Text error correction method, text error correction device, computer program product, training method, and asr method

Abstract

To improve the text error correction effect by avoiding a situation where the more errors are corrected, the more errors increase when correcting errors in text.SOLUTION: A text error correction method is provided, comprising causing a first neural network model to: generate an error-corrected text of a processing target text and an error correction label indicative of whether the error-corrected text is better than the processing target text on the basis of the processing target text; and select and output either the error-corrected text or the processing target text as a correction result on the basis of the error correction label.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to text processing, and more particularly to a text error correction method, a text error correction device, a computer program product, a non-transitory computer-readable recording medium, a method for Automatic Speech Recognition (ASR) utilizing the text error correction method, and a method for training a neural network model for use in text error correction. [Background technology]

[0002] In the past, it was necessary to spend a lot of time and effort to find and correct errors in text, such as grammatical errors and spelling mistakes.With the development of artificial intelligence technology, text error correction tools have emerged that can automatically identify and correct errors in text by learning from large amounts of language data.

[0003] Advances in data processing technology and the rapid spread of the mobile internet have resulted in the generation of vast amounts of data. ASR technology is a type of automatic speech recognition technology based on artificial intelligence and speech signal processing technology, and is used to convert human speech into text. Automatic speech recognition technology supports many services closely related to our daily lives, such as the speech-to-text conversion function on mobile phones and the automatic generation of subtitles on video sites. Summary of the Invention

[0004] The present disclosure relates to a text error correction method, a text error correction device, a computer program product, a non-transitory computer-readable storage medium, and a method for automatic speech recognition. An indication is provided as to whether the error-corrected text is an improvement over the target text, and either the error-corrected text or the target text can be selected and output as a correction result, thereby avoiding a situation where the more errors corrected, the more errors are increased. A method for training a neural network model for text error correction is also proposed, which can generate error-corrected text and an indication as to whether the error-corrected text is an improvement over the target text.

[0005] According to one aspect of the present disclosure, there is provided a text error correction method, the method including: a step of using a first neural network model to generate, based on a target text, an error-corrected text of the target text and an error correction label indicating whether the error-corrected text is an improvement over the target text; and a step of selecting and outputting, as a result of the correction, the error-corrected text or the target text based on the error correction label.

[0006] According to an embodiment of the present disclosure, the first neural network model is trained such that the second neural network model obtains a group of output error-corrected texts based on a group of predetermined input texts corresponding to a group of predetermined reference texts, calculates a training error correction label for each output error-corrected text in the group of output error-corrected texts based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts to obtain a group of training error correction labels, takes the group of predetermined input texts as inputs of the first neural network model, and takes the group of predetermined reference texts or the group of output error-corrected texts and the group of training error correction labels as outputs of the first neural network model.

[0007] According to an embodiment of the present disclosure, calculating a training error correction label for each output error-corrected text of the group of output error-corrected texts based on a group of output error-corrected texts, a group of predetermined input texts, and a group of predetermined reference texts to obtain a group of training error correction labels includes: obtaining a first predetermined input text of the group of predetermined input texts, a first output error-corrected text corresponding to the first predetermined input text, and a first predetermined reference text; calculating a first character error rate (CER) between the first predetermined input text and the first predetermined reference text, and a second CER between the first output error-corrected text and the first predetermined reference text; comparing the first CER and the second CER to obtain a first training error correction label corresponding to the first output error-corrected text.

[0008] According to one embodiment of the present disclosure, the step of comparing the first CER with the second CER to obtain a first training error correction label corresponding to the first output error-corrected text includes the steps of: setting the first training error correction label to indicate that the first output error-corrected text has been improved for the first predetermined input text in response to the first CER being greater than the second CER; and setting the first training error correction label to indicate that the first output error-corrected text has not been improved for the first predetermined input text in response to the first CER being less than or equal to the second CER.

[0009] According to one embodiment of the present disclosure, the first neural network model and the second neural network model are implemented in the same neural network.

[0010] According to one embodiment of the present disclosure, the first neural network model includes an autoregressive model (AR model).

[0011] According to another aspect of the present disclosure, there is provided a text error correction apparatus, the apparatus including a processor and a memory having one or more computer programs stored therein, the one or more computer programs, when executed by the processor, causing the processor to perform any of the above-described methods of text error correction.

