Guided writing feedback method and system based on large model

Abstract

The invention discloses a guide type writing feedback method and system based on a large model, and the method comprises the steps: carrying out the preprocessing and deep diagnosis of a user text, generating an internal report containing error details, only outputting a weak feedback for prompting an error position to a user at the initial stage, so as to induce the user to carry out the autonomous correction, and on this basis, carrying out the deep diagnosis of the user text. The system evaluates a user correction fragment in real time, dynamically upgrades a feedback level based on a current bias error record only when correction is incorrect, and gradually provides medium feedback containing meta-language interpretation or strong feedback of a direct correction scheme. Through the layered teaching support, the active reflection and self-monitoring ability of a learner are effectively stimulated, the cognitive requirements of different levels of users are accurately met while excessive correction of the model is avoided, and the internalization efficiency of language rules and the pertinence of writing teaching are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of intelligent feedback, and more specifically, to a guided writing feedback method and system based on a large model. Background Technology

[0002] In the field of computer-assisted language teaching, the core value of writing feedback lies not only in accurately correcting errors, but also in stimulating learners' cognitive processing through effective instructional scaffolding, prompting them to internalize language rules through self-reflection and correction.

[0003] However, existing intelligent writing feedback solutions mainly rely on general-purpose large language models, and their feedback strategies primarily focus on directly presenting correct text, i.e., immediately providing the corrected answer after identifying errors. This approach has significant drawbacks: on the one hand, directly providing the corrected result replaces the learner's necessary self-monitoring and reflection during the writing process, leading to students being in a passive state of acceptance and failing to motivate them to actively discover problems and explore correction paths, thus hindering substantial improvement in writing skills; on the other hand, general-purpose large language models are mainly trained on general corpora, and in writing feedback for specific languages ​​(such as German), they often suffer from insufficient stability, are prone to creating illusions, or over-correction, making it difficult to provide gradual guidance in accordance with pedagogical principles and failing to meet the needs of learners at different levels for differentiated feedback levels.

[0004] Therefore, we look forward to an optimized guided writing feedback method based on a large model. Summary of the Invention

[0005] To address the aforementioned technical problems, embodiments of this application provide a guided writing feedback method and system based on a large model.

[0006] According to one aspect of this application, a guided writing feedback method based on a large model is provided, comprising: Obtain the original user input; The original user input is preprocessed to obtain preprocessed text; Potential errors are identified and comprehensively diagnosed in the preprocessed text to obtain an internal diagnostic report. The internal diagnostic report includes a list of deviations, and each element in the list includes the location of the error, the type of error, and a suggested correction. Based on the location of errors in the internal diagnostic report, a weak feedback response is generated. Receive user-submitted correction fragments in response to the weak feedback response, and evaluate the user-submitted correction fragments and the suggested correction schemes in the internal diagnostic report in real time to obtain the evaluation results; If the evaluation result is incorrect, a hierarchical decision is made on the current error record and the current feedback level to obtain a medium or strong feedback response.

[0007] According to another aspect of this application, a large-model-based guided writing feedback system is provided, comprising: The original user input acquisition module is used to acquire original user input. The preprocessing module is used to preprocess the original user input to obtain preprocessed text; The potential error identification and comprehensive diagnosis module is used to identify and comprehensively diagnose potential errors in preprocessed text to obtain an internal diagnostic report. The internal diagnostic report includes a deviation list, and each element in the deviation list includes the error location, error type, and suggested correction scheme. The weak feedback response module is used to generate a weak feedback response based on the location of the error in the internal diagnostic report; The evaluation module is used to receive user-submitted correction fragments in response to the weak feedback response, and to evaluate the user-submitted correction fragments and the suggested correction schemes in the internal diagnostic report in real time to obtain the evaluation results. The feedback hierarchy progressive decision module is used to make a feedback hierarchy progressive decision on the current error record and the current feedback level in response to an incorrect evaluation result, so as to obtain a medium feedback response or a strong feedback response.

