A manuscript review method and system

Abstract

The present application is about a manuscript review method and system, which is applied to the manuscript review system, including obtaining the text to be reviewed input by the user; calling the common error checking algorithm to check whether there are similar-sounding characters and similar-looking characters in the text to be reviewed; calling the mechanical checking algorithm and the state machine algorithm to check whether there are custom error-prone words and target vocabulary in the text to be reviewed; outputting the review results corresponding to the text to be reviewed. This application uses a large language model as the technical foundation, and greatly improves the intelligence level of Chinese proofreading and polishing capabilities by using high-quality data to learn a variety of special sub-tasks. In professional fields such as news publishing, media manuscripts, and government documents, it can assist professional users to improve proofreading quality, increase proofreading speed, and reduce error rates, focusing on solving wording and expression problems such as inappropriate wording and mixed sentence patterns, making sentence expression more fluent, and empowering the high-quality development of current and future language and writing work.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent manuscript review, and in particular to a manuscript review method and system. Background Art

[0002] In the era of information explosion, the amount of content released by media companies has shown exponential growth, and content is released and updated frequently. In order to ensure the accuracy of the content, manual proofreading and inspection are indispensable. The process is tedious and boring. However, insufficient staffing or insufficient professional knowledge reserves result in content errors that often slip through the net.

[0003] In order to assist manual proofreading and reduce the error rate, proofreading systems came into being. They analyze and correct texts based on natural language processing, knowledge graphs and other technologies. They can automatically detect typos, grammatical errors, punctuation problems, etc. in manuscripts, and provide correction suggestions, thereby greatly improving the accuracy and readability of manuscripts.

[0004] At present, most of the proofreading software systems on the market can only correct common errors. When it is necessary to proofread professional documents in a specific industry, they are unable to recognize their proper nouns or target vocabulary, making it difficult to correct proprietary knowledge. Summary of the invention

[0005] In order to solve or partially solve the problems existing in the relevant technologies, this application provides a manuscript review method and system. Based on advanced technologies such as natural language processing, knowledge graph and image recognition, it introduces the AI ​​large language model proofreading function, uses high-quality large-scale corpus training, and greatly improves the proofreading ability of content published in specific industries.

[0006] The first aspect of the present application provides a manuscript review method, which is applied to a manuscript review system and is characterized by comprising:

[0007] Get the text to be reviewed entered by the user;

[0008] Calling common error checking algorithms to check whether there are similar-sounding and similar-looking characters in the text to be reviewed;

[0009] Call the mechanical checking algorithm and the state machine algorithm to compare the words in the text to be reviewed with the proper nouns in the custom word library, and check whether there are custom error-prone words and target words in the text to be reviewed;

[0010] Output the review results corresponding to the text to be reviewed.

[0011] Among them, calling common error checking algorithms to check whether there are similar-sounding characters and similar-looking characters in the text to be reviewed includes:

[0012] Divide the text to be reviewed into sentences to obtain original sentences;

[0013] Perform sentence conversion on the original sentence to obtain a control sentence and wrong word suggestions, calculate the control sentence and the original sentence, and determine whether to adopt the wrong word suggestions.

[0014] Among them, calling the mechanical inspection algorithm and the state machine algorithm includes:

[0015] Preliminary correction of the text to be reviewed based on the mechanical checking algorithm, identification of possible errors and provision of suggested words to obtain the preliminary review text;

[0016] The initial text is corrected again based on the state machine algorithm, and errors are judged and reported through the state transfer and matching of the state machine.

[0017] Among them, the control sentences obtained by sentence conversion of the original sentence include:

[0018] According to the model vocabulary, the original sentence is converted into a sentence integer vector, and the sentence integer vector is segmented and filtered to obtain the original word;

[0019] When the original word is a single word, it is directly output as the word to be processed;

[0020] When the original word is not a single word, the word to be processed is determined;

[0021] Get the pinyin corresponding to the word to be processed, and replace the original word with the pinyin to obtain a new sentence integer vector;

[0022] Input the new sentence integer vector into the preset model, output the new sentence and typo suggestions, and convert the new sentence into a control sentence.

