A method and system for mining automobile health record information based on HanLP and BERT models

Abstract

The application provides a kind of based on HanLP and BERT model's automobile health record information mining method and system, first obtain automobile health record information dataset and divide into training set and test set, adopt HanLP to filter training set data and obtain automobile information keyword dataset, then each data is respectively marked and labeled according to attribute and data format is converted into jsonlines format, then the text data in jsonlines format is divided and marked using BIO marking method to obtain BIO word table, transition matrix is generated according to the marking in BIO word table, BERT model is obtained by training transition matrix, and the optimal BERT model is obtained by optimizing BERT model using evaluation index, finally, the automobile maintenance fault data in test set is input into the optimal BERT model, and the automobile maintenance fault keyword is output to match the automobile maintenance project data and the automobile maintenance spare parts data, to establish the standardized automobile health record information database, to reduce the potential maintenance or repair cost of automobile.

Description

Technical Field

[0001] This invention relates to the fields of information extraction technology and intelligent transportation technology, and in particular to a method and system for mining vehicle health record information based on HanLP and BERT models. Background Technology

[0002] According to data from the Highway Research Institute of the Ministry of Transport, my country's auto repair industry is developing rapidly, with more than 31 provinces and cities achieving full coverage of the electronic health record system for auto repair. However, during the recording process of the electronic health record system, there are many problems such as information loss and classification errors, which seriously affect the recovery and utilization of auto repair data and reduce the economic benefits brought to enterprises.

[0003] The existing methods for extracting key information from vehicle health records mainly include the following: The first method relies on the analysis of fault repair reports made by repair personnel on vehicles returned for repair to extract key information. However, this method has different accuracy rates for different vehicle models and different types of reports from different repair shops, resulting in significant differences in the results and low efficiency. The second method uses existing vehicle health record information databases to cluster and extract vehicle repair data. However, this method is limited by the types of vehicle health information databases and the effectiveness of clustering.

[0004] Therefore, there is an urgent need for a data mining method for automotive health records to solve the problems existing in current methods for extracting key information from automotive health records. Summary of the Invention

[0005] To address the problems of inconsistent results, low efficiency, and limitations imposed by the types of vehicle health information databases and clustering effectiveness in current methods for extracting key information from vehicle health records, this invention provides a vehicle health record information mining method based on HanLP and BERT models. It utilizes basic vehicle repair information data to standardize vehicle fault descriptions, employs the HanLP natural language processing package to quickly establish a matching mechanism between vehicle repair faults, repair items, and repair parts, and uses BIO tagging for data segmentation and labeling. A semantic recognition model for vehicle health information records is proposed to establish accurate vehicle information records, effectively reducing potential vehicle maintenance costs and bringing economic benefits to enterprises. This invention also relates to a vehicle health record information mining system based on HanLP and BERT models.

[0006] The technical solution of the present invention is as follows:

[0007] A method for mining vehicle health record information based on HanLP and BERT models, characterized by the following steps:

[0008] Component dependency analysis extraction steps: Obtain the vehicle health record information dataset, and divide the vehicle health record information dataset into training set and test set according to a preset ratio. Then, use HanLP component dependency analysis to extract vehicle repair fault data, vehicle repair project data and vehicle repair parts data from the training set data to obtain the vehicle information keyword dataset.

[0009] Data labeling and format adjustment steps: Use a text annotation tool to label each data in the car information keyword dataset according to its attributes, and convert the labeled data to jsonlines format. Then, use the BIO tagging method to segment the text data in the jsonlines format to obtain multiple groups of individual words, and tag each group of individual words according to the tag type to obtain the BIO vocabulary.

