Error correction large model iterative training method fusing user semantics and past error correction records

Abstract

The invention provides an error correction large model iterative training method fusing user semantics and past error correction records, and the method comprises the following steps: constructing error correction metadata: constructing the error correction metadata for data including chapters and sentences by utilizing a construction mode of a large model instruction data set; inputting a to-be-corrected text specified by a user; the dynamic feedback module is called to send reviewing feedback to the user decision module by receiving output of the large reviewing model; and the user decision module receives the output result, performs satisfaction feedback of error correction suggestions and error correction results, records and reassembles satisfaction feedback results of the user, and returns the results to the dynamic feedback module. According to the invention, errors in the composition can be identified and corrected more accurately.

Description

Technical Field

[0001] The present application relates to the technical field of text error correction in large language models, and in particular to an iterative training method for a large error correction model that integrates user semantics and past error correction records. Background Art

[0002] In recent years, with the rapid development of artificial intelligence (AI) and natural language processing (NLP) technologies, automated essay grading technology has made significant progress.

[0003] Early automated article correction systems mainly relied on rule-based methods. These systems detect and correct spelling errors, grammatical errors, and punctuation errors through predefined grammar rules and dictionaries. Although these methods perform well in dealing with simple language errors, they have limitations in dealing with complex language structures and contextual understanding. With the introduction of statistical and machine learning methods, automated article correction technology has been further developed. These methods analyze large amounts of text data and learn the statistical characteristics of language, so that they can detect and correct errors more accurately. For example, methods such as hidden Markov models (HMM) and conditional random fields (CRF) are widely used in the detection of spelling and grammatical errors. In recent years, deep learning technology, especially large models based on neural networks (such as GPT-3, BERT, T5, etc.), has performed well in natural language processing tasks. These models are able to capture the complex characteristics of language and contextual information through large-scale pre-training, so that they perform well in automated article correction. For example, generative models such as GPT-3 can generate high-quality text and can understand and correct complex language errors. Modern automated article grading systems not only focus on grammatical and spelling errors, but also extend to more complex language features such as punctuation, style, coherence, and structure. These systems can provide multi-dimensional grading suggestions to help individuals improve their writing skills in an all-round way. For example, some systems can detect the coherence and logical structure of sentences and provide improvement suggestions. Large models and diverse system engineering provide grading technology with a richer interactive experience and technical accuracy, which can meet the multi-faceted experience of multiple graders.

[0004] Although large models can generate high-quality text and correction suggestions, there are still some technical problems and shortcomings in using large models to provide personalized technical feedback. In order to deal with these problems, most researchers have turned their technical issues to the reliability of review production results and the generation of dynamic review results that integrate user information, assisting review by personalized review libraries for specific users and integrating more fine-grained knowledge information. However, how to make full use of the open large language model (LLM) to control the reliability of review results and provide personalized feedback still faces the following challenges.

[0005] Technical Issue 1: Although article review technology can quickly generate review results, the correctness and reliability of the review results cannot be guaranteed, resulting in cascading step-by-step review errors and evaluations.

[0006] Large models perform well in handling local grammatical and spelling errors, but they are still limited in understanding the context and coherence of the entire article. The model may not accurately understand the logical structure and thematic coherence of long articles, resulting in incomplete or inaccurate correction suggestions. In addition, due to the limitations of the training data and algorithms of large models, the model may have biases in semantic understanding. This bias may lead to incorrect correction suggestions, especially when dealing with complex sentence structures and polysemous words. Therefore, incorrect correction suggestions generated by the model at one stage may affect subsequent correction steps, resulting in cascading step-by-step correction errors and evaluations.

[0007] Technical issue 2: Each user has different writing styles and common errors. In fact, the result generation of a large model needs to integrate the user's actual information to achieve controllable generation and improvement suggestions.

[0008] Although large models can generate high-quality texts and correction suggestions, they are still insufficient in providing personalized feedback. Each individual's writing style and common errors are different, and existing large models may not be able to provide customized improvement suggestions for each individual's specific needs. The model needs to integrate the user's actual information, such as writing style, common errors, and learning progress. However, how to effectively collect and use this information while ensuring data privacy and security is an important challenge. Large models need to have controllable generation capabilities to provide customized correction suggestions based on the user's actual needs. This requires the model to be able to flexibly adjust parameters and strategies during the generation process to adapt to the writing styles and improvement needs of different users. Summary of the invention

[0009] The present application provides an iterative training method for a large error correction model that integrates user semantics and past error correction records, which can be used to solve the technical problem of not targeting individual characteristics in the error correction process.

