Automobile document semantic extraction method and device, and storage medium

Abstract

The application relates to the technical field of document processing and question answering, in particular to a car document semantic extraction method and device and a storage medium, wherein the method comprises the following steps: constructing a car document and question answering dataset; generating a lightweight Transform semantic understanding model by using the car document and question answering dataset, so as to construct a car question answering platform according to the lightweight Transform semantic understanding model; evaluating the field extraction accuracy and the question answering accuracy of the car question answering platform based on a preset hierarchical quantitative evaluation system; and extracting core field information in a target car document by using the car question answering platform in the case that the field extraction accuracy and the question answering accuracy both meet corresponding index conditions. Thus, the problems that related technologies rely on historical question answering texts, are difficult to cover the full-scene semantic requirements of car documents, and result in low data labeling and utilization efficiency are solved.

Description

Technical Field

[0001] This application relates to the field of document processing and question answering technology, and in particular to a method, apparatus and storage medium for semantic extraction of automotive documents. Background Technology

[0002] Among related technologies, there are methods to obtain target question-and-answer text from historical question-and-answer text, determine the target part-of-speech corresponding to the candidate words in the historical question-and-answer text, calculate the word frequency of the target part-of-speech in the target question-and-answer text, calculate the inverse document frequency corresponding to the target part-of-speech, and then calculate the target weight of the candidate words in the historical question-and-answer text, thereby determining the keywords corresponding to the target question-and-answer text.

[0003] However, related technologies rely on historical question-and-answer texts. However, the acquisition, organization, and annotation of historical question-and-answer texts require a lot of manpower and resources, and it is difficult to cover the semantic needs of automotive documents in all scenarios, resulting in low efficiency in data annotation and utilization, which urgently needs to be solved. Summary of the Invention

[0004] This application provides a method, apparatus, and storage medium for semantic extraction of automotive documents, in order to solve the problem that related technologies rely on historical question-and-answer text, which makes it difficult to cover the full-scene semantic needs of automotive documents, resulting in low efficiency in data annotation and utilization.

[0005] The first aspect of this application provides a method for semantic extraction from automotive documents, comprising the following steps: constructing an automotive document and question-and-answer dataset; generating a lightweight Transformer semantic understanding model using the automotive document and question-and-answer dataset, and constructing an automotive question-and-answer platform based on the lightweight Transformer semantic understanding model; evaluating the field extraction accuracy and question-and-answer accuracy of the automotive question-and-answer platform based on a preset hierarchical quantitative evaluation system, so that, when both the field extraction accuracy and the question-and-answer accuracy meet the corresponding indicator conditions, the core field information in the target automotive document is extracted using the automotive question-and-answer platform.

[0006] Based on the above technical means, this application embodiment constructs an automotive document and question-and-answer dataset and generates a lightweight Transformer semantic understanding model. This model enables the model to understand automotive terminology and complex query intents. While ensuring high accuracy, it significantly reduces model complexity and computational resource consumption, and improves inference speed. By utilizing an automotive question-and-answer platform to extract core field information from target automotive documents, it supports functions such as natural language questioning, quick answers to preset questions, question-and-answer history and export, improving practicality and realizing one-stop, implementable intelligent processing of automotive documents, effectively supporting the R&D, testing, and after-sales processes in the automotive industry.

[0007] Optionally, in one embodiment of this application, the construction of the automotive document and question-and-answer dataset includes: generating a simulated document based on a preset automotive crash test report template; generating questions with multiple expression forms and their corresponding field labels; and constructing the automotive document and question-and-answer dataset based on the simulated document and the questions and their corresponding field labels.

[0008] Based on the above technical means, the embodiments of this application provide the model with rich learning samples that fit the automotive field by constructing an automotive document and question-answer dataset, which is the foundation for improving the model's domain adaptability and semantic understanding accuracy. In addition, data augmentation strategies can be introduced, such as adding slang and colloquialisms, to further enhance the model's generalization ability.

