Text checking method based on large language model

Abstract

The invention discloses a text review method based on a large language model, and relates to the technical field of text review, and the method comprises the steps: carrying out the preliminary review of target text information needing to be subjected to text review, obtaining a second to-be-tested text feature, and obtaining a second to-be-tested text feature according to a first to-be-tested text feature of a composer to which the target text information belongs; counting weight receivers according to a change trend stable period and a change trend orientation synchronization condition between the to-be-tested text features 1 and 2, counting a target receiver 1 of an audience to which the target text information belongs, obtaining a target receiver 2 required by a composer, and sending the target receiver 2 to the to-be-tested text feature 1 and the to-be-tested text feature 2; if the second target receiving party is different from the first target receiving party and the second target receiving party is partially different from the first target receiving party, counting difference correlation trend text features, checking target text information according to the first to-be-tested text features and the difference correlation trend text features, and comparing and checking obtained text checking results. According to the text checking method based on the large language model, the text checking efficiency can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of text proofreading, in particular to a text proofreading method based on a large language model. BACKGROUND

[0002] In recent years, with the development of deep learning technology, the emergence of pre-training language models such as Transformer architecture has brought major breakthroughs in natural language processing. These models learn the general rules and characteristics of language through pre-training on large-scale corpus, and can have a deeper understanding and representation of text.

[0003] Traditional text proofreading mainly relies on manual work, which not only requires a lot of time and effort, but also the professional level, experience and familiarity of the proofreader with the text content, etc. will affect the quality of proofreading. And traditional text proofreading is for standard correction, for text content with individualization, it is necessary to analyze the individual characteristics possessed by the individual writing party, and whether the receiving party has obvious individual characteristics is gradually analyzed, so as to correctly express the original intention of the text content, and at the same time meet the receiving degree of the receiving party, thereby improving the efficiency of text proofreading and user satisfaction. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the present application provides a text proofreading method based on a large language model.

[0005] The text proofreading method based on a large language model provided by the present application comprises:

[0006] Step S1, a text proofreading model is established according to historical text correction data, a target text information to be proofread is preliminarily proofread to obtain a text proofreading result one, popular writing text features are obtained and the popular writing text features are taken as a to-be-tested text feature two, a weight receiving party is statistically obtained according to a to-be-tested text feature one of a writing party to which the target text information belongs and a synchronous condition of a change trend stability period and a change trend direction between the to-be-tested text feature one and the to-be-tested text feature two.

[0007] Step S2, a target receiving party one of the audience of the text features belonging to the same category as the target text information is matched from the to-be-tested text feature two, a target receiving party two of the writing party's demand is obtained, if the target receiving party two is different from the target receiving party one and there is a difference with the weight receiving party, a difference correlation trend text feature is statistically obtained.

[0008] Step S3, based on the to-be-tested text feature, adjust the correction mode of the text correction model to obtain an adjusted text correction model one, use the adjusted text correction model one to correct the target text information to obtain a text correction result two, according to the difference correlation trend text feature, adjust the correction mode of the text correction model to obtain an adjusted text correction model two, use the adjusted text correction model two to correct the target text information to obtain a text correction result three, and check the difference part information of the text correction result one, the text correction result two and the text correction result three to obtain a text correction result.

[0009] Preferably, the target text information to be corrected is obtained, and the target text information is input into the text correction model for testing to obtain a text correction result one.

[0010] The individualized writing text features of the writing party in the historical period to which the target text information belongs are obtained to obtain a to-be-tested text feature one, and the popular writing text features are obtained to obtain a to-be-tested text feature two.

[0011] The statistical results of the different period change trends of the to-be-tested text feature one are obtained to obtain a change trend one, and the statistical results of the different period change trends of the to-be-tested text feature two are obtained to obtain a change trend two, and the change trends of the same period are compared to obtain a comparison result.

[0012] Preferably, if the two change trends in the comparison result are in different stable periods and the change trends are not synchronized, it is judged that the text correction result one is the correction result of the final target text information, and the target text correction result is output.

[0013] The part result in which the two change trends are in the same stable period and have the same change trend synchronization is extracted from the comparison result, a screening result is output, and according to the screening result, the weight receiving party of the text feature corresponding to the audience of the screening result in the to-be-tested text feature two is counted.

[0014] Preferably, according to the target text information, the target receiving party one of the audience of the text feature belonging to the same category as the target text information is matched from the to-be-tested text feature two, and the target receiving party two of the writing party demand is obtained.

