causes the computer to implement the error entity detection method in the second aspect or any of the embodiments of the second aspect, or the training method of the error entity detection model in the third aspect or any of the embodiments of the third aspect.

[0021] According to the above technical solutions, the error entity detection device, method and training method of the error entity detection model provided in the embodiments of the present application can detect the error entity probability corresponding to each word segmentation in the to-be-detected text by using the error entity detection model, and label the error entity and the classification field to which the error entity belongs according to the error entity probability, to obtain the error entity labeling result and the classification field labeling result. The error entity detection model in the embodiments of the present application introduces the classification field to which the error entity belongs, so that the error entity detection model can pay more attention to the error entity itself and try to ignore the field difference, thereby alleviating the exposure bias problem and improving the generalization detection capability for error entity detection. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present application or the implementation manners in the related art, the drawings needed to be used in the embodiments or related art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0023] Figure 1 An example of the hardware and software configuration block diagram of the error entity detection device provided in the embodiments of the present application is shown;

[0024] Figure 2 An example of the flowchart of the error entity detection method provided in the embodiments of the present application is shown;

[0025] Figure 3 An example of the network structure diagram of the error entity detection model provided in the embodiments of the present application is shown;

[0026] Figure 4 An example of the flowchart of another error entity detection method provided in the embodiments of the present application is shown;

[0027] Figure 5An exemplary flowchart of a method for error entity detection is shown in FIG. 1.

[0028] Figure 6 An exemplary flowchart of another method for error entity detection is shown in FIG. 2.

[0029] Figure 7 An exemplary network structure of an error entity detection model is shown in FIG. 3.

[0030] Figure 8 An exemplary flowchart of a method for error entity detection is shown in FIG. 4.

[0031] Figure 9 An exemplary flowchart of another method for error entity detection is shown in FIG. 5.

[0032] Figure 10 An exemplary flowchart of a training method for an error entity detection model is shown in FIG. 6. DETAILED DESCRIPTION

[0033] For the purpose of making the objectives and embodiments of the present application clearer, the following will clearly and completely describe the exemplary embodiments of the present application with reference to the accompanying drawings. Obviously, the described exemplary embodiments are only a part of the embodiments of the present application, but not all of them.

[0034] It should be noted that the brief description of the terms in the present application is only for the convenience of understanding the following described embodiments, but is not intended to limit the embodiments of the present application. Unless otherwise specified, these terms should be understood according to their ordinary and general meanings.

[0035] The terms "comprising" and "having" and any variations thereof, are intended to cover but not limited to inclusive, for example, a product or device containing a series of components does not have to be limited to all the components listed clearly, but can include other components that are not listed clearly or inherent to these products or devices.

[0036] Applicant found through research that at present, error entity detection is mainly integrated in an end-to-end error correction model. When text is input into the end-to-end error correction model, the text after correction can be obtained. However, due to insufficient diversity of training samples and the need for timely updating of error-prone words in actual projects, most end-to-end error correction models use a traditional phased error correction framework, mainly a segmented structure of error detection, candidate recall, and error correction. It can be seen that correct detection of error entities is the basis for subsequent error entity recall, candidate sorting, and error entity correction. For error entity detection, traditional feature-based and rule-based methods have the characteristics of strong controllability and high accuracy, but require large and continuous maintenance investment and are difficult to have good generalization. The end-to-end error entity detection method only focuses on whether the error entity can be correctly identified, greatly relies on the guidance of sample true values during training, and has good training effect, but has poor effect during actual prediction without true value guidance. At the same time, due to the lack of attention to context semantic information, the exposure bias problem is further amplified.

[0037] To solve the above technical problems, the present application provides an error entity detection device, method and training method of error entity detection model. Next, the error entity detection device is described in detail.

[0038] In some embodiments, the error entity detection device comprises an error entity detection model, wherein the error entity detection model comprises an error entity probability detection network configured to detect an error entity probability corresponding to each word segmentation in the text to be detected; and a labeling network configured to label an error entity and a classification field to which the error entity belongs according to the error entity probability, to obtain an error entity labeling result and a classification field labeling result.

[0039] The error entity detection device can have various implementation forms, for example, it can be a smart speaker, a television, a refrigerator, a washing machine, an air conditioner, a smart curtain, a router, a set-top box, a mobile phone, a personal computer (PC) smart television, a laser projection device, a monitor, an electronic bulletin board, a wearable device, a vehicle-mounted device, an electronic table, etc. But not limited thereto.

[0040] The error entity detection model included in the error entity detection device is a trained error entity detection model.

[0041] Figure 1 An exemplary hardware and software configuration block diagram of the error entity detection device provided by the embodiments of the present application is shown. As shown in the figure, the error entity detection device comprises a processor 1, a memory 2, a communication interface 3, and a power supply 4. Figure 1As shown, the error entity detection device 100 can include an error entity detection model 110, a communicator 120, a detector 130, a display 140, a user interface 150, and a memory 160.

[0042] The error entity detection model 110 includes an error entity probability detection network and a labeling network, so that the error entity detection model can output the error entity probability corresponding to each wordpiece in the text to be detected, the error entity labeling result corresponding to the text to be detected, and the classification field labeling result.

[0043] The communicator 120 is a component for communicating with external devices or servers according to various communication protocol types. For example, the communicator 120 can include at least one of a Wifi module, a Bluetooth module, a wired Ethernet module, and other network communication protocol chips or near field communication protocol chips, and an infrared receiver. The error entity detection device 100 can establish transmission and reception of data signals with external devices through the communicator 120.

