CN112800752B - Error correction method, apparatus, device and storage medium - Google Patents
Error correction method, apparatus, device and storage medium Download PDFInfo
- Publication number
- CN112800752B CN112800752B CN202011627609.3A CN202011627609A CN112800752B CN 112800752 B CN112800752 B CN 112800752B CN 202011627609 A CN202011627609 A CN 202011627609A CN 112800752 B CN112800752 B CN 112800752B
- Authority
- CN
- China
- Prior art keywords
- text
- corrected
- character
- graph
- convolution layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000012937 correction Methods 0.000 title claims abstract description 151
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000012545 processing Methods 0.000 claims abstract description 49
- 238000012986 modification Methods 0.000 claims description 18
- 230000004048 modification Effects 0.000 claims description 18
- 238000012549 training Methods 0.000 claims description 17
- 238000005096 rolling process Methods 0.000 claims description 4
- 238000003672 processing method Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 10
- 230000008569 process Effects 0.000 description 10
- 239000011159 matrix material Substances 0.000 description 6
- 239000013598 vector Substances 0.000 description 5
- 230000008878 coupling Effects 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000012217 deletion Methods 0.000 description 3
- 230000037430 deletion Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 230000000873 masking effect Effects 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000000926 separation method Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/232—Orthographic correction, e.g. spell checking or vowelisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The application discloses an error correction method, a device, equipment and a storage medium, wherein the error correction method comprises the following steps: acquiring a text to be corrected; performing text processing on the text to be corrected by using the correction model to obtain text processing information, wherein the text processing information comprises context semantic information of the text to be corrected and similar character information in the text to be corrected; and predicting the text processing information by using the error correction model to obtain a prediction error correction result of the text to be corrected. According to the scheme, the error correction accuracy can be improved to a certain extent.
Description
Technical Field
The present application relates to the field of artificial intelligence, and in particular, to an error correction method, apparatus, device, and storage medium.
Background
With the continuous development and progress of social technology, various electronic devices have been popularized in people's lives. The input method is used as an interaction entrance between people and electronic equipment, and has the characteristics of convenience, high efficiency, accuracy and the like. However, many error correction methods for text are created during the input process because writing is not normal or a key error causes an input text error. The general error correction method has the following flow: and finding suspected error words through word segmentation, and determining a final error correction result by using homophonic or similar words for the error words according to the statistical language model. Such error correction using only a single pinyin to determine the error correction result may result in inaccurate error correction results.
Disclosure of Invention
The application provides at least one error correction method, device, equipment and storage medium.
The first aspect of the present application provides an error correction method, comprising: acquiring a text to be corrected; performing text processing on the text to be corrected by using the correction model to obtain text processing information, wherein the text processing information comprises context semantic information of the text to be corrected and similar character information in the text to be corrected; and predicting the text processing information by using the error correction model to obtain a prediction error correction result of the text to be corrected.
The similar character information in the text to be corrected comprises similar characters of the original characters in the text to be corrected and the similarity between the original characters and the similar characters.
The error correction model comprises a first sub-network and a second sub-network; the first sub-network is used for acquiring context semantic information of the text to be corrected, and the second sub-network is used for acquiring similar character information in the text to be corrected.
Performing text processing on the text to be corrected by using the correction model to obtain text processing information, wherein the text processing information comprises: performing graph convolution operation on the text to be corrected by using at least one convolution layer in the second sub-network to obtain similar character information; the graph convolution operation adopts a similar graph corresponding to a convolution layer, wherein the similar graph comprises a plurality of confusable characters and similar weights among the confusable characters.
Performing a graph convolution operation on the text to be corrected by using at least one convolution layer in the second sub-network to obtain similar character information, including: taking each convolution layer as a current convolution layer; carrying out convolution operation on input information of the current convolution layer and a similar graph by using the current convolution layer to obtain a first graph convolution result of the current convolution layer; the input information of the first layer of convolution layer is a text to be corrected, and the input information of the non-first layer of convolution layer is an output result of the previous convolution layer; based on the first graph convolution result of the current convolution layer and the output results of all convolution layers or part of convolution layers before the current convolution layer, obtaining the output result of the current convolution layer; and obtaining similar character information based on the output result of the last convolution layer.
The method for obtaining the first graph convolution result of the current convolution layer comprises the following steps of: respectively carrying out convolution operation on the input information of the current convolution layer and at least one similar graph by using the current convolution layer to obtain a second graph convolution result corresponding to each similar graph of the current convolution layer; wherein the confusable characters in different similar graphs of each convolution layer are aimed at different input modes; carrying out weighted summation on the second graph convolution result corresponding to each similar graph of the current convolution layer to obtain a first graph convolution result of the current convolution layer; the weight of the second graph convolution result corresponding to the similar graph is determined based on the input mode adopted by the text to be corrected.
Before performing a graph convolution operation on the text to be corrected by using at least one convolution layer in the second sub-network to obtain similar character information, the method further comprises: acquiring a pruned text corpus, wherein the pruned text corpus comprises character pairs consisting of characters before and after modification; selecting a plurality of character pairs according to the similarity identification standard corresponding to the input mode adopted by the character pairs; based on the selected pairs of characters, confusable characters are determined.
The input mode comprises at least one of a handwriting input mode, a stroke input mode and a five-stroke input mode; selecting a plurality of character pairs according to a similar identification standard corresponding to an input mode adopted by the character pairs, wherein the character pairs comprise: if the input mode adopted by the character pairs is a handwriting input mode, selecting the character pairs with the comprehensive similarity of the character structure and the character components being greater than or equal to a preset similarity threshold value; if the input mode adopted by the character pairs is a stroke input mode, selecting the character pairs with the same preset number of strokes; if the input mode adopted by the character pairs is a five-stroke input mode, selecting the character pairs with the editing distance of the five-stroke codes smaller than or equal to a preset distance threshold value.