[0012] According to another aspect of the present disclosure, there is provided a computer program product including a computer program which, when executed by a processor, causes the processor to perform any of the above-mentioned text error correction methods.

[0013] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, cause the processor to perform any of the text error correction methods described above.

[0014] According to another aspect of the present disclosure, there is provided a method for automatic speech recognition (ASR), comprising: an ASR model converting speech into a target text; a first neural network model generating, based on the target text, an error-corrected text of the target text and an error correction label indicating whether the error-corrected text is an improvement over the target text; and selecting and outputting, as a result of the ASR, either the error-corrected text or the target text based on the error correction label.

[0015] According to another aspect of the present disclosure, there is provided a method for training a first neural network model, the method including: obtaining, by a second neural network model, a group of output error-corrected texts based on a group of predetermined input texts corresponding to a group of predetermined reference texts; calculating, based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts, a training error correction label for each output error-corrected text of the group of output error-corrected texts to obtain a group of training error correction labels; and training the first neural network model so that the group of predetermined input texts is an input of the first neural network model, and the group of predetermined reference texts or the group of output error-corrected texts and the group of training error correction labels are outputs of the first neural network model.

[0016] According to an embodiment of the present disclosure, whether the error-corrected text is an improvement over the text to be processed is indicated, and the error-corrected text or the text to be processed can be selected and output as the result of the correction, thereby avoiding a situation where the more errors are corrected, the more errors are generated, and improving the ability of text processing.

[0017] The forms, features and advantages of the present disclosure will be more clearly and readily understood from the following description of examples taken in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 shows a schematic diagram of the overall technical concept of the present disclosure. [Figure 2] FIG. 2 shows a flowchart of a text error correction method according to an embodiment of the present disclosure. [Figure 3] FIG. 3 illustrates a method for training a first neural network model according to an embodiment of the present disclosure. [Figure 4]FIG. 4 shows details of the method for training shown in FIG. [Figure 5] FIG. 5 shows a schematic diagram of a text error correction device according to an embodiment of the present disclosure. [Figure 6] FIG. 6 shows a flowchart of a method for automatic speech recognition according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0019] Hereinafter, embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although several embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure may be implemented in various forms and should not be construed as being limited to the embodiments described herein, but rather, these embodiments are provided to enable a more thorough and complete understanding of the present disclosure. It should be understood that the accompanying drawings and embodiments of the present disclosure are merely for illustrative purposes and are not intended to limit the scope of protection of the present disclosure.

[0020] It should be understood that the steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel, and other steps may be included and / or some steps may be omitted in the method embodiments.

[0021] As used herein, the term "comprises" and variations thereof are open-ended, i.e., "including, but not limited to." The term "based on" means "based at least in part on." The term "in one embodiment" means "at least one embodiment," the term "another embodiment" means "at least one other embodiment," and "some embodiments" means "at least some embodiments." Definitions of other terms are provided in the description below.

[0022] It should be understood that the concepts of "first", "second", etc. referred to in the present disclosure are merely intended to distinguish different devices, modules or units, and are not intended to limit the order of functions performed by these devices, modules or units or their dependencies on each other.

[0023] It should be noted that the modifications "a" and "a plurality" referred to in this disclosure are exemplary rather than limiting, and it should be understood by those skilled in the art that they should be understood as "one or more" unless the context clearly dictates otherwise.

[0024] There are many scenarios requiring text error correction. For example, when conducting a search, a user may enter a search request containing spelling mistakes, incorrect word order, grammatical errors, and other errors. Using an error correction model, the search request can be automatically corrected and presented to the user, providing search results that better meet the user's needs. However, depending on the size of the dataset used to train the error correction model and the field of the training data, the error correction results output by the error correction model may not always be satisfactory, and the more errors corrected, the more errors may increase. As another example, in the field of automatic speech recognition (ASR), existing ASR models can receive speech signals as input and convert them into text to output. However, the ASR conversion results still contain errors that are obvious to humans, so an error correction model for the ASR model is needed to correct the recognized erroneous text. ASR models are typically obtained by training using large amounts of speech and text data and are relatively stable in terms of generality. However, the amount of data used to train an ASR error correction model is relatively small. Therefore, in practice, the more errors an ASR error correction model corrects, the more errors it may generate. For example, the text output by the ASR model may be correct, but after being corrected by the ASR error correction model, the correct text may be incorrectly corrected; or the text output by the ASR model may contain few errors, but after being corrected by the ASR error correction model, the number of errors in the text may increase.

