Field text proofreading method, device and equipment and storage medium

Abstract

The application discloses a domain text proofreading method and device, equipment and a storage medium, and relates to the technical field of text processing. The method comprises the following steps: in the case that a domain text to be proofread in a target domain is received, a domain dictionary corresponding to the target domain is called; the domain dictionary comprises a first dictionary in which exclusive expressions that need to be forcibly exempted in the target domain are defined, and a second dictionary in which disabled expressions that need to be forcibly corrected in the target domain are defined; the domain dictionary is implanted into a prompt word, and a large model is called to perform proofreading processing on the domain text, thereby generating a primary proofreading result; based on the domain text and the primary proofreading result, the large model is called to optimize the primary proofreading result, thereby obtaining a final proofreading result. Through model fine-tuning by using the domain dictionary, domain adaptation is realized in a low-cost manner. In combination with a proofreading mode in which primary proofreading and secondary verification optimization are combined, false detection and missed detection caused by adaptation deviation of the domain dictionary are corrected, and the proofreading performance on the domain text is comprehensively improved.

Description

Technical Field

[0001] This application relates to the field of text processing technology, and in particular to a method, apparatus, device and storage medium for proofreading text in the field. Background Technology

[0002] In highly regulated and standardized professional fields, such as finance and law, text review is a crucial step in ensuring the compliance, accuracy, and professionalism of document content. Currently, existing text review solutions all have significant shortcomings, resulting in poor performance for reviewing domain-specific texts. For example, commonly used rule-based review tools and those based on large language models differ in their capabilities. The former can only identify basic language errors such as typos and punctuation, making it difficult to adapt to the specific standards of professional fields and lacking professional adaptability. The latter generally relies on domain-specific fine-tuning of the basic large language model to achieve domain adaptation. However, domain-specific fine-tuning requires building large-scale, high-quality domain datasets, leading to high computational consumption, long technology implementation cycles, and high implementation costs. Furthermore, the fine-tuned large model lacks flexibility when facing rapidly changing business scenarios. In addition, while some solutions introduce word filtering mechanisms, they cannot fundamentally solve problems such as misjudgment of proper nouns and missed detection of prohibited expressions, ultimately resulting in poor accuracy and reliability of domain-specific text review results.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a domain text review method, apparatus, device, and storage medium, aiming to solve the technical problem that existing text review schemes have significant defects, resulting in poor performance in reviewing domain text.

[0005] To achieve the above objectives, this application proposes a domain-specific text review method, which includes: Upon receiving domain text to be reviewed in the target domain, the domain thesaurus corresponding to the target domain is invoked; the domain thesaurus includes a first thesaurus and a second thesaurus, the first thesaurus defines proprietary expressions that the target domain must be exempted from, and the second thesaurus defines prohibited expressions that the target domain must be corrected. The domain thesaurus is embedded into the prompt words, and the large model is called to review the domain text based on the embedded prompt words to generate the initial review result. Based on the domain text and the initial review result, the large model is invoked to optimize the initial review result, resulting in the final review result.

[0006] In one embodiment, the initial review result includes error correction annotations for erroneous expressions in the domain text; the step of optimizing the initial review result by calling the large model based on the domain text and the initial review result to obtain the final review result includes: Based on the domain text and the initial review results, the large model is invoked; The domain text and the initial review result are used as input parameters of the large model. The large model is used to verify the initial review result to determine the execution status of the domain thesaurus. Based on the execution results, the initial review results are optimized, and the number of newly annotated erroneous statements is determined based on the optimized review results. If the number of erroneous statements exceeds a preset threshold, the optimized review result is used as the initial review result, and the step of calling the large model based on the domain text and the initial review result is returned and executed. If the number of erroneous statements is less than or equal to the number threshold, the optimized review result will be determined as the final review result.

[0007] In one embodiment, the error correction annotation includes error type, error level, and modification suggestion; in the step of using the large model to verify the initial review result, the large model is configured as follows: In the initial review results, exemption verification is performed on expressions matching the first thesaurus, and annotation correctness verification is performed on expressions matching the second thesaurus. For the initial review results, the criteria for judging error descriptions with the same error type and error level across the entire text of the domain are checked for consistency. The reasonableness of the suggested modifications in the error correction annotations of the initial review results is verified.

[0008] In one embodiment, after the step of optimizing the initial review result by calling the large model based on the domain text and the initial review result to obtain the final review result, the method further includes: The final review results are displayed visually in a preset interactive interface; Receive feedback from the user on the final review result on the interactive interface; The domain terminology is updated based on the feedback.

[0009] In one embodiment, the step of updating the domain lexicon based on the feedback operation includes: An operation instruction is generated based on the feedback operation; if the feedback operation includes a dictionary addition operation, the operation instruction includes a dictionary addition instruction. In response to the lexicon addition instruction, the target expression specified by the feedback operation is added to the first lexicon or the second lexicon to update the domain lexicon.

[0010] In one embodiment, the interactive interface includes a first display area and a second display area; the step of visually displaying the final review result in the preset interactive interface includes: The original text of the domain text is displayed in the first display area, and the erroneous text fragments corresponding to each erroneous statement recorded in the final review result are highlighted in the original text. The error correction annotations of the erroneous text fragments are displayed in the second display area; wherein the error correction annotations and the corresponding erroneous text fragments are interactively linked.

[0011] In one embodiment, after the step of receiving feedback from the user on the final review result on the interactive interface, the method further includes: If the target representation is determined to be an erroneous representation based on the final review result, a target erroneous representation in the domain text with the same error type as the target representation is determined based on the error type of the target representation; wherein, the target representation is the representation specified by the feedback operation; If the feedback operation includes a modification operation, the target representation and the target error representation are synchronously corrected according to the modification operation.

[0012] Furthermore, to achieve the above objectives, this application also proposes a domain text review device, which includes: The calling module is used to call the domain lexicon corresponding to the target domain when it receives domain text to be reviewed in the target domain; the domain lexicon includes a first lexicon and a second lexicon, the first lexicon defines proprietary expressions that the target domain needs to be exempted, and the second lexicon defines prohibited expressions that the target domain needs to be corrected. The initial review module is used to embed the domain thesaurus into the prompt words, and call the large model to review the domain text based on the embedded prompt words to generate the initial review result. The verification and optimization module is used to optimize the initial review result by calling the large model based on the domain text and the initial review result, so as to obtain the final review result.