[0010] The present application provides an iterative training method for a large error correction model that integrates user semantics and past error correction records, the method comprising:

[0011] Step 1: Error correction metadata construction: Use the construction method of the large model instruction data set to construct error correction metadata for data including chapters and sentences; input the text to be corrected specified by the user;

[0012] Step 2: By accepting the output of the large correction model, the dynamic feedback module is called to send the review feedback to the user decision module; the user decision module accepts the output results, provides correction opinions and satisfaction feedback on the correction results, and records and reassembles the user's satisfaction feedback results and returns them to the dynamic feedback module.

[0013] Furthermore, step 1, error correction metadata construction: using the construction method of the large model instruction data set, error correction metadata is constructed for data including chapters and sentences; including:

[0014] Step 11: The user calls the composition construction module and establishes a correction prompt template to construct the correction input of the large model; the correction prompt template is used to receive the original text of the chapter to be corrected, call the semantic analysis module to generate the semantic data of the correction chapter, call the user matching module to match the user, and query the past correction types and correction records;

[0015] Step 12: Use the error correction metadata template generated in step 11 to fine-tune the open source large language parameter model.

[0016] Further, step 11: the user calls the composition construction module to establish the error correction prompt template to construct the error correction input of the large model, including:

[0017] Step 111, receiving the original text of the article to be corrected, for receiving the user's article review request, the original text includes a user review text to be reviewed and the user identity ID;

[0018] Step 112, calling the semantic analysis module to generate semantic data of the correction chapter for processing the original correction text input by the user, on the one hand by receiving external semantic information fusion requests, on the other hand by calling external large model API and tools, semantic decomposition, semantic tree construction and semantic vector generation of the correction text;

[0019] Step 113, calling the user matching module to match the user ID to query the user's past error correction type and error correction record, and constructing a user error correction request template;

[0020] The user matching module is used to integrate the user's exclusive error correction type and error correction record; first, the user matching module accepts the user ID, queries the internal knowledge base to find all the user's past error correction records, and extracts the error correction type and error correction record in the error correction record. Among them, the error correction type includes part of speech errors, verb tense errors, and noun singular and plural errors; the error correction record includes the sentence to be corrected and the corrected text. After extraction, the text is constructed according to the following error correction request template:

[0021] Error correction request template: "Based on your past error correction records [error correction records and error types extracted from the user's past error correction system], the problems you often encounter include [list the common error correction types of users] Please pay attention to these aspects and check other possible errors at the same time";

[0022] Step 114, the generated sub-text data is packaged for all samples according to the following error correction metadata template;

[0023] Input:"Please correct the error according to the following information:

[0024] Sentence to be corrected: [Original text to be corrected entered by the user]

[0025] Semantics: [`I have apple.ROOT aux det obj']

[0026] Semantic vector: [0.12, 0.31, 0.52, 0.71, 0.12, 0.25, 0.48, 0.63, 0.81]

[0027] Common error types of users: [types of errors that users have corrected in the past]

[0028] Error correction request template data: "Based on your past error correction records [Example 1,…], the problems you often encounter include []; please pay attention to these aspects and check for other possible errors"

[0029] OutputOutPut: The modified sentence and its error type.

[0030] Furthermore, the error correction text is semantically decomposed, a semantic tree is constructed, and a semantic vector is generated as follows:

[0031] Step 1121, semantic decomposition refers to using the existing summary model to extract the core meaning of the text to be corrected and generate an abbreviated sentence; the user inputs the task query text to the summary model, and the model generates a result; the task query text is: "According to our understanding, the most concise meaning of this sentence is: [here is the text to be corrected]";

[0032] Step 1122, semantic tree construction: according to the above sentence, call the semantic tree generation tool to generate the semantic tree in the text chapter to be corrected; wherein the semantic dependency tree is generated by the syntax tool and grammar analyzer in the StandFordNLP tool;

[0033] Step 1123, semantic vector generation: the text and dependency relationships generated by the semantic decomposition process and the semantic tree construction process are re-represented by digital vectorization, and the generated text is converted into a digital representation with specific semantic relationships.

[0034] Further, step 12, using the error correction metadata template generated in step 11, fine-tuning the open source large language parameter model; including:

[0035] Given a large language model, use the error correction metadata template constructed in step 11, take the input Input and output Output as the training input and training target output of the large language model respectively, and through parameter learning and training, the large language model has the error correction capability, and finally becomes a corrected large language model; the method is as follows:

[0036] Output = LLM (error correction metadata) = LLM (original text, error correction semantic information, user error correction record information) (1)

[0037] Furthermore, in step 2, by receiving the output of the large grading model, the dynamic feedback module is called to send the grading feedback to the user decision module; the user decision module receives the output result, provides feedback on the correction opinions and the satisfaction of the correction results, and records and reassembles the user's satisfaction feedback results and returns them to the dynamic feedback module, including:

[0038] Step 21, receiving the output of the correction model after the current sentence to be corrected is corrected, and calling the dynamic feedback module to provide feedback on the user's current review result; including obtaining the result of the correction of the sentence to be corrected from the correction model; assuming that the output of the correction model is 0 m , output O m Contains preliminary error correction information and suggestions;