Wherein determining the confusable character based on the selected plurality of character pairs comprises: determining the smoothness of the characters before modification and the characters after modification in the selected character pairs in the text corpus; and if the smoothness corresponding to the character before modification is smaller than the preset smoothness and smaller than the smoothness corresponding to the character after modification, the character in the character pair is used as the confusing character.
The error correction model is obtained by training at least the following steps: pre-training the first sub-network by using the first sample text and the actual error condition of the first sample text; training the first sub-network and the second sub-network by using the second sample text and the actual error correction result of the second sample text.
The first sample text is obtained by randomly replacing at least one character of a normal text corpus; and/or the second sample text is obtained based on the back-deleted text corpus, wherein the back-deleted text corpus is used for collecting and obtaining confusable characters, and the error correction model is used for carrying out graph convolution by using a similarity graph obtained based on the confusable characters to obtain similar character information of the input text.
The prediction error correction result of the text to be corrected comprises character error correction results of all characters in the text to be corrected and the prediction probability of the character error correction results; after the text processing information is predicted by using the error correction model to obtain a predicted error correction result of the text to be corrected, the method further comprises the following steps: the character error correction result with the prediction probability exceeding the preset probability threshold value is used as a final error correction result; and correcting the text to be corrected by utilizing the final correction result.
A second aspect of the present application provides an error correction apparatus comprising: the acquisition module is used for acquiring the text to be corrected; the processing module is used for carrying out text processing on the text to be corrected by utilizing the correction model to obtain text processing information, wherein the text processing information comprises context semantic information of the text to be corrected and similar character information in the text to be corrected; and the result prediction module is used for predicting the text processing information by using the error correction model to obtain a prediction error correction result of the text to be corrected.
A third aspect of the present application provides an electronic device comprising a memory and a processor for executing program instructions stored in the memory to implement the error correction method described above.
A fourth aspect of the present application provides a computer readable storage medium having stored thereon program instructions which, when executed by a processor, implement the above error correction method.
According to the scheme, the text to be corrected is corrected by combining the context semantic information of the text to be corrected and the similar character information of the text to be corrected, so that the accuracy of correction is improved compared with the case that single similar character information is used.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application as claimed.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description, serve to explain the principles of the application.
FIG. 1a is a flow chart of an embodiment of an error correction method of the present application;
FIG. 1b is a similar pictorial representation of an embodiment of the error correction method of the present application;
FIG. 2 is a schematic diagram of an embodiment of an error correction apparatus according to the present application;
FIG. 3 is a schematic diagram of an embodiment of an electronic device of the present application;
FIG. 4 is a schematic diagram of an embodiment of a computer readable storage medium of the present application.
Detailed Description
The following describes embodiments of the present application in detail with reference to the drawings.
In the following description, for purposes of explanation and not limitation, specific details are set forth such as the particular system architecture, interfaces, techniques, etc., in order to provide a thorough understanding of the present application.
The term "and/or" is herein merely an association relationship describing an associated object, meaning that there may be three relationships, e.g., a and/or B, may represent: a exists alone, A and B exist together, and B exists alone. In addition, the character "/" herein generally indicates that the front and rear associated objects are an "or" relationship. Further, "a plurality" herein means two or more than two. In addition, the term "at least one" herein means any one of a plurality or any combination of at least two of a plurality, for example, including at least one of A, B, C, and may mean including any one or more elements selected from the group consisting of A, B and C.
Referring to fig. 1a, fig. 1a is a flow chart illustrating an embodiment of an error correction method according to the present application. Specifically, the method may include the steps of:
step S11: and acquiring a text to be corrected.
The text to be corrected may be obtained by transmitting the text to the device through other devices, or may be obtained by obtaining a text generated by the device, for example, the text to be corrected may be a text generated by the device in real time. For example, in the process of editing a document, a user can acquire a text edited by the user in real time and correct the text in real time, or when the user clicks a correction button in the document, the document being used is acquired as the text to be corrected. Therefore, the method for acquiring the text to be corrected is numerous, and the embodiment of the disclosure does not specifically specify the method for acquiring the text to be corrected. The device is a device capable of executing the error correction method according to the embodiment of the disclosure.
Step S12: and carrying out text processing on the text to be corrected by using the correction model to obtain text processing information.
The text processing information comprises context semantic information of the text to be corrected and similar character information in the text to be corrected. There are various ways to obtain the context semantic information, for example, a general deep semantic model may obtain the context semantic information of the text to be corrected, where the deep semantic model may be BERT (Bidirectional Encoder Representations from Transformer). The manner in which the BERT model extracts the corresponding contextual semantic information of the text to be error corrected may be generally generic, and embodiments of the present disclosure do not specifically specify this. The similar character information of the text to be corrected may be obtained by searching for similar character information corresponding to each character in the text to be corrected, and so on.
Step S13: and predicting the text processing information by using the error correction model to obtain a prediction error correction result of the text to be corrected.
And predicting the context semantic information of the text to be corrected and the similar character information in the text to be corrected, which are obtained through comprehensive text processing, so as to obtain a final prediction correction result of the text to be corrected. The prediction mode may be to multiply the context semantic information and the similar character information of the text to be corrected to obtain a final prediction correction result. Of course, in other embodiments, the text processing information may be predicted in other ways as well.
According to the scheme, the text to be corrected is corrected by combining the context semantic information of the text to be corrected and the similar character information of the text to be corrected, so that the accuracy of correction is improved compared with the case that single similar character information is used.
In some disclosed embodiments, the error correction model includes a first subnetwork and a second subnetwork. The first sub-network is used for acquiring context semantic information of the text to be corrected, and the second sub-network is used for acquiring similar character information in the text to be corrected. Wherein the first subnetwork may be a deep semantic model and the second subnetwork may be a deep neural network model. The first sub-network model and the second sub-network model are relatively independent, namely, no execution relationship exists, namely, the text to be corrected is input into the two sub-networks at the same time, and then the two sub-networks respectively and independently acquire the context semantic information of the text to be corrected and the similar character information in the text to be corrected.
