A text correction method and system based on multi-element semantic features and mask strategy

Abstract

The application provides a text error correction method and system based on a mask strategy and multi-element semantic feature coding, the method comprising: taking a voice transcription text to be corrected as input, using a pre-constructed confusion dictionary to determine candidate error positions in the voice transcription text; using an encoder in a trained self-encoder based on multi-element semantic feature extraction to encode the voice transcription text to obtain a continuous feature vector; inputting the obtained feature vector into a trained error detection model based on a mask strategy to obtain an error occurrence confidence for each of the determined candidate error positions; and for each candidate error position, in response to the error occurrence confidence of the candidate error position being higher than a predefined threshold, using the confusion dictionary to correct the text of the candidate error position to output a corrected voice transcription text.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and more particularly, to a text error correction method based on multi-element semantic features and mask strategy. BACKGROUND

[0002] With the continuous development of artificial intelligence big data technology, China Telecom big data analysis and mining related business is increasing, and now China Telecom has become an important user of big data artificial intelligence technology. In the actual business of telecommunications, artificial intelligence and other related technologies have been widely used for information analysis and data mining of big data. Today, relying on the extensive call business of telecommunications, call data analysis has great significance in marketing and other businesses. The analysis of call record data often depends on the text record of voice transcription. However, due to the limitations of voice transcription and the interference of call device noise, the transcribed text often has many errors, so in this case, the text needs to be corrected.

[0003] To solve this problem, the traditional solution is to use a confusion dictionary, and technicians use homophonic words, near-homophonic words, and homographic words to construct a confusion dictionary for error correction. The perplexity of the text sentence is calculated based on the n-gram language model of the Markov process to detect errors, and the detected errors are corrected through the confusion dictionary. Then, it is determined whether the perplexity is reduced, and the error correction goal is achieved. This method mainly has the following two problems that affect the accuracy of text error correction:

[0004] (1) Using a confusion dictionary to detect errors in text requires serializing each word for inspection, and also depends on the accuracy of the word segmentation tool. When the performance of the word segmentation tool on the text to be detected is average, the word segmentation tool itself will divide the error, thereby affecting the error correction results of the language model and the confusion dictionary, and the correct words will be incorrectly modified.

[0005] (2) Using an n-gram language model for error detection can only consider the previous 1-3 words, and it is difficult to obtain the global information of the text to be detected. Simply expanding the language model window size will produce a large number of out-of-vocabulary words and low-frequency words, and after smoothing, it will still cause the overall perplexity of the sentence to be too large, affecting the judgment of the transcription error. At the same time, the language model is difficult to explicitly learn the grammatical and semantic information, and these shortcomings will lead to large errors in error positioning and verification.

[0006] Therefore, in order to improve the performance of speech transcription error character detection and correction while ensuring high explainability and business flexibility based on the confusion dictionary correction, it is expected to provide an improved speech transcription text correction method to improve the correction accuracy while speeding up the processing speed of single sentence text. SUMMARY

[0007] The summary is provided to introduce some concepts in a simplified form that will be further described in the detailed description below. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used to determine the scope of the claimed subject matter.

[0008] In view of the deficiencies in the prior art, the present application provides a method for improving the accuracy of confusion dictionary text correction based on random masking and combined with multi-element semantic feature encoding. While ensuring high explainability and business flexibility based on confusion dictionary correction, the performance of speech transcription error character detection and correction is improved using random masking strategy and multi-element semantic feature encoding, which improves the accuracy while speeding up the processing speed of single sentence text. The present application trains an efficient error detection model in a supervised manner, intelligently identifies possible error points and performs probability evaluation. For a piece of transcribed text, only one round of forward operation is needed to obtain the possible error positions.

[0009] According to a first aspect of the present application, a method for speech transcription text correction is provided, the method comprising: taking a speech transcription text to be corrected as input, using a pre-constructed confusion dictionary to determine alternative error positions in the speech transcription text; using an encoder in a trained autoencoder based on multi-element semantic feature extraction to encode the speech transcription text to obtain a continuous feature vector; inputting the obtained feature vector into a trained error detection model based on a masking strategy to obtain an error occurrence confidence for each of the determined alternative error positions; and for each alternative error position, in response to the error occurrence confidence of the alternative error position being higher than a predefined threshold, using the confusion dictionary to correct the text of the alternative error position to output a corrected speech transcription text.