[0008] Compared to existing technologies, this application provides a guided writing feedback method and system based on a large model. It generates an internal report containing error details by preprocessing and deeply diagnosing user text. Initially, it only provides weak feedback indicating the location of errors to induce self-correction. Building on this, the system provides real-time evaluation of the corrected segments, dynamically escalating the feedback level only when the correction is incorrect, based on the current error record. This gradually provides medium-level feedback including meta-language explanations or strong feedback with direct correction solutions. Through tiered instructional scaffolding, it effectively stimulates learners' active reflection and self-monitoring abilities. While avoiding over-correction by the model, it accurately adapts to the cognitive needs of users at different levels, significantly improving the internalization efficiency of language rules and the relevance of writing instruction. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1A flowchart of a guided writing feedback method based on a large model according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow for a large-model-based guided writing feedback method according to an embodiment of this application. Figure 3 This is a block diagram of a large-model-based guided writing feedback system according to an embodiment of this application; Figure 4 This is a schematic diagram illustrating a weak feedback application according to an embodiment of this application; Figure 5 This is a schematic diagram of a feedback application according to an embodiment of this application; Figure 6 This is a schematic diagram of a strong feedback application according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] In the technical solution of this application, a guided writing feedback method based on a large model is proposed. Figure 1 This is a flowchart of a guided writing feedback method based on a large model according to an embodiment of this application. Figure 2 This is a system architecture diagram of a guided writing feedback method based on a large model according to an embodiment of this application. Figure 1 and Figure 2 As shown, the guided writing feedback method based on a large model according to an embodiment of this application includes the following steps: S1, obtaining original user input; S2, preprocessing the original user input to obtain preprocessed text; S3, performing potential error identification and comprehensive diagnosis on the preprocessed text to obtain an internal diagnostic report, the internal diagnostic report including a deviation list, each element in the deviation list including error location, error type, and suggested correction scheme; S4, generating a weak feedback response based on the error location in the internal diagnostic report; S5, receiving user-submitted correction fragments for the weak feedback response, and performing real-time evaluation of the user-submitted correction fragments and the suggested correction schemes in the internal diagnostic report to obtain an evaluation result; S6, in response to an incorrect evaluation result, performing a feedback hierarchy decision on the current error record and the current feedback level to obtain a medium or strong feedback response.

[0016] Specifically, S1 involves acquiring raw user input. This step aims to capture language materials generated by learners to initiate subsequent error identification and instructional scaffolding construction processes. Raw user input refers to the initial data carrier submitted by the user on the system's front-end interface, before being processed by the system kernel. It encompasses all language materials that the user wants the system to perform grammatical checks, polishing, or evaluation. This concept includes both sequences of natural language text directly typed by the user and electronic documents containing text content uploaded by the user (such as .doc, .docx, .txt, etc.). Its quality and completeness directly determine the accuracy of subsequent internal diagnostic reports. In the context of computer-assisted language instruction (CALL), raw user input is not only a text carrier containing potential grammatical or pragmatic errors but also the starting point for triggering multiple rounds of interactive guidance from the system. By accurately and completely acquiring the content of the user's intended expression, the system can provide effective prompts for errors in the user's submitted essays through a three-tiered feedback model of weak, medium, and strong feedback.

[0017] In practice, the system first provides users with an interactive environment through a model dialogue module. This environment supports multi-turn dialogue and aims to simulate real-world writing guidance scenarios. Users have two main ways to input content on this interface: one is by directly editing and writing text through the text input box, suitable for immediate short text exercises or modifications based on specific feedback; the other is by importing completed documents through the file upload function, with the system explicitly supporting Word or TXT file uploads. After the user completes content input or file selection on the front-end interface and clicks the send button, the front-end encapsulates this data into a request and sends it to the back-end. The back-end processing module, as the core logic layer of the system, is responsible for receiving various requests from the front-end and handling data storage and task scheduling. It is worth noting that during this step, the system not only receives the data stream but also performs preliminary labeling of the data source at the receiving level to distinguish between direct text strings and document objects requiring further file stream parsing, thus providing a data foundation for subsequent format parsing. In addition, this step also involves contextual association of user identity. That is, when the system obtains input, it will bind the input to specific user historical conversation records (such as independent dialogue records under a large model) based on the status of the account information module and the model selection module, so as to ensure data isolation and consistency of user experience.

[0018] Specifically, S2 involves preprocessing the original user input to obtain preprocessed text. It should be understood that original user input often exhibits heterogeneity and non-standardization. Specifically, as mentioned above, users can directly input content through text input boxes or upload Word or TXT files as input. This multi-source input method results in the original data potentially containing file header information, invisible control characters, non-pure text formatting marks, and inconsistent encoding formats. Directly inputting such unprocessed, noisy data into the subsequent writing feedback model not only increases the model's resource consumption but may also interfere with the model's understanding of text semantics, leading to deviations in the generation of internal diagnostic reports. Furthermore, to transform the continuous text stream into discrete units with logical structures (such as sentences or paragraphs) for accurate error location, the technical solution of this application preprocesses the original user input to transform the heterogeneous original input into structured, standardized, and clean preprocessed text, laying a data foundation for subsequent error identification and comprehensive diagnosis.

[0019] In practice, the first step is to parse the original user input to obtain the raw text string. Since users may upload Word or TXT files, this step first uses file stream recognition technology to determine the MIME type of the input data. For Word documents (.doc / .docx), the system calls the document parser to strip away the XML structure or binary format information and extract the plain text content; for directly input text or TXT files, the character stream is read directly and ultimately converted into the raw text string in memory.