[0039] ,Step 22, the dynamic feedback module will judge the execution of the error correction result of the large model according to the current state, or generate corresponding feedback information U according to the user's current review feed ; Assume that the feedback currently reviewed by the user is Fu, then the output of the dynamic feedback module is expressed as:

[0040] F feed =State(O m ,U feed )

[0041] Among them, F feed represents the function used by the dynamic feedback module to determine the current state, U feed Indicates the user's current review status; feedback information O m Used to adjust and optimize the output results of the error correction model;

[0042] Step 23, the review feedback module is sent to the user decision module, and the user receives the output of the large model in the decision module, and performs secondary feedback and decision-making; the user decision module receives the feedback information from the dynamic feedback module, and combines the preliminary output of the error correction large model to perform comprehensive analysis and judgment; the output of the large model of the user decision module is secondary evaluated and decided, and the final feedback result is recorded; let the output of the user decision module be U feed , expressed as:

[0043] U feed =F user (O m ,KE)

[0044] Among them, F user represents the function of user's secondary feedback, and KE represents the user's expertise and experience.

[0045] Step 24: The final feedback result is the output U of the user decision module feed The error correction feedback module in step 22 records and reassembles the output of the user decision module, and compares U feed and correct the output of the large model m The output difference is used to identify the difference points; the data organization of the dynamic feedback module is expressed as:

[0046] Dr=DIFF(U feed ,O m ))

[0047] Wherein, DIFF represents the data organization function of the retraining module, i.e., the difference point identification data, and Dr represents the difference point identification data used for retraining;

[0048] Difference point identification data is used to detect differences between the original essay and the model answer or correct version;

[0049] Step 25, after the user performs the error correction result feedback operation multiple times, he / she continuously interacts with the large model, forming some error correction intermediate process data; the intermediate process data implicitly includes the chapter article originally submitted by the user and the feedback record data context, constructing the error correction alignment data.

[0050] The error correction alignment data is used to align the identified differences with the corresponding correction suggestions;

[0051] In step 26, the difference point recognition data and the error correction alignment data in the dynamic feedback module are inputted into the retraining module together, and finally the parameter training and update operations of the large model are called to execute; the method adopted to adjust the large model is small parameter update, that is, fixing the original parameters of the model to maintain the error correction capability, and opening some adjustable parameters at the same time to support the improvement of the model in various sub-capabilities.

[0052] Further, the difference point identification data is used to detect the difference between the original composition and the standard answer or the correct version; including:

[0053] Original large model output text: the error correction answer output by the large model;

[0054] User feedback answer: the user’s error correction answer output by the big model, and the adjusted ideal error correction feedback text data;

[0055] Differences: Differences identified by comparing the original text with the standard answer; Differences include spelling errors, grammatical errors, punctuation errors, vocabulary errors, etc.

[0056] Difference type: Categorize each difference point, including spelling errors, grammatical errors, and punctuation errors;

[0057] Difference location: The specific location of the difference in the original text.

[0058] Furthermore, the error correction alignment data includes:

[0059] Error type: Specific error type, including spelling errors, grammatical errors, and punctuation errors;

[0060] Error location: the specific location of the error in the original text;

[0061] Correction suggestions: Correction suggestions provided for each error type;

[0062] Alignment rounds: the number of rounds of error correction;

[0063] Contextual information: contextual information of the sentence where the error occurs;

[0064] Confidence Rating: A confidence rating for each correction suggestion.

[0065] The difference between this application and the prior art is that:

[0066] Multi-step error correction context awareness: This application introduces the concept of multi-step error correction context awareness. By analyzing and utilizing the context information in the multi-step error correction process, it improves the performance of the large error correction model in complex error correction scenarios; it can better understand and handle continuous error correction tasks, and improve the accuracy and consistency of error correction.

[0067] Personalized error correction and semantic perception: This application emphasizes the importance of personalized error correction perception and error correction semantic understanding. Through the user's historical error correction records and preferences, the error correction strategy of the large model is adjusted to make it more in line with the user's personalized needs. At the same time, semantic fragments are obtained by combining semantic perception trees and other methods for large model training. The error correction method with personalized and semantic fusion can improve user satisfaction and usage experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A flow chart of the method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0070] The concepts involved in the method provided in this application are defined as follows:

[0071] Definition 1: A large language model is a language representation model that has a large amount of training data, contains many parameters, and has strong understanding ability;

[0072] Definition 2: Large language model reasoning, that is, using an open large language model as the core cognitive engine to build various reasoning tasks in complex real-world scenarios and directly call the process of generating appropriate response queries;

[0073] Definition 3: Text to be corrected, i.e. the text of the article to be reviewed and modified submitted by the user;

[0074] Definition 4: Error Correction Big Model: The big correction model is a process of fine-tuning a large number of parameter models using a large number of essay error-correct corpus pairs and a big model fine-tuning method, which can ultimately achieve the effect of essay correction.