[0025] The present disclosure has been made in light of the above-mentioned problems. A text error correction method, a text error correction device, a computer program product, a non-transitory computer-readable storage medium, and a method for automatic speech recognition according to embodiments of the present disclosure indicate whether the error-corrected text (i.e., the text after error correction) is an improvement over the target text (i.e., the text before error correction), and can select and output the error-corrected text or the target text as a correction result, thereby avoiding a situation where the more errors corrected, the more errors are increased. A method for training a text error correction neural network model is also proposed, which can generate error-corrected text and an indication of whether the error-corrected text is an improvement over the target text.

[0026] The idea of the present disclosure is to obtain a label indicating whether the error-corrected text is an improvement over the target text, and then determine whether to correct the errors in the target text before outputting it, or outputting the target text as is, thereby avoiding a situation where the more errors are corrected, the more errors are increased. Figure 1 schematically illustrates the overall technical idea of the present disclosure. The invention of the present disclosure relies on artificial intelligence, specifically, various neural network models. The neural network model involves two stages: a training stage and a data generation stage. Figure 1(a) illustrates the training stage, and Figure 1(b) illustrates the data generation stage.

[0027] As shown in FIG. 1(a), in the training phase, a first neural network model is trained using a group of known or annotated data pairs (predetermined input text, predetermined reference text). For example, the predetermined input text may correspond to a recognized text output by an ASR model as a result of speech recognition, and the predetermined reference text may correspond to a correct text corresponding to the recognized text, i.e., a correct text of the speech corresponding to the recognized text. The predetermined reference text may be obtained by manual annotation. As shown in FIG. 1(a), the training data for the first neural network model may include a group of training error correction labels and a group of output error-corrected texts. The group of output error-corrected texts is generated by a second neural network model based on the group of predetermined input texts. The second neural network model may be, for example, an ASR error correction model, which performs error correction on the group of recognized texts (i.e., predetermined input texts) output by the ASR model and generates a corresponding group of output error-corrected texts. The group of output error-corrected texts and the group of predetermined reference texts respectively correspond to the group of predetermined input texts, where the group of predetermined input texts corresponds to the text before error correction, and the group of output error-corrected texts corresponds to the text after error correction. After obtaining the group of output error-corrected texts, the group of training error-correction labels may be calculated based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts. FIG. 1(a) schematically shows that the calculation is performed by a calculation unit. It should be understood that the calculation unit can be implemented using any computer, processor, central processing unit (CPU), microcontroller unit (MCU), etc.Each training error correction label may indicate whether an output error-corrected text (i.e., the text after error correction) is an improvement over a corresponding predetermined input text (i.e., the text before error correction). A first neural network model may then be trained using the group of predetermined input texts as its input and the group of predetermined reference texts and the group of training error correction labels as its output. Alternatively, a first neural network model may be trained using the group of predetermined input texts as its input and the group of output error-corrected texts and the group of training error correction labels as its output. The first and second neural network models may be implemented using various existing generative models, including, for example, autoregressive and non-autoregressive models (e.g., transformers).

[0028] As shown in FIG. 1(b), in the data generation stage, a first neural network model is used to correct errors in a target text. The target text may be, for example, a recognized text output by an ASR model as a result of speech recognition, and errors may exist in the text. For example, the target text may be a single sentence, and the sentence may contain errors such as extra characters, omissions, typos, and homonyms. The trained first neural network model may generate corresponding error-corrected text and an error correction label associated with the error-corrected text based on the target text. The error correction label indicates whether the error-corrected text is an improvement over the target text. An output unit may select and output the error-corrected text or the target text, i.e., the text after error correction or the text before error correction, as a correction result for the target text based on the error correction label. The output unit mentioned here may be implemented using a general I / O module.

[0029] If the error correction label indicates that the error-corrected text is not an improvement over the target text, this indicates that the model's error correction effect is poor. Therefore, instead of using the error-corrected text, the original text (i.e., the target text) without error correction is output as is. This avoids a situation in which the more errors are corrected, the more errors are increased, thereby improving the effectiveness of text error correction. Furthermore, the method proposed in this disclosure reduces the need to annotate more data pairs (predetermined input text, predetermined reference text) to train the error correction model and improve the error correction model's algorithm, thereby reducing the consumption of human resources. Furthermore, reducing the complexity of the error correction model's algorithm also reduces the demands on processor performance and memory space.