[0010] BERT model training steps: Generate a transition matrix based on the combination of tags in the BIO vocabulary, use the transition matrix as training samples to train the BERT model to obtain the trained BERT model, and use evaluation metrics to optimize the trained BERT model to obtain the optimal BERT model;

[0011] The steps for establishing the information database are as follows: Input the vehicle repair fault data in the test set into the optimal BERT model, output vehicle repair fault keywords, match vehicle repair project data and vehicle repair parts data according to the vehicle repair fault keywords, and establish a vehicle health record information database.

[0012] Preferably, in the data labeling and format adjustment step, the open-source text annotation tool docano is used to label each data in the car information keyword dataset according to its attributes, and an NLP annotation project is created. The annotation includes full annotation and partial annotation.

[0013] In the BERT model training step, evaluation metrics including accuracy, recall, and the harmonic value of accuracy and recall are used to optimize the trained BERT model.

[0014] Preferably, in the BERT model building step, optimizing the BERT model using evaluation metrics includes:

[0015] The scores of precision, recall, and the harmonic sum of precision and recall are calculated based on the evaluation metrics for full and partial labeling, respectively. The influence of low-scoring fuzzy labels on model recognition is removed, and high-scoring label settings are saved to improve model accuracy. The optimal BERT model is then output.

[0016] Preferably, in the BERT model building step, before training the transition matrix to obtain the BERT model, each group of marked individual characters is converted into character vectors, thereby generating the original position vectors of the character vectors. Based on the character vectors and position vectors, the transition matrix is ​​trained to obtain the BERT model.

[0017] Preferably, in the data calibration and format adjustment steps, the label types include accessory names, verbs, general adjectives, locations, brands, and colors.

[0018] A vehicle health record information mining system based on HanLP and BERT models is characterized by comprising, in sequence, a component dependency analysis extraction module, a data calibration and format adjustment module, a BERT model training module, and an information database establishment module.

[0019] The component dependency analysis extraction module acquires the vehicle health record information dataset and divides it into a training set and a test set according to a preset ratio. Then, HanLP component dependency analysis is used to extract vehicle repair fault data, vehicle repair project data, and vehicle repair parts data from the training set data to obtain the vehicle information keyword dataset.

[0020] The data labeling and format adjustment module uses a text annotation tool to label each data in the car information keyword dataset according to its attributes, and converts the labeled data into jsonlines format. Then, the BIO tagging method is used to segment the text data in the jsonlines format to obtain multiple groups of individual words, and each group of individual words is tagged according to the tag type to obtain the BIO vocabulary.

[0021] The BERT model training module generates a transition matrix based on the combination of tags in the BIO vocabulary. The transition matrix is ​​used as training samples to train the BERT model and obtain the trained BERT model. The evaluation metric is then used to optimize the trained BERT model to obtain the optimal BERT model.

[0022] The information database establishment module inputs automotive repair fault data from the test set into the optimal BERT model, outputs automotive repair fault keywords, and matches automotive repair project data and automotive repair parts data based on the automotive repair fault keywords to establish an automotive health record information database.

[0023] Preferably, the evaluation metrics include precision, recall, and the harmonic value of precision and recall; the annotation includes full annotation and partial annotation.

[0024] Preferably, in the BERT model building module, optimizing the BERT model using evaluation metrics includes:

[0025] The scores of precision, recall, and the harmonic sum of precision and recall are calculated based on the evaluation metrics for full and partial labeling, respectively. The influence of low-scoring fuzzy labels on model recognition is removed, and high-scoring label settings are saved to improve model accuracy. The optimal BERT model is then output.

[0026] Preferably, in the BERT model building module, before training the transition matrix to obtain the BERT model, each group of marked individual characters is converted into character vectors, thereby generating the original position vectors of the character vectors. Based on the character vectors and position vectors, the transition matrix is ​​trained to obtain the BERT model.

[0027] Preferably, the label types include accessory name, verb, general adjective, location, brand, and color.