[0013] In addition, to achieve the above objectives, this application also proposes a domain text review device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the domain text review method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the domain text review method described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the domain text review method described above.

[0016] One or more technical solutions proposed in this application have at least the following technical effects: Upon receiving domain text to be reviewed in the target domain, the process first involves calling the corresponding domain lexicon and embedding it into the prompt words of the large model. This guides the large model to review the domain text, yielding an initial review result. Then, based on the initial review, optimization is performed to correct false positives and false negatives caused by lexicon adaptation biases, resulting in the final review result. By introducing a domain lexicon and dynamically embedding it into the large model's review process through prompt word engineering, this approach replaces the costly and time-consuming domain fine-tuning of the large model itself. This significantly reduces computational consumption, shortens the technology implementation cycle, saves implementation costs, and improves the flexibility of the large model in responding to changes in business scenarios. Furthermore, the domain thesaurus explicitly defines the rule execution logic for mandatory exemption and mandatory error correction, guiding large models to comply with the thesaurus rules during the review process. This ensures that proper nouns are systematically exempted from review and avoids misjudgment, while forcibly detecting all prohibited expressions within a specific domain to prevent omissions. Based on the initial review, optimization is carried out by combining the initial review with secondary review and optimization to correct misjudgments and omissions caused by domain thesaurus adaptation bias, fundamentally improving the review performance of professional domain texts. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an embodiment of the text review method in this application field; Figure 2 This is a schematic diagram of the structure of the text review system provided in the embodiments of this application; Figure 3 A schematic diagram illustrating the initial review process of the domain text provided in this application embodiment; Figure 4 A schematic diagram of the verification and optimization process for the initial review results provided in this application embodiment; Figure 5 A simplified flowchart illustrating the domain text review method provided in this application embodiment; Figure 6 This is a schematic diagram of the module structure of a text review device in the field of this application. Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the domain text review method in this application embodiment.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] The main solution of this application embodiment is: intelligent text review for professional fields. It relies on a large language model that has been pre-trained on a large scale, and adopts the method of embedding prompt words of the large model into the domain lexicon configuration to replace the domain fine-tuning of the large model. After the large model performs the initial review of the text to be reviewed, the initial review result is verified and optimized, and the final review result is displayed visually. Based on the user's feedback on the visualized review result, the domain lexicon is updated and the full text is adapted in sync, so as to realize the dynamic updating and continuous optimization of the review capability.

[0024] In this embodiment, for ease of description, the domain text review device will be used as the execution subject in the following description.

[0025] Existing text review solutions all have shortcomings. For example, solutions based on fixed rules lack adaptability to specific domains, and rule updates are lagging, making them unable to cope with dynamically changing text content requirements within a domain, resulting in poor flexibility. Review solutions based on large-scale model fine-tuning consume a lot of computing power, have long technology implementation cycles, high implementation costs, and insufficient flexibility in adapting to business scenarios. Review solutions using word filtering mechanisms cannot simultaneously solve the problems of misjudging proper nouns and missing prohibited expressions. Furthermore, most existing text review solutions are single-round outputs without secondary review and verification mechanisms, resulting in weak ability to detect hidden errors. They also lack dynamic optimization mechanisms driven by user feedback, failing to transform user adjustments during use into system knowledge. Long-term use still requires a lot of manual intervention, making it difficult to continuously improve review efficiency.

[0026] Based on this, this application provides a solution that achieves professional domain adaptation in a low-cost manner by fine-tuning the replacement model with a configurable domain lexicon, balancing adaptation efficiency and flexibility. The domain lexicon defines proprietary expressions that must be exempted from professional domain requirements and prohibited expressions that must be corrected, which can protect domain proper nouns and prohibited expressions that require correction. It can simultaneously solve the problems of misjudgment of proper nouns and missed detection of prohibited expressions, thereby improving the proofreading accuracy of professional domains.

[0027] Based on this, a combination of initial review and secondary review is adopted. The secondary review corrects the deviations in the initial review results, reduces missed and false detections, and improves the overall review performance of professional texts.

[0028] Furthermore, by visually displaying the final proofreading results, user feedback is obtained, allowing for updates to the domain thesaurus and full-text adaptation based on this feedback. This achieves "one-point feedback, full-text adaptation," transforming user needs into knowledge accumulation during use, thus enabling dynamic optimization and continuous improvement of proofreading capabilities. Moreover, the proofreading results and user feedback are logically linked, simplifying user operations and resolving the issues of poor interactivity and cumbersome operation in existing solutions.

[0029] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or domain text proofreading device capable of performing the above functions. The following description uses a domain text proofreading device as an example to illustrate this embodiment and the subsequent embodiments.

[0030] This application provides a domain-specific text review method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the text review method in this application.

[0031] In this embodiment, the domain text review method includes steps S10 to S30: Step S10: Upon receiving the domain text to be reviewed in the target domain, the domain thesaurus corresponding to the target domain is invoked; the domain thesaurus includes a first thesaurus and a second thesaurus, wherein the first thesaurus defines proper nouns that the target domain must be exempted from mandatory exemption, and the second thesaurus defines prohibited expressions that the target domain must be corrected. Step S20: The domain terminology is embedded into the prompt words. Based on the embedded prompt words, the large model is called to perform review processing on the domain text and generate the initial review result. Step S30: Based on the domain text and the initial review result, call the large model to optimize the initial review result and obtain the final review result.

[0032] First, upon receiving the domain text to be reviewed in the target domain, the domain thesaurus corresponding to the target domain is invoked. The domain thesaurus includes a first thesaurus and a second thesaurus. The first thesaurus defines proprietary expressions that the target domain needs to be exempted from mandatory review. These proprietary expressions include, but are not limited to, proper nouns, technical terms, fixed expressions, and compliant terms in the target domain. The second thesaurus defines prohibited expressions that the target domain needs to be corrected. These prohibited expressions include, but are not limited to, prohibited words and phrases, sensitive expressions, and prohibited wording.