[0010] In the technical solution of the embodiments of the present application, the encoder based on multi-element semantic features and the error detection model based on the masking strategy are used to correct the speech transcription text, which can improve the accuracy while speeding up the processing speed of single sentence text.

[0011] In some embodiments, the confusion dictionary is pre-constructed by collecting historical analysis data and common error-prone words, homophones, near-homophones and homographs, wherein the collected error words are recorded as error word groups-corrected word groups in the confusion dictionary.

[0012] In some embodiments, determining the alternative error positions in the speech transcription text by utilizing the pre-constructed confusion dictionary further comprises: utilizing a forward maximum matching algorithm to mark the words in the confusion dictionary in the speech transcription text to retrieve the possible errors collected in the confusion dictionary, so as to determine the alternative error positions in the speech transcription text.

[0013] In some embodiments, encoding the speech transcription text by utilizing the encoder in the trained autoencoder based on multi-semantic feature extraction to obtain a continuous feature vector further comprises: performing multi-semantic feature extraction on the speech transcription text to obtain discrete semantic features, wherein the multi-semantic feature extraction comprises Chinese pinyin feature extraction, Chinese part-of-speech feature extraction, and text word position feature extraction; and mapping the obtained discrete semantic features into a low-dimensional continuous feature space to obtain a continuous feature vector. By utilizing the autoencoder based on multi-semantic feature extraction to encode the multi-semantic features extracted from the speech transcription text into a continuous feature vector, the accuracy and interpretability of error detection and correction can be improved.

[0014] In some embodiments, the autoencoder and the error detection model are trained by utilizing multi-task learning based on the following operations, wherein the training data set comprises artificially annotated speech transcription texts, wherein the parts in the determined alternative error positions in the speech transcription texts are annotated as actually having errors or no errors: inputting each training sample in the training data set into the autoencoder to extract and encode discrete semantic features; inputting each training sample in the training data set into the error detection model to mask the alternative error positions using a mask based on a mask strategy, and to predict the actual characters at the masked positions by utilizing context; and calculating and accumulating the loss of the autoencoder and the loss of the error detection model, and updating the parameters of the autoencoder and the error detection model. By training the autoencoder based on multi-semantic feature extraction and the error detection model based on the mask strategy by utilizing multi-task learning, the interpretability and prediction accuracy of the model can be further enhanced.

[0015] According to a second aspect of the present application, there is provided a system for speech transcription text error correction, the system comprising: an alternative error determination module configured to take a speech transcription text to be corrected as input, determine alternative error positions in the speech transcription text by utilizing a pre-constructed confusion dictionary; a semantic feature encoding module configured to encode the speech transcription text by utilizing an encoder in a trained multi-semantic feature extraction based autoencoder to obtain a continuous feature vector; a transcription error detection module configured to input the obtained feature vector into a trained mask strategy based error detection model to obtain an error occurrence confidence for each of the determined alternative error positions; and

[0016] a transcription error correction module configured to, for each alternative error position, in response to that alternative error position’s error occurrence confidence being higher than a predefined threshold, correct the text at that alternative error position by utilizing the confusion dictionary to output a corrected speech transcription text.

[0017] In some embodiments, the confusion dictionary is pre-constructed by collecting historical analysis data and common error-prone words, homophones, near-homophones and homographs, wherein the collected error words are recorded as error word group-corrected word group in the confusion dictionary.

[0018] In some embodiments, the semantic feature encoding module is further configured to: perform multi-semantic feature extraction on the speech transcription text to obtain discrete semantic features, wherein the multi-semantic feature extraction comprises Chinese pinyin feature extraction, Chinese part-of-speech feature extraction and text word position feature extraction; and map the obtained discrete semantic features into a low-dimensional continuous feature space to obtain the continuous feature vector.

[0019] In some embodiments, the autoencoder and the error detection model are trained by utilizing multi-task learning based on the following operations, wherein a training data set comprises human-annotated speech transcription texts, wherein the determined alternative error positions in the speech transcription texts are annotated with parts that actually have errors and parts that do not have errors: input each training sample in the training data set into the autoencoder to extract discrete semantic features for encoding; input each training sample in the training data set into the error detection model to mask the alternative error positions using masks based on a mask strategy, and utilize context to predict actual characters at the masked positions; and calculate and accumulate losses of the autoencoder and the error detection model, and perform parameter updates of the autoencoder and the error detection model.