[0023] Among them, the preliminary correction of the text to be reviewed based on the mechanical inspection algorithm includes:

[0024] Select the initialization action based on the number of words in the original sentence;

[0025] Choose to build a candidate word state machine, an original word state machine, or build a mechanical check matrix for word matching and verification to obtain possible errors and suggested words.

[0026] The calculation of the control sentence and the original sentence and the determination of whether to adopt the suggestion for correcting the wrong word include:

[0027] Determine whether the control sentence is the same as the original sentence based on the credibility calculation;

[0028] If they are the same, no errors are found;

[0029] If they are different, the comparison sentence and the original sentence are segmented to obtain the segmentation results, the segmentation results are input into the Kenlm model to calculate the sentence scores, and the score differences of the sentence scores are compared;

[0030] When the score difference exceeds the preset threshold, the wrong word suggestion is adopted.

[0031] Among them, the words to be processed obtained by judgment include:

[0032] Judge whether the original word exists;

[0033] If it exists, it is judged as a normal word and does not need to be processed. If it does not exist, judge whether it is a proper noun;

[0034] If it is, it is judged as a normal word and does not need to be processed. If not, judge whether the first and last characters of the original word are in the error-prone word dictionary library. When confirmed, output the first and last characters of the original word as the word to be processed;

[0035] In addition, when the original word starts with "elephant", output "elephant" as the word to be processed.

[0036] Among them, the initialization actions selected based on the number of characters in the original sentence include:

[0037] Obtain the number of characters in the original sentence;

[0038] When the number of characters in the original sentence is equal to 2, construct a candidate word state machine and an original word state machine;

[0039] When the number of characters in the original sentence is greater than 2, construct a mechanical inspection matrix.

[0040] Among them, the secondary correction of the preliminary review text based on the state machine algorithm includes:

[0041] Construct an initial tree;

[0042] After segmenting the preliminary review text, add it to the initial tree as a node to obtain a node tree;

[0043] Based on the node tree, construct a state machine, and convert the information of the nodes and edges in the node tree into the states and transition rules of the state machine.

[0044] Based on the state machine, perform traversal matching to perform lexical matching on the given string.

[0045] The second aspect of this application provides a manuscript proofreading system applicable to the manuscript proofreading method shown in the first aspect, including:

[0046] A foreground presentation layer for realizing user interface presentation, including an editor, check result annotation, and a custom word library;

[0047] A core business layer connected to the foreground presentation layer through an API for providing business processing capabilities;

[0048] A data layer for providing system business processing capabilities and realizing inspection logics of multiple dimensions and multiple objectives;

[0049] A management support layer serving the core business layer and the data layer for providing system management and control capabilities.

[0050] The technical solution provided by this application may have the following beneficial effects:

[0051] This application provides a manuscript review method and system, which uses a large language model as the technical foundation, and greatly improves the intelligence level of Chinese proofreading and polishing capabilities by using high-quality data to learn a variety of special sub-tasks. In professional fields such as news publishing, media manuscripts, and government documents, it can assist professional users to improve proofreading quality, increase proofreading speed, and reduce error rates, focusing on solving wording and expression problems such as inappropriate wording and mixed sentence patterns, making sentences more fluent, and empowering the high-quality development of current and future language and writing work.

[0052] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and other objects, features and advantages of the present application will become more apparent through a more detailed description of exemplary embodiments of the present application in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the present application.

[0054] Figure 1 It is a flowchart of a manuscript review method shown in an embodiment of the present application;

[0055] Figure 2 It is a structural diagram of the manuscript review system shown in the embodiment of the present application. DETAILED DESCRIPTION

[0056] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present application more thorough and complete, and to fully convey the scope of the present application to those skilled in the art.

[0057] It should be understood that although the terms "first", "second", "third", etc. may be used in this application to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of this application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0058] In the description of the present application, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0059] Unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0060] The technical solution of the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0061] Embodiment 1:

[0062] like Figure 2 The manuscript review methods shown include:

[0063] The front-end presentation layer is used to implement the user interface presentation, including editors, inspection result annotations, and custom vocabulary.