The context semantic information and detailed character information of the text to be corrected are independently acquired through the two sub-networks, so that the situation that information is split due to the complexity of the process of acquiring the two information in stages is reduced.
The error correction model is obtained by training at least the following steps: the first sub-network is pre-trained using the first sample text and the actual error conditions of the first sample text. Specifically, the first sample text may be any normal text, for example, may be a plurality of articles. Alternatively, the first text sample is obtained by a random substitution of at least one character for the normal text corpus. Alternatively, the replacement of normal text corpus includes replacing 80% of the masked characters with [ mask ], 10% with other characters and another 10% with original characters. In the pre-training process, the first sub-network does not know the character position replaced by 10% of original characters, so that the first sub-network is forced to rely on the context semantic information more to predict words, and the first sub-network is endowed with certain error correction capability. Of course, in other embodiments, the manner of masking the characters may be other manners, and is not limited to the masking ratio. Inputting the first text sample after the shielding treatment into a first sub-network to obtain a prediction result, adjusting parameters of the first sub-network by back propagation according to the error condition between the prediction result and an actual character label, and iteratively executing the steps of predicting and back propagation adjustment parameters until the error between the given prediction result of the first sub-network and the actual character label is smaller than a preset error, and considering that the pre-training is completed. The training speed of the subsequent network model can be increased by pre-training the first sub-network by using the normal text.
After the pre-training is completed, training the first sub-network and the second sub-network by using the second sample text and the actual error correction result of the second sample text. Specifically, the second sample text may be obtained by obtaining the second sample text in a device performing the error correction method. For example, the second sample text is derived based on the pruned text corpus. The method comprises the steps of deleting corpus texts in a back-deleting mode, wherein the back-deleting corpus texts are used for collecting and obtaining confusable characters, and the error correction model carries out graph convolution by utilizing a similarity graph obtained based on the confusable characters to obtain similar character information of an input text. The method for acquiring the pruned text corpus comprises the steps of acquiring the pruned text corpus through a user log of the equipment. The pruned text corpus includes character pairs composed of characters before and after modification. Specifically, the expression form of the back-deleted text corpus is an on-screen text corpus, a back-deleted modified corpus and a modified final text corpus. The on-screen text corpus refers to text information displayed on a screen of the device, and the deletion-back modification information refers to deleted information. That is, in the embodiment of the present disclosure, different confusable characters may be constructed by acquiring different deleted corpus texts for different users, and the characters may be used to train an error correction model corresponding to the user, i.e. more targeted error correction may be implemented, and the accuracy of error correction may be improved.
After the corpus is retrieved, a plurality of character pairs are selected according to the similarity identification standard corresponding to the input mode adopted by the character pairs. The input mode may be at least one of a handwriting input mode, a stroke input mode and a five-stroke input mode.
And if the input mode adopted by the character pairs is a handwriting input mode, selecting the character pairs with the comprehensive similarity of the character structure and the character components being greater than or equal to a preset similarity threshold value. The preset similarity threshold may be 50%. Because in handwriting input, the input process may be more random, the strokes are stuck more, the stroke sequence is not clear, and the like due to different personal writing habits, the details are not clear, and the similarity of character pairs cannot be judged through the stroke sequence, and the like. Therefore, in the embodiment of the present disclosure, the judgment is made with the comprehensive similarity of the character result and the character component. For example, the results of "sunny" and "sunny" are the same, both are of left-right structure, and the right component parts of the two words are the same, both are "green", i.e., the structures are the same, and the content of one of the two components is the same, and the similarity of the two words is greater than 50% through calculation, which indicates that the fonts of the two words are similar.
If the input mode adopted by the character pairs is a stroke input mode, selecting the character pairs with the same preset number of strokes. The stroke input mainly obtains a similarity result according to a stroke sequence. Unlike handwriting input, stroke input does not pay attention to character overall result information, and is mainly determined according to character strokes and sequence thereof, and does not depend on personal input habits. The pre-set number may be determined based on the common strokes of the respective characters, e.g., the number of strokes of one character in a pair is 12 and the number of strokes of the other character is 13, and the pre-set number may be determined to be the first 8 strokes. Of course, this is by way of example only, and in other embodiments the pre-set number may be on the order of forty to eighty percent of the number of strokes of one character that is less than the number of strokes of the pair of characters. And if the preset number of strokes are the same, recognizing that the character fonts in the character pair are similar. The character pairs described in the embodiments of the present disclosure may have two characters, or of course, may have multiple characters, and the number of characters in the character pairs is not specifically defined herein.
If the input mode adopted by the character pairs is five-stroke input, selecting the character pairs with the editing distance of the five-stroke input codes smaller than or equal to a preset distance threshold value. The editing distance is related to the repetition rate, the higher the repetition rate is, the lower the editing distance is, and similarly, the lower the repetition rate is, the higher the editing distance is. Five-stroke input mainly comprises a series of radicals and word separation rules distributed on 26 keys, chinese characters are separated into radical sequences according to the word separation rules, and corresponding editing distances can be obtained when corresponding key sequences are obtained. If the positions of two adjacent keys in the key sequence are exchanged, one key is considered to be the repeated code. For example, the five-stroke code "jge" of "fine" and the five-stroke code "jeg" of "bright" are repeated two by one in three keys, and the edit distance is considered to be 1, and the edit distance is considered to be similar to the font, that is, the edit distance is considered to be the number of non-duplicate codes. In the embodiment of the present disclosure, if the edit distance is less than or equal to 1, the glyphs of the characters can be considered to be similar.
Since the user input process of the deletion is not always to find errors to correct errors, and the input intention is possibly changed, the non-correction corpus is filtered by judging the similarity of the content fonts before and after the deletion. But also is based on the input of the font, and different input modes often have different characteristics. Different input characteristics of handwriting input, stroke input, five-stroke input and the like determine that the handwriting input, the stroke input, the five-stroke input and the like have different standards on the character pattern similarity. Therefore, the embodiment of the disclosure makes the recognition of the character pattern similarity before and after the deletion more accurate by formulating different similarity recognition standards for different input modes.