[0075] Definition 5: Error Correction Prompt Template, a template and method for calling a large language model;

[0076] Definition 6: Open large language model fine-tuning, that is, large language models with public code and training data;

[0077] Definition 7: Semantic encoding, that is, through a pre-trained language representation model, the function is to convert the semantic information of the text into a random discrete vector;

[0078] Definition 8: Query, a query statement or text required to be input when calling a question-answering or generating a class model;

[0079] Definition 9: Large model instruction fine-tuning is the process of combining multiple instructions of different types, constructing input-output corpus pairs based on expected output results, and fine-tuning the model.

[0080] Step 1: Error correction metadata construction: Use the construction method of the large model instruction data set to construct error correction metadata for data including chapter sentences; input the text to be corrected specified by the user; specifically: for each chapter text to be corrected,;

[0081] Step 11: The user calls the composition construction module and establishes the error correction prompt template to construct the error correction input of the large model; because the traditional error correction chapter contains too little information, such as the lack of user error type information and semantic information, the composition construction module is used to reassemble the user's text to be reviewed to enhance the semantic information and personalized information of the text. In this process, the error correction prompt template is used to accept the original text of the chapter to be corrected, call the semantic analysis module to generate the semantic data of the error correction chapter, and call the user matching module to perform user matching to query the past error correction type and error correction record;

[0082] Step 111, receiving the original text of the article to be corrected, for receiving the user's article review request, the original text includes a user review text to be reviewed and the user identity ID;

[0083] Step 112, calling the semantic analysis module to generate semantic data of the correction chapter for processing the original correction text input by the user, on the one hand by receiving external semantic information fusion requests, on the other hand by calling external large model API and tools, semantic decomposition, semantic tree construction and semantic vector generation of the correction text;

[0084] Step 1121, semantic decomposition refers to using the existing summary model to extract the core meaning of the text to be corrected and generate an abbreviated sentence; the user only needs to input the task query text to the summary model, and the model generates the result; the task query text is: "According to our understanding, the most concise meaning of this sentence is: [here is the text to be corrected]"; for example, input a text to be corrected: 'I has a apple.'. The task query text is: "According to our understanding, the most concise meaning of this sentence is: [I has a apple.]". The refined semantic text generated by the summary model is: `I have apple'.

[0085] Step 1122, semantic tree construction: based on the above sentence, call the semantic tree generation tool to generate the semantic tree in the text section to be corrected; wherein the semantic dependency tree is generated by the syntactic tool and syntax analyzer in the StandFordNLP tool (a common natural language processing tool).

[0086] For example, enter a text to be corrected: 'I has a apple.'

[0087] The semantic tree generated by the tool is: ROOT

[0088] ├──(part of speech: word)nsubj:I

[0089] ├──(part of speech: word)aux:have

[0090] ├──(part of speech: word)det:an

[0091] └──(part of speech: word)obj:apple

[0092] Step 1123, semantic vector generation: the text and dependency relationships generated by the semantic decomposition process and the semantic tree construction process are represented again by digital vectorization, and the generated text is converted into a digital representation with a specific semantic relationship; the main purpose of this step is to model the external association information implicit in the large-scale corpus relationship and convert these texts into a quantitative digital representation. The number represents the amount of natural information possessed by the text generated in step 1121 and step 1122.

[0093] For example, the digitized vector representation of `I have apple.ROOT aux det obj' is: [0.12, 0.31, 0.52, 0.71, 0.12, 0.25, 0.48, 0.63, 0.81]

[0094] By calling the semantic analysis module to generate semantic data of the correction passage (calling the process of step 1121, step 1122, and step 1123), the above process of generating semantic sentences, grammatical dependency trees, and semantic vectors completes the semantic generation of the text to be corrected.

[0095] Step 113, calling the user matching module to match the user ID to query the user's past error correction type and error correction record, and constructing a user error correction request template;

[0096] The user matching module is used to integrate the user's exclusive error correction type and error correction record; unlike existing methods, incorporating this information into the subsequent fine-tuning training of the large model can potentially enhance the model's ability to perform personalized error correction and achieve differentiated error correction prompts. Specifically: First, the user matching module accepts the user ID, queries the internal knowledge base to query all the user's past error correction records, and extracts data on the error correction type and error correction record in the error correction record. Among them, the error correction types include part of speech errors, verb tense errors, and noun singular and plural errors; the error correction record includes the sentence to be corrected and the corrected text. After extraction, the text is constructed according to the following error correction request template:

[0097] Error correction request template: "Based on your past error correction records [error correction records and error types extracted from the user's past error correction system, etc.], the problems you often encounter include [list common error correction types encountered by users, such as verb tense errors, noun singular and plural errors, etc.] Please pay attention to these aspects and check for other possible errors at the same time."