[0044] The detector 130 is used to collect signals of the external environment or interaction with the outside. For example, the detector 130 includes an image collector such as a camera, which can be used to collect images containing text, or the detector 130 includes a sound collector such as a microphone, etc., for receiving external sounds.

[0045] The sound collector can be a microphone, also known as a "microphone", "sound collector", which can be used to receive user's voice and convert the sound signal into an electrical signal. The error entity detection device 100 can be provided with at least one microphone. In some embodiments, the error entity detection device 100 can be provided with two microphones, which can not only collect sound signals, but also achieve noise reduction functions. In some embodiments, the error entity detection device 100 can also be provided with three, four or more microphones, which can collect sound signals, reduce noise, and also identify the source of the sound to achieve directional recording functions, etc.

[0046] In addition, the microphone can be built-in on the error entity detection device 100, or the microphone can be connected to the error entity detection device 100 in a wired or wireless manner. Of course, the position of the microphone on the error entity detection device 100 is not limited in the embodiments of the present application. Alternatively, the error entity detection device 100 can not include a microphone, i.e., the microphone is not provided in the error entity detection device 100. The error entity detection device 100 can externally connect a microphone (also referred to as a microphone) through an interface (such as a USB interface). The externally connected microphone can be fixed on the error entity detection device 100 by an external fixing member (such as a camera support with a clamp).

[0047] The user interface 150 can be configured to receive external data signals. For example, the user interface 150 includes a touch screen, a button, and the like, which can be configured to receive a text input by a user.

[0048] In some embodiments, the error entity probability detection network comprises: a semantic vector determination layer configured to determine a semantic vector corresponding to the text to be detected; and an error entity probability detection layer configured to determine an error entity probability corresponding to each wordpiece according to the semantic vector.

[0049] In some embodiments, the semantic vector determination layer is further configured to receive a wordpiece result and a wordpiece label result corresponding to the text to be detected; and wherein the semantic vector corresponding to the text to be detected is determined according to the wordpiece result and the wordpiece label result.

[0050] In some embodiments, the semantic vector determination layer is further configured to determine a mask vector corresponding to the text to be detected.

[0051] In some embodiments, the labeling network comprises: a fusion layer configured to fuse the semantic vector and the mask vector corresponding to the text to be detected according to the error entity probability, to obtain a fusion vector; and a labeling layer configured to label an error entity and a classification field to which the error entity belongs according to the fusion vector, to obtain an error entity labeling result and a classification field labeling result.

[0052] In some embodiments, the labeling layer comprises: a multi-head attention layer configured to determine an error detection vector according to the fusion vector; a residual connection layer configured to concatenate the fusion vector and the error detection vector to obtain a concatenated vector; an error entity labeling layer configured to label the error entity according to the concatenated vector to obtain the error entity labeling result; and a classification field labeling layer configured to label the classification field to which the error entity belongs according to the concatenated vector to obtain the classification field labeling result.

[0053] In some embodiments, the error entity detection device further comprises: a post-processing module configured to extract the error entity from the text to be detected according to the error entity labeling result.

[0054] In some embodiments, the error entity detection device further comprises: a pre-processing module configured to perform wordpiece processing on the text to be detected to obtain a wordpiece result, and perform label processing on the wordpiece result based on a knowledge graph to obtain a wordpiece label result.

[0055] In summary, the error entity detection device can detect the error entity probability corresponding to each word segmentation in the to-be-detected text by using the error entity detection model, and label the error entity and the classification field to which the error entity belongs according to the error entity probability, to obtain the error entity labeling result and the classification field labeling result. The error entity detection model in the embodiment of the present application introduces the classification field to which the error entity belongs, so that the error entity detection model can pay more attention to the error entity itself and try to ignore the field difference, thereby alleviating the exposure bias problem and improving the generalization detection capability for error entity detection. Moreover, the word segmentation result and the word segmentation labeling result can be input to the error entity detection model, so that more semantic information is contained in the information input to the error entity detection model, thereby enabling the error entity detection model to better combine the context semantics to determine the error entity probability corresponding to the word segmentation. In this way, it is beneficial to further alleviate the exposure bias problem and further improve the accuracy of the error entity detection model in detecting error entities. Moreover, the semantic vector and the mask vector corresponding to the to-be-detected text can be fused to obtain a fusion vector according to the error entity probability, and the error entity can be labeled according to the fusion vector, so that the error entity detection model can better utilize the error entity probability to capture the error entity and improve the accuracy of the error entity detection model in detecting error entities. Moreover, the error detection vector can be determined according to the fusion vector, and the fusion vector and the error detection vector are spliced to obtain a spliced vector, so that the error information representation capability of the error entity detection model is stronger. In this way, when the error entity is labeled according to the spliced vector, the accuracy of labeling the error entity can be improved. Moreover, the error entity can be extracted, so that the error entity can be separated from the to-be-detected text, thereby making the error entity more intuitive and more understandable. Moreover, the to-be-detected text can be preprocessed to obtain the word segmentation result and the word segmentation labeling result, so that the error entity detection device does not need to rely on external equipment and can complete the preprocessing by itself. Compared with preprocessing relying on external equipment, the problem of being unable to perform error entity detection due to the inability to communicate with external equipment can be avoided.

[0056] Figure 2 An exemplary flowchart of a method for detecting error entities is shown. As shown in Figure 2 The method for detecting error entities provided by the embodiment of the present application includes the following steps:

[0057] S210, detecting the error entity probability corresponding to each word segmentation in the to-be-detected text.

