Industrial injury intelligent auxiliary identification method and system based on large language model

Abstract

The invention provides an industrial injury intelligent auxiliary identification method and system based on a large language model, and establishes an efficient industrial injury automatic identification method and system based on deep semantic analysis and clause matching capability of the large language model, precise classification and judgment of multiple illness condition combinations are achieved, the injury condition description text content input by a user can be analyzed, core information is accurately extracted from the injury condition description text content, classification is automatically carried out according to the complexity of illness conditions, then corresponding industrial injury rules and terms are automatically matched, corresponding industrial injury levels are automatically judged, and manual intervention is not depended on. Moreover, through the multi-level semantic understanding capability of the large language model, multi-illness description can be effectively processed, and accurate identification suggestions can also be output for complex description of multiple illness. Compared with an existing method based on rule or symptom matching, the method and system can remarkably improve the efficiency and accuracy of industrial injury identification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and medical assistance technology, and in particular to a method and system for intelligent auxiliary identification of work-related injuries based on a large language model. Background Art

[0002] Work injury assessment is an important part of the labor security system and is directly related to the rights and interests of workers with work injuries. The current assessment process is mainly based on manual analysis, and the injured need to go to the corresponding department of the hospital for examination. The doctor determines the injury registration based on the actual injury of the injured and the work injury assessment regulations issued by the government, and finally submits it to the government human resources and social security department for unified registration and management. The entire process requires close cooperation between employees, human resources and social security departments, and different departments of the hospital, resulting in huge communication and coordination costs. At the same time, because injury assessment is highly dependent on manual analysis, it is not only time-consuming but also easily affected by subjective factors and misjudgment. It often requires multiple assessments to finally determine the level.

[0003] As a typical closed scenario that requires a lot of background knowledge support, the work injury identification scenario has a very wide scope, involving hundreds of thousands of detailed regulations. On the other hand, the injuries of the injured vary, and it is difficult to form a unified expression, especially for complex injuries. Therefore, work injury identification cannot be processed by intelligent systems in the medical field with data management as the core. With the rapid development of large language models (such as GPT and Qwen, etc.) in the field of natural language processing, especially in knowledge-intensive scenarios such as medicine and law, the application has gradually shown significant advantages, greatly expanding the capabilities of intelligent systems, making it possible to have an auxiliary work injury identification system that can help doctors greatly improve the efficiency and accuracy of identification. However, the current intelligent systems in the medical field are mostly focused on general diagnosis and symptom management. There is no intelligent system specifically for work injury identification. The existing intelligent systems in the medical field are also difficult to be used directly or through simple improvements for auxiliary work injury identification. Specifically, the existing intelligent systems in the medical field mainly include the following categories:

[0004] Rule-based medical auxiliary diagnosis system: This type of system is based on a series of decision rules and conditions in the medical knowledge base, and provides doctors with preliminary diagnostic support by matching with symptom descriptions. This system is more suitable for the identification of simple conditions and can provide auxiliary reference for routine diagnosis. However, due to the complexity of work injury identification and its dependence on legal provisions, simple rule matching is difficult to meet the refined needs of subdivided fields.

[0005] Diagnostic systems based on symptom matching: Many such systems compare the patient's description of the condition with a predefined symptom library through keyword matching. This method is more effective when dealing with basic symptoms, but when faced with complex or multiple symptom combinations, its recognition and judgment capabilities are limited due to the lack of in-depth understanding of semantics, making it difficult to apply to the sophisticated scenario of work injury identification.

[0006] Intelligent medical systems based on natural language processing: With the development of natural language processing technology, some intelligent medical systems have begun to use machine learning models to conduct deeper semantic analysis of the symptoms described by patients. These systems generate diagnostic results by extracting key features, but they often remain in the field of general medicine and lack deep adaptation to the sub-fields of work-related injuries, making it difficult to effectively apply them in the specific rules and clauses of work-related injury identification.