[0033] It should be noted that in some embodiments, both proprietary expressions and disabled expressions of the target domain can be defined simultaneously in the domain thesaurus, without any specific limitations.

[0034] It should also be noted that the target domain can be any specific professional field, and the domain thesaurus for each professional field is configurable.

[0035] Upon receiving domain text to be reviewed, the domain thesaurus associated with the target domain is invoked based on the target domain to which the neighboring text belongs. Invoking the domain thesaurus includes, but is not limited to, static loading. In one embodiment, the domain thesaurus is stored in a database, and upon receiving domain text to be reviewed in the target domain, the domain thesaurus is dynamically retrieved from the database and loaded based on the metadata of the domain text.

[0036] The metadata of the domain text includes, but is not limited to, filename extensions, file header information, and domain tags selected by the user. The domain text can be uploaded or entered by the user, and the domain tags in the metadata can be generated by the user during the process of uploading the domain text to be reviewed.

[0037] For the domain lexicon of the target domain, it is embedded into the prompt words of the large model. Based on the prompt words after the domain lexicon is embedded, the large model is called to perform the review process on the domain text to be reviewed, and the initial review result is obtained.

[0038] It should be noted that the domain lexicon defines the review benchmark rules for the large model. These review benchmark rules include the first lexicon rules defined by the first lexicon and the second lexicon rules defined by the second lexicon.

[0039] In one feasible implementation, for the domain thesaurus to be invoked, it is dynamically embedded into the prompt words of the large model through a preset template structure. Then, using the prompt words containing specific review benchmark rules, the corresponding large model is invoked to review the domain text to be reviewed.

[0040] During the review process, the large model is configured to: perform review tasks based on review instructions and lexicon rules in the prompt words, identify expression problems in the domain text, and thus generate an initial review result. This initial review result reflects the exemption of expressions matching the first lexicon and the mandatory correction of expressions matching the second lexicon.

[0041] The initial review results are structured, meaning that the review task is performed based on the review instructions and lexicon rules in the prompt words, the expression problems existing in the domain text are identified, and thus the structured initial review results are generated.

[0042] In one embodiment, a "layered instruction" strategy is adopted for the implantation of the domain lexicon. First, a basic instruction is sent to the large model, which instructs the large model to act as a domain review expert, thus defining the role of the large model. Then, the content of the domain lexicon is appended as a review rule appendix to the domain text to be reviewed, and the large model is instructed to prioritize and compelly refer to this review rule appendix before making any linguistic logic judgments, thereby clearly separating the role instruction from the review rules.

[0043] In another embodiment, before the review process, the text to be reviewed can be preprocessed by temporarily replacing all text expressions matched by the first thesaurus (e.g., replacing them with specific tags or encrypted hash values). The preprocessed text and the complete hints containing domain terms are then input into the large model. After the large model completes the review, the domain text is post-processed in the output stage, replacing the temporary tags back with the original text expressions. This approach can shield proprietary expressions during the review process, preventing the large model from making any grammatical or semantic misjudgments, thus reducing misjudgments of the domain text.

[0044] Based on the domain text and the initial review results, the large model is invoked again to optimize the initial review results, resulting in the final review result. Utilizing the large model's reflection and verification capabilities, the initial review results are optimized. Specifically, the original domain text and the initial review results are used as input to the large model, which is then invoked again to instruct it to review and correct the initial review results from different dimensions, resulting in a more accurate and reliable final review result.

[0045] In this embodiment, the large model called for optimizing the initial review result is the same as the large model called for the initial review. In another embodiment, the large model called for optimizing the initial review result is not the same as the large model called for the initial review. That is, the original domain text and the initial review result are used as input to the large model, and the same or another large model is called again, instructing the called large model to review and correct the initial review result from different dimensions, so as to obtain a more accurate and reliable final review result.

[0046] In one feasible implementation, the optimization of the initial review results employs a role-based debate mechanism. Two large model instances with different initial instructions are invoked to play different roles, or the same large model is prompted to play two different roles. One role is "Reviewer A," whose task is to rigorously review the domain text and identify potential errors. The other role is "Reviewer B," whose task is to carefully examine the original domain text and the initial review results to identify potential omissions—errors not detected by Reviewer A. After both roles submit their review results, a pre-designated logic execution module acts as the referee, merging and resolving conflicts to form the final review result.

[0047] In another feasible implementation, the optimization of the initial review results focuses on confidence level calibration. Specifically, prompts are used to instruct the large model to assign a confidence score to each item in the initial review results. Then, for review items with confidence scores below a preset threshold (e.g., below 80%), a second review process is triggered. During the second review, detailed context related to the review item is provided to the large model, requiring it to conduct a thorough review. By selectively optimizing the initial review results, the consumption of computational resources can be effectively controlled while ensuring the quality of the review results.

[0048] As an example, the domain text review method provided in this application is applied to a text review system, the overall architecture of which is as follows: Figure 2 As shown, the text review system includes an interaction layer, an application layer, and a storage layer.

[0049] Furthermore, such as Figure 2As shown, the interaction layer provides a user interface for users to upload domain text to be reviewed; the application layer includes a rule configuration module, an initial review module, and a verification and optimization module; the storage layer contains a database and a large-scale pre-trained model. The rule configuration module configures the domain lexicon. The initial review module retrieves the domain lexicon from the storage layer's database and embeds it into the prompt words, then uses the large model in the storage layer to perform review processing based on these prompt words, obtaining the initial review result. The verification and optimization module uses the large model in the storage layer to review and correct the initial review result against the original content of the domain text, obtaining the final review result.

[0050] In one embodiment, the rule configuration module may further include a first configuration submodule and a second configuration submodule. The first configuration submodule is used to configure proprietary expressions that need to be forcibly exempted to obtain a corresponding first lexicon. The second configuration submodule is used to configure disabled expressions that need to be forcibly corrected to obtain a corresponding second lexicon.