[0058] In the embodiment of the present application, the error entity detection device includes a trained error entity detection model, and the error entity probability detection network in the error entity detection model can detect the error entity probability corresponding to each word segmentation in the to-be-detected text.

[0059] Specifically, the text to be detected is text that needs to be detected whether it includes an error entity.

[0060] There are various specific ways to obtain the text to be detected, for example, the text to be detected can be obtained through automatic speech recognition (ASR), optical character recognition (OCR), etc. of course, the text to be detected input by the user through a touch screen, keyboard, etc. user interface. But not limited to this.

[0061] Specifically, the word segmentation is the token obtained after the word segmentation processing of the text to be detected.

[0062] For Chinese text to be detected, each character in the text to be detected can be segmented out through word segmentation processing, that is, each character is a word segmentation. For example, the text to be detected is "play network movie dream in no one", and after word segmentation processing of the text to be detected, ten word segments (i.e. word segmentation results) can be obtained, which are "broadcast", "broadcast", "network", "network", "electricity", "film", "dream", "in", "no", and "person".

[0063] The text to be detected can be processed by a machine learning model capable of word segmentation processing or manually. But not limited to this.

[0064] Specifically, the error entity probability corresponding to the word segmentation is the probability that the word segmentation is included in the error entity.

[0065] In the process of obtaining the text to be detected, due to ASR recognition error, OCR recognition error, user input error character through touch screen, keyboard, etc. reasons, it may cause the text to be detected to include error entity. For example, the text to be detected is "play network movie dream in no one", and the correct text should be "play network movie in no one", and the error entity is "dream in no one".

[0066] In some embodiments, S210 can include: S211, determining the semantic vector corresponding to the text to be detected; S212, determining the error entity probability corresponding to each word segmentation according to the semantic vector.

[0067] Figure 3 An exemplary network structure diagram of the error entity detection model provided by the embodiments of the present application is shown. As shown in Figure 3 The error entity probability detection network 310 can include a semantic vector determination layer and an error entity probability detection layer. The semantic vector determination layer can determine the semantic vector corresponding to the text to be detected, and the error entity probability detection layer can determine the error entity probability corresponding to each word segmentation according to the semantic vector.

[0068] Specifically, the semantic vector determination layer can include a BERT layer. But it is not limited thereto.

[0069] Specifically, the semantic vector is to represent the to-be-detected text by an Embedding vector.

[0070] It can be understood that by setting the error entity probability corresponding to each wordpiece according to the semantic vector, the error entity detection model can determine the error entity probability corresponding to the wordpiece in combination with the context semantics, instead of only focusing on the error entity itself. In this way, it is beneficial to alleviate the exposure bias problem and improve the accuracy of the error entity detection model in detecting error entities.

[0071] S220, according to the error entity probability, labeling the error entity and the classification field to which the error entity belongs, to obtain an error entity labeling result and a classification field labeling result.

[0072] In the embodiments of the present application, the labeling network in the error entity detection model can label the error entity and the classification field to which the error entity belongs according to the error entity probability, to obtain an error entity labeling result and a classification field labeling result.

[0073] Specifically, the error entity labeling result is a result obtained by labeling the wordpiece corresponding to the error entity (i.e., the wordpiece contained in the error entity) after error labeling.

[0074] Optionally, different labeling symbols can be used to label the wordpieces in the to-be-detected text to distinguish the wordpieces corresponding to the error entity.

[0075] For example, the wordpieces corresponding to the error entity are labeled with a first marking symbol (e.g., 1), and the wordpieces corresponding to the non-error entity are labeled with a second marking symbol (0).

[0076] Specifically, the classification field labeling result is a result obtained by labeling the classification field to which the error entity belongs.

[0077] Optionally, the classification field to which the error entity belongs can be labeled itself, or the ID of the classification field to which the error entity belongs can be labeled. But it is not limited thereto.

[0078] In some examples, as shown in Figure 3 The labeling network 320 can include an error entity labeling layer and a classification field labeling layer. The error entity labeling layer can label the error entity according to the error entity probability to obtain an error entity labeling result, and the classification field labeling layer can label the classification field to which the error entity belongs according to the error entity probability to obtain a classification field labeling result.

[0079] It can be understood that, by setting the error entity detection model to determine the error entity probability in the first stage and to perform labeling in the second stage, that is, the error entity detection model adopts a phased processing architecture, the error entity detection capability of the error entity detection model can be improved.

[0080] The error entity detection model in the embodiments of the present application can be used to detect the error entity probability corresponding to each word segmentation in the to-be-detected text, and label the error entity and the classification field to which the error entity belongs according to the error entity probability, to obtain the error entity labeling result and the classification field labeling result. The error entity detection model in the embodiments of the present application introduces the classification field to which the error entity belongs, so that the error entity detection model can pay more attention to the error entity itself and try to ignore the field difference, thereby alleviating the exposure bias problem and improving the generalization detection capability for error entity detection.

[0081] As an extension and refinement of the above-mentioned embodiments, the embodiments of the present application provide another error entity detection method, Figure 4 An exemplary flowchart of another error entity detection method provided by the embodiments of the present application is shown. As shown in Figure 4 The error entity detection method provided by the embodiments of the present application includes the following steps:

[0082] S410, receiving a word segmentation result and a word segmentation labeling result corresponding to the to-be-detected text.

[0083] In the embodiments of the present application, the semantic vector determination layer can receive the word segmentation result and the word segmentation labeling result corresponding to the to-be-detected text.

[0084] Specifically, the word segmentation labeling result is the result obtained by performing labeling processing on the word segmentation corresponding to the to-be-detected text.