[0028] The beneficial effects of this invention are as follows:

[0029] This invention provides a method for mining vehicle health record information based on HanLP and BERT models. First, a vehicle health record information dataset is acquired and divided into training and testing sets according to a preset ratio. Then, HanLP component dependency analysis is used to extract the training set data to obtain a vehicle information keyword dataset. Next, a text annotation tool is used to label each data point in the vehicle information keyword dataset according to its attributes, and the labeled data is converted to JSONLines format for easy import and processing of subsequent model data. Then, the BIO tagging method is used to segment the text data in the JSONLines format to obtain multiple groups of individual characters, and each group of individual characters is labeled according to tag type to obtain a BIO vocabulary. Finally, based on the BIO... The tokens in the O-vocabulary are combined to generate a transition matrix. This transition matrix is ​​used as training samples to train the BERT model, resulting in a trained BERT model. Evaluation metrics are then used to optimize the trained BERT model, yielding the optimal BERT model. Finally, automotive repair fault data from the test set is input into this optimal BERT model, which is a semantic recognition model for automotive health information records. This optimal BERT model outputs automotive repair fault keywords. Based on these keywords, it matches automotive repair item data and automotive repair parts data to establish an automotive health record information database. In essence, the accuracy of the BERT training model is judged based on the test set data, and the optimal model and automotive health record information database are output. Through this optimal BERT model—the semantic recognition model for automotive health information records—accurate vehicle information records are established. This invention utilizes basic automotive repair information data to standardize automotive fault descriptions, quickly establishes a matching mechanism between fault descriptions, repair items, and repair parts, and proposes a semantic recognition model for automotive health information records to establish accurate vehicle information records. This effectively reduces potential maintenance or repair costs for automobiles, bringing economic benefits to enterprises. This invention, without adding any additional equipment or data, fully extracts the effective semantic information from the data. Based on the HanLP and BERT models, it mines and matches the relationships between automotive repair faults, projects, and parts. It also uses the BIO tagging method for data segmentation and labeling, effectively avoiding the inaccuracy of file information caused by the colloquial language of the data or the differences in wording among repair manufacturers in traditional methods. It can quickly and accurately extract keyword information from the original data and accurately establish an effective automotive health record information matching mechanism, which is conducive to the establishment of a complete and comprehensive automotive health record.

[0030] This invention also relates to a vehicle health record information mining system based on HanLP and BERT models. This system corresponds to the aforementioned vehicle health record information mining method based on HanLP and BERT models. It can be understood as a system that implements the aforementioned vehicle health record information mining method based on HanLP and BERT models. It includes a component dependency analysis extraction module, a data labeling and format adjustment module, a BERT model training module, and an information database establishment module connected in sequence. Each module works collaboratively to standardize vehicle fault descriptions using basic vehicle maintenance information data. It uses HanLP to quickly establish a matching mechanism between vehicle maintenance faults, maintenance items, and maintenance parts. It uses BIO tagging to segment and label data, and proposes a semantic recognition model for vehicle health information records to establish accurate vehicle information records. This can effectively reduce potential vehicle maintenance or repair costs and bring economic benefits to enterprises. Attached Figure Description

[0031] Figure 1 This is a flowchart of the vehicle health record information mining method based on HanLP and BERT models of the present invention.

[0032] Figure 2 This is a schematic diagram illustrating the principle of training the BERT model in this invention. Detailed Implementation

[0033] The present invention will now be described with reference to the accompanying drawings.

[0034] This invention relates to a method for mining vehicle health record information based on HanLP and BERT models. The flowchart of this method is as follows: Figure 1 As shown, the steps are as follows:

[0035] Component dependency analysis extraction steps: Obtain the vehicle health record information dataset and divide it into training and test sets according to a preset ratio. Use HanLP component dependency analysis to extract vehicle repair fault data, vehicle repair item data, and vehicle repair parts data from the training set data to obtain the vehicle information keyword dataset. That is, Figure 1 The diagram shows the collection of vehicle health record information, the segmentation of the dataset, and the extraction of keywords based on HanLP.