[0098] Step 114, the generated sub-text data is packaged for all samples according to the following error correction metadata template;

[0099] Input:"Please correct the error according to the following information:

[0100] Sentence to be corrected: [Original text to be corrected entered by the user]

[0101] Semantics: [`I have apple.ROOT aux det obj']

[0102] Semantic vector: [0.12, 0.31, 0.52, 0.71, 0.12, 0.25, 0.48, 0.63, 0.81]

[0103] Common error types of users: [types of errors that users have corrected in the past]

[0104] Error correction request template data: "Based on your past error correction records [Example 1,...], the problems you often encounter include [list common types of errors that users often encounter, such as verb tense errors, noun singular and plural errors, etc.]; please pay attention to these aspects and check for other possible errors at the same time"

[0105] OutputOutPut: the modified sentence and its error type;

[0106] Since the error correction metadata template consists of multiple information parts, the process of using these data for model training can improve the model error correction ability, semantic understanding ability and user personalized information analysis ability of the existing large batch model.

[0107] Step 12, using the error correction metadata template generated in step 11, by fine-tuning the open source big language parameter model, it becomes an error correction big language model with specific capabilities, and in this process enhances the big language model's ability to understand the error correction metadata, and enables the big language model to automatically have more accurate and personalized passage error correction capabilities.

[0108] Specifically, given an open source large language model, use the error correction metadata template constructed in step 11, use the input Input and output Output as the training input and training target output of the large language model respectively, and through parameter learning and training, the large language model has the ability to correct errors, and finally becomes a corrected large language model; the method is as follows:

[0109] Output = LLM (error correction metadata) = LLM (original text, error correction semantic information, user error correction record information) (1)

[0110] Among them, the open source large language model LLM refers to a commonly used large-parameter language model in the field of natural language processing. It is a model used to handle various natural language processing tasks and can generally be retrained and developed according to actual needs.

[0111] Combined with step 11 and step 12, the text to be corrected input by the original user is trained by the open source big language model. During the training process, the model will accept a series of sample data constructed by a specific correction metadata template and learn the implicit textual relations in the sample data. Finally, the open source big prediction model will become a correction big model and will be able to learn multi-granular semantic understanding capabilities and personalized correction capabilities.

[0112] Step 2: By accepting the output of the large grading model, the dynamic feedback module is called to send the grading feedback to the user decision module; the user decision module accepts the output results, provides correction opinions and satisfaction feedback on the correction results, and records and reassembles the satisfaction feedback results of these users, and returns them to the dynamic feedback module.

[0113] Since the correction model trained in step one often has problems with error correction reliability and historical dependence when performing the error correction process on new sentences to be reviewed, the model results cannot be highly dependent on past semantic information and user correction records. This requires users to provide further feedback on the error correction results and incorporate these feedback results into the improvement of the correction model. This is used to further adjust and align the correction model so that the model can be aligned for learning. The introduction of step two is mainly used to make up for the multi-step modification capability of the correction model constructed in step one.

[0114] Step 21, receiving the output of the correction model after the current sentence to be corrected is corrected, and calling the dynamic feedback module to provide feedback on the user's current review result; including obtaining the result of the correction of the sentence to be corrected from the correction model; assuming that the output of the correction model is 0 m , output O m Contains preliminary error correction information and suggestions;

[0115] ,Step 22, the dynamic feedback module will judge the execution of the error correction result of the large model according to the current state, or generate corresponding feedback information U according to the user's current review feed ; Assume that the feedback currently reviewed by the user is Fu, then the output of the dynamic feedback module is expressed as:

[0116] F feed=State(O m ,U feed )

[0117] Among them, F feed represents the function used by the dynamic feedback module to determine the current state, U feed Indicates the user's current review status; feedback information O m Used to adjust and optimize the output results of the error correction model;

[0118] Step 23, the review feedback module is sent to the user decision module, and the user receives the output of the large model in the decision module, and performs secondary feedback and decision-making; the user decision module receives the feedback information from the dynamic feedback module, and combines the preliminary output of the error correction large model to perform comprehensive analysis and judgment; the output of the large model of the user decision module is secondary evaluated and decided, and the final feedback result is recorded; let the output of the user decision module be U feed , expressed as:

[0119] U feed =F user (O m ,KE)

[0120] Among them, F user represents the function of user's secondary feedback, and KE represents the user's expertise and experience.

[0121] Step 24: The final feedback result is the output U of the user decision module feed The error correction feedback module in step 22 records and reassembles the output of the user decision module, and compares U feed and correct the output of the large model m The output difference is used to identify the difference points; the data organization of the dynamic feedback module is expressed as:

[0122] Dr=DIFF(U feed ,O m ))

[0123] Among them, DIFF represents the data organization function of the retraining module, namely the difference point recognition data, and Dr represents the difference point recognition data used for retraining; through the above steps, the system can effectively utilize the user's feedback information and difference point data to continuously optimize and improve the performance of the error correction model.