[0030] 2 shows a flowchart of a text error correction method 200 according to an embodiment of the present disclosure. The method 200 may correspond to the data generation phase described above. As shown in FIG. 2, the method 200 according to an embodiment of the present disclosure may include steps S210 to S220.

[0031] In step S210, the first neural network model may generate error-corrected text and an error correction label for the target text based on the target text. The first neural network model may generate corresponding error-corrected text for each target text, thereby correcting errors in the target text. At the same time, the first neural network model may generate the error-corrected text and an error correction label corresponding to the target text. The error correction label may indicate the effect of the error correction in the error-corrected text, i.e., whether the error-corrected text has improved relative to the target text. Such improvement may include, for example, a reduction in typos or extraneous characters in the text. The error correction label may take the form of a number, such as "0" or "1," where "0" indicates that the error-corrected text has not improved relative to the target text, and "1" indicates that the error-corrected text has improved relative to the target text. It should be understood that the error correction label may be other numerical values indicating improvement or may have other forms, as long as it indicates whether the error-corrected text is an improvement over the processed text.

[0032] In step S220, based on the error correction label generated in step S210, the error-corrected text generated in step S210 or the target text is selected and output as the correction result. If the error correction label indicates that the error-corrected text is an improvement over the target text, the error-corrected text may be adopted as the correction result, i.e., error correction is performed on the target text. If the error correction label indicates that the error-corrected text is not an improvement over the target text, the target text may be used as the correction result, i.e., no error correction is performed on the target text. This ensures that the output correction result corresponds to the error-corrected text or the target text that is relatively better or has fewer errors, avoiding a situation where the more errors are corrected, the more errors are increased, and improving the effectiveness of text error correction.

[0033] As described above, the first neural network model is a model that has been trained by the training process shown in FIG. 1(a). FIG. 3 illustrates a method for training the first neural network model according to an embodiment of the present disclosure. According to one embodiment, the first neural network model used in method 200 may be trained by method 300 shown in FIG. 3. As shown in FIG. 3, method 300 according to an embodiment of the present disclosure may include steps S310 to S330.

[0034] In order for the first neural network model to be able to generate error correction labels and error-corrected texts based on the input target texts in the data generation stage, the first neural network model needs to be trained using related training data. In an embodiment of the present disclosure, a second neural network model may generate training error-corrected texts corresponding to the training input texts. As shown in FIG. 3, in step S310, the second neural network model may generate a group of output error-corrected texts based on a group of predetermined input texts. The second neural network model may be, for example, a pre-trained error correction model, and may generate a group of output error-corrected texts corresponding to a group of predetermined input texts. The group of predetermined input texts may also have a corresponding group of predetermined reference texts. For example, the group of predetermined input texts may be a group of texts containing errors such as extra characters, omissions, misspellings, and homonyms, and the group of predetermined reference texts may be correct texts in which the extra characters, omissions, misspellings, and homonyms have been eliminated. In an ASR application field, the predetermined input texts may correspond to recognized texts resulting from speech recognition output by an ASR model, and the predetermined reference texts may correspond to correct texts corresponding to the recognized texts, i.e., correct texts of speech corresponding to the recognized texts. The group of predetermined reference texts may be obtained by manual annotation. The group of output error-corrected texts and the group of predetermined reference texts may correspond to the group of predetermined input texts, respectively, where the group of predetermined input texts corresponds to texts before error correction, the group of output error-corrected texts corresponds to texts after error correction, and the group of predetermined reference texts corresponds to texts that are the desired result of error correction, i.e., correct texts.

[0035] After the group of output error-corrected texts is obtained, in step S322, a training error correction label for each output error-corrected text in the group of output error-corrected texts is calculated based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts to obtain a group of training error correction labels. Each training error correction label may indicate whether an output error-corrected text (i.e., text after error correction) is improved relative to a corresponding predetermined input text (i.e., text before error correction). The error correction label may take the form of a number such as "0" or "1," where "0" indicates that the error-corrected text is not improved relative to the target text, and "1" indicates that the error-corrected text is improved relative to the target text. It should be understood that the error correction label may take other forms or indicate whether or not the error-corrected text is improved relative to the target text.