[0051] In the text review system, the components work together. First, the domain lexicon is configured and initialized through the rule configuration module. After the domain lexicon is called and embedded into the prompt words of the large model, the large model can be guided to perform review.

[0052] In some embodiments, the text review system further includes a preprocessing module for preprocessing the domain text to be reviewed. After preprocessing, the domain text is input into the initial review module for further processing, resulting in an initial review result output by the initial review module. Further, the initial review result is validated and optimized by a validation and optimization module to obtain the final review result.

[0053] The initial proofreading module, the core functional module for text proofreading, is configured to rely on a large model. Through a combination of fixed templates and dynamic variables, a prompt word engineering process embeds the domain lexicon as dynamic variables into the prompt words, guiding the large model to proofread domain-specific text. The prompt words instruct the large model to act as a domain text proofreading expert, strictly adhering to industry standards, regulatory requirements, and rules of the neighboring lexicon, and being familiar with the grammatical rules and writing conventions of the target domain, as well as proficient in the normative and standard requirements for various types of text in the target domain. This clear role positioning activates the knowledge related to text proofreading and industry standards within the large model, ensuring that the error correction results meet professional standards and industry requirements.

[0054] The core task of the initial review module is review. The prompts from the large model instruct it to review and analyze the original content of the domain text, focusing on errors in the text. These errors include, but are not limited to, typos, extra or missing words, inappropriate word choice, incorrect numbering, contradictory data, duplicate content, punctuation errors, incomplete sentences, grammatical errors, formatting errors, unclear expression, overly colloquial language, and imprecise expression. The initial review results should list all identified errors; if no errors are found, an empty list should be output, ensuring the large model doesn't overlook key error-correction dimensions and comprehensively covers all possible problems in the domain text.

[0055] Furthermore, during the review process, for the domain lexicon dynamically embedded in the prompt words, the large model is explicitly instructed to exempt expressions matching the first lexicon from error correction and not generate review records. For expressions matching the second lexicon, their language logic is ignored, and error correction annotation is enforced to avoid misjudging proprietary expressions or introducing domain-disabled expressions, thereby improving the system's adaptability to business scenarios.

[0056] In this embodiment, upon receiving domain text to be reviewed in the target domain, the process begins by calling the domain lexicon corresponding to the target domain and embedding it into the prompt words of the large model. This guides the large model to review the domain text, yielding a preliminary review result. Then, based on the initial review, optimization is performed to correct false positives and false negatives caused by lexicon adaptation biases, resulting in the final review result. By introducing a domain lexicon and dynamically embedding it into the large model's review process through prompt word engineering, the costly and time-consuming domain fine-tuning of the large model itself is replaced by prompt word insertion. This significantly reduces computational power consumption, shortens the technology implementation cycle, saves implementation costs, and improves the flexibility of the large model in responding to changes in business scenarios. Furthermore, the domain thesaurus explicitly defines the rule execution logic for mandatory exemption and mandatory error correction, guiding large models to comply with the thesaurus rules during the review process. This ensures that proper nouns are systematically exempted from review and avoids misjudgment, while forcibly detecting all prohibited expressions within a specific domain to prevent omissions. Based on the initial review, optimization is carried out by combining the initial review with secondary review and optimization to correct misjudgments and omissions caused by domain thesaurus adaptation bias, fundamentally improving the review performance of professional domain texts.

[0057] In one feasible implementation, the initial review result includes error correction annotations for erroneous expressions in the domain text, and the optimization of the initial review result is iterative. Based on this, step S30 may include steps S31 to S35: Step S31: Based on the domain text and the initial review result, call the large model; Step S32: Use the domain text and the initial review result as input parameters of the large model, and use the large model to perform a verification operation on the initial review result to determine the execution status of the domain thesaurus; Step S33: Optimize the initial review result based on the execution status, and determine the number of newly added erroneous statements based on the optimized review result; Step S34: If the number of erroneous statements is greater than a preset threshold, the optimized review result is used as the initial review result, and the step of calling the large model based on the domain text and the initial review result is returned and executed. Step S35: If the number of erroneous statements is less than or equal to the number threshold, then the optimized review result is determined as the final review result.

[0058] Based on the domain text and the initial review results, a large model is invoked, using the domain text and the initial review results as input parameters to the invoked large model. The large model is then used to perform a verification operation on the initial review results, which is performed iteratively.

[0059] Specifically, in each iteration, the execution status of the domain thesaurus is determined, and the initial review results are optimized based on the execution status to obtain the optimized review results.

[0060] Furthermore, based on the optimized review results, the number of newly labeled erroneous expressions is determined. As one implementation method, the optimized review results in the current iteration are compared with the optimized review results in the previous iteration to determine the number of newly added labels marked as erroneous expressions in this iteration. It should be noted that in the first iteration, the optimized review results are compared with the initial review results to determine the number of newly labeled erroneous expressions added in the first iteration optimization.

[0061] If the number of newly added erroneous annotations exceeds a preset threshold, the optimized review result is used as the initial review result, and step S32 is re-executed. If the number of newly added erroneous annotations is less than or equal to the preset threshold, the optimized review result is determined as the final review result. When the optimized review result is used as the initial review result, in each iteration of optimization, the optimized review result is compared with the initial review result to determine the number of newly added erroneous annotations in the optimized review result.

[0062] As an example, the error correction annotation for the erroneous statement includes the error type, error level, and modification suggestions. In step S32, during the verification operation of the initial review results using the large model, the large model is configured as follows: In the initial review results, exemption verification is performed on expressions matching the first thesaurus, and annotation correctness verification is performed on expressions matching the second thesaurus. For the initial review results, the criteria for judging error descriptions with the same error type and error level across the entire text of the domain are checked for consistency. The reasonableness of the suggested modifications in the error correction annotations of the initial review results is verified.

[0063] In other words, during the optimization process of the initial review results, the large model is configured to perform exemption verification on expressions matching the first thesaurus and annotation correctness verification on expressions matching the second thesaurus in the initial review results; perform consistency verification on the judgment criteria of similar erroneous expressions with the same error type and error level in the full text of the domain text in the initial review results; and perform rationality verification on the modification suggestions in the error correction annotations of the initial divine results, thereby obtaining the execution status of the domain thesaurus.