[0064] The system front-end and back-end are connected through API. The back-end provides business processing capabilities, and the front-end realizes the display of various user interfaces. It mainly includes: editor, inspection result annotation, vocabulary customization and other functions.

[0065] The core business layer is connected to the front-end presentation layer through an API to provide business processing capabilities, including: Chinese inspection module, NLP semantic algorithm package, character-related inspection module, English inspection module, text inspection core algorithm package, target vocabulary inspection module, phonetic-similar character matching module, and graphic-similar character matching module.

[0066] The data layer is used to provide system business processing capabilities and implement multi-dimensional and multi-target inspection logic, including artificial intelligence models, big data corpus models, core vocabulary libraries, target vocabulary libraries, and user-defined vocabulary libraries.

[0067] The management support layer serves the core business layer and the data layer, and is used to provide system management and control capabilities, and implement functions such as user login and authentication, vocabulary management, operation and scheduling.

[0068] The core business layer and data layer are the providers of the system's business processing capabilities, and implement inspection logic for multiple dimensions and multiple targets. It supports user-defined vocabulary and strengthens the inspection capabilities and scope according to the user's own characteristics. The management support layer serves the core business layer and data layer, and provides system management and control capabilities.

[0069] This system supports providing text checking functions for customers' third-party systems through an open API system, achieving seamless integration between systems. The system background integrates a natural language processing algorithm package to check common errors and knowledge errors.

[0070] The system has built-in dictionaries of similar-sounding characters and similar-looking characters, which can be upgraded as needed.

[0071] Related modules and files for setting up common error checking algorithms include: machine learning model, model vocabulary file, pyltp word segmentation, pyltp word segmentation annotation, kenlm language model, easy-to-error character dictionary library, and Chinese pinyin library.

[0072] The system background checks words related to culture, society and other industry fields through a custom word library. Currently, the custom word libraries supported are: (1) person positions; (2) person rankings; (3) special terms; (4) target vocabulary; (5) easily confused words; (6) attention words; (7) attention personnel; (8) words exempt from inspection. The system has developed its own mechanical inspection algorithm and state machine algorithm to match the words defined in the word library. If a word that meets the conditions is found, it will be reported.

[0073] The goal of a custom word library is to solve industry-related problems. It needs to be continuously adjusted and optimized according to the development of the industry and the deepening of system use. In the early stage of system use, the number of word library entries is small, and the probability of missed reports is high. As the number of word library entries increases and the accuracy improves, the probability of missed reports will decrease. After the content of the entry is replaced by a new expression, the probability of use decreases, the related missed reports will be fewer, and the false alarms caused may increase. According to judgment, deleting outdated entries will result in very low-concept missed reports, and the related false alarms will also be reduced. Real-time adjustment and optimization of entries are achieved based on the word library management module.

[0074] Embodiment 2

[0075] like Figure 1 The manuscript review method shown in the embodiment is applied to the manuscript review system shown in the first embodiment, and includes:

[0076] S1. Obtain the text to be reviewed input by the user;

[0077] S2. Call a common error checking algorithm to check whether there are characters with similar pronunciations or similar appearances in the text to be reviewed.

[0078] The main steps of common error checking are described as follows.

[0079] S201.分句 the article to be checked. Try to control each sentence within 32 characters. An overly long sentence that cannot be split will affect the model's calculation performance.

[0080] S202.Convert the sentence into a sentence integer vector according to the model vocabulary. According to the model vocabulary, convert the characters at the corresponding positions of the sentence into integers corresponding to the subscripts in the vocabulary to form an integer vector. For characters not existing in the model vocabulary, use spaces to replace them.

[0081] S203.Perform word segmentation on the sentence. The word segmentation is operated using the pyltp tool library.

[0082] S204.Screen the characters to be processed in the word segmentation result. Only process Chinese characters and exclude words with non-Chinese characters. For single-character cases, regard them as characters to be processed. For cases with more than one character, if it starts with

象

象

象

[0083] S205.Replace the sentence integer vector with the single character to be processed to obtain a new sentence vector. Find the corresponding pinyin of the single character to be processed in the error-prone character dictionary. If not found, use the pinyin defined in the Chinese character pinyin library. Search for this pinyin in the model dictionary and replace the value at the corresponding character position in the sentence integer vector to obtain a new sentence integer vector.