[0020] According to a third aspect of the present application, a system for speech transcription text error correction is provided, the system comprising: a memory storing a trained multi-semantic feature extraction based autoencoder and a mask strategy based error detection model and computer executable instructions; and at least one processor, the computer executable instructions, when executed, causing the at least one processor to perform the method of any one of the preceding first aspect.

[0021] Compared with the prior art, the method and system for speech transcription text error correction provided by the present application have at least the following advantages:

[0022] (1) The present application uses a mask strategy based error detection model instead of an n-gram language model for error detection. The words in the confusion dictionary are replaced by masks instead of original characters. The mask positions are predicted to determine whether the text transcription error actually occurs. The possible error points are intelligently identified and probability evaluated. For a piece of transcribed text, only one round of forward operation is needed to obtain the possible error positions, which improves the accuracy and processing speed of error correction.

[0023] (2) The present application is based on the pinyin features, part-of-speech features, word order features and other multi-discrete features of the text to be detected. The encoder maps to a low-dimensional continuous feature space as an explicit feature for model training, which improves the accuracy and interpretability of error detection and correction.

[0024] (3) The present application introduces multi-task learning to enhance the interpretability and prediction accuracy of the model by simultaneously learning multi-semantic encoding and mask error prediction.

[0025] These and other features and advantages will be apparent from a reading of the following detailed description and a review of the associated drawings. It is to be understood that both the foregoing general description and the following detailed description are meant only to illustrate and not to limit the various aspects claimed. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more thoroughly understand the manner in which the above-recited features of the present application are to be used, a more detailed description is made by reference to the various embodiments, some of which are illustrated in the drawings. It is emphasized that, according to common practice, the drawings are not to scale. Rather, the various features of the application are illustrated and described as relative to one another.

[0027] Figure 1 An example architecture diagram of a system for speech transcription text error correction according to an embodiment of the present application is shown;

[0028] Figure 2An example flowchart of a method for training an autoencoder and an error detection model using multi-task learning is shown according to an embodiment of the present application;

[0029] Figure 3 An example flowchart of a method for speech-to-transcript text error correction is shown according to an embodiment of the present application;

[0030] Figure 4 An example flowchart of a multi-element semantic feature encoding part is shown according to an embodiment of the present application; and

[0031] Figure 5 An example architecture diagram of a system for speech-to-transcript text error correction is shown according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] The present application will be described in detail below with reference to the accompanying drawings.

[0033] Figure 1 An example architecture diagram of a system 100 for speech-to-transcript text error correction is shown according to an embodiment of the present application. As shown in Figure 1 The system 100 of the present application includes at least an alternative error determination module 101, a semantic feature encoding module 102, a transcript error detection module 103, and a transcript error correction module 104.

[0034] The alternative error determination module 101 can be configured to take the speech-to-transcript text to be corrected as input, and determine the alternative error positions in the speech-to-transcript text by using a pre-constructed confusion dictionary. In some cases, the confusion dictionary supporting error correction can be pre-constructed by collecting historical call analysis data, department business experience of outbound marketing, and common error-prone words and homophones, and the collected error words are recorded in the form of "error word group-corrected word group" in the confusion dictionary. Further, the alternative error determination module 101 can be configured to use a forward maximum matching algorithm to mark the words appearing in the confusion dictionary in the speech-to-transcript text to be corrected to retrieve the possible errors collected in the confusion dictionary, thereby marking the alternative error positions in the speech-to-transcript text.

[0035] The semantic feature encoding module 102 can be configured to encode the speech transcription text to be corrected using an encoder in the trained autoencoder based on multi-semantic feature extraction to obtain a continuous feature vector. Since the pronunciation of a word, the position in a sentence, and the part of speech possessed in the sentence are often considered in the process of human intervention in text correction, the semantic feature encoding module 102 can be further configured to perform multi-semantic feature extraction on the speech transcription text to be corrected to obtain discrete semantic features, wherein the multi-semantic feature extraction includes Chinese pinyin feature extraction, Chinese part of speech feature extraction, and text word position feature extraction, and then map the obtained discrete semantic features to a low-dimensional continuous feature space to obtain a continuous feature vector.

[0036] The transcription error detection module 103 can be configured to input the obtained feature vector into the trained error detection model based on the mask strategy to obtain an error occurrence confidence for each of the determined alternative error positions. In some cases, the above-mentioned autoencoder based on multi-semantic feature extraction and the error detection model based on the mask strategy are trained using multi-task learning, and the specific training process is shown in Figure 2 and is described in further detail below.