[0064] In one embodiment, reference is made to Figure 3 The initial review process shown involves the large model performing an initial review of the domain text. This review uses neighboring text as input and employs a prompt word engineering technique to process the domain text, yielding the initial review result as the output of the large model. The prompt words in the prompt word engineering technique are primarily used to guide the large model's role positioning, review task, error correction annotation, review benchmark rules, and the output format of the review result.

[0065] Specifically, role positioning defines the domain reviewer's identity for the large model; review tasks guide the large model to perform text review as a domain reviewer; error correction annotation guides the large model to annotate detected erroneous expressions, defining the information and format to be annotated; review benchmark rules are defined by the domain lexicon, mainly including exemption rules for proprietary expressions from the first lexicon and prohibition rules defined by prohibited expressions from the second lexicon; and the output format defines the large model's... Optionally, the initial review results can be output in JSON format. The standardized output format defines several key fields, including error type, error level, suggested corrections, original text, and revised text. The error level should include at least high, medium, and low ratings to help users quickly identify the severity or importance of the error and prioritize addressing critical issues.

[0066] Furthermore, referring to Figure 4The verification and optimization process for the initial review results is shown. In the process of verifying and optimizing the initial review results, the original content of the domain text to be reviewed and the initial review results are used as inputs to the large model. The large model is guided to perform the verification operation through prompt word engineering, and the optimized review results output by the large model are obtained, as well as the judgment results that determine whether further optimization is needed.

[0067] Among them, the prompt words in the prompt word project are mainly used to guide the large model to perform rule adaptation verification, error judgment standard consistency verification, modification suggestion rationality verification, and determine the number of erroneous expressions in newly added annotations.

[0068] Specifically, rule adaptation verification is mainly used to verify expressions in the domain text that match the domain thesaurus. This includes exemption verification for expressions matching the first thesaurus and annotation correctness verification for expressions matching the second thesaurus, determining whether erroneous expressions requiring mandatory correction have been correctly annotated to avoid false positives and false negatives. Error judgment standard consistency verification is mainly used to guide the large model to verify the consistency of judgment standards for similar erroneous expressions across the entire domain text. Similar erroneous expressions specifically refer to those with the same error type and error level. The judgment standards for similar expressions are compared across paragraphs to ensure logical consistency throughout the text. Modification suggestion rationality verification is mainly used to guide the large model to verify the rationality of modification suggestions in the error correction annotations, including but not limited to verifying the compliance of the modification suggestions' content, whether they conform to language norms and industry practices, to avoid new deviations.

[0069] By verifying rule adaptation, consistency of error judgment criteria, and rationality of modification suggestions, the execution status of the domain thesaurus is obtained. The number of newly labeled erroneous expressions in the prompt word project is used to guide the large model to correct and optimize the initial review results based on the execution status of the domain thesaurus, obtain the optimized review results, and compare them with the initial review results to determine the number of newly labeled erroneous expressions, thereby determining whether the initial review results need to be further optimized.

[0070] Finally, the optimized review result and the conclusion of whether further optimization is needed are used as the output of the large model. Based on the judgment of whether further optimization is needed in the output of the large model, if further optimization is needed, the optimized review result is used as the initial review result, and the verification and optimization of the initial review result is retried. If further optimization is not needed, the optimized review result is the final review result.

[0071] The initial review results are validated and optimized using a multi-round iterative optimization approach guided by instructions and multi-dimensional validation. Based on a large model, the initial review results are iteratively optimized through a multi-dimensional instruction system. In each round of optimization, unidentified erroneous statements are supplemented, misjudged review records are deleted, and unreasonable modification suggestions are adjusted until there are no new error correction annotations or the number of newly added error annotations is less than or equal to the preset threshold, at which point the final review result is obtained.

[0072] In one feasible implementation, steps S40 to S60 may be included after step S30: Step S40: The final review result is displayed visually in a preset interactive interface; Step S50: Receive feedback from the user on the final review result on the interactive interface; Step S60: Update the domain terminology based on the feedback operation.

[0073] The final review results are displayed visually in a pre-defined interactive interface. Based on the visual display of the final review results, the system receives user feedback on the final review results in the interactive interface and updates the domain terminology database accordingly.

[0074] As an example, refer to Figure 5 The review process shown begins with configuring a domain-specific thesaurus. This allows for the management and exemption of specialized expressions within a particular domain, as well as mandatory warnings for prohibited expressions. The domain thesaurus configuration supports single-entry addition, batch import, editing, deletion, and categorized management in text or table formats. It also supports vocabulary annotations, such as marking domain attributes, usage scenarios, and reasons for disabling. The configured domain thesaurus is synchronized to the database and serves as the review baseline rule. During the initial review, the domain thesaurus is invoked via API interactions, embedding warning words. Exemption logic for the domain thesaurus has higher priority than basic review.

[0075] Furthermore, the system receives domain text uploaded by the user for review. Optionally, this can be received through a user interface. Based on the domain text, the system calls the corresponding domain lexicon, embeds it into prompt words, and calls the large model. The prompt words guide the large model to perform an initial review of the domain text, resulting in the initial review result.

[0076] The initial review results and domain text are used as input parameters for the large model. The large model is then called again to verify and optimize the initial review results of the domain text. This includes exemption verification for expressions matching the first thesaurus in the initial review results, correctness verification for annotations matching the second thesaurus, consistency verification for the judgment criteria of erroneous expressions with the same error type and error level throughout the entire domain text, and reasonableness verification for the modification suggestions in the error correction annotations of the initial review results. The execution status of the domain thesaurus is then determined.

[0077] Based on the execution of the domain thesaurus, the initial review result is optimized to obtain an optimized review result. The number of newly annotated erroneous expressions in the optimized review result compared to the initial review result is determined. If the number of newly annotated erroneous expressions is greater than a preset threshold, the optimized review result is used as the initial review result and, together with the domain text, as input parameters of the large model to re-trigger the optimization of the initial review result until the number of newly annotated erroneous expressions is less than or equal to the preset threshold, thus obtaining the final review result.