[0084] S206.Use the model to calculate the newly obtained sentence vector. To improve efficiency, sort all sentences by length. To record the sentence positions, a comparison table is used to record the position relationship before and after sorting. To improve efficiency, the model uses the concept of batch calculation, calculating 128 sentences at a time, and using blank spaces (the model dictionary subscript value of

补

[0085] S207. Obtain a new sentence based on the model calculation result and perform credibility calculation. Convert the model calculation result into a Chinese sentence. If the converted sentence is the same as the original sentence, it means that no error is found. If they are different, calculate the scores of the converted sentence and the original sentence. Use pyltp to segment the sentence, and use kenlm to calculate the scores of these segmented words as the sentence score. Compare the difference in sentence scores. If the difference exceeds the threshold, adopt this wrong character suggestion.

[0086] S3, calling the mechanical checking algorithm and the state machine algorithm, comparing the words in the text to be reviewed with the proper nouns in the custom word library, and checking whether there are custom error-prone words and target words in the text to be reviewed;

[0087] S31. The main steps of mechanical inspection are described as follows.

[0088] S311. Initialization. Accept a list or dictionary as input. For the scenario of 2 characters, enumerate the characters corresponding to each character in the Chinese Pinyin library and the Chinese glyph library as wrong character candidates, and form candidate words with another character to construct candidate word DFA and original word DFA; for the scenario of 3 characters or more, construct a mechanical check matrix to save the combination of any two characters in the word. When there are repeated combinations in multiple sentences, discard the latter combination.

[0089] S312. Match the sentence. When the sentence length is two words, use DFA for checking. Use candidate word DFA to match the sentence, and the matched result is marked as wrong; use original word DFA to match the sentence, and the matched result is marked as correct. Return the correct and wrong matching results. When the sentence length exceeds 2 words, use mechanical check matrix for checking.

[0090] S313. Mechanical check matrix check process. Slide and cut the sentence using a window size of 3, sliding one word each time to obtain a group of 3-word window words. For each 3-word window word obtained by cutting, check the words composed of 12, 13 and 23 bits. If they exist in the mechanical check matrix, obtain the corresponding words in the mechanical check matrix, record the position of the words and the window words obtained by the current cutting, and obtain the mechanical check result matrix.

[0091] S314. Mechanical inspection result matrix inspection process. For each word, calculate the maximum number of different characters. When the word length is 7 characters or less, the maximum allowed difference is 1 character; when the word length is 8-12 characters, the maximum allowed difference is 2 characters; when the word length is 13 or more, the maximum allowed difference is 25% of the characters of the word length. Calculate the minimum matching length, which is twice the word length minus the maximum number of different characters. If the number of occurrences of a word in the mechanical inspection result matrix is ​​less than the minimum matching length, discard the word. Get the number of characters that the word matches in the sentence, and count one character at most once. If the number of characters is less than the word length minus the maximum number of different characters, discard the word. Then calculate the length of this word as a suggested word. If the number of characters exceeds half of the sentence length, the word length is extended by 1. If it is less than half, the word maintains its original length as the suggested word length. In this way, the suggested word and its fuzzy matching length are obtained.

[0092] S315. Check the correctness of the suggested word in the current sentence. If the word exists in the sentence, the check passes. Calculate the maximum number of difference words, which is 1 when the word length is 6 words or less, and 2 otherwise. If the word carries the target tag, that is, it exists in the target vocabulary array, then calculate the maximum number of difference words in the same way as in the previous step. Use this length to divide the sentence length to get the maximum difference threshold. If the sentence length is less than the length of the suggested word, calculate the Li distance of the sentence relative to the suggested word. If it is greater than or equal to the maximum number of difference words, discard the suggested word. Otherwise, record the suggested word and mark whether the suggestion is correct based on whether the suggested word exists in the sentence. If the sentence length is greater than or equal to the length of the suggested word, start from the maximum number of difference words added to the length of the suggested word and end when the maximum number of difference words is subtracted, and find the longest optional suggested word from the sentence. For each length, for the suggested word, use this length as the sliding window length on the sentence for matching. If the unique maximum number of identical words is not matched, discard the inspection result of this length. If the unique maximum number of identical characters is matched, a substring is taken from the sentence with this length of characters, and the Li distance of the sentence substring relative to the suggested word is calculated. If it is greater than or equal to the maximum number of different characters, the substring is expanded by one character to the left and right, and the Li distance is recalculated. If it is still greater than or equal to the maximum number of different characters, the suggested word is discarded. Otherwise, the suggested word and its related position and length are recorded, and whether the suggestion is correct is marked based on whether the suggested word exists in the sentence.