[0020] Next, the original text strings are standardized and cleaned to obtain clean text strings. It should be understood that original text strings often contain a large amount of non-linguistic noise. Specifically, multi-source input inevitably introduces elements such as file header BOM markers, invisible control characters, non-standard whitespace symbols (such as full-width spaces and tabs), and inconsistent encoding formats. Directly injecting this uncleaned raw data into the writing feedback model would, on the one hand, needlessly consume the model's budget and reduce processing efficiency; on the other hand, the interference from non-linguistic symbols might lead to biases in the model's understanding of text semantics, thus affecting the accuracy of error identification in the internal diagnostic report. Therefore, in the technical solution of this application, a standardized cleaning process is used to eliminate heterogeneity and noise in the data, providing high-quality text input for subsequent logical segmentation and deep diagnostics. Specifically, First, the character encoding is standardized. Since user-uploaded TXT files may use various encodings such as ASCII, GBK, or UTF-8, the system first detects the encoding format of the original string and forcibly converts it to the unified UTF-8 format to prevent garbled characters in subsequent processing. Second, invisible characters are removed. During this process, the system scans the string, using regular expressions to match and remove control characters that do not contribute to the semantics of the text (such as characters in the Unicode range U+0000 to U+001F, excluding newline characters). The system first removes hidden characters such as line breaks (because newlines need to be retained for subsequent paragraph splitting) and zero-width spaces. Then, it normalizes whitespace, replacing consecutive spaces, tabs (`\t`), or full-width spaces in the user input with single standard half-width spaces and removing redundant whitespace at the beginning and end of the string. This process ensures consistency in word separators, preventing ambiguity caused by varying whitespace lengths. Next, it filters special characters. Based on a pre-defined whitelist (containing German letters, common punctuation marks, etc.), the system filters out illegal special characters or binary gibberish that could cause parsing errors. After these processes, the resulting string is a clean text string that retains the complete semantic information of the original input while possessing machine readability and standardized characteristics.

[0021] Furthermore, context-aware logical segmentation is performed on the clean text string to obtain preprocessed text. It's understandable that while large language models are capable of processing long texts, to achieve a weak feedback mechanism—that is, to only indicate the specific location of errors in the text—the system must be able to pinpoint the exact sentence or even word level. If only a continuous, unstructured clean text string is input, the model's feedback may only vaguely indicate the error in the entire paragraph, without providing precise index positions (such as "paragraph 1, sentence 2"). Therefore, context-aware logical segmentation of the clean text string transforms the linear string into a preprocessed text object with a hierarchical structure. This allows subsequent error identification and comprehensive diagnosis to be based on finer-grained context, thereby improving the accuracy of identifying specific grammatical or pragmatic errors and avoiding the attention divergence problem common in long text processing.

[0022] In this process, the clean text string is first segmented into a paragraph array using strong paragraph delimiters. During this process, the system scans the input clean text string, searching for strong delimiters that represent the end of logical paragraphs. Typically, these delimiters include newline characters (such as...). , The system uses regular expressions or string splitting algorithms to divide long text into several independent paragraph strings, using these symbols as the splitting points, and stores them as an ordered paragraph array.

[0023] Secondly, the paragraphs in the paragraph array are segmented into sentences to obtain the preprocessed text. During this process, the system needs to identify sentence end markers (such as `.`, `?`, and `!` in German), but must exclude punctuation marks used outside of sentence ends (such as the period in the abbreviation "zB"). Specifically, each paragraph can be further subdivided into sentence sequences by applying rule-based or statistical model-based sentence boundary detection algorithms. The final preprocessed text is a structured data object (e.g., a JSON object) containing paragraph indices, sentence indices, and corresponding text content.

[0024] Specifically, in step S3, latent error identification and comprehensive diagnosis are performed on the preprocessed text to obtain an internal diagnostic report. This report includes a deviation list, where each element includes the error location, error type, and suggested correction scheme. It should be understood that general-purpose large language models often suffer from instability, semantic illusion, or overcorrection when directly generating feedback, and cannot simply present the model output directly to the user. Instead, a controlled diagnostic process is necessary. Therefore, in the technical solution of this application, latent error identification and comprehensive diagnosis of the preprocessed text enable the system to strategically choose whether to only display the location (weak feedback), provide a meta-language explanation (medium feedback), or directly give the answer (strong feedback) based on the user's ability level and interaction progress. It is worth noting that by using a specialized model with supervised fine-tuning (SFT) instead of a general-purpose model, the aim is to improve the professionalism of German grammar recognition and ensure that the diagnostic results conform to teaching logic.