[0015] The similarities and differences between the target receiving party two, the target receiving party one and the weight receiving party are compared, if the target receiving party two, the target receiving party one and the weight receiving party are all different, it is judged that the text correction result one is the target text correction result, and the target text correction result is output.

[0016] Preferably, if the target receiver two is different from the target receiver one and is partially different from the weight receiver, then the partially different receivers are extracted from the target receiver two, and a difference receiver one is outputted, and the partially different receivers are extracted from the weight receiver, and a difference receiver two is outputted;

[0017] According to the difference receiver one, a preprocessed text feature one is obtained by counting the favorite text features of the difference receiver one;

[0018] According to the difference receiver two, a preprocessed text feature two is obtained by corresponding screening from the to-be-tested text feature two;

[0019] The preprocessed text feature one and the preprocessed text feature two are subjected to difference variation correlation trend statistics to obtain a difference correlation trend text feature.

[0020] Preferably, according to the to-be-tested text feature one, an adjusted text proofreading model one is obtained by adjusting the proofreading mode of the text proofreading model;

[0021] According to the difference correlation trend text feature, an adjusted text proofreading model two is obtained by adjusting the proofreading mode of the text proofreading model;

[0022] The target text information is inputted into the adjusted text proofreading model one for testing to obtain a text proofreading result two;

[0023] The target text information is inputted into the adjusted text proofreading model two for testing to obtain a text proofreading result three.

[0024] Preferably, the same information part in the text proofreading result one, the text proofreading result two and the text proofreading result three is extracted, and a reserved proofreading text information is outputted;

[0025] If the three kinds of residual difference part information in the text proofreading result one, the text proofreading result two and the text proofreading result three except the reserved proofreading text information are all different, then the target text information is subjected to repeated proofreading, and a repeated proofreading information is outputted;

[0026] If the three kinds of residual difference part information in the text proofreading result one, the text proofreading result two and the text proofreading result three except the reserved proofreading text information exist two kinds of residual difference part information, and the same sub-information of the two kinds of residual difference part information is the same, then the same sub-information is extracted, and the difference sub-information of the two kinds of residual difference part information except the same sub-information is subjected to difference integration to obtain integrated sub-information;

[0027] The reserved proofreading text information, the same sub-information and the integrated sub-information are combined to obtain a target text proofreading result.

[0028] Compared with the prior art, the present application has the following characteristics and beneficial effects:

[0029] By correcting data according to historical text, a text proofreading model is established to preliminarily correct the target text information that needs to be proofread, so as to obtain a text proofreading result one. In order to reduce the over-correction or error correction of the proofreading result, further information verification is performed. According to the past individualized text features of the writing party to which the target text information belongs and the statistics of popular text features, i.e., the first to-be-tested text feature and the second to-be-tested text feature, the two to-be-tested text features are compared in terms of variation trend, so as to know the receiving party of the text provided by the writing party, i.e., the weight receiving party, and facilitate the judgment of whether the receiving party of the target text information of the writing party needs to be changed, which helps to verify the accuracy of the text proofreading result one. In order to improve the real-time performance of the information and the development variability of the information, the receiving party of the target text information is judged according to the text features of the target text information, i.e., the first target receiving party is obtained, the second target receiving party is obtained, and the three receiving parties are compared, so as to obtain the difference correlation trend text feature according to the variation of the receiving party in the comparison result, so as to adjust the proofreading mode of the text proofreading model. Then, the first to-be-tested text feature is used to adjust the text proofreading model, so as to perform second and third text proofreading processing on the target text information. Finally, the three text proofreading results obtained are verified in terms of the same part information and the difference part information. Through the above processing method, the traditional technology of only one-time text correction processing based on a general text proofreading model is avoided. Finally, whether the text proofreading result exists over-correction or does not conform to the text features of the receiving party required by the writing party to provide, which leads to the unclear understanding of the receiving party to the part information of the text proofreading result, etc. Through the above processing process, the text feature variation trend of the writing party and the text feature variation trend of the receiving party are combined to further verify the text proofreading result, improve the efficiency of the text proofreading work, and improve the user satisfaction. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a step block diagram of a text proofreading method based on a large language model according to the embodiment. DETAILED DESCRIPTION

[0031] The application will be further described in detail below in combination with the following embodiments.