[0009] Optionally, in one embodiment of this application, the step of generating a lightweight Transformer semantic understanding model using the automotive document and question-answer dataset includes: acquiring automotive domain technical document corpus; using the automotive domain technical document corpus to further pre-train a preset BERT-base pre-trained model to update the weights of the preset BERT-base pre-trained model to obtain a pre-trained optimized model; and using the automotive document and question-answer dataset to fine-tune the pre-trained optimized model to generate the lightweight Transformer semantic understanding model.

[0010] Based on the above technical means, the embodiments of this application effectively reduce the complexity of the model through lightweight model architecture design. In addition, through domain adaptive pre-training and targeted multi-task fine-tuning, the model fully learns the semantic features and question-field mapping relationship of the automotive domain, thereby achieving high-accuracy intent recognition and field classification.

[0011] Optionally, in one embodiment of this application, the construction of the car question-and-answer platform includes: constructing a semantic understanding module, a document retrieval module, an answer generation module, and an interaction optimization module. The semantic understanding module receives user natural language questions and uses the lightweight Transformer semantic understanding model to output the field type and confidence level corresponding to the user's natural language question. The document retrieval module performs document retrieval based on the field type and confidence level to obtain structured document information. The answer generation module generates structured answer content based on the structured document information. The interaction optimization module interacts with the user and stores the answer content. The car question-and-answer platform is established based on the semantic understanding module, the document retrieval module, the answer generation module, and the interaction optimization module.

[0012] Based on the above technical means, the embodiments of this application can integrate a high-precision OCR (Optical Character Recognition) engine through the document retrieval module, which effectively ensures the accurate extraction of original information from complex automotive documents. In addition, through the semantic understanding module, it can combine a pre-trained lightweight Transformer semantic understanding model to perform semantic parsing on users' natural language questions. The answer generation module effectively ensures that the answers are both accurate and comprehensive. Furthermore, the interaction optimization function of the interaction optimization module improves ease of use and practicality.

[0013] Optionally, in one embodiment of this application, the formula for calculating the field extraction accuracy may be, but is not limited to, the following: , in, Extract the accuracy for the field. The number of fields correctly extracted by the model. This represents the total number of fields in all documents within the test set.

[0014] Based on the above technical means, the embodiments of this application can directly reflect the accuracy of the model in locating and extracting target fields in automotive documents by calculating the field extraction accuracy, and objectively measure the consistency of the model's parsing and classification of text semantics, table structure and field boundaries.

[0015] Optionally, in one embodiment of this application, the formula for calculating the question-and-answer accuracy rate may be, but is not limited to, the following: , in, The accuracy rate of the question and answer is... The number of questions the model answered correctly. This represents the total number of issues in the test set.

[0016] Based on the above technical means, the embodiments of this application can directly quantify the overall correctness of the model's response to users' natural language questions by calculating the question-answering accuracy rate, and comprehensively reflect the overall execution accuracy of the model in understanding user intent, locating fields, and generating answers.

[0017] A second aspect of this application provides a semantic extraction device for automotive documents, comprising: a first construction module for constructing an automotive document and question-and-answer dataset; a second construction module for generating a lightweight Transformer semantic understanding model using the automotive document and question-and-answer dataset, and constructing an automotive question-and-answer platform based on the lightweight Transformer semantic understanding model; and an extraction module for evaluating the field extraction accuracy and question-and-answer accuracy of the automotive question-and-answer platform based on a preset hierarchical quantitative evaluation system, so as to extract core field information from a target automotive document using the automotive question-and-answer platform when both the field extraction accuracy and the question-and-answer accuracy meet the corresponding indicator conditions.

[0018] Based on the above technical means, this application embodiment constructs an automotive document and question-and-answer dataset and generates a lightweight Transformer semantic understanding model. This model enables the model to understand automotive terminology and complex query intents. While ensuring high accuracy, it significantly reduces model complexity and computational resource consumption, and improves inference speed. By utilizing an automotive question-and-answer platform to extract core field information from target automotive documents, it supports functions such as natural language questioning, quick answers to preset questions, question-and-answer history and export, improving practicality and realizing one-stop, implementable intelligent processing of automotive documents, effectively supporting the R&D, testing, and after-sales processes in the automotive industry.

[0019] A third aspect of this application provides an automobile, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the automobile document semantic extraction method as described in the above embodiments.