[0036] Specifically, first, divide the dataset of automotive health record information obtained by the user into a training set and a test set according to a preset ratio. The ratio of the training set to the test set can be set to 7:3. For the training set data, use the HanLP constituency dependency analysis in Python to analyze the syntactic relationships between sentences and words in the training set data, generate a tree structure, that is, generate an initial dependency syntax tree. By extracting the virtual root node, analyze the syntactic relationships of the original automotive repair fault information, repair item information, and repair parts information in the training set data, and exclude fuzzy words such as "的", "得", "有时", "抖", etc. Extract the automotive information keywords of the automotive repair fault data, automotive repair item data, and automotive repair parts data to obtain an automotive information keyword dataset. This step normalizes the repair information, excludes fuzzy semantic words in the syntax, and uses HanLP to uniformly name the automotive repair fault data, automotive repair parts data, and automotive repair item data.

[0037] It should be noted that for the part names containing plural components and some complex information that is difficult for HanLP to recognize or is recognized inaccurately, manual screening and extraction are performed. Manually check whether the keywords automatically extracted by HanLP conform to normal grammar and exclude abnormal data. That is to say, based on HanLP, the initial data is screened and complex data is manually screened according to the data.

[0038] Data calibration and format adjustment steps: Use a text annotation tool to label each data in the automotive information keyword dataset according to its attributes respectively, convert the format of the labeled data to the jsonlines format, then use the BIO tagging method to segment the text data in the jsonlines format into multiple single characters, and use the BIO tagging method to mark each group of single characters according to the label types to obtain a BIO vocabulary.

[0039] Such as Figure 1The data annotation shown is based on the doccano platform. You can access doccano's official text annotation platform, select a text category, create an NLP annotation project, add a corpus filtered using HanLP, and design labels based on car repair fault information, repair project information, and repair parts information. Then, annotate and export the data according to the required labels. Specifically, first, use doccano's open-source text annotation tools to label each data point in the car information keyword dataset (i.e., car repair fault data, car repair project data, and car repair parts data) according to attributes. Set the label types to include: n (part name), v (verb), adj (general adjective), position (location), brand (brand), and color (color), and store the labeled data by category. Then, use Python to batch convert the labeled data to jsonlines format. The jsonlines format data content includes the required information such as the ID of the car repair fault, car repair project, and car repair parts data, the text content "text," and the label "label." Then, the BIO tagging method is used to segment the jsonlines format text data into the input format for the BERT model. This involves segmenting the "text" content (i.e., text data) in the jsonlines format from words into multiple groups of individual characters, and then using BIO tagging to label each group of individual characters according to the aforementioned label categories. This results in a BIO vocabulary, with the tagging rules as follows: B corresponds to the beginning of an entity word, I corresponds to the middle of an entity word, and O corresponds to no entity word. Preferably, the annotation includes full annotation and partial annotation.

[0040] BERT model training steps: Generate a transition matrix based on the combination of tags in the BIO vocabulary, use the transition matrix as training samples to train the BERT model to obtain the trained BERT model, and use evaluation metrics to optimize the trained BERT model to obtain the optimal BERT model.

[0041] The BERT model used is based on a bidirectional language model and is an unsupervised training algorithm. Internally, it uses a bidirectional Transformer encoder module as its main component. Single-character inputs are mapped to character vectors for model training. Precision, Recall, and F1 Score (the harmonic sum of precision and recall) are used as evaluation metrics. The F1 Score is maximized, and the transition matrix of the BIO vocabulary is output. The accuracy of the current output is determined by combining the character vectors before and after the CRF layer. Full and partial annotations are considered, and data labels are selected to optimize the BERT model.