[0025] The internal diagnostic report refers to a core data object (usually in JSON or XML format) generated and maintained by the system in the background. It is the source of truth for all subsequent user interactions and feedback.

[0026] In practice, the preprocessed text is first injected into the prompt word template to obtain the final prompt word. During this process, the system calls the prompt word engineering component in the background processing module to load the carefully designed prompt word template. This template not only contains the preprocessed text to be analyzed, but also embeds domain expert knowledge and the Chain-of-Thought (CoT) strategy, which aims to guide the model to simulate the reasoning process of an expert teacher, rather than simply predicting the next word. The generated final prompt word is a text block containing explicit instructions (such as requiring the error location and type to be output in a specific format).

[0027] Next, the final prompt words are input into the trained interactive German intelligent writing feedback model to obtain the model's raw response. During this process, the constructed final prompt words are input into the interactive German intelligent writing feedback model, a specialized model trained using SFT based on a large corpus of real writing data from Chinese German learners, enabling it to more accurately identify Chinglish errors. Upon receiving the input, the model generates a raw response stream containing natural language descriptions and underlying structured data.

[0028] Furthermore, based on the preprocessed text, the original model response is parsed and standardized to obtain an internal diagnostic report. During this process, the system's parser scans the model's original output, extracts key information, and maps it to a standardized data structure, forming an internal object containing a list of deviations. The parser verifies the integrity of each deviation element, ensuring it contains precise coordinates (e.g., paragraph ID, sentence ID), specific error classification (e.g., grammar, vocabulary), and a standard correction answer. Specifically, the deviation list is an array of objects, where each element represents a specific identified language error. The error location typically refers to the logical paragraph and sentence index where the error occurs, used for highlighting in weak feedback; the error type indicates whether the error belongs to a grammatical declension, verb conjugation, or vocabulary collocation category, used for the meta-language explanation in the generated feedback; and the suggested correction is the standard correct text generated by the model, used in strong feedback or when evaluating user corrections.

[0029] Specifically, S4 generates a weak feedback response based on the error location in the internal diagnostic report. It should be understood that traditional feedback systems often directly provide the corrected text, replacing students' self-reflection and self-monitoring during the learning process. This results in students being in a passive receiving state, which is not conducive to stimulating their active learning and self-reflection abilities. To address this deficiency, this application adopts a three-tiered feedback model, where weak feedback serves as the initial intervention, aiming to utilize an error code strategy—that is, only indicating the specific location of the error in the text without directly providing an explanation or correction. This approach targets high-level learners with a certain level of knowledge and self-monitoring ability, guiding them to explore the causes of errors and correction methods by providing only location clues, thereby deepening their understanding of language rules through in-depth internal processing.

[0030] In practice, the first step is to extract the current error record from the internal diagnostic report. Since the internal diagnostic report may contain multiple deviation elements (i.e., multiple errors), the system usually selects a current error record to be processed based on a preset teaching strategy (such as by error severity or text order). This record is a complete data structure that includes the error location, error type, and suggested correction scheme.

[0031] Next, the error location is extracted from the current error record and assigned to the feedback payload object to obtain a weak feedback response. During this process, the system ignores the error type and suggested correction scheme in the current record, only reading the error location information (such as paragraph and sentence indices). Then, the system calls a pre-set weak feedback dialogue template (e.g., "I found a problem in paragraph X, sentence Y; could you try to correct it?"), dynamically injecting the extracted location parameters into the template's placeholders to construct the final feedback payload object. This object is then sent to the front-end interface through the model dialogue module and presented to the user, thus completing the weak feedback generation process. This process technically masks sensitive answer information, ensuring the strict enforcement of the error code policy.

[0032] Taking the scheme in this application as an example, suppose the internal diagnostic report generated by the system backend contains a record about a German preposition error. The data content of this record is: `{ "Position": "Paragraph 1, Sentence 2", "Type": "Preposition Error", "Correction": "zur"}`. When the system executes this step, it first locks this record as the current error record. Subsequently, the algorithm logic will only extract the position information of `"Paragraph 1, Sentence 2"`, while disabling the fields `"Preposition Error"` and `"zur"`. The system then assigns this position information to a response object, and the generated content may be: "Dr. Write: Please note that I detected an inaccurate expression in the second sentence of the first paragraph. Please check it." After receiving this message, the user needs to return to the second sentence of the first paragraph of the original text, independently check whether it is a vocabulary, grammar, or spelling problem, and try to correct it, rather than directly obtaining the answer. This mechanism forces learners to transform from passive recipients to active problem solvers.