Based on the selected pairs of characters, confusable characters are determined. And the text corpus where the character pairs meeting the similarity identification standards are located is reserved, and the rest of the text corpus where the character pairs not meeting the similarity identification standards are located is deleted. For example, a single pruned text corpus is a complete sentence, a sentence is retained if there are pairs of characters in the sentence that meet the similarity recognition criteria, and the sentence is deleted from the pruned text corpus if there are no pairs of characters in the sentence that meet the similarity recognition criteria. For example, the user inputs "today's weather doze" first under the stroke input, then deletes "doze" back, modifies to "fine", and finally the text corpus is "today's weather fine", wherein "doze" and "fine" conform to the font similarity standard in the stroke input, so the back-deleted text corpus pair can be reserved.
Further, to improve the quality of the back-deleted text corpus, the embodiment of the disclosure further obtains the final back-deleted text corpus by calculating the smoothness of the back-deleted text corpus before and after the back-deleted text corpus so as to ensure that errors do exist in the error-correction pre-deleted text corpus, the correctness of the error-correction post-deleted text corpus is obviously improved, and the quality of the error-correction corpus is improved compared with the smoothness of the error-correction pre-deleted text corpus. The specific mode comprises the steps of determining the smoothness of the characters before modification and the characters after modification in the text corpus. Meaning that the smoothness of the text corpus of the stricken text where the characters before modification are located is determined respectively. The sentence probability can be calculated through a language model to measure the specific sentence smoothness. Sentence prosity is the expression of whether the sentence accords with the human expression habit, and sentence probability reflects the prosity of the sentence from a statistical perspective by sampling the occurrence frequency of the sentence in human expression. In the embodiment of the disclosure, the probability of the text corpus is calculated according to the joint probability of each word in the text corpus, and the specific calculation method is shown as the formula:
Wherein P (S) represents the probability of pruning text corpus S, w 1 w 2 …w N Representing word sequences that make up the pruned text corpus S, N representing the pruned text corpusNumber of words included, w i Represents the i-th word, p (w i |w 1 w 2 …w i-1 ) The conditional probability of the i-th word given the first i-1 words is represented. And if the smoothness corresponding to the character before modification is smaller than the preset smoothness and smaller than the smoothness corresponding to the character after modification, the character in the character pair is used as the confusing character. The preset smoothness is used for representing normal specific smoothness, and the preset smoothness can be set according to requirements.
Wherein a plurality of confusing characters may form a confusing character set. And extracting the confusable character set corresponding to the input mode according to the input mode of the character pair. For example, if some character pairs in the text corpus are input through handwriting, a confusing character set corresponding to the handwriting input mode will be constructed, some character pairs are input through strokes, a confusing character set corresponding to the stroke input mode will be constructed, and some characters are input through five strokes, a confusing character set corresponding to the five-stroke input mode will be constructed. And taking the reserved back-deleted text corpus as a second sample text.
The second sample text and the confusing character set are input into the error correction model to train the first sub-network and the second sub-network.
The flow of predicting the second sample text, namely the pruned text corpus, by the error correction model to obtain a corresponding predicted result is the same as the flow of obtaining the predicted error correction result of the text to be corrected. Therefore, the process of obtaining the predicted result of the second sample text in the error correction model refers to the following process of obtaining the predicted error correction result of the text to be corrected.
The text to be corrected is input into the correction model. Optionally, the correction model firstly acquires word vectors of the text to be corrected, and the word vectors of the text to be corrected are respectively input into the first sub-network and the second sub-network. The text processing method for the text to be corrected by using the correction model comprises the following steps of: and performing graph convolution operation on the text to be corrected by using at least one convolution layer in the second sub-network to obtain similar character information. The graph convolution operation adopts a similar graph corresponding to a convolution layer. Wherein the similarity graph includes a number of confusable characters and similar weights between the confusable characters. Specifically, the similarity graph is constructed according to the confusable character sets, and the confusable character sets corresponding to different input modes can construct the similarity graph corresponding to the input modes. The similarity graph is a binary adjacent matrix of NxN, N is the number of characters in the confusion character set, wherein the value of the position in the character pair corresponding matrix is 1, and the value of the position in the non-character pair corresponding matrix is 0. In the similar diagram, the character arrangement order of the rows and columns is the same. Referring also to fig. 1b, fig. 1b is a similar schematic representation of an embodiment of the error correction method of the present application. As shown in FIG. 1b, the character arrangements for the rows and columns of the similarity map are "Jing", "Shu", "Qing" and "Qing". As can be derived from the similarity map, the confusable character set corresponding to the similarity map includes (jinshi) and (qing) two character pairs. And, the characters corresponding to the first row and the first column are "all" and "all" because the two characters of "all" and "all" are dissimilar, the value corresponding to the first row and the first column is 0, and similarly, the characters corresponding to the first row and the third column are "all" and "all" because the two characters are similar characters, the value corresponding to the first row and the third column is 1. Of course, the similar diagrams herein are merely examples, and the number of rows and columns of the similar diagrams in the real application scenario may be plural.
Specifically, the formula corresponding to the convolution operation in the convolution layer is as follows:
wherein, the parameter A represents a similarity graph,representing a regularized version of a; h l Representing the input of the first layer, the input H of the first layer convolution layer 0 The text to be corrected can be a word vector E of the text to be corrected; w (W) l Representing a layer i learnable network weight parameter matrix. Wherein the matrix of network weight parameters for each convolutional layer may be different. Of course, the arrangement of the similarity graphs of different convolution layers may be different, as wellMay be the same. Alternatively, the more common is the similar characters represented in the similarity graph according to the increase in the number of layers, wherein the characters in the similarity graphs of different layers may partially overlap or weight overlap or not overlap at all.
And carrying out convolution operation on the input information of the current convolution layer and the similar graph by using each convolution layer as the current convolution layer to obtain a first graph convolution result of the current convolution layer. The input information of the first layer of convolution layer is the text to be corrected, and the input information of the non-first layer of convolution layer is the output result of the previous convolution layer.