[0085] The word segmentation corresponding to the to-be-detected text can be labeled by a machine learning model capable of labeling processing, a word segmentation labeling system based on a knowledge graph, or manually. But it is not limited thereto.

[0086] For example, the to-be-detected text is "play network movie dream no one", and the labeling result corresponding to the to-be-detected text is "play: actionPlay; network: videoTopic, jokeType, newsRole, broadcastType, musicTopic; movie: videoKey, videoType, showType; dream: musicFeeble; no one: musicFeeble".

[0087] S420, determining a semantic vector of the to-be-detected text according to the word segmentation result and the word segmentation labeling result.

[0088] In the embodiment of the present application, after receiving the word segmentation result and the word segmentation labeling result corresponding to the text to be detected, the semantic vector determination layer can determine the semantic vector of the text to be detected according to the word segmentation result and the word segmentation labeling result.

[0089] S430, determining an error entity probability corresponding to each word segmentation according to the semantic vector.

[0090] Specifically, S430 is similar to S212, which will not be repeated here.

[0091] S440, labeling error entities and classification fields to which the error entities belong according to the error entity probability, to obtain error entity labeling results and classification field labeling results.

[0092] Specifically, S430 is similar to S220, which will not be repeated here.

[0093] In the embodiment of the present application, the word segmentation result and the word segmentation labeling result are input into the error entity detection model, so that more semantic information is contained in the information input into the error entity detection model, thereby enabling the error entity detection model to better determine the error entity probability corresponding to the word segmentation in combination with the context semantics. In this way, it is beneficial to further alleviate the exposure bias problem and further improve the accuracy of the error entity detection model in detecting error entities.

[0094] As an extension and refinement of the above-mentioned embodiment, the present application provides another error entity detection method, Figure 5 An exemplary flowchart of another error entity detection method provided by the embodiment of the present application is shown. As shown in the figure, Figure 5 The error entity detection method provided by the embodiment of the present application includes the following steps:

[0095] S510, receiving a word segmentation result and a word segmentation labeling result corresponding to a text to be detected.

[0096] Specifically, S510 is similar to S410, which will not be repeated here.

[0097] S520, determining a semantic vector of the text to be detected according to the word segmentation result and the word segmentation labeling result.

[0098] Specifically, S520 is similar to S420, which will not be repeated here.

[0099] S530, determining an error entity probability corresponding to each word segmentation according to the semantic vector.

[0100] Specifically, S530 is similar to S430, which will not be repeated here.

[0101] S540, fusing the semantic vector and the mask vector corresponding to the text to be detected according to the error entity probability, to obtain a fusion vector.

[0102] In the embodiment of the present application, the labeling network comprises a fusion layer, which can fuse the semantic vector corresponding to the to-be-detected text and the mask vector according to the error entity probability to obtain a fusion vector.

[0103] Specifically, a preset proportion of the word segmentation is randomly selected from the to-be-detected text, the word segmentation of the preset proportion is erased from the to-be-detected text, and the erased word segmentation is predicted based on the remaining word segmentation in the to-be-detected text. The remaining word segmentation in the to-be-detected text and the predicted word segmentation are represented by an Embedding vector, that is, the mask vector can be obtained.

[0104] Optionally, the mask vector can be determined by the semantic vector determination layer according to the word segmentation result and the word segmentation labeling result.

[0105] For example, the semantic vector determination layer can comprise a BERT layer, so that the mask vector can be obtained by using the Mask mechanism in Bert. In other words, the mask vector is a Mask-Token vector preset in the BERT layer.

[0106] Specifically, the soft-mask mechanism can be used to fuse the semantic vector corresponding to the to-be-detected text and the mask vector according to the error entity probability. However, it is not limited thereto.

[0107] S550, according to the fusion vector, labeling the error entity and the classification field to which the error entity belongs to obtain the error entity labeling result and the classification field labeling result.

[0108] Specifically, S550 is similar to S440, and will not be described here.

[0109] The embodiment of the present application can fuse the semantic vector corresponding to the to-be-detected text and the mask vector according to the error entity probability to obtain a fusion vector, and label the error entity according to the fusion vector, so that the error entity detection model can better utilize the error entity probability to capture the error entity, and improve the accuracy of the error entity detection model in detecting the error entity.

[0110] As an extension and refinement of the above-mentioned embodiment, the embodiment of the present application provides another error entity detection method, Figure 6 An exemplary flowchart of another error entity detection method provided by the embodiment of the present application is shown. As shown in Figure 6 The error entity detection method provided by the embodiment of the present application comprises the following steps:

[0111] S610, receiving a word segmentation result and a word segmentation labeling result corresponding to a to-be-detected text.

[0112] Specifically, S610 is similar to S510, and will not be described here.

[0113] S620, determining a semantic vector of the to-be-detected text according to the word segmentation result and the word segmentation tagging result.

[0114] Specifically, S620 is similar to S520, and details are not repeated here.

[0115] Figure 7 An exemplary network structure diagram of the error entity detection model provided by the embodiment of the application is shown. As shown in the figure, Figure 7 The error entity detection model includes a semantic vector determination layer 710, an error entity probability detection layer 720, a fusion layer 730, a multi-head attention layer 740, a residual connection layer 750, an error entity labeling layer 760, and a classification field labeling layer 770.

[0116] The word segmentation result and the word segmentation tagging result of the to-be-detected text can be organized as a serialized text input into the semantic vector determination layer 710, so that the error entity detection model uses the semantic vector determination layer 710 to realize semantic vector expression of the to-be-detected text, and obtains a semantic vector of the to-be-detected text, i.e., semantic encoding vector information, thereby improving semantic expression capability.