[0032] REFERENCE Figure 1 A text proofreading method based on a large language model, the method comprising the following steps:

[0033] Step S1, a text proofreading model is established according to historical text correction data, a target text information to be proofread is preliminarily proofread to obtain a text proofreading result one, popular writing text features are obtained and taken as a to-be-tested text feature two, a weight receiving party is obtained according to a to-be-tested text feature one of a writing party of the target text information, and a synchronous condition of a change trend stability period and a change trend direction between the to-be-tested text feature one and the to-be-tested text feature two.

[0034] Step S2, a target receiving party one of a text feature belonging to the same category as the target text information is matched from the to-be-tested text feature two, a target receiving party two of a writing party demand is obtained, if the target receiving party two is different from the target receiving party one and different from the weight receiving party, a difference correlation trend text feature is obtained.

[0035] Step S3, an adjustment of a proofreading mode of the text proofreading model is performed based on the to-be-tested text feature one to obtain an adjusted text proofreading model one, the target text information is proofread by using the adjusted text proofreading model one to obtain a text proofreading result two, an adjustment of the proofreading mode of the text proofreading model is performed to obtain an adjusted text proofreading model two according to the difference correlation trend text feature, the target text information is proofread by using the adjusted text proofreading model two to obtain a text proofreading result three, a difference part information of the text proofreading result one, the text proofreading result two and the text proofreading result three is verified to obtain a text proofreading result.

[0036] Specifically, by correcting the historical text data, a text proofreading model is established to preliminarily correct the target text information to obtain a text proofreading result one. In order to reduce the over-correction or error correction of the proofreading result, further information verification is performed. According to the past personalized text features of the writing party to which the target text information belongs and the statistics of popular text features, that is, the first to-be-tested text feature and the second to-be-tested text feature, the two to-be-tested text features are compared in terms of variation trend, so as to know the receiving party of the text provided by the writing party, that is, the weight receiving party, so as to judge whether the receiving party of the target text information of the writing party exists change, which is helpful to verify the accuracy of the text proofreading result one. In order to improve the real-time performance of the information and the development variability of the information, the receiving party of the target text information is judged according to the text features of the target text information, that is, the first target receiving party is obtained, the second target receiving party is obtained, and the writing party demand is obtained. The three receiving parties are compared, so as to compare the receiving party variation condition according to the comparison result, and the difference correlation trend text feature is obtained by statistically comparing the corresponding text feature variation correlation trend, so as to adjust the proofreading mode of the text proofreading model. The text proofreading model is adjusted by the first to-be-tested text feature, and the target text information is subjected to second and third text proofreading processing. Finally, the three text proofreading results are verified in terms of the same part information and the difference part information. Through the above processing mode, the traditional technology is avoided, which only performs one-time text correction processing on the general text proofreading model, and finally the text proofreading result is obtained. Whether there is over-correction or the text features of the receiving party provided by the writing party are not required, which leads to the problem that the receiving party does not understand part of the information of the text proofreading result. Through the above processing process, the text feature variation trend of the writing party and the text feature variation trend of the receiving party are combined to further verify the text proofreading result, improve the efficiency of the text proofreading work, and improve the user satisfaction.

[0037] The specific step S1 includes the following sub-steps:

[0038] The target text information to be proofread is obtained, and the target text information is input into the text proofreading model for testing to obtain a text proofreading result one.

[0039] The personalized writing text features of the writing party to which the target text information belongs in the historical period are obtained to obtain the first to-be-tested text feature, and the popular writing text features are obtained to obtain the second to-be-tested text feature.

[0040] The first to-be-tested text feature is statistically compared in terms of variation trend in different periods to obtain variation trend one, the second to-be-tested text feature is statistically compared in terms of variation trend in different periods to obtain variation trend two, and variation trend one and variation trend two are compared in terms of variation trend in the same period to obtain a comparison result.

[0041] If it is judged that the two kinds of variation trends in the comparison result are in different stable periods and the variation trends are out of sync, it is judged that the text proofreading result is the proofreading result of the final target text information, and the target text proofreading result is output.

[0042] From the comparison result, the part result in which the two kinds of variation trends are in the same stable period and there is the same variation trend synchronization is extracted, the screening result is output, and according to the screening result, the weight receiver of the audience corresponding to the text feature belonging to the screening result in the second to-be-tested text feature is counted.