[0037] The transcription error correction module 104 can be configured to, for each alternative error position, in response to the error occurrence confidence of the alternative error position being higher than a predefined threshold, correct the text at the alternative error position using a pre-constructed confusion dictionary, and output the corrected speech transcription text after correcting all positions that need to be corrected.

[0038] Those skilled in the art can understand that the system of the present application and its modules can be implemented in hardware or software, and the modules can be combined or combined in any suitable manner.

[0039] Figure 2 An example flowchart of a method 200 for training an autoencoder and an error detection model using multi-task learning according to one embodiment of the present application is shown.

[0040] Before training the autoencoder based on multi-semantic feature extraction and the error detection model based on the mask strategy, a human-labeled training data set needs to be constructed. Specifically, the training data set can be constructed by first using a forward maximum matching algorithm to mark the words in the confusion dictionary on the existing speech transcription corpus, and retrieve the possible errors collected in the confusion dictionary. Then a batch of training data sets are constructed by human annotation to identify the parts of the alternative error positions retrieved that actually have errors and do not have errors.

[0041] Subsequently, multi-task learning is employed in the training process to simultaneously train the autoencoder based on the multi-semantic feature extraction and the error detection model based on the mask strategy. Multi-task learning refers to the simultaneous and parallel learning of multiple related tasks, the gradient is simultaneously back-propagated, and the multiple tasks help each other learn through the shared representation at the bottom layer, thereby improving the generalization effect. In short, multi-task learning is to learn multiple related tasks together, and share, complement, and promote learning of the domain-related information learned by each other through a shared representation at the shallow layer in the learning process, thereby improving the generalization effect.

[0042] Specifically, in the training process, each sample in the constructed training data set can be input into the autoencoder to extract and encode the discrete semantic features, where the autoencoder includes an encoder and a decoder. In addition, each sample in the constructed training data set can be input into the error detection model to use a mask to cover the annotated alternative error positions based on the mask strategy, and to predict the actual characters of the mask positions by using the context. Subsequently, the artificially annotated error occurrence positions are taken as the true labels, multi-task learning is employed to simultaneously focus on the above two tasks of multi-semantic feature encoding and mask error prediction, the losses (e.g., l1 and l2) of the two tasks are calculated, the losses are accumulated to update the parameters of the autoencoder and the error detection model, and then the steps of the above training process are repeated to continuously iterate and optimize until convergence. In some cases, the accumulation of the losses refers to accumulating the losses based on the weights assigned to each task (e.g., the weight of task 1 is w1, and the weight of task 2 is w2) to obtain the total loss l, where l = w1l1 + w2l2.

[0043] Figure 3 An example flowchart of a method 300 for speech transcription text error correction according to one embodiment of the present application is shown. The method 300 starts at step 301, the alternative error determination module 101 can input the speech transcription text to be corrected as input, and determine the alternative error positions in the speech transcription text by using the pre-constructed confusion dictionary. In some cases, the confusion dictionary can be pre-constructed by collecting historical analysis data and common error-prone words, homophones, near-homophones, and homographs, where the collected error words are recorded as error word groups-corrected word groups in the confusion dictionary. Further, the alternative error determination module 101 can use the forward maximum matching algorithm to mark the words in the speech transcription text to be corrected that appear in the confusion dictionary to retrieve the possible errors collected in the confusion dictionary to identify the alternative error positions in the speech transcription text to be corrected.

[0044] At step 302, the semantic feature encoding module 102 can utilize the encoder in the trained autoencoder based on multi-semantic feature extraction to encode the speech transcription text to be corrected to obtain a continuous feature vector. Further, the semantic feature encoding module 102 can perform multi-semantic feature extraction on the speech transcription text to be corrected to obtain discrete semantic features, and map the obtained discrete semantic features into a low-dimensional continuous feature space to obtain a continuous feature vector. In some cases, the multi-semantic feature extraction can include Chinese pinyin feature extraction, Chinese part-of-speech feature extraction, and Chinese word order feature extraction.

[0045] At step 303, the transcription error detection module 103 can input the obtained continuous feature vector into the trained error detection model based on the mask strategy to obtain an error occurrence confidence for each of the determined alternative error positions. The error detection model based on the mask strategy can be used to cover the words at the identified alternative error positions with a mask, predict the actual characters at the mask positions using the context, to determine the error occurrence confidence of the positions.