[0117] S630, determining an error entity probability corresponding to each word segmentation according to the semantic vector.

[0118] Specifically, S630 is similar to S530, and details are not repeated here.

[0119] Continuing to refer to Figure 7 The semantic vector is input into the error entity probability detection layer 720, which can perform error detection on each word segmentation (i.e., token) in the to-be-detected file, thereby outputting a probability of an error entity corresponding to each word segmentation (i.e., error entity probability).

[0120] S640, fusing the semantic vector and the mask vector corresponding to the to-be-detected text according to the error entity probability, to obtain a fusion vector.

[0121] Specifically, S640 is similar to S540, and details are not repeated here.

[0122] Continuing to refer to Figure 7 The semantic vector, the mask vector, and the error entity probability are input into the fusion layer 730, which can fuse the semantic vector (i.e., semantic expression vector information) and the mask vector (i.e., Mask-Token vector information preset in Bert) based on the error entity probability, using a soft-mask mechanism, to obtain a fusion vector (i.e., a semantic feature vector containing mask information).

[0123] S650, determining an error detection vector according to the fusion vector.

[0124] In the embodiment of the present application, the marking layer can further include a multi-head self-attention layer, and the multi-head self-attention layer can determine the error detection vector according to the fusion vector.

[0125] Specifically, the multi-head attention is to establish different projection information in multiple different projection spaces. After the input matrix is projected differently to obtain many output matrices, the output matrices are spliced together. In this way, each attention mechanism can optimize different feature parts of each segmented word, thereby balancing the possible deviation of the same attention mechanism, allowing the word meaning to have more diverse expressions, thereby improving the effect of the error entity detection model.

[0126] S660, splicing the fusion vector and the error detection vector to obtain a spliced vector.

[0127] In the embodiment of the present application, the marking layer can further include a residual connection module, and the residual connection module can splice the fusion vector and the error detection vector to obtain a spliced vector.

[0128] Optionally, the splicing process is to add the fusion vector and the error detection vector.

[0129] Continuing to refer to Figure 7 The fusion vector is input into the multi-head attention layer 740, the multi-head attention layer 740 can obtain multiple error detection vectors by using the multi-head attention mechanism, the multiple error detection vectors are input into the residual connection layer 750, and the residual connection layer 750 can connect the multiple error detection vectors with the fusion vector to obtain a spliced vector (i.e., a model error detection expression vector). By using the multi-head attention mechanism and the residual connection, the error information representation capability of the error entity detection model can be improved.

[0130] S670, marking the error entity according to the spliced vector to obtain an error entity marking result, and marking the classification field to which the error entity belongs according to the spliced vector to obtain a classification field marking result.

[0131] Continuing to refer to Figure 7 The spliced vector is taken as an input and transmitted into the error entity marking layer 760 and the classification field marking layer 770 respectively to obtain the error entity marking result and the classification field marking result. By using the classification field marking result, the exposure bias problem of the entity detection model can be alleviated.

[0132] In the embodiment of the present application, the error detection vector can be determined according to the fusion vector, and the fusion vector and the error detection vector can be spliced to obtain a spliced vector, so that the error information representation capability of the error entity detection model is stronger. In this way, when the error entity is marked according to the spliced vector, the accuracy of marking the error entity can be improved.

[0133] As an extension and refinement of the above-mentioned embodiments, the embodiments of the present application provide a method for detecting error entities, Figure 8 An exemplary flow diagram of a method for detecting error entities provided by the embodiments of the present application is shown in FIG. 8. As shown in FIG. 8, the method for detecting error entities provided by the embodiments of the present application includes the following steps: Figure 8

[0134] S810, detecting an error entity probability corresponding to each word segmentation in the text to be detected.

[0135] Specifically, S810 is similar to S210, and thus is not described here again.

[0136] S820, labeling error entities and classification fields to which the error entities belong according to the error entity probability, to obtain error entity labeling results and classification field labeling results.

[0137] Specifically, S810 is similar to S220, and thus is not described here again.

[0138] S830, extracting error entities from the text to be detected according to the error entity labeling results.

[0139] In the embodiments of the present application, the error entity detection device can further include a post-processing module. The post-processing module can receive error entity labeling results output by the error entity detection model, so as to extract error entities from the text to be detected according to the error entity labeling results.

[0140] Of course, when the classification field result includes an ID of a classification field, the post-processing module can also determine the classification field to which the error entity belongs according to the ID.

[0141] The embodiments of the present application can extract error entities. In this way, the error entities can be separated from the text to be detected, so that the error entities are more intuitive and more understandable.

[0142] As an extension and refinement of the above-mentioned embodiments, the embodiments of the present application provide another method for detecting error entities, Figure 9 An exemplary flow diagram of another method for detecting error entities provided by the embodiments of the present application is shown in FIG. 9. As shown in FIG. 9, the method for detecting error entities provided by the embodiments of the present application includes the following steps: Figure 9

[0143] S910, performing word segmentation processing on the text to be detected to obtain word segmentation results, and performing tagging processing on the word segmentation results based on a knowledge graph to obtain word segmentation tagging results.

[0144] ​​In the embodiment of the present application, the error entity detection device can further comprise a preprocessing module. When receiving the text to be detected, the preprocessing module can perform word segmentation processing on the text to be detected to obtain a word segmentation result, and perform tagging processing on the word segmentation result based on the knowledge graph to obtain a word segmentation tagging result, so as to input the word segmentation result and the word segmentation tagging result into the error entity detection model.