[0033] Specifically, in step S5, the system receives user-submitted corrected fragments in response to the weak feedback and immediately evaluates the corrected fragments and suggested corrections from the internal diagnostic report to obtain an evaluation result. It should be understood that, according to pedagogical principles, after the system issues weak feedback that only indicates the location of the error, the learner submits corrected content through self-monitoring and reflection. The system must immediately and accurately determine the effectiveness of this correction. Without this step, the system cannot know whether the learner has successfully crossed the zone of proximal development, and therefore cannot decide whether to maintain the current state or escalate the feedback level. Traditional evaluation models often rely on a single, static reference answer, which is prone to false negatives when faced with the ambiguity of natural language (e.g., the user uses an expression that is different from the preset answer but equally correct). Therefore, in the technical solution of this application, an adaptive evaluation mechanism based on a dynamically corrected equivalent set and confidence weighting is introduced to ensure the robustness and fairness of the evaluation.

[0034] In practice, the first step is to evaluate the user's corrected text and the suggested corrections in the current error record in real time to obtain a similarity score. It's understandable that traditional evaluation mechanisms, in determining the correctness of user corrections, suffer from a rigid dependence on the evaluation reference point. This limitation stems from their evaluation model, which compares the user's corrected text one-to-one with a single, absolute suggested correction. This model reveals a significant technical weakness when dealing with natural language scenarios with multiple correct expressions: it fails to effectively handle multiple equivalent relationships in language expression. This means that a specific language error often has multiple grammatically correct and semantically equivalent corrections. For example, in the context of German learning, for the original incorrect expression *Ich gehe zu Schule*, the system-generated suggested correction might be *Ich gehe zur Schule*, while the user might independently correct it to *Ich gehe in die Schule*. In practical use, both correction methods are perfectly correct and idiomatic expressions. However, due to their different literal forms, the original mechanism, based on a single reference point, is highly likely to judge the user's valid corrections as incorrect, resulting in a large number of false negatives. This technical limitation not only reduces the accuracy of the assessment but may also negatively impact learners' motivation and provide incorrect learning guidance.

[0035] To address the limitations of the assessment caused by the aforementioned technical reasons, this application proposes an adaptive assessment mechanism based on dynamically modified equivalent sets and confidence weighting, which replaces the original static single reference point by introducing a multi-dimensional, dynamically generated reference standard.

[0036] In this process, firstly, a corrected equivalent set is generated based on the current error record and the user-submitted correction fragment. That is, by constructing a set containing multiple potential correct answers, a more inclusive and realistic benchmark is provided for evaluation. Specifically, the system receives the original error context contained in the error record to be processed, as well as the user-submitted correction fragment, and automatically constructs a secondary query suggestion word based on this information. This suggestion word is delivered to a supervised fine-tuned large language model, which is instructed to dynamically generate a corrected equivalent set. This set is a structured list containing multiple grammatically correct and semantically equivalent correction schemes for the original error, and the model assigns a confidence score to each scheme as the standard answer. In the embodiments of this application, this generation process is expressed by the formula: in, This represents the final generated modified equivalent set. This represents a large language model that has been supervised and fine-tuned and used as a generator. This represents a quadratic query suggestion function specifically constructed to generate this equivalent set. This represents the error context information extracted from the error log. This represents the user-submitted correction fragment. In this way, by expanding the evaluation benchmark from a fixed point to a dynamic surface, and leveraging the richness and diversity of language expression, a dynamic and multi-dimensional evaluation reference standard is established, thus laying a solid data foundation for subsequent accurate and fair evaluations.

[0037] Furthermore, based on the corrected equivalent set, an adaptive best-match scoring is performed on the user-corrected fragment to obtain an adaptive evaluation score as the similarity score. It should be understood that merely determining whether the user correction exists in the equivalent set is still coarse; a more refined scoring mechanism is needed to quantify the closeness of the user correction to the optimal answer in the set, and to intelligently adjust the score using confidence information generated by the model. Specifically, the system iterates through each candidate correction scheme in the corrected equivalent set generated in the previous step. For each scheme, it calculates the lexical similarity and semantic similarity between the user-corrected fragment and the scheme in parallel, and performs a weighted summation using a preset weight to obtain a basic similarity score. Subsequently, this basic similarity score is multiplied by the confidence score of the candidate scheme itself to obtain a confidence-adjusted matching score. Finally, the highest matching score among all candidate schemes is selected as the final adaptive evaluation score. In the embodiments of this application, this scoring process is expressed by the formula: in, This represents the calculated adaptive evaluation score. This represents a maximum value function used to find the maximum value among the calculation results of N candidate solutions. It is a weighted hyperparameter used to balance the importance of lexical and semantic similarity. and These represent functions for calculating lexical similarity (e.g., based on normalized Levenstein distance) and semantic similarity (e.g., based on sentence vector cosine similarity), respectively. On behalf of the user, correct the fragment. The text representing the i-th candidate correction scheme in the modified equivalent set. This represents the confidence score corresponding to the proposed solution.