[0007] As mentioned above, compared with other medical diagnosis scenarios, work injury identification has the characteristics of high identification accuracy requirements, a wide range of injury symptoms, and clear legal or clause restrictions. This makes the traditional medical intelligent diagnosis system have huge limitations in this scenario, which can be summarized as follows:

[0008] First, there is a lack of refined support for work-related injury identification: the current medical system lacks specificity in the field of work-related injury identification, and its diagnosis and reasoning mechanisms are usually only applicable to basic medical diagnosis. The judgment criteria, legal provisions and complexity of the disease involved in work-related injury identification require more sophisticated analytical capabilities, and the existing system cannot meet this demand.

[0009] Second, universal symptom matching cannot meet the needs of work-related injury identification: existing systems mostly use methods such as keyword matching and symptom retrieval, but because the description of work-related injury symptoms often involves comprehensive judgments of multiple medical conditions and legal clauses, universal matching methods lack sufficient flexibility and are difficult to accurately parse and distinguish complex situations of multiple medical conditions, resulting in potentially large deviations in identification results.

[0010] Third, there is a lack of deep semantic understanding of work-related injury clauses: Work-related injury identification requires not only a medical understanding of the symptoms, but also compliance with relevant laws and labor protection clauses. Existing systems usually lack natural language understanding capabilities and find it difficult to achieve deep semantic matching of clauses, making it difficult to provide reliable identification recommendations in practical applications.

[0011] Therefore, there is a need to develop an auxiliary work injury identification system based on a large language model based on the characteristics of work injury identification. Summary of the invention

[0012] The present invention is made to solve the above problems, and aims to provide a method and system that can more efficiently and accurately perform automatic work injury identification and legal rules and regulations matching. The present invention adopts the following technical solutions:

[0013] The present invention provides an intelligent auxiliary identification method for work-related injuries based on a large language model, which has the following technical features and includes the following steps: step S1, using injury data to train a large language model to obtain the trained large language model as a work-related injury identification model; step S2, building a rule support library based on standard clause data for work-related injury identification as a knowledge retrieval basis for the work-related injury identification model; step S3, inputting an injury description text into the work-related injury identification model, the model outputting an identification suggestion, the identification suggestion including a work-related injury determination result, and when the work-related injury determination result is a work-related injury, the identification suggestion further includes a corresponding standard clause for work-related injuries; step S4, inputting the injury description text and the corresponding identification suggestion into a multi-dimensional scoring model, the model outputting a corresponding work-related injury level, wherein in step S3, the work-related injury identification model determines the complexity of the injury description text through a classification function to determine whether it is a single condition description text or a multiple condition description text, and when the injury description text is a multiple condition description text, the work-related injury identification model decomposes it into a plurality of single condition description texts.

[0014] The intelligent auxiliary identification method for work-related injuries based on a large language model provided by the present invention may also have such technical features, wherein step S1 includes the following sub-steps: step S1-1, performing data cleaning on the injury data, and performing noise processing on the text in the injury data using a predetermined regular expression set to ensure data consistency; step S1-2, performing keyword extraction and word segmentation processing on the cleaned text to obtain a set of main feature keywords; step S1-3, using the original text and the corresponding set of main feature keywords as training data, training the large language model, and obtaining the work-related injury identification model.

[0015] The method for intelligent auxiliary identification of work-related injuries based on a large language model provided by the present invention may also have such a technical feature, wherein, in step S1-1, the text in the injury data is subjected to noise processing using the regular expression set, which is expressed as:

[0016]

[0017] In the formula, ε is the set of regular expressions, d i is the i-th text in the injury data. In step S1-2, the Jieba word segmentation algorithm is used to extract the main feature keyword set where w j The main feature keywords.