[0043] Specifically, as historical text correction data (such as historical text data including texts that need to be corrected in historical periods, corrected texts, etc., for example, calculating text probability by statistical language model (such as N-gram, HMM), selecting correction results more consistent with language habits. N-gram model (based on historical text statistical word sequence probability), hidden Markov model (HMM) captures error patterns), text feature one to be tested (such as personalized writing text features (here refers to word difficulty level, text arrangement, writing mode, word usage frequency, sentence length distribution, syntax complexity, punctuation usage, paragraph structure, etc. Features can be classified into comprehensive feature categories: such as common words, single text arrangement, writing structure according to time sequence, then classified as a1 category, for example, uncommon words, single text arrangement, writing structure according to time sequence, then classified as a2 category, and so on, and the optimization weight level of the category is increased in turn), trend one (if the text feature category of the writing party in historical period T1 is a1, the text feature category in historical period T2 is a2, the text feature category in historical period T3 is a5, the text feature category in historical period T4 is a3, and the text feature category in historical period T5 is a4 (here only five periods are selected for illustration, which can be extended), wherein T2 is earlier than T1, and so on, then a is taken as the y-axis, T is taken as the x-axis, and a variable curve graph is drawn, that is, trend one is Q1), text feature two to be tested (the same as the explanation of text feature one to be tested, it is necessary to explain that text feature two to be tested includes text feature one to be tested), trend two (if the popular text feature category in historical period T1 is a2, the text feature category in historical period T2 is a3, the text feature category in historical period T3 is a6, the text feature category in historical period T4 is a3, and the text feature category in historical period T5 is a2, trend two is Q2), if the two variable trends in the judgment comparison result are in different stable periods and the variable trends are different (if the popular text feature category in historical period t1 is a2, the text feature category in historical period t2 is a1, the text feature category in historical period t3 is a4, the text feature category in historical period T4 is a3, and the text feature category in historical period t5 is a4, trend two is Q21, wherein t is a historical period earlier or later than T, then Q1 and Q21 are compared, it can be seen that the curve variable direction is different in different periods, which is a characteristic, then it can be judged that the writing party has obvious irregular individualization, so it is not necessary to perform subsequent receiver statistical steps, and the text proofreading result one is directly selected as the final text proofreading result), weight receiver (if trend two is Q2, then Q1 and Q2 are compared, then the two curves have the same period curve variable direction synchronization characteristic, then it can be judged that the writing party has obvious regular individualization,The characteristics of the to-be-tested text features in the case of synchronizing the curve variation within the same period of the two curves are extracted, i.e., a2, a3, and a6, the receiver refers to which types of users the author wants to provide the authoring text for reading, etc. (Here, the user types are, for example, users in different fields, users in different age groups, etc.), the weight receiver refers to: if multiple types of users are involved, at least one type of user with a proportion greater than 50% is selected as the weight receiver.

[0044] The specific step S2 includes the following sub-steps:

[0045] According to the target text information, the target receiver one of the audience to which the text features belonging to the same category as the target text information are matched from the to-be-tested text feature two, and the target receiver two of the author's demand is obtained.

[0046] Compare the target receiver two with the target receiver one and the weight receiver. If the target receiver two, the target receiver one, and the weight receiver are all different, it is judged that the text proofreading result one is the target text proofreading result, and the target text proofreading result is output.

[0047] If the target receiver two is different from the target receiver one, and there are some different receivers with the weight receiver, the different receivers are extracted from the target receiver two, and the difference receiver one is output. The different receivers are extracted from the weight receiver, and the difference receiver two is output.

[0048] According to the difference receiver one, the love text features of the difference receiver one are counted to obtain the pre-processing text features one.

[0049] According to the difference receiver two, the pre-processing text features two are correspondingly filtered from the to-be-tested text features two.

[0050] The difference variation correlation trend statistics of the pre-processing text features one and the pre-processing text features two are performed to obtain the difference correlation trend text features.