[0038] This step establishes an adaptive scoring model that not only assesses the closeness of a user's correction to multiple correct answers but also uses confidence information for intelligent weighting. This means that corrections highly similar to high-confidence standard answers receive high scores, while maintaining a reasonable score range for matches with low-confidence (but still valid) non-mainstream answers. In this way, the generated comprehensive evaluation score accurately reflects the true quality of the user's correction in a multi-solution equivalent environment.

[0039] Subsequently, a threshold-based evaluation decision is made on the similarity score to obtain the evaluation result. That is, the continuous evaluation score generated in the previous step is transformed into a discrete Boolean judgment result that can be directly used by subsequent processes in the system. Specifically, the calculated adaptive evaluation score is compared with a system-preset, configurable evaluation threshold. In the embodiments of this application, this decision-making process is expressed by the formula: in, Represents the final Boolean evaluation result. TT represents the adaptive evaluation score, and TT represents the preset evaluation threshold. This step acts as a decision gate, making a final correct or incorrect judgment based on the more robust and intelligent evaluation score produced by the previous steps, thus completing the closed loop of the entire evaluation process and outputting a deterministic conclusion, thereby ensuring the integrity and operability of the entire improvement mechanism.

[0040] In summary, this mechanism overcomes the inaccuracies caused by existing assessment mechanisms' reliance on a single, static reference answer, especially in natural language correction scenarios with multiple equivalence relationships. Specifically, by dynamically generating an equivalent set containing multiple correct correction schemes and employing a confidence-weighted adaptive scoring algorithm, it can accurately identify and correctly evaluate user-submitted corrections that differ literally from the standard answer but are actually completely effective, thus significantly reducing the false negative rate. Secondly, the system's intelligence and adaptability are enhanced, transforming the assessment process from a rigid template matching to a dynamic analysis process that understands linguistic diversity and simulates expert teachers' multi-dimensional considerations. Finally, these technological effects collectively contribute to improving the overall effectiveness and user experience of language learning support systems, providing learners with more accurate and encouraging feedback, thereby building a more efficient and humanized intelligent writing guidance environment.

[0041] Specifically, in step S6, in response to an incorrect evaluation result, a progressive decision is made regarding the current error record and the current feedback level to obtain a medium or strong feedback response. That is, if a user's submitted correction is still incorrect after receiving weak feedback that only indicates the location, it means that location cues alone are insufficient to support self-correction. In this case, the system must dynamically intervene to provide higher-level cognitive support. Without this step, the system will remain in an ineffective loop, failing to effectively promote the learner's knowledge internalization. In the embodiments of this application, if the user's correction is incorrect, the three-tiered feedback model needs to provide further medium feedback, pointing out the error location while providing a meta-language explanation; if the user still cannot correct independently, strong feedback is required, directly providing the correct correction method. This approach protects the learner's enthusiasm for exploration while ensuring the final learning outcome.

[0042] In practice, given that the current feedback level is weak, the error type and deviation location are extracted from the current error record. A specific example is as follows: Figure 4 As shown. Furthermore, a meta-language explanation is generated based on the error type. Subsequently, the meta-language explanation and error location are combined to obtain a feedback response, as illustrated in the example below. Figure 5As shown. In the embodiments of this application, after the user's first correction attempt fails, the system first extracts the error type and error location from the current error record. The error location is used to pinpoint the error, while the error type (e.g., grammatical error, inappropriate word choice, logical coherence problem, etc.) serves as the basis for generating explanatory content. Next, the system generates a meta-language explanation based on the error type. Here, the meta-language explanation refers to a knowledge-based description or rule explanation about the language itself. For example, for an "error in article and preposition collocation," the meta-language explanation might be, "In German, the preposition 'zu' combined with the definite articles 'der,' 'die,' 'das,' etc., requires conjugation to 'zur' or 'zum.'" This step relies on a pre-built knowledge base or rule engine, which uses the error type as an index and links it to the corresponding explanatory text template. Finally, the meta-language explanation and error location are combined to obtain a feedback response. In other words, the generated meta-language explanation text is integrated with the markers pointing to the specific location in the original text (such as highlighting, underlining, or paragraph / sentence numbering) into a structured response message, thereby forming a mid-feedback response. This response includes the error location, the reason for the error, and relevant language rules to guide the user to make a second, more theoretically grounded correction attempt. Furthermore, in response to the current feedback level being medium feedback, a suggested correction scheme is extracted from the current error record, and this suggested correction scheme is used as the strong feedback response. In the embodiments of this application, after the user's second correction attempt based on the medium feedback response (i.e., after receiving an error explanation) is still judged as "incorrect," the system believes that the user may still be unable to complete the correct correction independently and needs more direct assistance. At this time, the decision logic shifts to providing a final suggested solution. Specifically, firstly, a suggested correction scheme is extracted from the current error record; this suggested correction scheme is the standard answer or recommended writing style stored in the deviation list. Then, the suggested correction scheme is output to the user as the strong feedback response, as illustrated in the example below. Figure 6 As shown. This strong feedback response directly demonstrates the correct way of expressing it, designed to ensure that the learning process can continue and allow users to learn by observing the correct answers.