[0042] The core ideas for training and optimizing the BERT model are as follows: ① Select a corpus containing car health information and set the ratio of training set to test set to 7:3; ② Split the data samples after component dependency processing into several words, use the BIO tagging method to segment the input file data, and convert the words into word vectors as the training vocabulary; ③ Split the original training data into single-character inputs to the BERT model, map the characters to word vectors, and start model training; ④ Output the transition matrix of the BIO vocabulary, learn the vocabulary transition matrix through the CRF layer, combine the previous and subsequent outputs to judge the accuracy of the current output, and adjust the vocabulary transition matrix to improve accuracy; ⑤ Use the Adam Optimizer optimization algorithm to find the optimal fitting solution for the dataset; ⑥ Combine the BIO vocabulary output results and evaluate the F1 score based on the Precision / Recall / F1Score (the harmonic value of precision and recall) to train the BERT model.

[0043] Specifically, such as Figure 2 As shown, before training the BERT model, the training parameters need to be fine-tuned. This involves adjusting the upstream training parameters based on changes in downstream training parameters, with upstream and downstream parameters mutually optimizing each other. Therefore, each group of individual characters labeled using the BIO tagging method is first input into the BERT model. The BERT model then maps each character to a character vector A. n Start model training and simultaneously generate the position vector B of the original data for the word vectors. nTo facilitate the construction of constraints during subsequent training, a transition matrix C is generated based on the tokens (B, I, O) in the BIO vocabulary. This transition matrix C is then output to the CRF layer of the BERT model to constrain the output conditions, including conditions such as text starting with "B-" or "O". Learning the transition matrix C improves accuracy. Finally, the trained BIO vocabulary is combined with the output results, and the BERT model is evaluated based on P / R / F1 scores. Specifically, the CRF layer learns the transition matrix C to obtain the trained BERT model. Then, the original unlabeled car repair fault data, car repair project data, and car repair parts data are used as input to the trained BERT model. The CRF layer combines the word vectors output before and after training the BERT model to determine the accuracy of the current output, considering both full and partial annotations, selecting data labels, and using evaluation metrics to optimize the BERT model to obtain the optimal BERT model. In essence, the Precision / Recall / F1 Score (P / R / F1 score) evaluation metrics are used to calculate the changes in Precision, Recall, and F1 Score under full and partial labeling conditions. The impact of low-scoring, ambiguous labels on model recognition is removed, while high-scoring labels are preserved to improve model accuracy. After training the model to convergence, the optimal BERT model is output. Figure 1 The diagram illustrates training a BERT model based on P / R / F1 scores, adjusting labels, and optimizing the model. For example, in a real-world case, the P / R / F1 scores for full labeling are 79.46 / 82.87 / 81.13, while the P / R / F1 scores for partial labeling are 81.92 / 81.02 / 81.47. The label settings under partial labeling are selected to improve model accuracy.

[0044] The steps for establishing the information database are as follows: Test and match information on the test set based on the trained model, input the car repair fault data in the test set into the optimal BERT model, output car repair fault keywords, match car repair project data and car repair parts data based on the car repair fault keywords, and establish a car health record information database.

[0045] Specifically, the vehicle repair fault data from the test set is input into the optimal BERT model to obtain vehicle repair fault keywords. These keywords are then used to match vehicle repair item data and vehicle repair parts data. The output results of the test set and the optimal BERT model are compared to obtain the accuracy of the optimal BERT model. Finally, the optimal BERT model is used to analyze the remaining data, outputting data files in ANSI and UTF-8 encoding formats to establish a vehicle health record information database. The model output is the best result under various label annotations, and the output can be a CSV data file containing the matched vehicle health record information entries. For example, in a practical case, in implementing the matching of vehicle repair item and vehicle repair parts information based on vehicle repair faults, the trained optimal BERT model achieved an accuracy of 83.6%, outputting a vehicle health record information database in XLS file format.