[0051] Specifically, as the target receiver one (the text feature a4 of the target text information, then the main two types of field receivers corresponding to a4 and a5 in the to-be-tested text feature two are counted, that is, the target receiver one), the target receiver two (refers to the at least one type of field receiver specified by the writing party), if the target receiver two is different from the target receiver one and the weight receiver (then it can be judged that the writing party has obvious irregular personalization, so there is no need to perform the subsequent receiver statistical step, and the text correction result one is directly selected as the final text correction result), the difference receiver one (the target receiver one is R1 and R2, the target receiver two is R3 and R4, and the weight receiver is R4 and R5, then R3 is the difference receiver one), the difference receiver two (that is, R5), the pre-processing text feature one (which can be obtained by questionnaire survey and other data statistical methods, if the pre-processing text feature one is a3), the pre-processing text feature two (if it is a5), the difference correlation trend text feature (for example, according to the writing structure in a3, which is purely in time sequence, and the writing structure in a5, which is purely in story coherence sequence, the two writing structures used in a3 and a5 are interleaved, and when the coherence of a part of the text information is not strong, it is inserted into the time sequence to connect, and the coherence strength is judged (which can be based on the connecting words: the connecting words with weak coherence strength, such as: the connecting words representing the selection relationship, the connecting words representing the enumeration not exhausted; the connecting words with high coherence strength, such as: the connecting words representing the cause-effect relationship, the connecting words representing the progressive relationship, and the connecting words representing the conditional relationship)).

[0052] The specific step S3 includes the following sub-steps:

[0053] According to the to-be-tested text feature one, the text correction model is adjusted to obtain an adjusted text correction model one.

[0054] According to the difference correlation trend text feature, the text correction model is adjusted to obtain an adjusted text correction model two.

[0055] The target text information is input into the adjusted text correction model one for testing to obtain a text correction result two.

[0056] The target text information is input into the adjusted text correction model two for testing to obtain a text correction result three.

[0057] The same information part in the text correction result one, the text correction result two and the text correction result three is extracted, and the reserved correction text information is output.

[0058] If the three remaining difference part information in the text correction result one, the text correction result two and the text correction result three, except for the reserved correction text information, are all different, then the target text information is repeatedly corrected, and the repeated correction information is output.

[0059] If the three remaining difference part information in the text proofreading result one, the text proofreading result two and the text proofreading result three is removed, and the two remaining difference part information has the same sub-information, the same sub-information is extracted, the difference sub-information in the two remaining difference part information is removed, and the integrated sub-information is obtained by difference integration.

[0060] The reserved proofreading text information, the same sub-information and the integrated sub-information are combined into the target text proofreading result.

[0061] Specifically, the text proofreading model one is adjusted (for example, after the to-be-tested text feature one is input into the text proofreading model, the original text proofreading step is optimized, for example, the range of text proofreading is limited, for example, the text features involved in the original text proofreading model are a1-a10, and if the to-be-tested text feature one is a1-a5, the target text information proofreading is changed from the range a1-a10 to a1-a5, so that the final text proofreading result is more in line with the receiver and the satisfaction is improved), the text proofreading model two is adjusted (for example, the limited condition of the difference correlation trend text feature is input into the text proofreading model, and when the target text information is corrected, the direction of the difference correlation trend text feature is mainly followed to correct, so that the final text proofreading result is more in line with the receiver and the satisfaction is improved), the reserved proofreading text information (if the text proofreading result one, the text proofreading result two and the text proofreading result three are B1, B2 and B3 respectively, the text information of B1, B2 and B3 is compared, and the part same in the three is extracted, if it is b0), the repeated proofreading information (if the other part text information of B1, B2 and B3 except b0 is not the same, there is a great possibility of proofreading error or overproofreading, and the text proofreading is performed again), the integrated sub-information (if the sub-information of the other part text information of B2 and B3 except b0 is the same, for example, the other part text information of B2 except b0 is b1, and the other part text information of B3 except b0 is b2, the information comparison of b1 and b2 is performed, if only part of the information is the same, if it is b01, the different information is extracted, if it is b11 and b21, the meaning neutralization processing of b11 and b21 is performed (for example, the original text is: “he has a unique insight into this problem”, “he has a deep understanding of this problem”, and the integrated text is: “he has a deep and unique understanding of this problem”), to obtain an adjusted information b02, which is the integrated sub-information), and the target text proofreading result (that is, b0, b01 and b02 are integrated according to the original sentence position to obtain the final target text proofreading result).

[0062] The above are all preferred embodiments of the present application, and do not limit the protection scope of the present application, so that: all equivalent changes made according to the structure, shape, principle of the present application should be covered in the protection scope of the present application.