[0145] Specifically, the knowledge graph is essentially a semantic network and can represent semantic relationships between entities. In the knowledge graph, entities are vertices or nodes, and relationships are edges. The knowledge graph can be constructed in various ways. The focus of the embodiment of the present application is not how to construct the knowledge graph, so it will not be described in detail.

[0146] S920, detecting the error entity probability corresponding to each word segmentation in the text to be detected according to the word segmentation result and the word segmentation tagging result.

[0147] S930, labeling the error entity and the classification field to which the error entity belongs according to the error entity probability, to obtain an error entity labeling result and a classification field labeling result.

[0148] For example, the text to be detected is "play network movie in the dream", the preprocessing module can obtain the word segmentation result and the word segmentation tagging result by inputting the text to be detected into the preprocessing module, and the error entity detection model can output the error entity probability corresponding to each word segmentation in the text to be detected, the error entity labeling result and the classification field labeling result corresponding to the text to be detected by inputting the word segmentation result and the word segmentation tagging result into the error entity detection model. The post-processing module can extract the detected error entity in the text to be detected by using the error entity labeling result. The error entity detection of the text to be detected is shown in Table 1:

[0149] Table 1

[0150]

[0151] The embodiment of the present application can preprocess the text to be detected to obtain the word segmentation result and the word segmentation tagging result, so that the error entity detection device does not need to rely on external equipment and can complete the preprocessing by itself. Compared with relying on external equipment for preprocessing, the problem of being unable to perform error entity detection due to being unable to communicate with external equipment can be avoided.

[0152] In some embodiments, the embodiment of the present application further provides a training method of an error entity detection model, which is applied to a model training device.

[0153] The model training device and the error entity detection device can be the same device, that is, after the training of the error entity detection model to be trained in the model training device is completed, the model training device becomes the error entity detection device.

[0154] The model training device and the error entity detection device can also be different devices, so that when the error entity detection model to be trained in the model training device completes training, the trained error entity detection model is deployed in other electronic devices to obtain the error entity detection device. At this time, the model training device can be a terminal device, such as a mobile phone, a computer, a personal computer, a smart television, etc. The model training device can also be a server. The server can be a server that provides various services. For example, the server can be a server cluster, or multiple server clusters, and can include one or more types of servers. The model training device can be connected to other electronic devices through a local area network (LAN), a wireless local area network (WLAN), etc. to send the trained error entity detection model to other electronic devices, so that the trained error entity detection model is deployed in the electronic device. Of course, the trained error entity detection model can also be stored in a U disk or other storage device first, and then read by other electronic devices from the U disk or other storage device, so that the trained error entity detection model is deployed in the electronic device. This is not limited.

[0155] Figure 10 An exemplary flowchart of a training method of an error entity detection model provided by an embodiment of the present application is shown. As shown in Figure 10 The training method of the error entity detection model provided by the embodiment of the present application includes the following steps:

[0156] S1010, obtaining a training sample, wherein the training sample includes a sample text, a real error entity probability corresponding to each word segmentation in the sample text, and a real error entity annotation result and a real classification field annotation result corresponding to the sample text.

[0157] Specifically, the sample text is a text containing an error entity.

[0158] There are various specific ways to obtain the sample text. For example, the text to be detected can be obtained by ASR, OCR, etc. Of course, the text to be detected can also be received by a user interface such as a touch screen or a keyboard. The sample text can also be extracted from the historical log information of a voice assistant. However, it is not limited thereto.

[0159] For each sample text, the sample text is subjected to word segmentation processing to obtain a word segmentation result, the real error entity probability corresponding to each word segmentation in the sample text is determined, the word segmentation corresponding to the error entity in the sample text is labeled to obtain a real error entity labeling result corresponding to the sample text, and the classification field to which the error entity belongs is labeled to obtain a real classification field labeling result corresponding to the sample text. Optionally, the word segmentation result corresponding to the sample text can also be labeled based on a knowledge graph to obtain a word segmentation labeling result. In this way, the training sample further includes the word segmentation labeling result.

[0160] In S1020, the training sample is input into the error entity detection model to be trained to obtain the error entity probability corresponding to each word segmentation in the sample text, the error entity labeling result and the classification field labeling result output by the error entity detection model to be trained.

[0161] Optionally, the error entity detection model to be trained can include an error entity probability detection network configured to detect the error entity probability corresponding to each word segmentation in the text to be detected; and a labeling network configured to label the error entity and the classification field to which the error entity belongs according to the error entity probability, to obtain the error entity labeling result and the classification field labeling result.

[0162] Optionally, the error entity probability detection network includes a semantic vector determination layer configured to determine the semantic vector corresponding to the text to be detected; and an error entity probability detection layer configured to determine the error entity probability corresponding to each word segmentation according to the semantic vector.

[0163] Optionally, the semantic vector determination layer is further configured to receive the word segmentation result and the word segmentation labeling result corresponding to the text to be detected; and wherein determining the semantic vector corresponding to the text to be detected includes determining the semantic vector of the text to be detected according to the word segmentation result and the word segmentation labeling result.

[0164] Optionally, the semantic vector determination layer is further configured to determine a mask vector corresponding to the text to be detected.

[0165] Optionally, the labeling network includes a fusion layer configured to fuse the semantic vector and the mask vector corresponding to the text to be detected according to the error entity probability, to obtain a fusion vector; and a labeling layer configured to label the error entity and the classification field to which the error entity belongs according to the fusion vector, to obtain the error entity labeling result and the classification field labeling result.