[0124] The following is a detailed description of the difference point identification data: The difference point identification data is used to detect the differences between the original essay and the standard answer or the correct version; including:

[0125] Original large model output text (Original Text): the error correction answer output by the large model;

[0126] User feedback answer (Reference Text): The user’s correction answer output by the big model, and the adjusted ideal correction feedback text data;

[0127] Differences: Differences identified by comparing the original text with the standard answer; Differences include spelling errors, grammatical errors, punctuation errors, vocabulary usage errors, etc.

[0128] Difference Types: Classify each difference point, including spelling errors, grammatical errors, and punctuation errors;

[0129] Difference Locations: The specific locations of the difference points in the original text (such as character positions or word positions).

[0130] Step 25, after the user performs the error correction result feedback operation multiple times, he / she continuously interacts with the big model, forming some error correction intermediate process data; the intermediate process data implicitly includes the chapter article originally submitted by the user and the feedback record data context (big model output, user feedback results, difference point identification data), and constructs error correction alignment data.

[0131] The error correction alignment data is used to align the identified difference points with the corresponding correction suggestions; the error correction alignment data includes:

[0132] Error Types: Specific error types, including spelling errors, grammatical errors, and punctuation errors;

[0133] Error Locations: The specific location of the error in the original text;

[0134] Correction Suggestions: Correction suggestions provided for each error type; for example, the correct spelling for a spelling error, the correct grammatical structure for a grammatical error, etc.

[0135] Alignment turns: the number of error correction rounds;

[0136] Context Information: The context of the sentence where the error occurs, so as to better understand and correct the error;

[0137] Confidence Scores: A confidence score for each correction suggestion, indicating the system's confidence in the correction suggestion.

[0138] Step 26, the difference point recognition data and the error correction alignment data in the dynamic feedback module are input together into the retraining module, and the parameter training of the large model is called to perform model fine-tuning. The fine-tuning process will update and correct the relevant parameters of the large model; the method used to adjust the large model is small parameter update, that is, fixing the original parameters of the model to maintain the error correction capability, and at the same time opening some adjustable parameters to support the improvement of the model in various sub-capabilities.

[0139] The present application is further described below in conjunction with specific embodiments.

[0140] Cases where users use this patent to correct errors in text:

[0141] Input: User-specified text with error correction

[0142] Sample text:

[0143] Original text: "I really has an apple?"

[0144] User ID: user123

[0145] 1. By calling the composition construction module, design a suitable error correction prompt template to construct the error correction input of the large model

[0146] 2. Receive the original correction text: Receive the user's article review request, including the user's review text to be reviewed and the user's identity ID.

[0147] 3. Semantic analysis module:

[0148] Semantic decomposition:

[0149] Prompt template: "According to our understanding, the most concise meaning of this sentence is: [semantic data text understood by the system]"

[0150] Sample output: "I have an apple."

[0151] ·Semantic tree construction: Use the StandFord syntax tool to generate a semantic dependency tree. Example output:

[0152]

[0153] 4. Semantic vector construction:

[0154] Convert the results of semantic decomposition and semantic tree construction into vector representation.

[0155] Example vector: [0.1, 0.3, 0.5, 0.7, 0.2, 0.4, 0.6, 0.8]

[0156] 5. User’s past error correction types and record tips:

[0157] ·Prompt template: "According to the above correction records, the problems you often encounter include verb tense errors, noun singular and plural errors, etc. Please pay attention to these aspects, and you can also check other possible errors."

[0158] The sample output is: Common types of errors by users: verb tense errors, noun singular and plural errors

[0159] 6. Construct error correction metadata template:

[0160] ·enter:

[0161] Input:"Please correct the error according to the following information:

[0162] Original sentence: 'I have an apple.'

[0163] Semantic meaning: 'I have an apple.'

[0164]

[0165] Example vector: [0.1, 0.3, 0.5, 0.7, 0.2, 0.4, 0.6, 0.8]

[0166] Common error types for this user: 'Verb tense error, noun singular and plural error'

[0167] Please correct this sentence to make sure it is correct in grammar, spelling, and word choice."

[0168] Output: The corrected sentence and its error type.

[0169] 7. Using the error correction metadata template generated in step 11, fine-tune the open source large language parameter model

[0170] 8. In actual use, receive the output of the error correction model:

[0171] Output = LLM (error correction metadata) = LLM (original text, error correction semantic information, user error correction record information)

[0172] 9. Call the dynamic feedback module to provide feedback on the user's current review

[0173] Receive the output of the error correction model: Om = "I have an apple."

[0174] Accept user feedback results: F feed ="I really have an apple."