Specifically, the input information of the current convolution layer and at least one similar graph are used for carrying out convolution operation respectively, and a second graph convolution result corresponding to each similar graph of the current convolution layer is obtained. That is, the similarity map possessed by a single convolution layer is not limited to 1, and there may be a plurality, for example, one similarity map for each confusable character set of the input mode. If the current convolution layer corresponds to a plurality of similar graphs, convolution operation is required to be carried out on the input information of the current convolution layer and all the similar graphs respectively. And carrying out weighted summation on the second graph convolution result corresponding to each similar graph of the current convolution layer to obtain the first graph convolution result of the current convolution layer. The weight of the second graph convolution result corresponding to the similar graph is determined based on the input mode adopted by the text to be corrected. For example, if the input mode of the text to be corrected is known, the weight of the second graph convolution result corresponding to the input mode is increased, and the weight of the second graph convolution result corresponding to the other input modes is reduced. If the input mode of the text to be corrected is not known, a preset weight can be adopted. The preset weight setting manner may be that the weights of the second graph convolution results corresponding to the similar graphs are the same, and of course, the weights of the second graph convolution results corresponding to the similar graphs may also be determined according to the use frequency of the user or the device for the input manner, for example, the user or the device frequently uses handwriting input and does not frequently use five strokes input, so that the weights occupied by the second graph convolution results corresponding to the similar graphs by handwriting input are increased, and the weights occupied by the second graph convolution results corresponding to the similar graphs by five strokes input are reduced. Thus, how to set the weights can be determined on demand.
For example, the formula for weighted summation of the similarity graphs for the second graph convolution result is as follows:
wherein C is l Representing the weighted result after convolution of the first layer graphs,second graph convolution result weighting weights representing a first layer kth similarity graph, f k (A k ，H l ) Second graph convolution operation result representing kth similarity graph, A k Representing the kth similarity graph. The weighting weights can be set according to the above manner, or can be determined by network learning.
By combining the first graph convolution results of the multiple similar graphs, convolution results under different input modes can be comprehensively considered, and the accuracy of the error correction result can be improved.
And obtaining the output result of the current convolution layer based on the first graph convolution result of the current convolution layer and the output results of all convolution layers or part of convolution layers before the current convolution layer. Of course, the number of partial convolution layers is greater than or equal to 0, i.e. the first graph convolution result of the current layer may be directly used as the output result of the current convolution layer. Optionally, the output result of the last layer of convolution layer is the sum of the first graph convolution result of the last layer of convolution layer and the output results of all previous convolution layers, so as to obtain the output result of the last layer, and the output result of the convolution layer other than the last layer is the first graph convolution result of each layer. Of course, in other embodiments, the output result of each convolution layer may be the sum of the first graph convolution result of each convolution layer and the output result of the previous convolution layer. And accumulating the first graph rolling result of the last layer and all output results of the previous layer to be used as the output result of the last second sub-network, so that the proportion of the text to be corrected in the output result is increased, and the original semantic information of the text to be corrected is better kept.
Wherein, the formula for adding the output results of each layer is as follows:
wherein H is l+1 Representing layer 1 input, i.e. layer l output, C l The weighted result after convolution of the first layer of graphs is shown.
And obtaining similar character information based on the output result of the convolution layer of the last layer. The output mode of the last layer may be:
in the disclosed embodiment, the graph convolution layers are assumed to be a common l layer. Of course, if some characters do not have a confusable character set, the output of the character at the last layer is its input in the error correction model. Wherein U is i A word vector representation representing the ith character after passing through the second sub-network,convolving the final output of the network for the ith character corresponding diagram, E i A word vector representation representing the i-th character.
The similar character information in the text to be corrected comprises similar characters of original character information of the text to be corrected and similarity between the original characters and the similar characters. Specifically, the similar character information in the text to be corrected includes information of characters of the text to be corrected and distances between the characters in the text to be corrected and the similar characters thereof in semantic space dimensions. Wherein the closer the distance, the higher the similarity of the character to the similar character is considered. The similar character herein may be any of the confusing characters. That is, the similar character information herein is not limited to the information of one similar character, and may be the similarity of characters in a plurality of texts to be corrected and a plurality of similar characters.
In some disclosed embodiments, the similarity graph used in performing the graph rolling operation on the text to be corrected by using at least one convolution layer in the second sub-network may be the similarity graph used in training the second sub-network directly, or may be an updated similarity graph before performing this step. That is, before executing this step, the pruned text corpus is re-obtained, a plurality of character pairs are selected according to the similarity recognition criteria corresponding to the input mode adopted by the character pairs consisting of the characters before and after modification in the pruned text corpus, and the confusing characters are determined based on the selected plurality of character pairs. Of course, only the newly added pruned text corpus after training the second subnetwork may be obtained for determining new confusing characters and then updating the original confusing characters. That is, the similarity map can be updated according to the daily use condition of the user or the device, so that the prediction result of the text to be corrected is more accurate. Of course, the procedure of re-acquiring the confusable character set is the same as the procedure of acquiring the confusable character set before training the second sub-network, and will not be described here again.
And multiplying the context semantic information of the text to be corrected output by the first sub-network with the similar character information of the text to be corrected output by the second sub-network to obtain each candidate prediction correction result. The candidate prediction error correction result comprises candidate character error correction results of all characters in the text to be corrected and probability of all candidate characters. And selecting the candidate prediction error correction result with the highest probability as the prediction error correction result of the position.
Further, the prediction probability of each prediction error correction result is calculated, wherein the formula for calculating the prediction probability of each prediction error correction result may be:
wherein the method comprises the steps ofRepresenting the predicted error correction result at position i, and X represents the text to be error corrected. The predicted probability exceeds the preset probabilityAnd the character error correction result of the threshold value is taken as a final error correction result. And correcting the text to be corrected by utilizing the final correction result. That is, if the prediction probability of the predicted error correction result of a certain character position is lower than the preset error correction result, the error correction is not performed on the character position. I.e. the input character at that position is kept unchanged. The text to be corrected is corrected under the condition that the prediction probability of the correction result is higher than the preset probability threshold, so that the probability of error correction of the text to be corrected is effectively reduced, and the accuracy of error correction is improved.