[0018] The method for intelligent auxiliary identification of work-related injuries based on a large language model provided by the present invention may also have such a technical feature, wherein, in step S2, for each clause text in the standard clause data for work-related injury identification, its structure is standardized into a tuple form (c, s, k), wherein c represents the work-related injury category corresponding to the clause, s is the subcategory, and k is the rule content, thereby constructing a quasi-rule support library

[0019] The intelligent auxiliary identification method for work-related injuries based on a large language model provided by the present invention may also have such a technical feature, wherein step S3 includes the following sub-steps: step S3-1, semantically analyzing the input injury description text; step S3-2, judging whether the injury description text contains multiple conditions based on the semantic analysis result and the classification function, and when the judgment is no, the injury description text is the single condition description text; when the judgment is yes, the injury description text is the multiple condition description text; step S3-3, when the judgment is yes in step S3-2, decomposing the injury description text into multiple single condition description texts; step S3-4, for each single condition description text, retrieving the closest category and subcategory from the rule support library; step S3-5, retrieving possible applicable clauses from the rule support library based on the retrieved category and subcategory; step S3-6, matching the standard terms of work-related injuries through semantic analysis and probability screening based on the possible applicable clauses.

[0020] The method for intelligent auxiliary identification of work-related injuries based on a large language model provided by the present invention may also have such a technical feature, wherein, in step S3-2, the work-related injury identification model performs semantic analysis on the input injury description text T, and uses the internal classification function The complexity of the injury description text T is determined. It is determined to be the multiple condition description text, and the injury description text T is decomposed until it is decomposed into multiple single condition description texts. In step S3-4, for each single condition description text, the most similar category and subcategory are retrieved from the rule support library using the maximum likelihood estimation method, which is expressed as:

[0021]

[0022] In the formula, c is the category, s is the subcategory, The closest category, The closest subcategory.

[0023] The method for intelligent auxiliary identification of work-related injuries based on a large language model provided by the present invention may also have such a technical feature, wherein, in step S3-5, the retrieval enhancement generation technology is used to retrieve the potentially applicable clauses from the rule support library based on the category and the subcategory, and in step S3-6, the potentially applicable clauses are input into the injury identification model as prompts to match the standard clauses of work-related injuries, and for the potentially applicable clauses, the work-related injury identification model is used to perform a probability judgment of yes or no on multiple clauses one by one, and the model embeds the embedding vector v of the injury description text T T and terms and conditions i The embedding vector Compare and judge the applicability of each clause through the probability output of the model.

[0024] The method for intelligent auxiliary identification of work-related injuries based on a large language model provided by the present invention may also have such a technical feature, wherein, in step S4, for the input identification suggestion, the multi-dimensional scoring model Through its inference scoring function f score Calculate the score of each candidate level, and select the candidate level with the highest score as the recommended work injury level:

[0025]

[0026] In the formula, g j is the candidate level, T is the injury description text, and k is the corresponding identification suggestion.

[0027] The intelligent auxiliary identification method for work-related injuries based on a large language model provided by the present invention may also have such a technical feature, the method also includes: step S5, generating explanatory information based on the injury description text, the identification suggestion and the work-related injury level, wherein the explanatory information includes rule tracing information and reasoning process backtracking information, the rule tracing information is the text content of the matched work-related injury standard clauses, and the reasoning process backtracking information is the reasoning logic description information of each step of generating the identification suggestion and generating the work-related injury level.

[0028] The present invention provides an intelligent auxiliary identification system for work-related injuries based on a large language model, which has the following technical features: a rule support library, which is constructed based on standard clause data for work-related injury identification; an input module, which is used for users to input injury description text; an injury identification and clause matching module, which includes a work-related injury identification model, which is a large language model trained using injury data, and the rule support library is used as a knowledge retrieval basis for the work-related injury identification model, and the work-related injury identification model is used to output corresponding identification suggestions based on the input injury description text, and the identification suggestion includes a work-related injury determination result. When the work-related injury determination result is a work-related injury, the identification suggestion further includes a corresponding standard clause for work-related injuries; and a work-related injury level determination module, which includes a multi-dimensional scoring model, which is used to determine the corresponding work-related injury level based on the identification suggestion.

[0029] Functions and Effects of the Invention

[0030] According to the intelligent auxiliary identification method and system of work-related injuries based on a large language model provided by the present invention, an efficient automatic identification method and system for work-related injuries are established based on the deep semantic parsing and clause matching capabilities of the large language model, wherein the natural language text of the work-related injury description is deeply semantically understood through the large language model, and accurate classification and judgment of multiple disease combinations are achieved. The text content of the injury description input by the user can be analyzed, and the core information can be accurately extracted from it, and the disease can be automatically classified according to the complexity of the disease, and then the corresponding work-related injury rules and clauses can be automatically matched, and the corresponding level of work-related injuries can be automatically determined without relying on manual intervention. In addition, through the multi-level semantic understanding capabilities of the large language model, multiple disease descriptions can be effectively processed, which is not only suitable for the processing of a single disease description, but also can output accurate identification suggestions for complex descriptions of multiple diseases. Compared with existing rule-based or symptom matching methods, this method and system can significantly improve the efficiency and accuracy of work-related injury identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of an intelligent auxiliary identification method for work-related injuries based on a large language model in an embodiment of the present invention;

[0032] Figure 2 is a detailed flow chart of an intelligent auxiliary identification method for work-related injuries based on a large language model in an embodiment of the present invention;

[0033] Figure 3 is a flow chart of step S1 in an embodiment of the present invention;

[0034] Figure 4 is a flow chart of step S3 in an embodiment of the present invention;

[0035] Figure 5 It is a logical block diagram of an intelligent auxiliary identification system for work-related injuries based on a large language model in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following is a detailed description of the intelligent auxiliary identification method and system for work-related injuries based on a large language model in combination with embodiments and drawings.

[0037] Figure 1 is a flow chart of the intelligent auxiliary identification method for work-related injuries based on a large language model in this embodiment, Figure 2 A detailed flow chart of the method for intelligent assisted identification of work-related injuries based on a large language model in this embodiment.

[0038] like Figure 1 and Figure 2 As shown, the method comprises the following steps:

[0039] Step S1, using injury data to train a large language model, and obtaining the trained large language model as a work injury identification model.

[0040] Step S2, constructing a rule support library based on the standard clause data of work-related injury identification as the knowledge retrieval basis of the work-related injury identification model.

[0041] Step S3, input the injury description text into the work injury assessment model, and the model outputs an assessment suggestion, which includes the work injury determination result. When the work injury determination result is a work injury, the assessment suggestion further includes the corresponding work injury standard clause.

[0042] Step S4, input the injury description text and the corresponding identification suggestions into a multi-dimensional scoring model, and the model outputs the corresponding work injury level.

[0043] Step S5, generating an explanation based on the injury description text, identification recommendations and work injury level.

[0044] The above steps will be described in detail below.

[0045] Step S1, using injury data to train a large language model, and obtaining the trained large language model as a work injury identification model.

[0046] In this embodiment, the injury data (training data) mainly comes from the injury data of Chongqing Human Resources and Social Security Bureau. The injury data is used to train (fine-tune) the large language model Qwen, and the fine-tuned Qwen model is obtained as the work injury identification model.

[0047] In order to ensure the high-precision performance of the model, the training data was fully preprocessed and standardized.

[0048] Figure 3 is a flow chart of step S1 in this embodiment.

[0049] like Figure 3 As shown, step S1 specifically includes the following sub-steps:

[0050] Step S1-1, clean the injury data, and use a predetermined regular expression set to perform noise processing on the text in the injury data to ensure data consistency, which is specifically formalized as follows:

[0051]

[0052] In the formula, ε is the set of regular expressions, d i It is the i-th text in the injury data.

[0053] Step S1-2, extract keywords and perform word segmentation processing on the cleaned text to obtain a set of main feature keywords.

[0054] Among them, for the cleaned text d′ i , use Jieba word segmentation algorithm to extract the main feature keywords w j Collection This process can optimize the model's lexical cognition of complex texts and lay a fine-grained semantic foundation for subsequent clause matching and category analysis.

[0055] Step S1-3, using the original text and the corresponding main feature keyword set as training data, training the large language model, and obtaining the trained large language model as a work injury identification model.

[0056] Step S2, constructing a rule support library based on the standard clause data of work-related injury identification as the knowledge retrieval basis of the work-related injury identification model.

[0057] Among them, for each clause text in the work injury assessment standard "New Standard for Labor Capacity Assessment (2024 Version)", the clause text structure is standardized into a tuple form (c, s, k), where c represents the work injury category corresponding to the clause, s is the subcategory, and k is the rule content, in order to build a rule support library

[0058] Step S3, input the injury description text into the work injury assessment model, and the model outputs an assessment suggestion, which includes the work injury determination result. When the work injury determination result is a work injury, the assessment suggestion further includes the corresponding work injury standard clause.

[0059] Figure 4 is a flow chart of step S3 in this embodiment.

[0060] like Figure 4 As shown, step S3 specifically includes the following sub-steps:

[0061] Step S3-1: the work injury assessment model performs semantic analysis on the input injury description text.

[0062] Among them, the fine-tuned large language model automatically performs semantic modeling on the input text and identifies the text structure and information density.

[0063] Step S3-2, the work injury identification model determines whether the input injury description text contains multiple conditions based on the semantic analysis results and the internal classification function. If it is judged as no, the injury description text is a single condition description text; if it is judged as yes, the injury description text is a multiple condition description text.

[0064] Step S3-3, when the judgment in step S3-2 is yes, the work injury assessment model decomposes the injury description text into multiple single condition description texts.

[0065] Among them, after receiving the injury description text T input by the user, the fine-tuned large language model Perform semantic analysis on the input text T and pass it through the internal classification function The complexity of the input text T is determined to determine whether the injury description text is a description text of a single disease condition or a description text containing multiple diseases. If it is determined to be a multiple condition description text, the model will decompose the text T until it is decomposed into multiple single condition description texts to ensure independent determination of each condition; otherwise, it will be determined to be a single condition description text, and the category of each single condition description text will continue to be analyzed.

[0066] Specifically, by setting appropriate large model prompts (Prompt), the large language model can be used to directly determine how many conditions (injury conditions) are contained in the injury description text. For example, in the prompt, the large language model is prompted that if there are multiple conditions in the input text, please break it down into single condition descriptions and separate multiple single condition descriptions with semicolons.

[0067] Step S3-4, the work injury identification model uses the maximum likelihood estimation method (MLE method) to retrieve the most similar categories and subcategories from the rule support library for each single condition description text, which can be expressed as:

[0068]

[0069] In the formula, c is the category, s is the subcategory, The closest category, The closest subcategory.

[0070] In this embodiment, the categories and subcategories refer to the official classification of the disease. For example, for the injury description "left patella fracture", its category is orthopedics, plastic surgery, and burns, and the subcategory is the upper and lower limbs subcategory.

[0071] Step S3-5, using the Retrieval Enhanced Generation (RAG) technology, retrieves potentially applicable clauses from the rule support library based on categories and subcategories.

[0072] Among them, for the first single condition description text, RAG technology is used to extract the Extract the matching clause content collection Then, based on the identified subcategories, irrelevant clauses were initially excluded, leaving only potentially applicable clauses, thus narrowing the screening scope.

[0073] Step S3-6, input the retrieved potentially applicable clauses as prompts into the injury assessment model, and match the standard clauses for work-related injuries through semantic analysis and probabilistic screening (prob).

[0074] Among them, for the set of retained terms, the reasoning ability of the large language model is used to analyze multiple terms k in it. i Make a "yes / no" probability judgment one by one. The model will input the embedding vector v of the text T T and terms and conditions i The embedding vector Compare and judge the applicability of each clause through the probability output of the model. If after a round of comparison, the probability judgment results of multiple clauses are "yes", the large model prompt is regenerated according to the judgment results, and the clause content is screened more finely in the same way of judging whether it is yes or no. Finally, a preliminary appraisal suggestion is generated.

[0075] For a single condition description text, the appraisal suggestion includes the determination result of whether it is a work-related injury. If the determination result is a work-related injury, the appraisal suggestion further includes the corresponding standard clauses for work-related injuries. For multiple condition description texts, the appraisal suggestion includes the determination result of whether each single condition description text is a work-related injury. Similarly, if the determination result is a work-related injury, the appraisal suggestion further includes the corresponding standard clauses for work-related injuries.

[0076] Step S4, input the injury description text and the corresponding identification suggestions into a multi-dimensional scoring model, and the model outputs the corresponding work injury level.

[0077] Among them, for the input identification suggestions, the multi-dimensional scoring model Through its inference scoring function f score Calculate the score of each candidate level and select the candidate level with the highest score as the recommended work injury level:

[0078]

[0079] In the formula, g j is the candidate level, T is the injury description text, and k is the corresponding identification suggestion.

[0080] Step S5, generating explanatory information based on the injury description text, identification recommendations and work injury level.

[0081] The generated explanation information includes rule tracing information and reasoning process backtracking information.

[0082] Regarding rule traceability information, the text content of the clause is obtained according to the matching work-related injury standard clause in the appraisal suggestion as rule traceability information. That is, the final selected rule clause and its content are displayed to the user (such as a doctor), so that the user can directly check the content of the clause and its scope of application to enhance the user's trust.

[0083] Regarding the backtracking information of the reasoning process, based on the injury description text, identification recommendations and work-related injury levels, concise reasoning logic description information is generated for each step of generating identification recommendations and determining the level of work-related injuries, so as to show the key steps from injury description text input, disease classification to rule clause screening, score calculation, etc. These contents are displayed to users to ensure that users can quickly understand the relevant judgment basis, thereby enhancing the interpretability of the method.

[0084] Specifically, the outputs of the work injury identification model (large language model) and the multi-dimensional scoring model in the above steps S3 and S4 are sorted and summarized to obtain the reasoning process backtracking information. For example, the input injury description text is "left patellar fracture". After generating the identification suggestion and the work injury level through steps S3 and S4 respectively, the outputs of the models in the process are sorted to obtain the following reasoning process backtracking information:

[0085] “The injury falls under the category of orthopedics, plastic surgery and burns, subcategory of upper and lower limbs.

[0086] The preliminary appraisal has the following relevant clauses:

[0087] Level 8, item 24: Traumatic arthritis caused by an intra-articular fracture of one of the large joints of the limbs, leaving mild functional impairment;

[0088] Level 9, item 23: After internal fixation or external fixation of long tubular bone fractures of the limbs;

[0089] Level 9, 24 items: after internal fixation of patella, calcaneus, talus, mandible or pelvis fractures;

[0090] Level 10, item 12: There is no functional impairment or mild functional impairment after the fractures in various parts of the body heal.

[0091] The final identification results are:

[0092] There are 12 items in the 10th level. "

[0093] This embodiment also provides a system corresponding to the above method.

[0094] Figure 5 It is a logical block diagram of the intelligent auxiliary identification system for work-related injuries based on a large language model in this embodiment.

[0095] like Figure 5 As shown, the intelligent auxiliary identification system 10 for work-related injuries based on a large language model includes a rule support library 11, an input module 12, an injury identification and clause matching module 13, a work-related injury level determination module 14, an explanation generation module 15, an output module 16, and a general control module 17.

[0096] The rule support library 11 is the rule support library constructed by the above step S2.

[0097] The input module 12 is used for the user to input injury description text, for example, it can be an interactive device that provides a user interface, or a communication module that communicates with a device held by the user.

[0098] The injury assessment and clause matching module 13 includes the above-mentioned injury assessment model, which can be trained through the above-mentioned step S1, and the injury assessment model processes the input text as described in the above-mentioned step S3. In addition, the injury assessment and clause matching module 13 provides the output result to the work injury level determination module 14 and the explanation generation module 15.

[0099] The work injury level determination module 14 includes the above-mentioned multi-dimensional scoring model, and the multi-dimensional scoring model processes the input text and the identification result as described in the above-mentioned step S4. In addition, the work injury level determination module 14 provides the output result to the explanation generation module 15.