[0175] 10. Record and reassemble user feedback results to identify differences:

[0176] {

[0177] "OriginalText":"I have an apple.",

[0178] "ReferenceText":"I really have an apple.",

[0179] "Differences":[

[0180] {

[0181] "Difference":"really",

[0182] "Type":"Vocabulary Addition",

[0183] "Location":"Between 'I' and 'have'"

[0184] } ]

[0186] }

[0187] 11. Construct error correction alignment data, which includes error type, error location, correction suggestion, alignment round number, context information, and confidence score.

[0188] Error Types: [Noun error]

[0189] Error Locations:[0-1,1-2]

[0190] Correction Suggestions: ["Add vocabulary between 'I' and 'have'"]

[0191] Alignment turns:1

[0192] Context Information:"I have an apple."

[0193] Confidence Scores: 0.95

[0194] 12. Based on the error correction alignment data, the error correction model is fine-tuned for a second time to make its output more accurate and complete.

[0195] The technical solution of this application can significantly improve the performance and user experience of the essay correction system. The specific technical effects are as follows:

[0196] Improve error correction accuracy

[0197] Through a multi-level feedback mechanism and multi-step error correction context perception, this patent can more accurately identify and correct errors in essays.

[0198] Improve user satisfaction

[0199] By combining users' professional knowledge and experience, the system can provide more accurate and personalized error correction suggestions, thereby improving user satisfaction.

[0200] Improve system flexibility and adaptability

[0201] By retraining modules and updating small parameters, the system can continuously optimize and improve the performance of the large error correction model to adapt to different user needs and scenarios.

[0202] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A large error correction model iterative training method integrating user semantics and past error correction records, characterized in that: The method comprises: Step 1: Error correction metadata construction: Use the construction method of the large model instruction data set to construct error correction metadata for data including chapters and sentences; input the text to be corrected specified by the user; Step 2: By accepting the output of the large correction model, the dynamic feedback module is called to send the review feedback to the user decision module; the user decision module accepts the output results, provides correction opinions and satisfaction feedback on the correction results, and records and reassembles the user's satisfaction feedback results and returns them to the dynamic feedback module.

2. The method according to claim 1, characterized in that Step 1: Error correction metadata construction: Use the construction method of the large model instruction data set to construct error correction metadata for data including chapters and sentences; including: Step 11: The user calls the composition construction module and establishes a correction prompt template to construct the correction input of the large model; the correction prompt template is used to receive the original text of the chapter to be corrected, call the semantic analysis module to generate the semantic data of the correction chapter, call the user matching module to match the user, and query the past correction types and correction records; Step 12: Use the error correction metadata template generated in step 11 to fine-tune the open source large language parameter model.

3. The method according to claim 2, characterized in that Step 11: The user calls the composition construction module and creates a correction prompt template to construct the correction input of the large model, including: Step 111, receiving the original text of the article to be corrected, for receiving the user's article review request, the original text includes a user review text to be reviewed and the user identity ID; Step 112, calling the semantic analysis module to generate semantic data of the correction chapter for processing the original correction text input by the user, on the one hand by receiving external semantic information fusion requests, on the other hand by calling external large model API and tools, semantic decomposition, semantic tree construction and semantic vector generation of the correction text; Step 113, calling the user matching module to match the user ID to query the user's past error correction type and error correction record, and constructing a user error correction request template; The user matching module is used to integrate the user's exclusive error correction type and error correction record; first, the user matching module accepts the user ID, queries the internal knowledge base to find all the user's past error correction records, and extracts the error correction type and error correction record in the error correction record. Among them, the error correction type includes part of speech errors, verb tense errors, and noun singular and plural errors; the error correction record includes the sentence to be corrected and the corrected text. After extraction, the text is constructed according to the following error correction request template: Error correction request template: "Based on your past error correction records [error correction records and error types extracted from the user's past error correction system], the problems you often encounter include [list the common error correction types of users] Please pay attention to these aspects and check other possible errors at the same time"; Step 114, the generated sub-text data is packaged for all samples according to the following error correction metadata template; Input:"Please correct the error according to the following information: Sentence to be corrected: [Original text to be corrected entered by the user] Semantics: [`I have apple.ROOT aux det obj'] Semantic vector: [0.12, 0.31, 0.52, 0.71, 0.12, 0.25, 0.48, 0.63, 0.81] Common error types of users: [types of errors that users have corrected in the past] Error correction request template data: "Based on your past error correction records [Example 1,…], the problems you often encounter include []; please pay attention to these aspects and check for other possible errors" OutputOutPut: The modified sentence and its error type.