[0078] The final review results are displayed visually. The interactive interface for this visualization can be the same as or different from the user interface for the domain text uploaded by the user for review; there are no specific limitations. Based on the visual display of the final review results, users can interact with the review results of the domain text and receive feedback from users on the final review results through the interactive interface. Driven by user feedback, the domain thesaurus is updated and kept up-to-date.

[0079] As one embodiment, based on the user's feedback operation, it is determined whether to add the user feedback to the domain thesaurus. If it is determined that the user feedback should be added to the domain thesaurus, the domain thesaurus is updated according to the user's feedback operation, and the review of the domain text ends. If it is determined that the user feedback does not need to be added to the domain thesaurus, the review of the domain text ends.

[0080] In one feasible implementation, the user's feedback operation on the final review result via the interactive interface includes at least one of the following: modification of the error correction annotation and addition of the dictionary. If the user's feedback operation includes adding the dictionary, then step S60 includes steps S61-S62: Step S61: Generate an operation instruction based on the feedback operation; if the feedback operation includes a dictionary addition operation, the operation instruction includes a dictionary addition instruction. Step S62: In response to the lexicon addition instruction, the target expression specified by the feedback operation is added to the first lexicon or the second lexicon to update the domain lexicon.

[0081] Based on user feedback, operation instructions are generated. If the feedback operation includes a dictionary addition operation, the operation instructions generated based on the user feedback operation include a dictionary addition instruction. In response to the dictionary addition instruction, the target expression specified by the user feedback operation is added to the first dictionary or the second dictionary to update the domain dictionary.

[0082] Specifically, the user's dictionary addition instruction includes the target dictionary to be executed when performing the user's dictionary addition operation. If the target dictionary is the first dictionary in the domain dictionary, the target expression is added to the first dictionary. If the target dictionary is the second dictionary in the domain dictionary, the target expression is added to the second dictionary.

[0083] In one feasible implementation, when the user's feedback operation includes a modification operation, step S60 includes steps S63-S64: Step S63: If the target representation is determined to be an erroneous representation based on the final review result, a target erroneous representation in the domain text with the same error type as the target representation is determined based on the error type of the target representation; wherein, the target representation is the representation specified by the feedback operation; Step S64: If the feedback operation includes a modification operation, the target representation and the target error representation are synchronously corrected according to the modification operation.

[0084] If the target statement is determined to be incorrect based on the final review results, a target incorrect statement with the same error type in the domain text is identified based on the error type of the target statement. Here, the target statement is the statement specified by the user's feedback operation. If the user's feedback operation includes a correction operation, both the target statement and the target incorrect statement are corrected simultaneously according to this correction operation. This achieves full-text adaptation with feedback in one place, reducing the number of times users provide feedback on similar statements and avoiding repetitive operations on similar statements.

[0085] In some embodiments, step S60 may include the following steps: An operation instruction is generated based on the feedback operation; if the feedback operation includes a dictionary addition operation, the operation instruction includes a dictionary addition instruction; if the feedback operation includes a correction operation, the operation instruction includes a correction instruction. In response to the lexicon addition instruction, the target expression executed by the feedback operation is added to the first lexicon or the second lexicon to update the domain lexicon; If the feedback operation includes a correction operation, it is determined whether the target statement is an incorrect statement based on the final review result; If the target representation is an incorrect representation, determine the target incorrect representation in the domain text that has the same error type as the target representation, based on the error type of the target representation. In response to the correction instruction, the target representation and the target error representation are corrected synchronously.

[0086] In other embodiments, the user's correction operation includes a correction type and a correction scope for the erroneous statement. The correction type specifically includes accepting modification, rejecting modification, and manual modification. The correction scope specifically includes single modification and full-text adaptation modification. When the correction type is accepting modification, the modification suggestion marked with the erroneous statement is adopted. When the correction type is rejecting modification, the modification suggestion marked with the erroneous statement is not adopted. When the correction type is manual modification, the modification suggestion manually marked by the user is adopted.

[0087] When simultaneously correcting the target representation and the target error representation, at least one of the target representation and the target error representation is corrected according to the correction type and correction scope corresponding to the correction instruction. The correction operation is to accept modification, reject modification, or manually modify. When the correction scope is a single correction, only the target representation is modified. When the correction scope is a full-text adaptation modification, the target representation and the target error representation are modified simultaneously.

[0088] Based on this, step S60 may also include: An operation instruction is generated based on the feedback operation; if the feedback operation includes a dictionary addition operation, the operation instruction includes a dictionary addition instruction; if the feedback operation includes a correction operation, the operation instruction includes a correction instruction. In response to the lexicon addition instruction, the target expression executed by the feedback operation is added to the first lexicon or the second lexicon to update the domain lexicon; If the feedback operation includes a correction operation, it is determined whether the target statement is an incorrect statement based on the final review result; If the target representation is an incorrect representation, determine the target incorrect representation in the domain text that has the same error type as the target representation, based on the error type of the target representation. In response to the correction instruction, the target representation and the target error representation are simultaneously corrected according to the correction type and correction scope contained in the correction instruction.

[0089] In one possible implementation, step S40 may include: Step S41: Display the original text of the domain text in the first display area, and highlight the erroneous text fragments corresponding to each erroneous statement recorded in the final review result in the original text; Step S42: Display the error correction annotation of the erroneous text fragment in the second display area; wherein the error correction annotation and the corresponding erroneous text fragment have interactive linkage.

[0090] The interactive interface for visually displaying the final review results includes a first display area and a second display area. Specifically, in the process of visually displaying the final review results, the original text of the domain text is displayed in the first display area, and the erroneous text fragments corresponding to each erroneous statement recorded in the final review results are highlighted in the original text.

[0091] Furthermore, the error correction annotations for the erroneous text fragments are displayed in the second display area, and these annotations are interactive with the corresponding erroneous text fragments.

[0092] It should be noted that the error correction annotation for the erroneous text fragment includes the error type, error level, and modification suggestions. The modification suggestions include the modification method for the erroneous text fragment and the modified text content. In some embodiments, the error correction annotation may also include the original content of the erroneous text fragment.