[0043] It is worth mentioning that the German intelligent writing feedback system proposed in this application firstly builds upon the Zhihai Qide large language model by collecting nearly 3,800 real writing samples from Chinese German learners and introducing domain experts for refined annotation, constructing a high-quality dedicated training dataset for supervised fine-tuning (SFT). During the data construction and training phases, carefully designed prompt word templates are employed, combined with domain expert knowledge, and a Chain-of-Thought (CoT) strategy is specifically introduced to enable the model to simulate the multi-stage reasoning process of expert teachers: "error localization, error classification, and error correction." This training method not only generates a teacher model that conforms to teaching logic, enabling it to accurately identify specific types of errors such as grammar, vocabulary, and pragmatics, but also significantly improves the model's professional capabilities in German grammar recognition and parsing, effectively avoiding the problem of general large models easily producing illusions or over-correction in specific language teaching scenarios. In actual operation, the system first performs deep preprocessing on the user's original input, including format parsing, character cleaning, and context-aware logical segmentation, and then generates an internal diagnostic report containing error details. Subsequently, the system executes a three-stage feedback process based on the tiered instructional scaffolding theory: initially, it sends only "weak feedback" indicating the location of errors, guiding advanced learners to self-correct using their self-monitoring abilities; when the system determines that the user's correction attempt is incorrect based on the dynamic correction equivalence set, it escalates to "medium feedback" providing meta-language explanations to assist in analyzing language phenomena; if the correction fails again, it finally provides "strong feedback" containing direct correction solutions. This mechanism, through multiple rounds of human-computer interaction and adaptive assessment, effectively stimulates learners' proactive reflection abilities, significantly improving the internalization efficiency of language rules while reducing writing anxiety.

[0044] In summary, the large-model-based guided writing feedback method according to the embodiments of this application is explained. It generates an internal report containing error details by preprocessing and deeply diagnosing user text. Initially, it only provides weak feedback indicating the location of errors to induce self-correction. Based on this, the system performs real-time evaluation of the user's corrected segments, dynamically escalating the feedback level only when the correction is incorrect, gradually providing medium feedback including meta-language explanations or strong feedback with direct correction solutions. Through tiered instructional scaffolding, it effectively stimulates learners' active reflection and self-monitoring abilities, avoiding model over-correction while accurately adapting to the cognitive needs of users at different levels, significantly improving the internalization efficiency of language rules and the relevance of writing instruction.

[0045] Furthermore, a guided writing feedback system based on a large model is also provided.

[0046] Figure 3 This is a block diagram of a guided writing feedback system based on a large model, according to an embodiment of this application. Figure 3As shown, the guided writing feedback system 300 based on a large model according to an embodiment of this application includes: a raw user input acquisition module 310 for acquiring raw user input; a preprocessing module 320 for preprocessing the raw user input to obtain preprocessed text; a potential bias identification and comprehensive diagnosis module 330 for identifying and comprehensively diagnosing potential biases in the preprocessed text to obtain an internal diagnostic report, the internal diagnostic report including a bias list, each element in the bias list including the bias location, bias type, and suggested correction scheme; a weak feedback response module 340 for generating a weak feedback response based on the bias location in the internal diagnostic report; an evaluation module 350 for receiving user-submitted correction fragments for the weak feedback response and performing real-time evaluation of the user-submitted correction fragments and the suggested correction schemes in the internal diagnostic report to obtain an evaluation result; and a feedback hierarchy progressive decision module 360 ​​for performing a feedback hierarchy progressive decision on the current bias record and the current feedback level in response to an incorrect evaluation result to obtain a medium or strong feedback response.

[0047] As described above, the large-model-based guided writing feedback system 300 according to embodiments of this application can be implemented in various wireless terminals, such as servers with large-model-based guided writing feedback algorithms. In one possible implementation, the large-model-based guided writing feedback system 300 according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the large-model-based guided writing feedback system 300 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the large-model-based guided writing feedback system 300 can also be one of many hardware modules of the wireless terminal.