And saving the document after error correction and/or the comparison document before and after error correction as a new document. Optionally, obtaining the similarity between the characters before and after the error correction of the document to be corrected, and if the characters before and after the error correction meet the similarity identification standard corresponding to the input mode, adding the characters before and after the error correction into the confusable character set corresponding to the input mode. If the input modes of the characters before and after error correction are not known, the corresponding similarity before and after error correction is judged according to the similarity identification standards corresponding to the three input modes, and if one or more similarity identification standards are met, the characters before and after error correction are respectively added into the corresponding confusing character set for updating the confusing character set. Further, the updated confusable character set is used to update the similarity graph in the second sub-network. The similarity graph is updated continuously, so that the subsequent error correction result of other texts to be corrected is more accurate.
So far, the error correction process of the text to be corrected is ended. And in the training process of the error correction model, after the final prediction error correction result is obtained, comparing the error between the prediction result and the real label of the second sample text through the objective function, and iteratively updating the error correction model parameters through an optimization algorithm. Wherein the optimization algorithm may be a random gradient descent algorithm or the like. The error correction model parameters here include parameters in the first sub-network and the second sub-network, for example, a network weight parameter matrix corresponding to each convolution layer in the second sub-network, and so on. And if the error between the predicted error correction result of the error correction model and the real label of the second sample text is smaller than the error threshold value, the training of the error correction model is considered to be completed, namely the error correction model can be put into use.
The main body of the error correction method may be an error correction apparatus, for example, the error correction method may be performed by a terminal device or a server or other processing device, where the terminal device may be a User Equipment (UE), a computer, a mobile device, a User terminal, a cellular phone, a cordless phone, a personal digital assistant (Personal Digital Assistant, PDA), a handheld device, a computing device, an in-vehicle device, a wearable device, or the like. In some possible implementations, the error correction method may be implemented by way of a processor invoking computer readable instructions stored in a memory.
Referring to fig. 2, fig. 2 is a schematic structural diagram of an error correction device according to an embodiment of the application. The error correction device 20 comprises an acquisition module 21, a processing module 22 and a result prediction module 23. Wherein, the obtaining module 21 is configured to obtain a text to be corrected; the processing module 22 is configured to perform text processing on the text to be corrected by using the correction model, so as to obtain text processing information, where the text processing information includes context semantic information of the text to be corrected and similar character information in the text to be corrected; and the result prediction module 23 is used for predicting the text processing information by using the error correction model to obtain a predicted error correction result of the text to be corrected.
According to the scheme, the text to be corrected is corrected by combining the context semantic information of the text to be corrected and the similar character information of the text to be corrected, so that the accuracy of correction is improved compared with the case that single similar character information is used.
The functions of each module may be described in the embodiments of the error correction method, which is not described herein.
Referring to fig. 3, fig. 3 is a schematic structural diagram of an electronic device according to an embodiment of the application. The electronic device 30 comprises a memory 31 and a processor 32, the processor 32 being arranged to execute program instructions stored in the memory 31 for implementing the steps of any of the error correction method embodiments described above. In one particular implementation scenario, electronic device 30 may include, but is not limited to: the microcomputer and the server, and the electronic device 30 may also include a mobile device such as a notebook computer and a tablet computer, which is not limited herein.
In particular, the processor 32 is adapted to control itself and the memory 31 to implement the steps of any of the error correction method embodiments described above. The processor 32 may also be referred to as a CPU (Central Processing Unit ). The processor 32 may be an integrated circuit chip having signal processing capabilities. The processor 32 may also be a general purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a Field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. In addition, the processor 32 may be commonly implemented by an integrated circuit chip.
According to the scheme, the text to be corrected is corrected by combining the context semantic information of the text to be corrected and the similar character information of the text to be corrected, so that the accuracy of correction is improved compared with the case that single similar character information is used.
Referring to fig. 4, fig. 4 is a schematic structural diagram of an embodiment of a computer readable storage medium according to the present application. The computer readable storage medium 40 stores program instructions 41 that can be executed by a processor, the program instructions 41 being for implementing the steps in any of the error correction method embodiments described above.
According to the scheme, the text to be corrected is corrected by combining the context semantic information of the text to be corrected and the similar character information of the text to be corrected, so that the accuracy of correction is improved compared with the case that single similar character information is used.
In some embodiments, functions or modules included in an apparatus provided by the embodiments of the present disclosure may be used to perform a method described in the foregoing method embodiments, and specific implementations thereof may refer to descriptions of the foregoing method embodiments, which are not repeated herein for brevity.
The foregoing description of various embodiments is intended to highlight differences between the various embodiments, which may be the same or similar to each other by reference, and is not repeated herein for the sake of brevity.
In the several embodiments provided in the present application, it should be understood that the disclosed method and apparatus may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of modules or units is merely a logical functional division, and there may be additional divisions of actual implementation, e.g., units or components may be combined or integrated into another system, or some features may be omitted, or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical, or other forms.