[0020] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described automotive document semantic extraction method.

[0021] A fifth aspect of this application provides a computer program product, including a computer program that, when executed, implements the above-described automotive document semantic extraction method.

[0022] This application's embodiments construct an automotive document and question-and-answer dataset and generate a lightweight Transformer semantic understanding model. This model enables the understanding of automotive terminology and complex query intents, significantly reducing model complexity and computational resource consumption while maintaining high accuracy, thus improving inference speed. By utilizing an automotive question-and-answer platform to extract core field information from target automotive documents, it supports functions such as natural language questioning, rapid answers to preset questions, question-and-answer history recording and export, enhancing practicality and achieving one-stop, implementable intelligent processing of automotive documents. This effectively supports the automotive industry's R&D, testing, and after-sales processes. Therefore, it solves the problem that related technologies rely on historical question-and-answer text, making it difficult to cover the full-scenario semantic needs of automotive documents, resulting in low data annotation and utilization efficiency.

[0023] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a semantic extraction method for automotive documents provided according to an embodiment of this application; Figure 2 This is a flowchart illustrating the construction of a car documentation and question-and-answer dataset according to an embodiment of this application; Figure 3 A flowchart illustrating the training and generation of a lightweight Transformer semantic understanding model according to an embodiment of this application; Figure 4 A flowchart illustrating the operation of a car Q&A platform according to an embodiment of this application; Figure 5(a) is a schematic diagram of the effect of a car Q&A platform provided according to an embodiment of this application; Figure 5(b) is a schematic diagram of the effect of a car Q&A platform provided according to an embodiment of this application; Figure 5(c) is a schematic diagram of the effect of a car Q&A platform provided according to an embodiment of this application; Figure 6 This is a schematic diagram of the question-and-answer history results of an automotive question-and-answer platform according to an embodiment of this application; Figure 7 This is a flowchart of a semantic extraction method for automotive documents according to an embodiment of this application; Figure 8 This is a block diagram of a semantic extraction device for automotive documents provided according to an embodiment of this application; Figure 9This is a schematic diagram of the structure of a car provided according to an embodiment of this application.

[0025] Figure label: 80 - Automotive document semantic extraction device; 100 - First building module, 200 - Second building module, 300 - Extraction module; 901 - Memory, 902 - Processor, 903 - Communication interface. Detailed Implementation

[0026] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0027] Currently, intelligent document processing and question-answering systems in the automotive industry mainly adopt the following technical solutions: (1) Constructing a knowledge base based on syntactic structure: By performing dependency parsing and semantic dependency parsing on car-related texts, we determine syntactic elements and their grammatical attributes (such as subject-predicate relations, attributive-head relations, etc.), and generate a "question-answer" knowledge base based on this.

[0028] (2) Determine the key words based on statistics: Using the term frequency (TF) and inverse document frequency (IDF) algorithms, and combining business indicators such as vehicle delivery volume and user satisfaction as weighting factors, core keywords are selected from historical question and answer texts to improve the targeting of the matching of the key words.

[0029] (3) Question answering based on GraphRAG (Graph Retrieval-Augmented Generation): By identifying entities (functions, components, parameters, etc.) and entity relationships (functional dependencies, spatial locations, etc.) in automotive knowledge documents, a knowledge graph is constructed, and LLM (Large Language Model) is combined to perform hierarchical community division, thereby realizing structured summary retrieval and answer generation.

[0030] While the aforementioned technologies have achieved certain results in specific fields, they still have the following significant shortcomings when processing highly specialized and unstructured automotive documents such as "car crash test reports": (1) Insufficient semantic understanding depth: The relevant syntactic analysis and keyword statistics methods are difficult to handle complex automotive professional terms and long and difficult sentences, resulting in limited accuracy in semantic mapping of core fields such as vehicle model, report number, and test parameters, and failing to cope with the diverse expressions of users.

[0031] (2) Imbalance between model complexity and inference efficiency: Although the related GraphRAG or large graph solutions have strong understanding capabilities, their model structures are complex and have extremely high requirements for computing resources. They are difficult to achieve fast response in resource-constrained environments, and their domain adaptability is poor due to the lack of targeted fine-tuning.