[0093] S316. Reduce false positives based on candidate words. If a suggestion word exists in a candidate word, and the original word of the candidate word is the same as the substring of the corresponding candidate word in the sentence, then discard the suggestion word. If a suggestion word exists in a candidate word, and the original word of the candidate word is a substring of the substring of the corresponding candidate word in the sentence, then discard the suggestion word. Remove the reported part of the sentence and continue checking until the end of the sentence is checked.

[0094] S32. The main steps of the state machine check are described as follows.

[0095] S321. Initialization. Accept a list or dictionary as input to build a DFA. For each word in the set, add it to the node tree. For the dictionary, accept the value of the dictionary as the positive word. The content passed in varies in different scenarios.

[0096] S322. Matching. Traverse the given string, use the DFA to check any added vocabulary contained, and return suggested words according to whether a dictionary is passed in. Try to match longer substrings. If the matched substring is a positive word, ignore this match.

[0097] S4. Output the proofreading result corresponding to the text to be proofread.

[0098] Specific inspection examples are as follows.

[0099] 1. Prepare custom error-prone words, add the correct word as [Zhang Sansan], and the error-prone word as [Zhang San];

[0100] 2. Input the inspection target: [The mosquito bite was very painful, and Zhang San endured it without moving.];

[0101] 3. Allocate an inspector thread to perform various inspections, including common error inspection and custom word library inspection.

[0102] 4. Example description of single-character error checking for common error checking. The model is initialized when the system starts. First, the input target is segmented into sentences, obtaining ["文字叮咬狠疼", "张三强忍着没有动"]. Taking the first sentence as an example, it is then split. Since it is not too long, the result is still ["文字叮咬狠疼"]. Then the sentence is converted into an integer vector to prepare for model calculation, obtaining [153, 457, 601, 3036, 1283, 3073]. Then pyltp and kenlm are used to check which characters need to be checked. After word segmentation, we get ["文字 / 叮咬 / 狠 / 疼"]. kenlm filtering removes ["文字 / 叮咬"], and through the error-prone character dictionary check, the character to be checked is ["狠"], with the pinyin ["hen"], and the corresponding index in the vocabulary is 12993. The replaced integer vector is [153, 457, 601, 3036, 12993, 3073]. The model performs calculations, omitting the intermediate batch alignment and length padding processes, and obtaining the result vector [452, 452, 452]. Aligning to position 4 in the original sentence, the character corresponding to index 452 in the vocabulary is ["很"]. After replacement, the new sentence is ["文字叮咬很疼"]. Calculate the scores of the original sentence and the new sentence, which are -23 and -18 respectively, and the difference is 5, meeting the threshold condition, so this calculation result is adopted. Other errors, such as ["文字"] should be ["蚊子"], are detected by the Chinese word and phrase checking module. The process structure of word and phrase checking is similar to that of single-character checking, with differences in data preprocessing and model weights, and the other processes are basically the same, so they will not be described again. The user gets the check result and is advised to change it to ["蚊子叮咬很疼"].

[0103] 5. Example description of error-prone word checking for the custom vocabulary. The input target is segmented into sentences, obtaining ["文字叮咬狠疼", "张三强忍着没有动"]. Taking ["张三强忍着没有动"] as an example, the system reads the latest error-prone word definitions from the error-prone word list, obtaining the error-prone word ["张三"]. The wrong word is defined as the object to be checked for DFA. When performing the match, ["张三"] is directly matched, and the positive word ["张三三"] is provided.