[0036] The training error correction label may be calculated in several ways. For example, the training error correction label may be calculated based on metrics such as ROUGE (Recall-Oriented Understudy for Gisting Evaluation), BLEU (Bilingual Evaluation Understudy), Character Error Rate (CER), Word Error Rate (WER), etc. These metrics are merely exemplary, and it should be understood that the present disclosure is not limited to using these metrics to calculate the training error correction label. In fact, any method capable of comparing differences between texts may be used to calculate the training error correction label. For simplicity, the following description will be given taking the example of calculating the training error correction label based on the Character Error Rate (CER).

[0037] In one embodiment, the training error correction label may be calculated based on the CER. In this case, the above step S320 may include sub-steps S321 to S323. Figure 4 shows these sub-steps of step S320.

[0038] In step S321, a first predetermined input text (corresponding to the text before error correction) from the group of predetermined input texts from step S310, a first output error-corrected text (corresponding to the text after error correction) corresponding to the first predetermined input text, and a first predetermined reference text (corresponding to the correct text) may be obtained. The term "first" here does not imply any preferred mode, priority, or dependency relationship between them. The "first predetermined input text" may be any text from the group of predetermined input texts. The operations for the first predetermined input text described herein are performed on any of the predetermined input texts from the group of predetermined input texts.

[0039] In step S322, a first CER between the first predetermined input text and the first predetermined reference text and a second CER between the first output error-corrected text and the first predetermined reference text may be calculated. The CER is a text quality assessment index that indicates the degree of difference between texts. The calculation method is to calculate the error rate by calculating the number of different characters between the two texts to be compared and dividing the calculated number by the total number of characters in the texts. Thus, the CER may indicate the degree of similarity between the two texts. The lower the CER, the more similar the two texts are. Therefore, the first CER may indicate the degree of similarity between the first predetermined input text and the first predetermined reference text, and the second CER may indicate the degree of similarity between the first output error-corrected text and the first predetermined reference text.

[0040] In step S323, the first CER and the second CER may be compared to obtain a first training error correction label corresponding to the first output error-corrected text. As described above, the first CER may indicate the degree of similarity between the first predetermined input text and the first predetermined reference text, and the second CER may indicate the degree of similarity between the first output error-corrected text and the first predetermined reference text, with a lower CER indicating a closer fit to the first predetermined reference text. Thus, a first training error correction label corresponding to the first predetermined reference text (and also corresponding to the first predetermined input text and the first predetermined reference text) may be determined based on a comparison of the first CER and the second CER. On the one hand, if the first CER is greater than the second CER, it indicates that for the first given input text, the first output error-corrected text is closer to the first predetermined reference text, and the first training error correction label may be set to indicate that the first output error-corrected text is an improvement for the first given input text. On the other hand, if the first CER is less than or equal to the second CER, it indicates that for the first output error-corrected text, the first given input text is closer to the first predetermined reference text, and the first training error correction label may be set to indicate that the first output error-corrected text is not an improvement for the first given input text.

[0041] For each predetermined input text in the group of predetermined input texts, a group of training error correction labels may be obtained by combining a corresponding output error correction text in the group of output error correction texts and a corresponding predetermined reference text in the group of predetermined reference texts, and performing the operations on the first predetermined input text described herein through steps S321 to S323.

[0042] After step S320, data for training the first neural network model is available. Therefore, in step S330, the first neural network model may be trained such that the group of predetermined input texts is used as the input of the first neural network model, and the group of predetermined reference texts or the group of output error-corrected texts and the group of training error-correction labels are used as the output of the first neural network model. Here, for training the first neural network model, the group of predetermined reference texts and the group of training error-correction labels may be used as the output of the first neural network model, and the group of output error-corrected texts and the group of training error-correction labels may be used as the output of the first neural network model. As described above, each predetermined input text has a corresponding output error-corrected text, and each output error-corrected text corresponds to a training error-correction label. Therefore, the first neural network model may be trained by using the group of predetermined input texts as input to the first neural network model, and using the group of output error-corrected texts and the group of training error correction labels as outputs of the first neural network model. Furthermore, as described above, each predetermined input text may have a corresponding predetermined reference text, and each predetermined reference text may correspond to a training error correction label. Since the predetermined reference text corresponds to the desired result text of error correction, i.e., the correct text, during the training process, the output error-corrected text may be replaced with the predetermined reference text, thereby achieving a better training effect. Therefore, on the other hand, the first neural network model may be trained by using the group of predetermined input texts as input to the first neural network model, and using the group of predetermined reference texts and the group of training error correction labels as outputs of the first neural network model.