In addition, each functional unit in the embodiments of the present application may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units. The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application may be embodied in essence or a part contributing to the prior art or all or part of the technical solution in the form of a software product stored in a storage medium, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute all or part of the steps of the methods of the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
Claims (11)
1. An error correction method, comprising:
acquiring a text to be corrected;
performing text processing on the text to be corrected by using an error correction model to obtain text processing information, wherein the text processing information comprises context semantic information of the text to be corrected and similar character information in the text to be corrected, and the error correction model comprises a first sub-network and a second sub-network; the first sub-network is used for acquiring context semantic information of the text to be corrected, and the second sub-network is used for acquiring similar character information in the text to be corrected;
predicting the text processing information by using the error correction model to obtain a prediction error correction result of the text to be corrected;
the text processing is performed on the text to be corrected by using the correction model to obtain text processing information, and the text processing method comprises the following steps: performing graph convolution operation on the text to be corrected by utilizing at least one convolution layer in the second sub-network to obtain the similar character information; the graph rolling operation adopts a similar graph corresponding to the convolution layer, wherein the similar graph comprises a plurality of confusable characters and similar weights among the confusable characters;
The performing a graph convolution operation on the text to be corrected by using at least one convolution layer in the second sub-network to obtain the similar character information includes: taking each convolution layer as a current convolution layer; performing convolution operation on the input information of the current convolution layer and a similar graph by using the current convolution layer to obtain a first graph convolution result of the current convolution layer; the input information of the first layer of the convolution layer is the text to be corrected, and the input information of the non-first layer of the convolution layer is the output result of the former convolution layer; obtaining an output result of the current convolution layer based on a first graph convolution result of the current convolution layer and output results of all convolution layers or part of convolution layers before the current convolution layer; obtaining the similar character information based on the output result of the last layer of the convolution layer;
the step of performing convolution operation on the input information of the current convolution layer and the similar graph by using the current convolution layer to obtain a first graph convolution result of the current convolution layer includes: respectively carrying out convolution operation on the input information of the current convolution layer and at least one similar graph by utilizing the current convolution layer to obtain a second graph convolution result corresponding to each similar graph of the current convolution layer; wherein confusable characters in different ones of the similarity graphs of each of the convolutional layers are for different input modes; carrying out weighted summation on the second graph convolution result corresponding to each similar graph of the current convolution layer to obtain a first graph convolution result of the current convolution layer; the weight of the second graph convolution result corresponding to the similar graph is determined based on the input mode adopted by the text to be corrected.
2. The method of claim 1, wherein the similar character information in the text to be corrected includes similar characters of original characters in the text to be corrected and similarities between the original characters and similar characters.
3. The method of claim 1, wherein prior to performing a graph convolution operation on the text to be error corrected using at least one convolution layer in the second sub-network to obtain the similar character information, the method further comprises:
acquiring a back-deleted text corpus, wherein the back-deleted text corpus comprises character pairs consisting of characters before and after modification;
selecting a plurality of character pairs according to similar identification standards corresponding to the input modes adopted by the character pairs;
and determining confusable characters based on the selected character pairs.
4. The method of claim 3, wherein the input mode includes at least one of a handwriting input mode, a stroke input mode, and a wubi input mode;
and selecting a plurality of character pairs according to the similarity identification standard corresponding to the input mode adopted by the character pairs, wherein the character pairs comprise:
if the input mode adopted by the character pairs is the handwriting input mode, selecting the character pairs with the comprehensive similarity of the character structure and the character components being greater than or equal to a preset similarity threshold;
If the input mode adopted by the character pairs is the stroke input mode, selecting the character pairs with the same preset number of strokes;
and if the input mode adopted by the character pairs is the five-stroke input mode, selecting the character pairs with the editing distance of the five-stroke codes smaller than or equal to a preset distance threshold value.
5. A method according to claim 3, wherein said determining confusable characters based on said selected ones of said pairs of characters comprises:
determining the smoothness of the characters before modification and the characters after modification in the selected character pairs in the pruned text corpus respectively;
and if the smoothness corresponding to the character before modification is smaller than the preset smoothness and smaller than the smoothness corresponding to the character after modification, taking the character in the character pair as the confusable character.
6. The method according to claim 1, wherein the error correction model is trained by at least the following steps:
pre-training the first sub-network by using a first sample text and an actual error condition of the first sample text;
training the first sub-network and the second sub-network by using the second sample text and the actual error correction result of the second sample text.
7. The method of claim 6, wherein the first sample text is obtained by randomly replacing at least one character of a normal text corpus; and/or the number of the groups of groups,
the second sample text is obtained based on a back-deleted text corpus, wherein the back-deleted text corpus is used for collecting and obtaining confusable characters, and the error correction model is used for carrying out graph convolution by utilizing a similarity graph obtained based on the confusable characters to obtain similar character information of the input text.
8. The method of claim 1, wherein the predicted error correction results for the text to be error corrected include character error correction results for each character in the text to be error corrected and a predicted probability for each character error correction result;
after the text processing information is predicted by using the error correction model to obtain a predicted error correction result of the text to be corrected, the method further comprises the following steps:
the character error correction result with the prediction probability exceeding a preset probability threshold is used as a final error correction result;
and correcting the text to be corrected by using the final correction result.
9. An error correction device, comprising:
the acquisition module is used for acquiring the text to be corrected;
The processing module is used for carrying out text processing on the text to be corrected by using an error correction model to obtain text processing information, wherein the text processing information comprises context semantic information of the text to be corrected and similar character information in the text to be corrected, and the error correction model comprises a first sub-network and a second sub-network; the first sub-network is used for acquiring context semantic information of the text to be corrected, and the second sub-network is used for acquiring similar character information in the text to be corrected;
the result prediction module is used for predicting the text processing information by using the error correction model to obtain a prediction error correction result of the text to be corrected;
the processing module performs text processing on the text to be corrected by using a correction model to obtain text processing information, and the processing module comprises: performing graph convolution operation on the text to be corrected by utilizing at least one convolution layer in the second sub-network to obtain the similar character information; the graph rolling operation adopts a similar graph corresponding to the convolution layer, wherein the similar graph comprises a plurality of confusable characters and similar weights among the confusable characters;
The performing a graph convolution operation on the text to be corrected by using at least one convolution layer in the second sub-network to obtain the similar character information includes: taking each convolution layer as a current convolution layer; performing convolution operation on the input information of the current convolution layer and a similar graph by using the current convolution layer to obtain a first graph convolution result of the current convolution layer; the input information of the first layer of the convolution layer is the text to be corrected, and the input information of the non-first layer of the convolution layer is the output result of the former convolution layer; obtaining an output result of the current convolution layer based on a first graph convolution result of the current convolution layer and output results of all convolution layers or part of convolution layers before the current convolution layer; obtaining the similar character information based on the output result of the last layer of the convolution layer;
the step of performing convolution operation on the input information of the current convolution layer and the similar graph by using the current convolution layer to obtain a first graph convolution result of the current convolution layer includes: respectively carrying out convolution operation on the input information of the current convolution layer and at least one similar graph by utilizing the current convolution layer to obtain a second graph convolution result corresponding to each similar graph of the current convolution layer; wherein confusable characters in different ones of the similarity graphs of each of the convolutional layers are for different input modes; carrying out weighted summation on the second graph convolution result corresponding to each similar graph of the current convolution layer to obtain a first graph convolution result of the current convolution layer; the weight of the second graph convolution result corresponding to the similar graph is determined based on the input mode adopted by the text to be corrected.