[0093] In other words, the interactive interface uses a dual-zone layout to visualize the final review results, which is divided into the original text display area and the error correction information area. The first display area is the original text display area, which fully presents the original content of the domain text. The second display area is the error correction information area, which is used to display the error correction annotations in the domain text. These error correction annotations can be displayed in the form of a list or other format.

[0094] Furthermore, users can view and manipulate all error statements in the second display area. During the user's operation on the error correction mark, there is an interactive linkage between the error correction mark and the corresponding error text fragment. For example, when the user's operation cursor hovers over one of the error text fragments or error correction marks, it will automatically locate the other item in the second display area to help the user quickly understand the details of the error statement.

[0095] In this embodiment, a domain-specific thesaurus configuration is used to embed prompt words into the large model, replacing the domain-specific fine-tuning of the large model. By defining proprietary expressions and disabled expressions for a professional domain through the domain thesaurus, there is no need to build a large-scale dataset required for fine-tuning training. This achieves accurate adaptation to professional domains in a low-cost manner, can be implemented quickly, consumes less computing power, and can adapt to rapidly changing business scenarios and different professional domains, thus improving the flexibility of text review.

[0096] Furthermore, by combining initial review with secondary verification and optimization, the deviations in the adaptation of domain-specific thesaurus rules are specifically corrected through verification and optimization, reducing issues such as missed reports and misjudgments, and significantly improving the accuracy and reliability of text review in professional domains.

[0097] By adopting a user feedback-driven preservation mechanism, we can not only achieve full-text adaptation with one point of feedback, but also directly transform user feedback into system knowledge, dynamically iterate the review capabilities during user use, and continuously adapt to changes in rules in the professional field.

[0098] By clearly defining proprietary and prohibited expressions in a professional field through a domain-specific thesaurus, the large model is guided to enforce exemptions for proprietary expressions and error correction annotations for prohibited expressions. This is done without being constrained or interfered with by language logic, accurately matching the review needs of the professional field and avoiding compliance risks from the source.

[0099] Based on the interactive linkage between the review results and the original content, the entire process of configuring the thesaurus, displaying the review results, and submitting user feedback is realized on the interactive interface, which realizes convenient operation, visualizes the results, and improves the efficiency of batch review of domain texts.

[0100] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the text review methods in this application field. Any simple modifications based on this technical concept are within the protection scope of this application.

[0101] This application also provides a domain-specific text review device; please refer to [reference needed]. Figure 6 The domain text review device includes: The calling module 10 is used to call the domain lexicon corresponding to the target domain when it receives domain text to be reviewed in the target domain; the domain lexicon includes a first lexicon and a second lexicon, the first lexicon defines proprietary expressions that the target domain needs to be exempted, and the second lexicon defines prohibited expressions that the target domain needs to be corrected. The initial review module 20 is used to embed the domain thesaurus into the prompt words, and call the large model to review the domain text based on the embedded prompt words to generate the initial review result. The verification optimization module 30 is used to optimize the initial review result by calling the large model based on the domain text and the initial review result, so as to obtain the final review result.

[0102] In one embodiment, the initial review result includes error correction annotations for erroneous expressions in the domain text; the verification optimization module 30 is further configured to: Based on the domain text and the initial review results, the large model is invoked; The domain text and the initial review result are used as input parameters of the large model. The large model is used to verify the initial review result to determine the execution status of the domain thesaurus. Based on the execution results, the initial review results are optimized, and the number of newly annotated erroneous statements is determined based on the optimized review results. If the number of erroneous statements exceeds a preset threshold, the optimized review result is used as the initial review result, and the step of calling the large model based on the domain text and the initial review result is returned and executed. If the number of erroneous statements is less than or equal to the number threshold, the optimized review result will be determined as the final review result.

[0103] In one embodiment, the error correction annotation includes error type, error level, and modification suggestions; in the verification optimization module 30, the large model is configured as follows: In the initial review results, exemption verification is performed on expressions matching the first thesaurus, and annotation correctness verification is performed on expressions matching the second thesaurus. For the initial review results, the criteria for judging error descriptions with the same error type and error level across the entire text of the domain are checked for consistency. The reasonableness of the suggested modifications in the error correction annotations of the initial review results is verified.

[0104] In one embodiment, the domain text review device further includes an interaction optimization module, used for: The final review results are displayed visually in a preset interactive interface; Receive feedback from the user on the final review result on the interactive interface; The domain terminology is updated based on the feedback.

[0105] In one embodiment, the interaction optimization module is further configured to: An operation instruction is generated based on the feedback operation; if the feedback operation includes a dictionary addition operation, the operation instruction includes a dictionary addition instruction. In response to the lexicon addition instruction, the target expression specified by the feedback operation is added to the first lexicon or the second lexicon to update the domain lexicon.

[0106] In one embodiment, the interactive interface includes a first display area and a second display area; the interaction optimization module is further configured to: The original text of the domain text is displayed in the first display area, and the erroneous text fragments corresponding to each erroneous statement recorded in the final review result are highlighted in the original text. The error correction annotations of the erroneous text fragments are displayed in the second display area; wherein the error correction annotations and the corresponding erroneous text fragments are interactively linked.

[0107] In one embodiment, the interaction optimization module is further configured to: If the target representation is determined to be an erroneous representation based on the final review result, a target erroneous representation in the domain text with the same error type as the target representation is determined based on the error type of the target representation; wherein, the target representation is the representation specified by the feedback operation; If the feedback operation includes a modification operation, the target representation and the target error representation are synchronously corrected according to the modification operation.

[0108] The domain text review apparatus provided in this application, employing the domain text review method described in the above embodiments, can solve the technical problem of poor performance in reviewing domain text due to defects in existing text review schemes. Compared with the prior art, the beneficial effects of the domain text review apparatus provided in this application are the same as those of the domain text review method provided in the above embodiments, and other technical features in the domain text review apparatus are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0109] This application provides a domain text review device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the domain text review method in Embodiment 1 above.