[0043] In one embodiment, the first neural network model and the second neural network model may be implemented using the same neural network, for example, an autoregressive neural network model. The second neural network model may be an existing error correction model, so its neural network is predetermined. Since the first neural network model and the second neural network model may be trained and implemented using the same neural network, the same model framework may be adopted, and the network architecture does not need to be changed.

[0044] In the text error correction method according to the embodiment of the present disclosure, it is indicated whether the error-corrected text is an improvement over the target text, and the error-corrected text or the target text is selected and output as the correction result, thereby avoiding a situation where the more errors are corrected, the more errors are generated, and improving the effect of text error correction.

[0045] According to an embodiment of the present disclosure, there is further provided a text error correction device. Figure 5 shows a schematic diagram of a text error correction device 500 according to an embodiment of the present disclosure.

[0046] 5, a text error correction device 500 according to an embodiment of the present disclosure includes a processor 510 and a memory 520. The memory 520 stores one or more computer programs.

[0047] The processor 510 is a program control device such as a microprocessor, and operates according to a program installed in, for example, the memory 520. The memory 520 is a storage element such as a ROM or RAM. The program executed by the processor 510 is stored in the memory 520. The text error correction device 500 shown in Figure 5 may be used to implement the text error correction method disclosed in the present application.

[0048] In the device according to the embodiment of the present disclosure, it is indicated whether the error-corrected text is an improvement over the text to be processed, and the error-corrected text or the text to be processed is selected and output as the result of the correction, thereby avoiding a situation where the more errors are corrected, the more errors there are, and improving the effectiveness of text error correction.

[0049] According to an embodiment of the present disclosure, there is further provided a computer program product, which includes a computer program for performing the text error correction method disclosed herein when the computer program is executed by a processor.

[0050] According to an embodiment of the present disclosure, there is further provided a non-transitory computer-readable storage medium having computer program instructions stored thereon, the computer program instructions, when executed by a processor, causing the processor to perform the text error correction method disclosed herein.

[0051] According to an embodiment of the present disclosure, there is further provided a method for training a first neural network model used in text error correction, which may include steps S310 to S330 described above in conjunction with FIG. 3, the details of which are omitted here for brevity.

[0052] The text error correction method according to the embodiment of the present disclosure may be applied to the field of automatic speech recognition (ASR). In this case, the processing target text may be the output text of an ASR model. Figure 6 shows a flowchart of a method 600 for ASR according to the embodiment of the present disclosure. The method 600 may include steps S610 to S630.

[0053] In step S610, the ASR model converts the speech to text to be processed. As described above, the ASR model may receive a speech signal as input, convert it to text, and output it. The text may contain errors, which require further processing to correct the errors.

[0054] In step S620, the first neural network model generates error-corrected text for the target text based on the target text and an error correction label indicating whether the error-corrected text is an improvement over the target text. In step S630, an output error-corrected text or the target text is selected as the ASR result based on the error correction label. For example, the target text may be a single sentence, which may contain errors such as extraneous characters, omissions, misspellings, and homonyms. Steps S620 and S630 may result in a sentence with improved semantic representation quality in which at least some of the errors in the sentence have been reduced, or if no improved sentence is obtained, the original sentence itself may be output. Steps S620 and S630 correspond to steps S210 and S220, respectively, described in conjunction with FIG. 2 above, and are therefore omitted here for brevity.

[0055] In the method used for ASR according to the embodiment of the present disclosure, whether the error-corrected text is an improvement over the target text is indicated, and the error-corrected text or the target text is selected and output as the correction result, thereby avoiding a situation where the more errors are corrected, the more errors are generated, and improving the effectiveness of text error correction in ASR scenarios.

[0056] In this disclosure, we have interpreted and explained applications of ASR scenes in several places as examples, but it should be understood that the invention disclosed herein can be applied to any field or scene other than ASR that requires text error correction.