Claims

1. A text review method based on a large language model, characterized in that: The following steps are involved: Step S1: Establish a text review model based on historical text correction data, perform preliminary review on the target text information to be reviewed to obtain text review result 1, obtain popular writing text features and use the popular writing text features as the test text feature 2, and statistically obtain a weighted receiver based on the test text feature 1 of the author of the target text information, as well as the synchronization status of the change trend stability period and change trend direction between the test text feature 1 and the test text feature 2; Step S2, matching the target recipient 1 of the audience of the text feature belonging to the same category as the target text information from the text feature 2 to be tested, obtaining the target recipient 2 required by the writer, and if the target recipient 2 is different from the target recipient 1 and has a difference with the weighted recipient, statistically obtaining the difference correlation trend text feature; Step S3, based on the text feature 1 to be tested, the review mode of the text review model is adjusted to obtain an adjusted text review model 1, the adjusted text review model 1 is used to review the target text information to obtain a text review result 2, according to the difference-related trend text feature, the review mode of the text review model is adjusted to obtain an adjusted text review model 2, the adjusted text review model 2 is used to review the target text information to obtain a text review result 3, the text review result 1, the text review result 2 and the text review result 3 are verified for the difference part information to obtain a text review result.

2. A text review method based on a large language model according to claim 1, characterized in that: Step S1 includes: Obtaining target text information to be reviewed, and inputting the target text information into the text review model for testing to obtain a first text review result; Obtaining personalized text features of the writer of the target text information in a historical period to obtain a first feature of the text to be tested, and obtaining popular text features to obtain a second feature of the text to be tested; The first change trend of the feature 1 of the text to be tested is statistically analyzed in different periods to obtain the first change trend, the second change trend of the feature 2 of the text to be tested is statistically analyzed in different periods to obtain the second change trend, and the first change trend and the second change trend are compared in the same period to obtain a comparison result.

3. A text review method based on a large language model according to claim 2, characterized in that: Step S1 further includes: If it is determined that the two change trends in the comparison result are in different stable periods and the change trends are not in the same direction, then the first text review result is determined to be the review result of the final target text information, and the target text review result is output; Extract the partial results in which the two change trends are in the same stable period and have the same change trend towards synchronization from the comparison results, output the screening results, and based on the screening results, calculate the weighted recipients of the audience corresponding to the text features of the screening results in the second text feature to be tested.

4. A text review method based on a large language model according to claim 3, characterized in that: Step S2 includes: According to the target text information, the target recipient 1 of the audience whose text feature belongs to the same category as the target text information is matched from the second text feature to be tested, and the target recipient 2 required by the writer is obtained; Compare the similarities and differences between target recipient 2, target recipient 1 and weighted recipient. If target recipient 2, target recipient 1 and weighted recipient are all different, determine that text review result 1 is the target text review result, and output the target text review result.

5. A text review method based on a large language model according to claim 4, characterized in that: Step S2 further includes: If the target receiver 2 is different from the target receiver 1 and has a partially different receiver from the weighted receiver, then extract the partially different receiver from the target receiver 2 and output the difference receiver 1, and extract the partially different receiver from the weighted receiver and output the difference receiver 2; According to the difference receiver 1, statistically analyzing the preference text features of the difference receiver 1 to obtain a preprocessed text feature 1; According to the second difference receiver, a second preprocessing text feature is correspondingly selected from the second text feature to be tested; The preprocessed text feature 1 and the preprocessed text feature 2 are subjected to difference change correlation trend statistics to obtain a difference correlation trend text feature.

6. A text review method based on a large language model according to claim 5, characterized in that: Step S3 includes: According to the first feature of the text to be tested, adjusting the text review model to obtain an adjusted text review model 1; According to the difference-related trend text features, the text review model is adjusted to obtain an adjusted text review model 2; Input the target text information into the adjusted text review model 1 for testing, and obtain the text review result 2; The target text information is input into the second adjustment text review model for testing to obtain the third text review result.

7. A text review method based on a large language model according to claim 6, characterized in that: Step S3 further includes: Extracting the same information parts among the first text review result, the second text review result, and the third text review result, and outputting the reserved review text information; If the three remaining difference information in the text review result 1, the text review result 2, and the text review result 3, excluding the reserved review text information, are all different, the target text information is reviewed again and the repeated review information is output; If two of the three types of remaining difference information excluding the reserved review text information in the first, second, and third text review results have the same sub-information, extract the same sub-information, and perform differential integration on the difference sub-information of the two types of remaining difference information excluding the same sub-information to obtain integrated sub-information; The reserved review text information, the same sub-information and the integrated sub-information are combined into the target text review result.

Images