[0046] At step 304, the transcription error correction module 104 can utilize a pre-constructed confusion dictionary to correct the text at the alternative error position when the error occurrence confidence of the alternative error position is higher than a predefined threshold, to output the corrected speech transcription text.

[0047] Figure 4 An example flowchart of a multi-semantic feature encoding part 400 according to an embodiment of the present application is shown. As shown in Figure 4 Figure 3 The multi-semantic feature extraction and encoding in step 302 in the

[0048] 302-1: Use a pinyin extraction tool to extract the pinyin of each sentence in the text to be corrected as discrete features for model learning;

[0049] 302-2: Use a sequence labeling tool (part-of-speech tagging) to extract the part-of-speech discrete features to which each word in each sentence in the text to be corrected belongs;

[0050] 302-3: Construct a discrete feature of the position of each word in each sentence in the text to be corrected; and

[0051] 302-4: Construct an autoencoder model to encode the discrete semantic feature information into a continuous feature vector.

[0052] Thus, the present application improves the model expression ability and prediction accuracy by fusing multi-semantic features and mapping these discrete information features in a continuous feature space through an autoencoder. ​

[0053] Figure 5 An example architecture diagram of a system 500 for speech-to-text error correction according to an embodiment of this application is shown. Figure 5 As shown, system 500 may include memory 501 and at least one processor 502.

[0054] Memory 501 may store a trained autoencoder and a masking-based error detection model. Memory 501 may include RAM, ROM, or a combination thereof. Memory 501 may store computer-executable instructions that, when executed by at least one processor 502, cause the at least one processor to perform the various functions described herein, including taking a speech-to-text text to be corrected as input; determining candidate error locations in the speech-to-text text using a pre-built confusion dictionary; encoding the speech-to-text text using an encoder in a trained autoencoder based on multivariate semantic feature extraction to obtain a continuous feature vector; inputting the obtained feature vector into a trained masking-based error detection model to obtain an error occurrence confidence for each of the determined candidate error locations; and for each candidate error location, correcting the text at that candidate error location using the confusion dictionary in response to the error occurrence confidence being higher than a predefined threshold, to output corrected speech-to-text text. In some cases, memory 501 may specifically include a BIOS that controls basic hardware or software operations, such as interaction with peripheral components or devices. Processor 502 may include intelligent hardware devices (e.g., general-purpose processors, DSPs, CPUs, microcontrollers, ASICs, FPGAs, programmable logic devices, discrete gate or transistor logic components, discrete hardware components, or any combination thereof).

[0055] The various illustrative blocks and modules described herein can be implemented or executed using a general-purpose processor, DSP, ASIC, FPGA or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices (e.g., a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors working in conjunction with a DSP core, or any other such configuration).

[0056] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Other examples and implementations are within the scope of the disclosure and appended claims. For example, due to the nature of software, functions described herein can be implemented using software executed by a processor, hardware, firmware, hardwiring, or combinations thereof. Features implementing functions can also be physically located at various positions, including being distributed such that portions of functions are implemented at different physical locations.

[0057] What has been described above includes examples of aspects of the claimed subject matter. It is, of course, not possible to describe every conceivable combination of components or methodologies for purposes of describing the claimed subject matter, but one of ordinary skill in the art will recognize that many further combinations and permutations of the described elements, methods, acts, and so on, are possible. Aspects of the disclosed subject matter are intended to embrace all such alterations, modifications, and variations falling within the spirit and scope of the appended claims.

Claims

1. A method for correcting errors in speech-to-text transcription, characterized in that, The method includes: The speech-to-text text to be corrected is used as input, and a pre-built confusion dictionary is used to determine the candidate error locations in the speech-to-text text. The speech-to-text is encoded using an encoder in a trained autoencoder based on multivariate semantic feature extraction to obtain a continuous feature vector. The obtained feature vectors are input into a trained error detection model based on a masking strategy to obtain the error occurrence confidence for each of the identified candidate error positions. The masking strategy instructs the error detection model to mask the words at the candidate error positions using a mask and to predict the actual character at the masked position using context. For each candidate error location, in response to the error occurrence confidence of the candidate error location being higher than a predefined threshold, the confusion dictionary is used to correct the text at the candidate error location to output the corrected speech-to-text.