[0057] The entire hardware computing device described in this disclosure, or parts thereof, may be implemented by various suitable hardware means, including, but not limited to, FPGA, ASIC, SoC, discrete gate or transistor logic, discrete hardware components, or any combination thereof. The devices, apparatus, methods, and systems disclosed herein are not limited to any particular hardware architecture or configuration. Components in the disclosed devices, apparatus, and systems may be separate or integrated, combined in different ways, and / or replaced or supplemented by other components. It should be understood that the teachings of this disclosure may be embodied in various forms of hardware, software, firmware, special purpose processors, or combinations thereof.

[0058] Block diagrams of devices, apparatus, methods, and systems according to the present disclosure are merely exemplary and are not intended to require or imply that they must be connected, arranged, or configured in the manner illustrated in the block diagrams. Those skilled in the art will recognize that these circuits, apparatus, devices, apparatus, and systems may be connected, arranged, or configured in any manner that achieves the desired purpose.

[0059] In the above description, the present invention has been described based on the embodiments. It should be understood by those skilled in the art that these embodiments are merely illustrative and that various modifications can be made to the combination of components and processes of the embodiments, and that such modifications are also within the scope of the present invention.

Claims

1. a first neural network model generating, based on the target text, an error-corrected text of the target text and an error correction label indicating whether the error-corrected text is an improvement over the target text; and selecting and outputting the error-corrected text or the processing target text as a result of the correction based on the error-correction label.

2. The first neural network model is a second neural network model for obtaining a group of output error-corrected texts based on a group of predetermined input texts corresponding to a group of predetermined reference texts; Calculate a training error correction label for each output error-corrected text in the group of output error-corrected texts based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts to obtain a group of training error correction labels; the group of predetermined input texts is an input of the first neural network model, and the group of predetermined reference texts or the group of output error-corrected texts and the group of training error-correction labels is an output of the first neural network model; 2. The method of claim 1, wherein the method is trained to:

3. Calculating a training error correction label for each output error-corrected text among the group of output error-corrected texts based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts to obtain a group of training error correction labels, obtaining a first predetermined input text from the group of predetermined input texts, a first output error-corrected text corresponding to the first predetermined input text, and a first predetermined reference text; calculating a first character error rate (CER) between the first predetermined input text and the first predetermined reference text, and a second CER between the first output error-corrected text and the first predetermined reference text; and comparing the first CER with the second CER to obtain a first training error correction label corresponding to the first output error-corrected text.

4. The step of comparing the first CER with the second CER to obtain the first training error correction label corresponding to the first output error correction text includes: responsive to the first CER being greater than the second CER, setting the first training error correction label to indicate that the first output error-corrected text is an improvement over the first given input text; and in response to the first CER being less than or equal to the second CER, setting the first training error correction label to indicate that the first output error-corrected text is not an improvement over the first given input text.

5. 5. The method of claim 2, wherein the first neural network model and the second neural network model are implemented in the same neural network.

6. The method of claim 1 , wherein the first neural network model comprises an autoregressive model.

7. an ASR model converting speech into text to be processed; a first neural network model generating, based on the target text, an error-corrected text of the target text and an error correction label indicating whether the error-corrected text is an improvement over the target text; and selecting and outputting the error-corrected text or the text to be processed as a result of ASR based on the error-correction label.

8. 1. A method for training a first neural network model for correcting errors in text, comprising: The first neural network model is a second neural network model for obtaining a group of output error-corrected texts based on a group of predetermined input texts corresponding to a group of predetermined reference texts; Calculate a training error correction label for each output error-corrected text in the group of output error-corrected texts based on the group of output error-corrected texts, the group of predetermined input texts, and the group of predetermined reference texts to obtain a group of training error correction labels; the group of predetermined input texts is an input of the first neural network model, and the group of predetermined reference texts or the group of output error-corrected texts and the group of training error-correction labels is an output of the first neural network model; This is how you train.

9. a processor; a memory having one or more computer programs stored therein; A text error correction device, wherein the text error correction method according to any one of claims 1 to 6 is performed by the processor when the one or more computer programs are executed by the processor.

10. A computer program product comprising a computer program, A computer program product, which, when executed by a processor, performs the method of any one of claims 1 to 6.

Images