10. An electronic device comprising a memory and a processor for executing program instructions stored in the memory to implement the method of any one of claims 1 to 8.
11. A computer readable storage medium having stored thereon program instructions, which when executed by a processor, implement the method of any of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011627609.3A CN112800752B (en) | 2020-12-31 | 2020-12-31 | Error correction method, apparatus, device and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011627609.3A CN112800752B (en) | 2020-12-31 | 2020-12-31 | Error correction method, apparatus, device and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112800752A CN112800752A (en) | 2021-05-14 |
| CN112800752B true CN112800752B (en) | 2023-12-01 |
Family
ID=75807918
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011627609.3A Active CN112800752B (en) | 2020-12-31 | 2020-12-31 | Error correction method, apparatus, device and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112800752B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114896965B (en) * | 2022-05-17 | 2023-09-12 | 马上消费金融股份有限公司 | Text correction model training method and device, text correction method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110442870A (en) * | 2019-08-02 | 2019-11-12 | 深圳市珍爱捷云信息技术有限公司 | Text error correction method, device, computer equipment and storage medium |
| CN111079412A (en) * | 2018-10-18 | 2020-04-28 | 北京嘀嘀无限科技发展有限公司 | Text error correction method and device |
| CN111523306A (en) * | 2019-01-17 | 2020-08-11 | 阿里巴巴集团控股有限公司 | Text error correction method, device and system |
| CN112016303A (en) * | 2020-09-07 | 2020-12-01 | 平安科技（深圳）有限公司 | Text error correction method, device and equipment based on graph neural network and storage medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8150161B2 (en) * | 2008-09-22 | 2012-04-03 | Intuit Inc. | Technique for correcting character-recognition errors |
-
2020
- 2020-12-31 CN CN202011627609.3A patent/CN112800752B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079412A (en) * | 2018-10-18 | 2020-04-28 | 北京嘀嘀无限科技发展有限公司 | Text error correction method and device |
| CN111523306A (en) * | 2019-01-17 | 2020-08-11 | 阿里巴巴集团控股有限公司 | Text error correction method, device and system |
| CN110442870A (en) * | 2019-08-02 | 2019-11-12 | 深圳市珍爱捷云信息技术有限公司 | Text error correction method, device, computer equipment and storage medium |
| CN112016303A (en) * | 2020-09-07 | 2020-12-01 | 平安科技（深圳）有限公司 | Text error correction method, device and equipment based on graph neural network and storage medium |
Non-Patent Citations (3)
| Title |
|---|
| Andreas Komninos 等.Designed with Older Adults to SupportBetter Error Correction in SmartPhone Text Entry: The MaxieKeyboard.MobileHCI '15: Proceedings of the 17th International Conference on Human-Computer Interaction with Mobile Devices and Services Adjunct.2015,第797–802页. * |
| 基于上下文语义的新闻人名纠错方法;杨越;黄瑞章;魏琴;陈艳平;秦永彬;;电子科技大学学报(第06期);全文 * |
| 基于细粒度词表示的命名实体识别研究;林广和;张绍武;林鸿飞;;中文信息学报(第11期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112800752A (en) | 2021-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108287858B (en) | Semantic extraction method and device for natural language | |
| CN107836000B (en) | Improved artificial neural network method and electronic device for language modeling and prediction | |
| CN109960726B (en) | Text classification model construction method, device, terminal and storage medium | |
| EP4027268A1 (en) | Training method for classification model, sample classification method and apparatus, and device | |
| WO2019084867A1 (en) | Automatic answering method and apparatus, storage medium, and electronic device | |
| CN107301865B (en) | Method and device for determining interactive text in voice input | |
| CN109032375B (en) | Candidate text sorting method, device, equipment and storage medium | |
| JP5744228B2 (en) | Method and apparatus for blocking harmful information on the Internet | |
| CN111046133A (en) | Question-answering method, question-answering equipment, storage medium and device based on atlas knowledge base | |
| US20060206313A1 (en) | Dictionary learning method and device using the same, input method and user terminal device using the same | |
| US20230385409A1 (en) | Unstructured text classification | |
| CN112906392B (en) | Text enhancement method, text classification method and related device | |
| CN110334179B (en) | Question-answer processing method, device, computer equipment and storage medium | |
| CN110163181B (en) | Sign language identification method and device | |
| KR20230040951A (en) | Speech recognition method, apparatus and device, and storage medium | |
| CN112395385B (en) | Text generation method and device based on artificial intelligence, computer equipment and medium | |
| US20210390370A1 (en) | Data processing method and apparatus, storage medium and electronic device | |
| CN111651986B (en) | Event keyword extraction method, device, equipment and medium | |
| CN112560452B (en) | Method and system for automatically generating error correction corpus | |
| CN109284358B (en) | Chinese address noun hierarchical method and device | |
| CN111506726B (en) | Short text clustering method and device based on part-of-speech coding and computer equipment | |
| CN111444905B (en) | Image recognition method and related device based on artificial intelligence | |
| CN110347806B (en) | Original text screening method, original text screening device, original text screening equipment and computer readable storage medium | |
| CN108803890B (en) | Input method, input device and input device | |
| CN112800752B (en) | Error correction method, apparatus, device and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |