A work injury identification research and judgment method based on natural language understanding

Abstract

The application provides an intelligent work injury identification research and judgment method based on natural language understanding, which can assist business personnel to obtain identification conclusions according to investigation results, search similar case files, give judgment results and legal basis, improve the accuracy and efficiency of work injury identification of business personnel, and avoid unnecessary litigation risks. The application has the following advantages: first, the method does not add any additional information to the original submitted information, which does not increase the burden of work injury identification personnel; second, the method can automatically assist business personnel to issue work injury identification conclusions and identification legal basis, and improve the work efficiency of business personnel; third, the method can search similar judicial judgment cases, and ensure the last link of work injury identification; and finally, the method is based on text information data, and is convenient for popularization and use in different institutions and different regions.

Description

Technical Field

[0001] This invention relates to an intelligent method for determining and judging work-related injuries based on natural language understanding, belonging to the field of smart government technology. Background Technology

[0002] Work-related injury determination is an administrative act authorized by law by the labor administration department to determine whether an employee's injury due to an accident (or occupational disease) constitutes a work-related injury or is deemed a work-related injury. Currently, government work-related injury determination mainly involves manual investigation using methods such as on-site investigation and evidence collection, and coordination with external units to collect evidence. Based on the investigation results, a determination is made in accordance with the "Regulations on Work-Related Injury Insurance" to determine whether it constitutes a work-related injury. Currently, there is a contradiction between the scarcity of professional personnel and the complexity of work-related injury cases, sometimes leading to legal proceedings and even losses in court.

[0003] In recent years, with the research and development of artificial intelligence in natural language processing (NLP) technology, NLP technology has been applied to many fields. Especially in recent years, with the proposal of smart government, many people have applied NLP technology to smart cities, social security, and other areas. Summary of the Invention

[0004] The purpose of this invention is to apply natural language processing technology to the determination of work-related injuries.

[0005] To achieve the above objectives, the technical solution of the present invention provides a method for determining and judging work-related injuries based on natural language understanding, characterized by comprising the following steps:

[0006] Step 1: Data Collection

[0007] Collect relevant data on previous work-related injury determinations, including information on the occurrence of work-related injuries, the results of work-related injury determinations, and the legal basis for work-related injury determinations;

[0008] Construct a work injury determination information database, and store information on the occurrence of work injury, the result of work injury determination, and the legal basis for work injury determination in the form of data tables in the work injury determination information database;

[0009] At the same time, by organizing information on the legal basis for injury determination, a legal case library is formed that includes typical cases corresponding to each legal clause in the relevant laws;

[0010] Step 2, Data Processing: This involves preprocessing the structured data in the work injury determination information database, deleting irrelevant fields, and selecting features. Specifically, this includes the following:

[0011] Data preprocessing: Convert the legal basis information for work-related injury determination into numerical data; based on the statistical results of the data, convert all legal clauses in the relevant laws into n categories and convert them into one-hot format; delete data samples with missing values ​​in each row of the work-related injury determination information database.

[0012] Irrelevant field deletion: Remove meaningless characters from the work injury occurrence information;

[0013] Information order conversion: Based on textual patterns, the order of textual information in the work injury situation information is converted, and the diagnostic information is placed at the beginning of the work injury situation information text;

[0014] Information text format is converted into vector format: The information on the occurrence of work-related injuries is segmented according to the granularity of characters, and then the text information is converted from text format to vector format through a character-label dictionary;

[0015] Step 3: Construct a model for work-related injury determination and legal basis, which includes the following:

[0016] The input to the model for determining work-related injuries and its legal basis is information about the occurrence of the work-related injury, with an input size of n. max ×m, where n max is the maximum number of words contained in a long text in unstructured data, and m is the number of data feature categories;

[0017] The model for determining work-related injuries and the legal basis for work-related injuries consists of two classification models and one retrieval model. The two classification models are the work-related injury determination model and the legal basis for work-related injury determination model, and the retrieval model is a matching-based retrieval model.

[0018] The classification model includes an information extraction part. After the information on the occurrence of work-related injuries is input into the information extraction part, the information extraction part extracts the sentence feature information. The information extraction part of both classification models is implemented using the BERT model.

[0019] In the work injury determination model, the sentence feature information obtained by the BERT model is passed through a fully connected layer to obtain a 1-dimensional vector, and then the sigmoid function is used to transform the vector into 0 and 1 results, where 0 represents not being recognized as a work injury and 1 represents being recognized as a work injury.

[0020] Simultaneously, information about the occurrence of work-related injuries is input into the legal basis model for work-related injury determination. In this model, after obtaining statement feature information through the BERT model, the statement feature information is then passed through a fully connected layer to obtain an n-dimensional vector. Subsequently, the legal basis model for work-related injury determination transforms the n-dimensional vector into an n-dimensional probability representation through a softmax activation function. Finally, the model obtains the corresponding legal basis clauses based on the obtained probabilities.

[0021] The retrieval model searches for standard cases in the legal case library based on the legal basis clauses obtained from the legal basis model for work-related injury determination.

[0022] Step 4, Model Training, includes the following steps:

[0023] Step 401: Initialize the model parameters of the work injury determination and legal basis model constructed in Step 3;

[0024] Step 402: Using the cross-entropy function as the loss function, the entire network is trained using sample data from the work injury determination information database after data information processing through the stochastic gradient descent algorithm.

[0025] During training, a sample set is obtained from the work injury determination information database and divided into a training set and a test set in a 7:3 ratio. The data processing method described in step 2 is used to process the data in the training set and the test set. The work injury determination and legal basis model is trained using the training set data, and then the model is tested using the test set data. If the test accuracy reaches the predetermined accuracy P_0, the model is saved; otherwise, the number of sample data in the sample set is increased, and the work injury determination and legal basis model is trained again until the accuracy meets the condition, and then the model is saved.

[0026] During the training process, the work injury determination model and the work injury determination legal basis model are trained independently.

[0027] Step 5: Display of Model Results

[0028] For newly added work-related injury information, after processing it according to the data processing method described in step 2 to obtain a data format consistent with the training samples, it is then input into the trained work-related injury determination and legal basis model to calculate the output result. If the work-related injury determination model outputs 1, the result is determined to be a work-related injury, and the work-related injury determination is output. At the same time, the legal basis for the work-related injury determination is derived using the work-related injury determination legal basis model. Then, the legal basis and similar cases are output synchronously after searching the legal clause case library through the retrieval model to obtain similar cases.

[0029] Preferably, in step 1: the work-related injury information includes the time, place, cause, and result of the work-related injury incident, wherein the result includes the on-site diagnosis result and the hospital diagnosis result; the work-related injury determination result information includes whether the injury is recognized as a work-related injury or not; the legal basis information for work-related injury determination includes the relevant clauses of the "Regulations on Work-Related Injury Insurance".

[0030] Preferably, in step 2: information about work-related injuries with fewer than 20 strings is considered invalid data and deleted.

[0031] Preferably, in step 2, the vector-form text information includes two parts: one is the ID information of the text, and the other is the ID information of the subordinate sentence.

[0032] Preferably, in step 3: the BERT model uses the improved ALBERT model as the information extraction model.

[0033] This invention provides an intelligent work injury determination and assessment model based on natural language understanding, mainly comprising two aspects: first, a work injury determination model; and second, a retrieval of similar cases related to work injury determination. In work injury determination, by adopting this invention, personnel can input a description of the injury process into the intelligent work injury determination and assessment model. The system can not only deduce the work injury determination conclusion but also search for relevant work injury precedents in terms of time, place, and cause, effectively improving the accuracy of work injury determination and avoiding unnecessary litigation risks.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] First, the method provided by this invention adds no extra information to the original submission information, thus not increasing the burden on personnel responsible for work injury determination. Second, this invention can automatically assist personnel in issuing work injury determination conclusions and legal basis, improving their work efficiency. Third, the method provided by this invention can search for similar judicial judgment cases, ensuring the final quality of work injury determination. Finally, the method provided by this invention is based on textual information data, making it easy to promote and use in different institutions and regions. Attached Figure Description

[0036] Figure 1 This is an overall flowchart of the present invention;

[0037] Figure 2 This illustrates the overall structure of the model used in the embodiments of the present invention. Detailed Implementation

[0038] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0039] like Figure 1 As shown, this embodiment takes the Qingdao Municipal Work Injury Determination Agency as an example, using the city's work injury determination data and work injury determination legal database as the original data source, and implements an auxiliary work injury determination method, specifically including the following steps:

[0040] Step 1: Data Collection

[0041] Data related to past work-related injury determinations was collected from the Human Resources and Social Security Bureau, including information on the occurrence of work-related injuries, the determination results, and the legal basis for the determinations. A work-related injury determination information database was constructed, with the information on the occurrence of work-related injuries, the determination results, and the legal basis for the determinations stored in separate data tables within the database.

[0042] In this embodiment, the data stored in the work-related injury determination information database has the following characteristics:

[0043] (1) Information on the occurrence of work-related injuries includes, but is not limited to, the time, place, cause and result of the work-related injury incident. Among them, the result mainly includes the on-site diagnosis result and the hospital diagnosis result.

[0044] (2) Information on the results of work injury determination includes, but is not limited to, information such as whether the injury is recognized as work injury or not.

[0045] (3) The legal basis for work-related injury determination includes, but is not limited to, the relevant clauses of the "Regulations on Work-Related Injury Insurance".

[0046] Step 1 constructed a dataset on work-related injury determination and its legal basis. Simultaneously, based on this information, and through the organization of professional personnel, a legal case library was created, containing typical cases corresponding to each legal clause. This data preparation is in place for the next stage of outputting similar cases.

[0047] Step 2, Data Processing: This involves preprocessing the structured data in the work injury determination information database, deleting irrelevant fields, and selecting features. Specifically, this includes the following:

[0048] (1) Data preprocessing: The legal basis information for work-related injury determination is converted into numerical data, and missing values ​​are deleted. In this embodiment, based on the statistical results of the data, the legal basis clauses in the "Regulations on Work-Related Injury Insurance" are converted into 7 categories, and then converted into one-hot format for model training.

[0049] When deleting missing values, data samples with missing values ​​in each row of the work injury determination information database are deleted. Based on practical experience, work injury information with fewer than 20 strings is considered invalid data and deleted.

[0050] (2) Removal of irrelevant fields: Remove meaningless characters such as spaces, &, *, #, and @ from the work injury information to reduce the impact of invalid characters on the results.

[0051] (3) Information order conversion: To prevent the inability to fully extract the diagnostic information in the work injury situation information because the model needs to extract text, the text information order in the work injury situation information is converted according to the text pattern, and the diagnostic information is placed at the beginning of the work injury situation information text.

[0052] In this embodiment, information on the occurrence of work-related injuries is obtained before the information sequence is converted.

[0053] (4) Converting information from text to vector form: The information on the occurrence of work-related injuries is segmented according to the word granularity, and then the text information is converted from text form to vector form through a word-label dictionary.

[0054] In this embodiment, vectorization is divided into two parts: one is the ID information of the text, and the other is the ID information of the subordinate sentence. The vectorized representation is shown in the table below:

[0055]

[0056] Step 3: Construct a model for work-related injury determination and legal basis, which includes the following:

[0057] The input to the work-related injury determination and legal basis model is information about the occurrence of the work-related injury, and the corresponding BERT model has an input size of n. max ×m, where n max denoted as , where m is the maximum number of words contained in a long text in unstructured data, and m is the number of data feature categories.

[0058] In this embodiment, the input size of the BERT model corresponding to the data is 150×2, where 150 is the text truncation length. This value is selected by calculating the character length of each data point in the work injury occurrence information, sorting them from largest to smallest, and then finding the character length at the 80th percentile position as the text truncation length. Each input contains two aspects of information: the text ID and the subordinate sentence ID. The data input in this step is the result obtained in part (4) of step 2, and the data length is 150.

[0059] The model for determining and legally binding work-related injuries consists of two classification models and one retrieval model. The two classification models are the work-related injury determination model and the work-related injury legal basis model, and the retrieval model is a matching-based retrieval model.

[0060] The classification model includes an information extraction component. After the information on the occurrence of work-related injuries is input into the information extraction component, the component extracts the sentence feature information. The information extraction components of the two classification models have the same structure but different parameters, and both are implemented using the BERT model.

[0061] In this embodiment, to reduce memory consumption during model training and decrease the size of the model during storage, the BERT model used is an improved version of BERT, ALBERT, as the information extraction model. ALBERT significantly reduces the number of model parameters by utilizing techniques such as parameter sharing and matrix factorization. It replaces NSP (Next Sentence Prediction) with SOP (Sentence Order Prediction) Loss, which improves performance on downstream tasks. Simultaneously, the reduced number of parameters does indeed speed up training. Furthermore, the ALBERT model can be scaled to larger models than BERT (ALBERT-xxlarge), thus offering advantages over the basic BERT model. The structure diagram of the ALBERT model is shown below. Figure 2 As shown.

[0062] In the work injury determination model, the statement feature information obtained by the BERT model is passed through a fully connected layer to obtain a 1-dimensional vector, and then the sigmoid function is used to transform the vector into a 0-1 vector, as shown in the following formula:

[0063]

[0064]

[0065] Where: x is a 1-dimensional vector; y = 1 or 0, where 0 represents not being recognized as a work-related injury and 1 represents being recognized as a work-related injury; S(x) represents the sigmoid function, used to convert a 1-dimensional vector into a result.

[0066] Simultaneously, information regarding the occurrence of the work-related injury is input into the legal basis model for work-related injury determination. Within this model, after obtaining statement feature information through the BERT model, this feature information is then passed through a fully connected layer to obtain an n-dimensional vector. Subsequently, the legal basis model transforms this n-dimensional vector into an n-dimensional probability representation using a softmax activation function. Finally, the legal basis model derives the corresponding legal provisions based on the obtained probabilities.

[0067] In this embodiment, the legal basis model for work-related injury determination obtains a 7-dimensional vector through the BERT model and fully connected layers. The 7-dimensional vector is then transformed into a 7-dimensional probability representation using the softmax activation function, as shown in the following equation:

[0068] y = softmax(x)

[0069] In the formula, x is a 7-dimensional vector and y is a 7-dimensional probability representation.

[0070] In this embodiment, the 7-dimensional vector is represented as [1, 1.445, 6.456, 7.35, 9.56, 7.33, 7.34], which, after softmax, is transformed into the following 7-dimensional probabilities:

[0071] [0.000139746,0.000218072,0.0327227,0.0800035,0.729289,0.0784193,0.0792074]

[0072] Among the above probabilities, 0.729289 is the largest. Therefore, the legal basis corresponding to the work injury occurrence information input into the work injury determination legal basis model is the clause of the fifth category in the work injury occurrence information. This clause is displayed on the interface along with the model results for work injury assessment personnel to further review the results.

[0073] The retrieval model searches for standard cases in the legal case library based on the legal basis clauses obtained from the work injury determination legal basis model.

[0074] Step 4, Model Training, includes the following steps:

[0075] Step 401: Initialize the model parameters of the work injury determination and legal basis model constructed in Step 3. In this embodiment, the initial parameters of the entire network are the parameters of the Google ALBERT pre-trained model.

[0076] Step 402: Using the cross-entropy function as the loss function, the entire network is trained using the stochastic gradient descent algorithm with sample data from the work injury determination information database after data information processing. The number of training samples in each batch is set to 900, and the maximum number of iterations is 150.

[0077] During training, a sample set is obtained from the work-related injury determination information database and divided into a training set and a test set in a 7:3 ratio. The data processing method described in step 2 is used to process the data in both the training and test sets. The work-related injury determination and legal basis model is trained using the training set data, and then tested using the test set data. If the test accuracy reaches a predetermined accuracy P0 (set to this in this embodiment), the model is saved; otherwise, the number of samples in the sample set is increased, and the work-related injury determination and legal basis model is trained again until the accuracy meets the condition, at which point the model is saved.

[0078] During the training process, the work injury determination model and the work injury determination legal basis model are trained independently.

[0079] In this embodiment, the predetermined accuracy P0 is set to 90%. The initial training uses 100,000 samples, divided into a training set and a test set in a 7:3 ratio. After preprocessing the data as described in step 2, the model is trained using the training set data, and then tested using the test set data. The results show that the accuracy of the test set is 95.3%, therefore the model is saved.

[0080] Step 5: Display of Model Results

[0081] For newly added work-related injury information, after processing it according to the data processing method described in step 2 to obtain a data format consistent with the training samples, it is then input into the trained work-related injury determination and legal basis model to calculate the output result. If the work-related injury determination model outputs 1, the result is determined to be a work-related injury, and the work-related injury determination is output. At the same time, the legal basis for the work-related injury determination is derived using the work-related injury determination legal basis model. Then, the legal basis and similar cases are output synchronously after searching the legal clause case library through the retrieval model to obtain similar cases.

Claims

1. A method for determining work-related injuries based on natural language understanding, characterized in that, Includes the following steps: Step 1: Data Collection Collect relevant data on previous work-related injury determinations, including information on the occurrence of work-related injuries, the results of work-related injury determinations, and the legal basis for work-related injury determinations; Construct a work injury determination information database, and store information on the occurrence of work injury, the result of work injury determination, and the legal basis for work injury determination in the form of data tables in the work injury determination information database; At the same time, by organizing information on the legal basis for injury determination, a legal case library is formed that includes typical cases corresponding to each legal clause in the relevant laws; Step 2, Data Processing: This involves preprocessing the structured data in the work injury determination information database, deleting irrelevant fields, and selecting features. Specifically, this includes the following: Data preprocessing: Convert the legal basis information for work-related injury determination into numerical data; based on the statistical results of the data, convert all legal clauses in the relevant laws into n categories and convert them into one-hot format; delete data samples with missing values ​​in each row of the work-related injury determination information database. Irrelevant field deletion: Remove meaningless characters from the work injury occurrence information; Information order conversion: Based on textual patterns, the order of textual information in the work injury situation information is converted, and the diagnostic information is placed at the beginning of the work injury situation information text; Information text format is converted into vector format: The information on the occurrence of work-related injuries is segmented according to the granularity of characters, and then the text information is converted from text format to vector format through a character-label dictionary; Step 3: Construct a model for work-related injury determination and legal basis, which includes the following: The input to the model for determining work-related injuries and its legal basis is information about the occurrence of the work-related injury, with an input size of n. max ×m, where n max is the maximum number of words contained in a long text in unstructured data, and m is the number of data feature categories; The model for determining work-related injuries and the legal basis for work-related injuries consists of two classification models and one retrieval model. The two classification models are the work-related injury determination model and the legal basis for work-related injury determination model, and the retrieval model is a matching-based retrieval model. The classification model includes an information extraction part. After the information on the occurrence of work-related injuries is input into the information extraction part, the information extraction part extracts the sentence feature information. The information extraction part of both classification models is implemented using the BERT model. In the work injury determination model, the sentence feature information obtained by the BERT model is passed through a fully connected layer to obtain a 1-dimensional vector, and then the sigmoid function is used to transform the vector into a 0-1 vector, where 0 represents not being recognized as a work injury and 1 represents being recognized as a work injury. Simultaneously, information about the occurrence of work-related injuries is input into the legal basis model for work-related injury determination. In this model, after obtaining statement feature information through the BERT model, the statement feature information is then passed through a fully connected layer to obtain an n-dimensional vector. Subsequently, the legal basis model for work-related injury determination transforms the n-dimensional vector into an n-dimensional probability representation through a softmax activation function. Finally, the model obtains the corresponding legal basis clauses based on the obtained probabilities. The retrieval model searches for standard cases in the legal case library based on the legal basis clauses obtained from the legal basis model for work-related injury determination. Step 4, Model Training, includes the following steps: Step 401: Initialize the model parameters of the work injury determination and legal basis model constructed in Step 3; Step 402: Using the cross-entropy function as the loss function, the entire network is trained using sample data from the work injury determination information database after data information processing through the stochastic gradient descent algorithm. During training, a sample set is obtained from the work injury determination information database and divided into a training set and a test set in a 7:3 ratio. The data processing method described in step 2 is used to process the data in the training set and the test set. The training set data is used to train the work injury determination and legal basis model, and then the test set data is used to test the model. If the test accuracy reaches the predetermined accuracy P_0, the model is saved. Otherwise, the number of sample data in the sample set is increased, and the work injury determination and legal basis model is trained again until the accuracy meets the condition, and then the model is saved. During the training process, the work injury determination model and the work injury determination legal basis model are trained independently of each other; Step 5: Display of Model Results For newly added work-related injury information, after processing it according to the data information processing method described in step 2 to obtain a data format consistent with the training samples, it is then input into the trained work-related injury determination and work-related injury legal basis model to calculate the output result. If the work-related injury determination model outputs 1, the result is determined to be a work-related injury, and the work-related injury determination is output. At the same time, the legal basis for the work-related injury determination is obtained using the work-related injury determination legal basis model. Then, the legal basis and similar cases are output synchronously after searching the legal clause case library through the retrieval model.

2. The work-related injury determination and assessment method based on natural language understanding as described in claim 1, characterized in that, In step 1: the information on the occurrence of the work-related injury includes the time, place, cause, and result of the work-related injury incident, wherein the result includes the on-site diagnosis result and the hospital diagnosis result; the information on the work-related injury determination result includes whether the injury is recognized as a work-related injury or not; the information on the legal basis for the work-related injury determination includes the relevant clauses of the "Regulations on Work-Related Injury Insurance".

3. The work-related injury determination and assessment method based on natural language understanding as described in claim 1, characterized in that, In step 2: Work injury information with fewer than 20 strings is considered invalid data and deleted.

4. The work-related injury determination and assessment method based on natural language understanding as described in claim 1, characterized in that, In step 2: the vector text information consists of two parts, one is the ID information of the text, and the other is the ID information of the subordinate sentence.

5. The work-related injury determination and assessment method based on natural language understanding as described in claim 1, characterized in that, In step 3: the BERT model uses the improved ALBERT model as the information extraction model.

Images