[0048] Alternatively, in another example, the large model-based guided writing feedback system 300 and the wireless terminal can also be separate devices, and the large model-based guided writing feedback system 300 can be connected to the wireless terminal via wired and / or wireless networks and transmit interactive information in accordance with an agreed data format.

[0049] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A guided writing feedback method based on a large model, characterized in that, include: Obtain the original user input; The original user input is preprocessed to obtain preprocessed text; Potential errors are identified and comprehensively diagnosed in the preprocessed text to obtain an internal diagnostic report. The internal diagnostic report includes a list of deviations, and each element in the list includes the location of the error, the type of error, and a suggested correction. Based on the location of errors in the internal diagnostic report, a weak feedback response is generated. Receive user-submitted correction fragments in response to the weak feedback response, and evaluate the user-submitted correction fragments and the suggested correction schemes in the internal diagnostic report in real time to obtain the evaluation results; If the evaluation result is incorrect, a hierarchical decision is made on the current error record and the current feedback level to obtain a medium or strong feedback response.

2. The guided writing feedback method based on a large model according to claim 1, characterized in that, The original user input is preprocessed to obtain preprocessed text, including: The original user input is parsed to obtain the original text string; The original text string is normalized and cleaned to obtain a clean text string; Context-aware logical segmentation is performed on the clean text string to obtain preprocessed text.

3. The guided writing feedback method based on a large model according to claim 2, characterized in that, Context-aware logical segmentation is performed on the clean text string to obtain preprocessed text, including: Use strong paragraph delimiters to divide the clean text string into layers to obtain a paragraph array; Sentence segmentation is performed on each paragraph in the paragraph array to obtain the preprocessed text.

4. The guided writing feedback method based on a large model according to claim 1, characterized in that, Potential errors are identified and comprehensively diagnosed in the preprocessed text to obtain an internal diagnostic report. This report includes a list of errors, with each element indicating the error location, type, and suggested corrections. The preprocessed text is injected into the prompt word template to obtain the final prompt word; Input the final prompt words into the trained interactive German intelligent writing feedback model to obtain the model's raw response; Based on the preprocessed text, the original response of the model is analyzed and the diagnostic report is standardized to obtain an internal diagnostic report.

5. The guided writing feedback method based on a large model according to claim 1, characterized in that, Based on the location of errors in the internal diagnostic report, generate a weak feedback response, including: Extract current error records from internal diagnostic reports; Extract the error location from the current error record and assign it to the feedback payload object to obtain a weak feedback response.

6. The guided writing feedback method based on a large model according to claim 5, characterized in that, Receive user-submitted correction fragments in response to the weak feedback response, and perform real-time evaluation of the user-submitted correction fragments and suggested correction schemes in the internal diagnostic report to obtain evaluation results, including: The user-corrected segments and suggested correction schemes in the current error record are evaluated in real time to obtain a similarity score; The similarity scores are evaluated and adjudicated based on a threshold to obtain the evaluation results.

7. The guided writing feedback method based on a large model according to claim 6, characterized in that, The user-corrected segment and the suggested corrections in the current error record are evaluated in real time to obtain a similarity score, including: Based on the current error record and the user-corrected fragments, generate a corrected equivalent set; Based on the modified equivalent set, an adaptive best-match score is applied to the user's modified fragment to obtain an adaptive evaluation score as the similarity score.

8. The guided writing feedback method based on a large model according to claim 1, characterized in that, In response to an incorrect evaluation result, a hierarchical decision is made regarding the current error record and the current feedback level to obtain a medium or strong feedback response, including: The response level is currently weak feedback. Extract the error type and deviation location from the current error record; Error-type-based meta-language interpretation; The meta-language interpretation and error location are combined to obtain the mid-feedback response; The response level is currently set to medium feedback. Extract suggested correction schemes from the current error records; The suggested revisions will be used as the strong feedback response.

9. A guided writing feedback system based on a large model, characterized in that, include: The original user input acquisition module is used to acquire original user input. The preprocessing module is used to preprocess the original user input to obtain preprocessed text; The potential error identification and comprehensive diagnosis module is used to identify and comprehensively diagnose potential errors in preprocessed text to obtain an internal diagnostic report. The internal diagnostic report includes a deviation list, and each element in the deviation list includes the error location, error type, and suggested correction scheme. The weak feedback response module is used to generate a weak feedback response based on the location of the error in the internal diagnostic report; The evaluation module is used to receive user-submitted correction fragments in response to the weak feedback response, and to evaluate the user-submitted correction fragments and the suggested correction schemes in the internal diagnostic report in real time to obtain the evaluation results. The feedback hierarchy progressive decision module is used to make a feedback hierarchy progressive decision on the current error record and the current feedback level in response to an incorrect evaluation result, so as to obtain a medium feedback response or a strong feedback response.

Images