[0166] Optionally, the label layer comprises: a multi-head attention layer configured to determine an error detection vector according to the fusion vector; a residual connection layer configured to splice the fusion vector and the error detection vector to obtain a spliced vector; an error entity label layer configured to label an error entity according to the spliced vector to obtain an error entity label result; and a classification field label layer configured to label a classification field to which the error entity belongs according to the spliced vector to obtain a classification field label result.

[0167] S1030, determining an error entity probability loss function according to the real error entity probability and the error entity probability output by the to-be-trained error entity detection model, determining an error entity label loss function according to the real error entity label result and the error entity label result output by the to-be-trained error entity detection model, and determining a classification field label loss function according to the real classification field label result and the classification field label result output by the to-be-trained error entity detection model.

[0168] Specifically, the error entity probability loss function is used to represent the accuracy of the probability that each word in the sample text is contained in the error entity.

[0169]

[0170] wherein L d is the error entity probability loss function, e i is a word contained in the real error entity, P d (g i =e i |X) is the probability of the word contained in the error entity predicted by the to-be-trained error entity detection model.

[0171] Specifically, the error entity label loss function is used to represent the accuracy of the label of the error entity in the sample text.

[0172]

[0173] wherein L s is the error entity label loss function, t i is the real error entity label result, P s (l i =t i |X) is the probability of the correct error entity label predicted by the to-be-trained error entity detection model.

[0174] Specifically, the classification field label loss function is used to represent the accuracy of the classification field to which the error entity belongs.

[0175]

[0176] wherein L cTo classify the field of loss function, o i To the real error entity classification field annotation, P c (y i = o i |X) is the probability of the correct classification field predicted by the error entity detection model to be trained.

[0177] S1040, according to the error entity probability loss function, the error entity annotation loss function, and the classification field annotation loss function, determine the model loss function.

[0178] Specifically, the error entity probability loss function, the error entity annotation loss function, and the classification field annotation loss function are weighted and added to obtain the model loss function.

[0179] For example, the model loss function is calculated by the following formula:

[0180] Loss(f) = γ1·L d + γ2·L s + γ3·L c

[0181] Where L d is the error entity probability loss function, L s is the error entity annotation loss function, L c is the classification field annotation loss function, γ1, γ2, γ3 is the manually labeled hyperparameter, that is, L d , L s , L c corresponding weight respectively.

[0182] S1050, according to the model loss function, adjust the parameters in the error entity detection model to be trained, until the model loss function meets the preset training end condition.

[0183] Specifically, the error entity detection model is trained iteratively until the model loss function meets the preset training end condition (i.e. reaches the convergence condition). At this time, the error entity detection model to be trained becomes the trained error entity detection model.

[0184] In order to explain the training process of the error entity detection model in detail, a detailed example is given below.

[0185] Based on the historical log information of the voice assistant, the training samples required for constructing the error entity detection model are extracted, each training sample includes sample text, sample text corresponding to the segmentation and labeling result, real error entity probability corresponding to each segmentation in the sample text, and real error entity annotation result and real classification field annotation result corresponding to the sample text. The relevant information of three training samples is shown in Table 2.

[0186] Table 2

[0187]

[0188]

[0189] The obtained training sample is input into the error entity detection model to obtain the error entity probability corresponding to each segmented word in the sample text output by the error entity detection model to be trained, the error entity annotation result, and the classification field annotation result, the error entity probability loss function, the error entity annotation loss function, and the classification field annotation loss function are calculated respectively, so as to determine the model loss function, and the error entity detection model to be trained is iteratively trained until the model loss function reaches the convergence condition.

[0190] The training method of the error entity detection model provided by the embodiments of the present application can obtain training samples, wherein the training samples include a sample text, a real error entity probability corresponding to each segmented word in the sample text, and real error entity annotation results and real classification field annotation results corresponding to the sample text. The training samples are input into the error entity detection model to be trained to obtain the error entity probability corresponding to each segmented word in the sample text output by the error entity detection model to be trained, the error entity annotation result, and the classification field annotation result. Then, the error entity probability loss function is determined according to the real error entity probability and the error entity probability output by the error entity detection model to be trained, the error entity annotation loss function is determined according to the real error entity annotation result and the error entity annotation result output by the error entity detection model to be trained, and the classification field annotation loss function is determined according to the real classification field annotation result and the classification field annotation result output by the error entity detection model to be trained. The model loss function is determined according to the error entity probability loss function, the error entity annotation loss function, and the classification field annotation loss function. Then, the parameters in the error entity detection model to be trained are adjusted according to the model loss function until the model loss function meets the preset training end condition. It can be seen that the training method of the error entity detection model introduces the error entity probability loss function, the error entity annotation loss function, and the classification field annotation loss function to construct the model loss function of the error entity detection model. Combined with the training strategy of multiple loss functions, the exposure bias problem caused in the training process can be alleviated, and the semantic understanding ability and generalization detection ability of the error entity detection model are improved.

[0191] In some embodiments, the electronic device further includes a memory configured to store a computer program; and a processor configured to, when executing the computer program, cause the electronic device to implement the error entity detection method or the training method of the error entity detection model of any of the above embodiments.

[0192] In some embodiments, the electronic device further includes a memory configured to store a computer program; and a processor configured to, when executing the computer program, cause the electronic device to implement the error entity detection method or the training method of the error entity detection model of any of the above embodiments.