[0046] This invention also relates to a vehicle health record information mining system based on HanLP and BERT models. This system corresponds to the aforementioned vehicle health record information mining method based on HanLP and BERT models, and can be understood as a system implementing the above method. The system includes a component dependency analysis extraction module, a data calibration and format adjustment module, a BERT model building module, and an information database building module connected in sequence. Specifically,

[0047] The component dependency analysis extraction module acquires the vehicle health record information dataset and divides it into a training set and a test set according to a preset ratio. Then, HanLP component dependency analysis is used to extract vehicle repair fault data, vehicle repair project data, and vehicle repair parts data from the training set data to obtain the vehicle information keyword dataset.

[0048] The data labeling and format adjustment module uses a text annotation tool to label each data in the car information keyword dataset according to its attributes, converts the labeled data to jsonlines format, and then uses the BIO tagging method to segment the text data in the jsonlines format into multiple groups of individual words. The BIO tagging method is then used to tag each group of individual words according to the tag type to obtain a BIO vocabulary.

[0049] The BERT model training module generates a transition matrix based on the combination of tags in the BIO vocabulary. The transition matrix is ​​used as training samples to train the BERT model and obtain the trained BERT model. The evaluation metric is then used to optimize the trained BERT model to obtain the optimal BERT model.

[0050] The information database establishment module inputs automotive repair fault data from the test set into the optimal BERT model, outputs automotive repair fault keywords, and matches automotive repair project data and automotive repair parts data based on the automotive repair fault keywords to establish an automotive health record information database.

[0051] Preferably, the evaluation metrics include precision, recall, and the harmonic value of precision and recall; the annotation includes full annotation and partial annotation.

[0052] Preferably, in the BERT model building module, optimizing the BERT model using evaluation metrics includes:

[0053] The scores of precision, recall, and the harmonic sum of precision and recall are calculated based on the evaluation metrics for full and partial labeling, respectively. The influence of low-scoring fuzzy labels on model recognition is removed, and high-scoring label settings are saved to improve model accuracy. The optimal BERT model is then output.

[0054] Preferably, in the BERT model building module, before training the transition matrix to obtain the BERT model, each group of individual characters after being labeled is converted into character vectors, and then the original position vectors of the character vectors are generated. Based on the character vectors and position vectors, the transition matrix is ​​trained to obtain the BERT model.

[0055] Preferably, the label types include accessory name, verb, general adjective, location, brand, and color.

[0056] This invention addresses the issues of inconsistent naming of repair parts information and inaccurate matching between repair faults, repair items, and parts in current vehicle health information records. It provides an objective and scientific method and system for mining vehicle health record information based on HanLP and BERT models. The method standardizes vehicle fault descriptions using basic vehicle repair information data, extracts and labels keywords using HanLP component dependency analysis, quickly establishes a matching mechanism between vehicle repair faults, repair items, and repair parts, analyzes semantic relationships under repair part information using the HanLP open-source toolkit, and uses the doccano open-source text annotation tool to annotate the raw data. BIO tagging is then used for data segmentation and labeling, and the annotated data is segmented into the input format specified by the BERT model. A semantic recognition model for vehicle health information records is proposed, and the trained model is evaluated and optimized based on P / R / F1 scores to establish an accurate and standardized vehicle health record information database. This effectively reduces potential vehicle maintenance or repair costs, bringing economic benefits to enterprises. Without adding any additional equipment or data, it can fully extract the effective semantic information from the data, and combine HanLP and BERT models to accurately establish an effective vehicle health record information matching mechanism and improve the vehicle health record information database.

[0057] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention patent.