4. The method according to claim 3, characterized in that The error correction text is semantically decomposed, a semantic tree is constructed, and a semantic vector is generated as follows: Step 1121, semantic decomposition refers to using the existing summary model to extract the core meaning of the text to be corrected and generate an abbreviated sentence; the user inputs the task query text to the summary model, and the model generates a result; the task query text is: "According to our understanding, the most concise meaning of this sentence is: [here is the text to be corrected]"; Step 1122, semantic tree construction: based on the above sentence, call a semantic tree generation tool to generate a semantic tree for the text chapter to be corrected; The semantic dependency tree is generated by the syntax tool and grammar analyzer in the StandFordNLP tool; Step 1123, semantic vector generation: the text and dependency relationships generated by the semantic decomposition process and the semantic tree construction process are re-represented by digital vectorization, and the generated text is converted into a digital representation with specific semantic relationships.

5. The method according to claim 2, characterized in that: Step 12, using the error correction metadata template generated in step 11, fine-tuning the open source large language parameter model; include: Given a large language model, use the error correction metadata template constructed in step 11, take the input Input and output Output as the training input and training target output of the large language model respectively, and through parameter learning and training, the large language model has the error correction capability, and finally becomes a corrected large language model; the method is as follows: OutPut=LLM (error correction metadata)=LLM (original text, error correction semantic information, user error correction record information) (1).

6. The method according to claim 1, characterized in that Step 2: By accepting the output of the large grading model, the dynamic feedback module is called to send the grading feedback to the user decision module; the user decision module accepts the output result, provides feedback on the correction opinions and the satisfaction of the correction results, and records and reassembles the user's satisfaction feedback results and returns them to the dynamic feedback module, including: Step 21, receiving the output of the correction model after the current sentence to be corrected is corrected, and calling the dynamic feedback module to provide feedback on the user's current review result; including obtaining the result of the correction of the sentence to be corrected from the correction model; assuming that the output of the correction model is 0 m , output O m Contains preliminary error correction information and suggestions; Step 22: The dynamic feedback module will determine whether to accept the error correction result of the large model according to the current state, or generate corresponding feedback information U according to the user's current review. feed ; Assume that the feedback currently reviewed by the user is Fu, then the output of the dynamic feedback module is expressed as: F feed =State(O m ,U feed ) Among them, F feed represents the function used by the dynamic feedback module to determine the current state, U feed Indicates the user's current review status; feedback information O m Used to adjust and optimize the output results of the error correction model; Step 23, the review feedback module is sent to the user decision module, and the user receives the output of the large model in the decision module, and performs secondary feedback and decision-making; the user decision module receives the feedback information from the dynamic feedback module, and combines the preliminary output of the error correction large model to perform comprehensive analysis and judgment; the output of the large model of the user decision module is secondary evaluated and decided, and the final feedback result is recorded; let the output of the user decision module be U feed , expressed as: You feed =F user (You m (I) Among them, F user represents the function of user's secondary feedback, and KE represents the user's expertise and experience. Step 24: The final feedback result is the output U of the user decision module feed The error correction feedback module in step 22 records and reassembles the output of the user decision module, and compares U feed and correct the output of the large model m The output difference is used to identify the difference points; the data organization of the dynamic feedback module is expressed as: Dr=DIFF(U feed ,O m )) Wherein, DIFF represents the data organization function of the retraining module, i.e., the difference point recognition data, and Dr represents the difference point recognition data used for retraining; Difference point identification data is used to detect differences between the original essay and the model answer or correct version; Step 25, after the user performs the error correction result feedback operation multiple times, he / she continuously interacts with the large model, forming some error correction intermediate process data; the intermediate process data implicitly includes the chapter article originally submitted by the user and the feedback record data context, constructing the error correction alignment data. The error correction alignment data is used to align the identified differences with the corresponding correction suggestions; In step 26, the difference point recognition data and the error correction alignment data in the dynamic feedback module are input into the retraining module, and finally the parameter training and update operations of the large model are called to execute; the method adopted to adjust the large model is small parameter update, that is, fixing the original parameters of the model to maintain the error correction capability, and at the same time opening some adjustable parameters to support the improvement of the model in various sub-capabilities.

7. The method according to claim 6, characterized in that Difference identification data is used to detect differences between the original essay and the model answer or correct version; including: Original large model output text: the error correction answer output by the large model; User feedback answer: the user’s error correction answer output by the big model, and the adjusted ideal error correction feedback text data; Differences: Differences identified by comparing the original text with the standard answer; Differences include spelling errors, grammatical errors, punctuation errors, vocabulary errors, etc. Difference type: classify each difference point, including spelling errors, grammatical errors, and punctuation errors; Difference location: the specific location of the difference point in the original text.

8. The method according to claim 6, characterized in that Error correction alignment data includes: Error type: Specific error type, including spelling errors, grammatical errors, and punctuation errors; Error location: the specific location of the error in the original text; Correction suggestions: Correction suggestions provided for each error type; Alignment rounds: the number of rounds of error correction; Contextual information: contextual information of the sentence where the error occurs; Confidence Rating: A confidence rating for each correction suggestion.

Images