[0110] The following is for reference. Figure 7 This document illustrates a structural diagram of a domain text review device suitable for implementing embodiments of this application. The domain text review device in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 7The illustrated domain text review device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0111] like Figure 7 As shown, the domain text proofreading device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the domain text proofreading device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the domain text proofreading device to communicate wirelessly or wiredly with other devices to exchange data. While the figures show domain text proofreading devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0112] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0113] The domain text review device provided in this application, employing the domain text review method described in the above embodiments, can solve the technical problem that existing text review schemes have significant defects, resulting in poor review performance of domain text. Compared with the prior art, the beneficial effects of the domain text review device provided in this application are the same as those of the domain text review method provided in the above embodiments, and other technical features in this domain text review device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0114] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0115] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0116] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the domain text review method in the above embodiments.

[0117] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0118] The aforementioned computer-readable storage medium may be included in the domain text review device; or it may exist independently and not assembled into the domain text review device.

[0119] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the domain text review device, cause the domain text review device to: Upon receiving domain text to be reviewed in the target domain, the domain thesaurus corresponding to the target domain is invoked; the domain thesaurus includes a first thesaurus and a second thesaurus, the first thesaurus defines proprietary expressions that the target domain must be exempted from, and the second thesaurus defines prohibited expressions that the target domain must be corrected. The domain thesaurus is embedded into the prompt words, and the large model is called to review the domain text based on the embedded prompt words to generate the initial review result. Based on the domain text and the initial review result, the large model is invoked to optimize the initial review result, resulting in the final review result.

[0120] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0121] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0122] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0123] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described domain text review method. This addresses the technical problem that existing text review schemes have significant defects, resulting in poor performance in reviewing domain text. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the domain text review method provided in the above embodiments, and will not be repeated here.

[0124] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the domain text review method described above.

[0125] The computer program product provided in this application can solve the technical problem that existing text review schemes have significant defects, resulting in poor review performance for domain text. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the domain text review method provided in the above embodiments, and will not be repeated here.

[0126] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A domain-specific text review method, characterized in that, The domain text review methods include: Upon receiving domain text to be reviewed in the target domain, the domain thesaurus corresponding to the target domain is invoked; the domain thesaurus includes a first thesaurus and a second thesaurus, the first thesaurus defines proprietary expressions that the target domain must be exempted from, and the second thesaurus defines prohibited expressions that the target domain must be corrected. The domain thesaurus is embedded into the prompt words, and the large model is called to review the domain text based on the embedded prompt words to generate the initial review result. Based on the domain text and the initial review result, the large model is invoked to optimize the initial review result, resulting in the final review result.

2. The domain text review method as described in claim 1, characterized in that, The initial review results include correction annotations for erroneous expressions in the text of the domain. The step of optimizing the initial review result by calling the large model based on the domain text and the initial review result to obtain the final review result includes: Based on the domain text and the initial review results, the large model is invoked; The domain text and the initial review result are used as input parameters of the large model. The large model is used to verify the initial review result to determine the execution status of the domain thesaurus. Based on the execution results, the initial review results are optimized, and the number of newly annotated erroneous statements is determined based on the optimized review results. If the number of erroneous statements exceeds a preset threshold, the optimized review result is used as the initial review result, and the step of calling the large model based on the domain text and the initial review result is returned and executed. If the number of erroneous statements is less than or equal to the number threshold, the optimized review result will be determined as the final review result.

3. The domain text review method as described in claim 2, characterized in that, The error correction annotation includes error type, error level, and modification suggestions; in the step of using the large model to verify the initial review results, the large model is configured as follows: In the initial review results, exemption verification is performed on expressions matching the first thesaurus, and annotation correctness verification is performed on expressions matching the second thesaurus. For the initial review results, the criteria for judging error descriptions with the same error type and error level across the entire text of the domain are checked for consistency. The reasonableness of the suggested modifications in the error correction annotations of the initial review results is verified.

4. The domain text review method as described in any one of claims 1 to 3, characterized in that, After the step of optimizing the initial review result by calling the large model based on the domain text and the initial review result to obtain the final review result, the method further includes: The final review results are displayed visually in a preset interactive interface; Receive feedback from the user on the final review result on the interactive interface; The domain terminology is updated based on the feedback.

5. The domain text review method as described in claim 4, characterized in that, The step of updating the domain terminology based on the feedback operation includes: An operation instruction is generated based on the feedback operation; if the feedback operation includes a dictionary addition operation, the operation instruction includes a dictionary addition instruction. In response to the lexicon addition instruction, the target expression specified by the feedback operation is added to the first lexicon or the second lexicon to update the domain lexicon.

6. The domain text review method as described in claim 4, characterized in that, The interactive interface includes a first display area and a second display area; the step of visually displaying the final review result in the preset interactive interface includes: The original text of the domain text is displayed in the first display area, and the erroneous text fragments corresponding to each erroneous statement recorded in the final review result are highlighted in the original text. The error correction annotations of the erroneous text fragments are displayed in the second display area; wherein the error correction annotations and the corresponding erroneous text fragments are interactively linked.

7. The domain text review method as described in claim 4, characterized in that, After the step of receiving user feedback on the final review result on the interactive interface, the method further includes: If the target representation is determined to be an erroneous representation based on the final review result, a target erroneous representation in the domain text with the same error type as the target representation is determined based on the error type of the target representation; wherein, the target representation is the representation specified by the feedback operation; If the feedback operation includes a modification operation, the target representation and the target error representation are synchronously corrected according to the modification operation.

8. A domain-specific text review device, characterized in that, The domain text review device includes: The calling module is used to call the domain lexicon corresponding to the target domain when it receives domain text to be reviewed in the target domain; the domain lexicon includes a first lexicon and a second lexicon, the first lexicon defines proprietary expressions that the target domain needs to be exempted, and the second lexicon defines prohibited expressions that the target domain needs to be corrected. The initial review module is used to embed the domain thesaurus into the prompt words, and call the large model to review the domain text based on the embedded prompt words to generate the initial review result. The verification and optimization module is used to optimize the initial review result by calling the large model based on the domain text and the initial review result, so as to obtain the final review result.

9. A domain-specific text review device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the domain text review method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the domain text review method as described in any one of claims 1 to 7.

Images