[0193] In some embodiments, the electronic device further includes a memory configured to store a computer program; and a processor configured to, when executing the computer program, cause the electronic device to implement the error entity detection method or the training method of the error entity detection model of any of the above embodiments.

[0194] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

[0195] In order to facilitate explanation, the above description has been made in combination with specific embodiments. However, the above exemplary discussion is not intended to exhaust or limit the embodiments to the specific forms disclosed above. Various modifications and variations can be derived according to the above teachings. The selection and description of the above embodiments are to better explain the principles and practical applications, so that those skilled in the art can better use the embodiments and various different modified embodiments suitable for specific use considerations.

Claims

1. An error entity detection apparatus characterized by comprising: The error entity detection model comprises an error entity probability detection network configured to detect an error entity probability corresponding to each wordpiece in the text to be detected. The error entity probability detection network is configured to detect an error entity probability corresponding to each wordpiece in the text to be detected. The labeling network is configured to label an error entity and a classification field to which the error entity belongs according to the error entity probability, to obtain an error entity labeling result and a classification field labeling result. The labeling network comprises a fusion layer configured to fuse a semantic vector and a mask vector corresponding to the text to be detected according to the error entity probability, to obtain a fusion vector. The labeling layer is configured to label the error entity and the classification field to which the error entity belongs according to the fusion vector, to obtain the error entity labeling result and the classification field labeling result. The labeling layer comprises a multi-head attention layer configured to determine an error detection vector according to the fusion vector. The residual connection layer is configured to splice the fusion vector and the error detection vector, to obtain a spliced vector. The error entity labeling layer is configured to label an error entity according to the spliced vector, to obtain the error entity labeling result. The classification field labeling layer is configured to label a classification field to which the error entity belongs according to the spliced vector, to obtain the classification field labeling result. The error entity probability detection network comprises a semantic vector determination layer configured to determine a semantic vector corresponding to the text to be detected.

2. The error entity detection device according to claim 1, characterized in that, The error entity probability detection layer is configured to determine an error entity probability corresponding to each wordpiece according to the semantic vector. The semantic vector determination layer is further configured to receive a wordpiece result and a wordpiece labeling result corresponding to the text to be detected. The semantic vector determination layer is further configured to determine a mask vector corresponding to the text to be detected.

3. The error entity detection device according to claim 2, characterized in that, Further comprising a post-processing module configured to extract the error entity from the text to be detected according to the error entity labeling result. Further comprising a preprocessing module configured to perform wordpiece processing on the text to be detected to obtain a wordpiece result, and to perform labeling processing on the wordpiece result based on a knowledge graph to obtain a wordpiece labeling result.

4. The error entity detection device according to claim 2, characterized by, The method comprises detecting an error entity probability corresponding to each wordpiece in the text to be detected.

5. The error entity detection device of claim 1, wherein, The error entity probability is used to label an error entity and a classification field to which the error entity belongs. The error entity probability is used to label an error entity and a classification field to which the error entity belongs.

6. The error entity detection device according to claim 3, characterized by The error entity probability is used to label an error entity and a classification field to which the error entity belongs. ​ 7. A method of false entity detection, the method comprising: ​ ​ ​ ​ ​ ​ The error entity and the classification field to which the error entity belongs are labeled according to the fusion vector, to obtain error entity labeling results and classification field labeling results, and the method comprises the following steps. The fusion vector and the error detection vector are spliced to obtain a spliced vector. The error entity is labeled according to the spliced vector to obtain the error entity labeling result. The classification field to which the error entity belongs is labeled according to the spliced vector to obtain the classification field labeling result.

8. A training method for an error entity detection model, applied to a model training device, characterized in that, The method comprises: Obtaining a training sample, wherein the training sample comprises a sample text, a real error entity probability corresponding to each word segmentation in the sample text, and real error entity labeling results and real classification field labeling results corresponding to the sample text; The training sample is input into a to-be-trained error entity detection model to obtain error entity probabilities corresponding to each word segmentation in the sample text, error entity labeling results and classification field labeling results output by the to-be-trained error entity detection model; An error entity probability loss function is determined according to the real error entity probability and the error entity probability output by the to-be-trained error entity detection model, an error entity labeling loss function is determined according to the real error entity labeling result and the error entity labeling result output by the to-be-trained error entity detection model, and a classification field labeling loss function is determined according to the real classification field labeling result and the classification field labeling result output by the to-be-trained error entity detection model; A model loss function is determined according to the error entity probability loss function, the error entity labeling loss function and the classification field labeling loss function; The parameters in the to-be-trained error entity detection model are adjusted according to the model loss function until the model loss function meets a preset training end condition; The to-be-trained error entity detection model detects error entity probabilities corresponding to each word segmentation in the sample text. Error entities and classification fields to which the error entities belong are labeled according to the error entity probabilities. The error entities and the classification fields to which the error entities belong are labeled according to the error entity probabilities, to obtain error entity labeling results and classification field labeling results, and the method comprises the following steps. A fusion vector is obtained by fusing a semantic vector and a mask vector corresponding to the sample text according to the error entity probability. The error entity and the classification field to which the error entity belongs are labeled according to the fusion vector, to obtain the error entity labeling result and the classification field labeling result. The error entity and the classification field to which the error entity belongs are labeled according to the fusion vector, to obtain the error entity labeling result and the classification field labeling result, and the method comprises the following steps. The fusion vector and the error detection vector are spliced to obtain a spliced vector. The error entity is labeled according to the spliced vector to obtain the error entity labeling result. The classification field to which the error entity belongs is labeled according to the spliced vector to obtain the classification field labeling result.

Images