[0104] Finally, it should also be noted that in this article, relationships such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variant is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0105] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0106] The embodiments of the present application have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.< / pad>

Claims

1. A manuscript review method, applied to a manuscript review system, characterized in that: Including: Obtain the text to be proofread input by the user; Call a common error checking algorithm to check whether there are homophonic and similar-looking characters in the text to be proofread, including: splitting the text to be proofread into original sentences; performing sentence conversion on the original sentences to obtain comparison sentences and typo suggestions, calculating the comparison sentences and the original sentences, and determining whether to adopt the typo suggestions; Call a mechanical checking algorithm and a state machine algorithm, compare the words in the text to be proofread with the proper nouns in the custom word library, and check whether there are custom error-prone words and target vocabulary in the text to be proofread; The calling of the mechanical checking algorithm and the state machine algorithm includes: initially correcting the text to be proofread based on the mechanical checking algorithm, judging possible errors and providing suggested words to obtain a preliminary proofreading text; performing secondary correction on the preliminary proofreading text based on the state machine algorithm, and judging errors and providing a report through the state transition and matching of the state machine; The initially correcting the text to be proofread based on the mechanical checking algorithm includes: Select an initialization action based on the number of characters in the original sentence; Select to construct a candidate word state machine, an original word state machine, or construct a mechanical checking matrix for word matching and verification to obtain the possible errors and the suggested words; The performing secondary correction on the preliminary proofreading text based on the state machine algorithm includes: Construct an initial tree; After splitting the preliminary proofreading text, add it to the initial tree as a node to obtain a node tree; Construct a state machine based on the node tree, and convert the information of the nodes and edges in the node tree into the states and transition rules of the state machine; Perform traversal and matching based on the state machine to perform lexical matching on a given string; Output the proofreading result corresponding to the text to be proofread.

2. The manuscript review method according to claim 1, characterized in that: The performing sentence conversion on the original sentence to obtain a comparison sentence includes: Convert the original sentence according to the model word list to obtain a sentence integer vector, split the sentence integer vector and filter to obtain original characters; When the original character is a single character, directly output it as the character to be processed; When the original character is not a single character, judge to obtain the character to be processed; Obtain the pinyin corresponding to the character to be processed, and replace the original character with the pinyin to obtain a new sentence integer vector; Input the new sentence integer vector into a preset model, output a new sentence and a typo suggestion, and convert the new sentence into a comparison sentence.

3. The manuscript review method according to claim 2, characterized in that: The judging to obtain the character to be processed includes: Judge whether the original character exists; If it exists, judge it as a normal word that does not need to be processed. If it does not exist, judge whether it is a proper noun; If it is, judge it as a normal word that does not need to be processed. If it is not, judge whether the first and last characters of the original character are in the error-prone character dictionary library. When confirmed, output the first and last characters of the original character as the character to be processed; And when the original character starts with "xiang", output "xiang" as the character to be processed.

4. The manuscript review method according to claim 1, characterized in that: The calculating the comparison sentence and the original sentence and determining whether to adopt the typo correction suggestion includes: Based on credibility calculation, judge whether the comparison sentence is the same as the original sentence; If they are the same, no error is found; If they are different, split the comparison sentence and the original sentence respectively to obtain a word segmentation result, input the word segmentation result into the kenlm model to calculate the sentence score, and compare the score difference of the sentence scores; When the score difference exceeds a preset threshold, the wrong character suggestion is adopted.

5. The manuscript review method according to claim 1, characterized in that: The selecting of the initialization action based on the number of words in the original sentence includes: Obtain the number of words in the original sentence; When the number of characters in the original sentence is equal to 2, constructing the candidate word state machine and the original word state machine; When the number of characters in the original sentence is greater than 2, a mechanical checking matrix is ​​constructed.

6. A manuscript review system, applicable to the manuscript review method according to claim 1, characterized in that: include: The front-end presentation layer is used to implement the user interface presentation, including editors, inspection result annotations, and custom vocabulary; The core business layer is connected to the front-end presentation layer through an API to provide business processing capabilities; The data layer is used to provide system business processing capabilities and implement inspection logic in multiple dimensions and for multiple purposes; The management support layer serves the core business layer and the data layer and is used to provide system management and control capabilities.

Images