2. The method as described in claim 1, characterized in that, The confusion dictionary is pre-constructed by collecting historical analysis data and common misspellings, homophones, near-homophones and homographs. The collected misspelled words are recorded as error phrases-corrected phrases in the confusion dictionary.

3. The method as described in claim 1, characterized in that, Determining alternative error locations in the speech-to-text using a pre-built obfuscation dictionary further includes: The forward maximum matching algorithm is used to mark words in the confusion dictionary in the speech-to-text to retrieve possible errors collected in the confusion dictionary, thereby determining the candidate error locations in the speech-to-text.

4. The method as described in claim 1, characterized in that, Encoding the speech-to-text using an encoder from a trained autoencoder based on multivariate semantic feature extraction to obtain a continuous feature vector further includes: The speech-to-text is subjected to multi-semantic feature extraction to obtain discrete semantic features, wherein the multi-semantic feature extraction includes Chinese Pinyin feature extraction, Chinese part-of-speech feature extraction, and text word position feature extraction; and The obtained discrete semantic features are mapped to a low-dimensional continuous feature space to obtain continuous feature vectors.

5. The method as described in claim 1, characterized in that, The autoencoder and the error detection model are trained using multi-task learning based on the following operation, wherein the training dataset includes manually annotated speech-to-text, wherein the speech-to-text is annotated with portions of the identified candidate error locations that actually contain errors and portions where no errors occur: Each training sample in the training dataset is input into the autoencoder to extract discrete semantic features for encoding; Each training sample in the training dataset is input into the error detection model to mask the candidate error positions using a masking strategy, and the actual character at the mask position is predicted using the context. as well as Calculate and sum the loss of the autoencoder and the loss of the error detection model, and update the parameters of the autoencoder and the error detection model.

6. A system for speech-to-text error correction, the system comprising: The alternative error determination module is configured to take the speech-to-text to be corrected as input and use a pre-built confusion dictionary to determine the locations of alternative errors in the speech-to-text. A semantic feature encoding module is configured to encode the speech-to-text using an encoder in a trained autoencoder based on multivariate semantic feature extraction to obtain a continuous feature vector. A transcription error detection module is configured to input the obtained feature vector into a trained error detection model based on a masking strategy to obtain the error occurrence confidence for each of the determined candidate error positions. The masking strategy is used to instruct the error detection model to mask the words at the candidate error positions using a mask and to use the context to predict the actual character at the mask position. as well as The transcription error correction module is configured to, for each candidate error location, in response to an error occurrence confidence level at the candidate error location being higher than a predefined threshold, use the confusion dictionary to correct the text at the candidate error location, so as to output the corrected speech-to-text.

7. The system as described in claim 6, characterized in that, The confusion dictionary is pre-constructed by collecting historical analysis data and common misspellings, homophones, near-homophones and homographs. The collected misspelled words are recorded as error phrases-corrected phrases in the confusion dictionary.

8. The system as described in claim 6, characterized in that, The semantic feature encoding module is further configured to: The speech-transcribed text is subjected to multi-semantic feature extraction to obtain discrete semantic features, wherein the multi-semantic feature extraction includes Chinese pinyin feature extraction, Chinese part-of-speech feature extraction, and text word position feature extraction; as well as The obtained discrete semantic features are mapped to a low-dimensional continuous feature space to obtain continuous feature vectors.

9. The system as described in claim 6, characterized in that, The autoencoder and the error detection model are trained using multi-task learning based on the following operation, wherein the training dataset includes manually annotated speech-to-text, wherein the speech-to-text is annotated with portions of the identified candidate error locations that actually contain errors and portions where no errors occur: Each training sample in the training dataset is input into the autoencoder to extract discrete semantic features for encoding; Each training sample in the training dataset is input into the error detection model to mask the candidate error positions using a masking strategy, and the actual character at the mask position is predicted using the context. as well as Calculate and sum the loss of the autoencoder and the loss of the error detection model, and update the parameters of the autoencoder and the error detection model.

10. A system for speech-to-text error correction, the system comprising: The memory stores a trained autoencoder based on multi-semantic feature extraction, an error detection model based on a masking strategy, and computer-executable instructions. as well as At least one processor, wherein the computer-executable instructions, when executed, cause the at least one processor to perform the method as described in any one of claims 1-5.

Images