Claims

1. A method for mining vehicle health record information based on HanLP and BERT models, characterized in that, Includes the following steps: Component dependency analysis extraction steps: Obtain the vehicle health record information dataset and divide the vehicle health record information dataset into training set and test set according to a preset ratio. Then, use HanLP component dependency analysis to analyze the syntactic relations of the original vehicle repair fault information, repair item information and repair parts information in the training set data. Extract the vehicle information keywords from the vehicle repair fault data, vehicle repair item data and vehicle repair parts data in the training set data to obtain the vehicle information keyword dataset. Data labeling and format adjustment steps: The open-source text annotation tool docano is used to label each data in the automotive information keyword dataset according to its attributes, and an NLP annotation project is created. The annotation includes full annotation and partial annotation. The annotated data format is converted to jsonlines format. Then, the BIO tagging method is used to segment the text data in the jsonlines format to obtain multiple groups of individual words. Each group of individual words is tagged according to the tag type to obtain the BIO vocabulary. BERT model training steps: A transition matrix is ​​generated based on the tag combinations in the BIO vocabulary. Then, each group of individual characters after tagging is converted into character vectors, and the original position vectors of these character vectors are generated. Based on the character vectors and position vectors, the transition matrix is ​​trained to obtain the BERT model. The transition matrix is ​​used as training samples to train the BERT model, resulting in the trained BERT model. The trained BERT model is then optimized using evaluation metrics including precision, recall, and the harmonic value of precision and recall. Specifically, the scores for precision, recall, and the harmonic value of precision and recall are calculated based on the evaluation metrics under full and partial labeling conditions. The influence of low-scoring, ambiguous labels on model recognition is removed, and high-scoring label settings are saved to improve model accuracy, resulting in the optimal BERT model. The steps for establishing the information database are as follows: Input the vehicle repair fault data in the test set into the optimal BERT model, output vehicle repair fault keywords, match vehicle repair project data and vehicle repair parts data according to the vehicle repair fault keywords, and establish a vehicle health record information database.

2. The method for mining vehicle health record information based on HanLP and BERT models according to claim 1, characterized in that, In the data calibration and format adjustment steps, the label types include accessory name, verb, general adjective, location, brand, and color.

3. A vehicle health record information mining system based on HanLP and BERT models, characterized in that, It includes a component dependency analysis extraction module, a data calibration and format adjustment module, a BERT model training module, and an information database establishment module, which are connected in sequence. The component dependency analysis extraction module acquires the vehicle health record information dataset and divides it into training and test sets according to a preset ratio. Then, HanLP component dependency analysis is used to analyze the syntactic relationships of the original vehicle repair fault information, repair item information and repair parts information in the training set data. The module extracts vehicle information keywords from the vehicle repair fault data, vehicle repair item data and vehicle repair parts data in the training set data to obtain the vehicle information keyword dataset. The data labeling and format adjustment module uses the open-source text annotation tool docano to label each data point in the automotive information keyword dataset according to its attributes, creating an NLP annotation project. The annotation includes full annotation and partial annotation. The annotated data is then converted to jsonlines format, and the text data in the jsonlines format is segmented into multiple groups of individual characters using the BIO tagging method. Each group of individual characters is then tagged according to the tag type to obtain a BIO vocabulary. The BERT model training module generates a transition matrix based on the combination of tags in the BIO vocabulary, then converts each group of individual characters after tagging into character vectors, and then generates the original position vectors of the character vectors. Based on the character vectors and position vectors, the transition matrix is ​​trained to obtain the BERT model. The transition matrix is ​​used as training samples to train the BERT model to obtain the trained BERT model. The trained BERT model is then optimized using evaluation metrics including precision, recall, and the harmonic value of precision and recall. Specifically, the scores of precision, recall, and the harmonic value of precision and recall are calculated according to the evaluation metrics under full labeling and partial labeling respectively. The influence of low-scoring fuzzy labels on model recognition is removed, and high-scoring label settings are saved to improve model accuracy, resulting in the optimal BERT model. The information database establishment module inputs automotive repair fault data from the test set into the optimal BERT model, outputs automotive repair fault keywords, and matches automotive repair project data and automotive repair parts data based on the automotive repair fault keywords to establish an automotive health record information database.

4. The vehicle health record information mining system based on HanLP and BERT models according to claim 3, characterized in that, The label types include accessory name, verb, general adjective, location, brand, and color.

Images