[0100] The explanation generation module 15 generates an explanation through the method of step S5, including rule tracing information and reasoning process backtracking information.

[0101] The output module 16 is used to output the identification results, the level of work-related injuries and explanations to the user, and may be, for example, a display device or a communication module for communicating with a device held by the user.

[0102] Functions and Effects of the Embodiments

[0103] According to the intelligent auxiliary identification method and system for work-related injuries based on a large language model provided in this embodiment, an efficient automatic identification method and system for work-related injuries are established based on the deep semantic parsing and clause matching capabilities of the large language model, wherein the natural language text of the work-related injury description is deeply semantically understood through the large language model to achieve accurate classification and judgment of multiple disease combinations, and can analyze the injury description text content input by the user, accurately extract core information from it, automatically classify it according to the complexity of the disease, and then automatically match the corresponding work-related injury rules and clauses, and automatically determine the corresponding work-related injury level without relying on human intervention. In addition, through the multi-level semantic understanding capabilities of the large language model, it is possible to effectively process multiple disease descriptions, which is not only suitable for the processing of a single disease description, but also can output accurate identification suggestions for complex descriptions of multiple diseases. Compared with existing rule-based or symptom matching methods, this method and system can significantly improve the efficiency and accuracy of work-related injury identification.

[0104] Furthermore, the method and system judge the complexity of the input injury description text according to the classification result of the classification function, thereby automatically distinguishing the input text as a single condition description text or a multiple condition description text, decomposing the multiple condition description text until it is decomposed into multiple single condition description texts, and determining the corresponding category and subcategory for each single condition description text and further matching the rule clauses. Therefore, it can not only automatically output identification suggestions for complex descriptions of multiple conditions, but also will not omit or ignore any of the conditions, and can effectively improve the identification accuracy of complex descriptions of multiple conditions.

[0105] Furthermore, in this method and system, during the rule clause screening and application stage, RAG technology is used to enhance the semantic comparison capability of the model, and irrelevant clauses are preliminarily excluded, retaining only potentially applicable clauses, thereby narrowing the screening scope. Therefore, not only can the matching rule clauses be made more accurate, but also the identification efficiency can be further improved.

[0106] In the embodiment, the explanation generation module is also used to provide the user with a clear reasoning process and the accompanying clause basis, so that the user can intuitively understand the reasons for the appraisal. This transparency enhances the application credibility of the method and system, and lays the foundation for the wide application of the method and system in actual work injury appraisal.

[0107] The above embodiments are only used to illustrate the specific implementation of the present invention, and the present invention is not limited to the description scope of the above embodiments. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for intelligently assisting identification of work-related injuries based on a large language model, characterized in that: The following steps are involved: Step S1, using injury data to train a large language model, and obtaining the trained large language model as a work injury identification model; Step S2, constructing a rule support library based on the standard clause data of work injury identification as a knowledge retrieval basis for the work injury identification model; Step S3, inputting the injury description text into the work injury assessment model, the model outputting an assessment suggestion, the assessment suggestion including the work injury assessment result, and when the work injury assessment result is a work injury, the assessment suggestion further includes corresponding work injury standard clauses; Step S4, inputting the injury description text and the corresponding identification suggestion into a multi-dimensional scoring model, which outputs the corresponding work injury level. Among them, in step S3, the work injury assessment model determines the complexity of the injury description text through a classification function to determine whether it is a single condition description text or a multiple condition description text. When the injury description text is a multiple condition description text, the work injury assessment model decomposes it into multiple single condition description texts.

2. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 1 is characterized in that: in, Step S1 includes the following sub-steps: Step S1-1, performing data cleaning on the injury data, and performing noise processing on the text in the injury data using a predetermined regular expression set to ensure data consistency; Step S1-2, extracting keywords and performing word segmentation processing on the cleaned text to obtain a set of main feature keywords; Step S1-3, using the original text and the corresponding main feature keyword set as training data, training the large language model to obtain the work injury identification model.

3. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 2 is characterized in that: in, In step S1-1, the text in the injury data is subjected to noise processing using the regular expression set, which is expressed as: In the formula, ε is the set of regular expressions, d i is the i-th text in the injury data, In step S1-2, the Jieba word segmentation algorithm is used to extract the main feature keyword set where w j The main feature keywords.

4. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 1 is characterized in that: in, In step S2, for each clause text in the standard clause data of work injury identification, its structure is standardized into a tuple form (c, s, k), where c represents the work injury category corresponding to the clause, s is the subcategory, and k is the rule content, so as to construct a pseudo-rule support library.

5. According to claim 1, the method for intelligently assisting identification of work-related injuries based on a large language model, Features: Wherein, step S3 includes the following sub-steps: Step S3-1, performing semantic analysis on the input injury description text; Step S3-2, judging whether the injury description text contains multiple conditions based on the semantic analysis result and the classification function, and if the judgment is no, the injury description text is the single condition description text; if the judgment is yes, the injury description text is the multiple condition description text; Step S3-3, when the judgment in step S3-2 is yes, the injury description text is decomposed into a plurality of single condition description texts; Step S3-4, for each of the single condition description texts, the most similar categories and subcategories are retrieved from the rule support library; Step S3-5, retrieving potentially applicable clauses from the rule support library based on the retrieved categories and subcategories; Step S3-6, based on the potentially applicable clauses, matching the standard clauses for work-related injuries through semantic analysis and probability screening.

6. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 5 is characterized in that: in, In step S3-2, the work injury identification model performs semantic analysis on the input injury description text T and uses the internal classification function Determine the complexity of the injury description text T, like It is determined to be the multiple condition description text, and the injury description text T is decomposed until it is decomposed into multiple single condition description texts. In step S3-4, for each of the single condition description texts, the most similar categories and subcategories are retrieved from the rule support library using the maximum likelihood estimation method, which is expressed as: In the formula, c is the category, s is the subcategory, The closest category, The closest subcategory.

7. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 5 is characterized in that: in, In step S3-5, the potentially applicable clauses are retrieved from the rule support library based on the category and the subcategory using the search enhancement generation technology. In step S3-6, the potentially applicable clauses are input into the injury assessment model as prompts to match the standard clauses for work-related injuries. For the potentially applicable clauses, the work-related injury assessment model is used to make a yes or no probability judgment on multiple clauses one by one. The model embeds the injury description text T into a vector v T and terms and conditions i The embedding vector Compare and judge the applicability of each clause through the probability output of the model.

8. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 1 is characterized in that: in, In step S4, for the input identification suggestion, the multi-dimensional scoring model Through its inference scoring function f score Calculate the score of each candidate level, and select the candidate level with the highest score as the recommended work injury level: In the formula, g j is the candidate level, T is the injury description text, and k is the corresponding identification suggestion.

9. The method for intelligently assisting identification of work-related injuries based on a large language model according to claim 1, characterized in that: Also includes: Step S5, generating explanation information based on the injury description text, the appraisal suggestion and the level of the work-related injury, The explanation information includes rule tracing information and reasoning process backtracking information. The rule tracing information is the text content of the matching standard clause of the work injury. The reasoning process backtracking information is the reasoning logic description information of each step of generating the identification recommendation and generating the work injury level.

10. An intelligent auxiliary identification system for work-related injuries based on a large language model, characterized in that: include: The rule support library is built based on the standard clause data of work-related injury appraisal; An input module, used for users to input injury description text; An injury assessment and clause matching module, which includes a work injury assessment model, wherein the work injury assessment model is a large language model trained using injury data, wherein the rule support library is used as a knowledge retrieval basis for the work injury assessment model, wherein the work injury assessment model is used to output corresponding assessment suggestions based on the input injury description text, wherein the assessment suggestions include work injury determination results, and when the work injury determination result is work injury, the assessment suggestions further include corresponding work injury standard clauses; as well as The work injury level determination module includes a multi-dimensional scoring model for determining the corresponding work injury level based on the identification suggestion.

Images