[0032] (3) Inefficient data annotation and utilization: Most of the related automation solutions rely on preset templates or historical question-and-answer pairs, and lack standardized and professional question-and-answer dataset construction methods for specific fields of automotive documents (such as specific test values, complex automotive configurations, etc.).

[0033] In summary, existing technical solutions for intelligent processing and question answering of automotive documents generally suffer from insufficient semantic understanding depth, weak domain adaptability, slow system response, poor ability to parse complex document structures, and difficulty in balancing lightweight deployment with high precision requirements. Particularly in scenarios requiring rapid and accurate location and answering of specialized questions from massive amounts of multi-format automotive documents, there is a lack of an integrated solution that deeply integrates lightweight semantic understanding, high-precision document parsing, and efficient retrieval and generation capabilities.

[0034] To address the aforementioned problems, this application proposes a method, apparatus, and storage medium for semantic extraction of automotive documents. The following description, with reference to the accompanying drawings, details the method, apparatus, and storage medium for semantic extraction of automotive documents according to embodiments of this application.

[0035] Figure 1 This is a flowchart of a semantic extraction method for automotive documents provided according to an embodiment of this application.

[0036] like Figure 1 As shown, the semantic extraction method for automotive documents includes the following steps: In step S101, a car documentation and question-and-answer dataset is constructed.

[0037] In some cases, embodiments of this application first call a preset car crash test report template, and generate a massive amount of highly realistic PDF documents by randomly filling in simulated data of 19 core fields such as car model, test parameters, report number, and test date. Then, based on the above 19 core fields, a preset rule engine generates questions with multiple forms of expression and their corresponding field tags to form a "question-field tag" mapping pair to construct a car document and question-and-answer dataset.

[0038] In step S102, a lightweight Transformer semantic understanding model is generated using automotive documents and question-answering datasets, and an automotive question-answering platform is built based on the lightweight Transformer semantic understanding model.

[0039] In some cases, embodiments of this application utilize automotive domain technical document corpora to further pre-train a pre-set BERT-based pre-trained model, updating the weights of the pre-set BERT-based pre-trained model to obtain a pre-trained optimized model. Then, using automotive documents and question-answering datasets, the pre-trained optimized model is jointly fine-tuned through multiple tasks to obtain a pre-trained fine-tuned model. Next, five-fold cross-validation is performed on the pre-trained fine-tuned model to obtain the five-fold cross-validation results. Based on the five-fold cross-validation results, model performance is evaluated. When the model performance evaluation passes, the optimal model is determined and identified as the lightweight Transformer semantic understanding model. Furthermore, embodiments of this application utilize the lightweight Transformer semantic understanding model to construct a semantic understanding module, combining it with a document retrieval module, an answer generation module, and an interaction optimization module to establish an automotive question-answering platform.

[0040] In step S103, based on a preset hierarchical quantitative evaluation system, the field extraction accuracy and question-answering accuracy of the car question-answering platform are evaluated. If both the field extraction accuracy and question-answering accuracy meet the corresponding indicator conditions, the core field information in the target car document is extracted using the car question-answering platform.

[0041] Specifically, the pre-defined tiered quantitative evaluation system is divided into two levels: field extraction evaluation and question-answering performance evaluation. These levels may include, but are not limited to, field extraction accuracy and single-question question-answering accuracy, providing a unified basis for platform performance verification.

[0042] The target vehicle document refers to the automotive technical document used for extracting core fields, such as a vehicle crash test report, which may include, but is not limited to, core fields such as vehicle model, report number, test date, and vehicle dimensions.

[0043] In some cases, embodiments of this application can use a pre-defined hierarchical quantitative evaluation system to first quantitatively evaluate the field extraction accuracy and single-question answering accuracy of the automotive Q&A platform. Then, by comparing the obtained field extraction accuracy and single-question answering accuracy with the pre-defined indicator conditions, it can be determined whether the platform performance meets the standards. Under the premise that both indicators meet the corresponding requirements, the automotive Q&A platform can then be used to extract the core field information of the target automotive document.

[0044] Optionally, in one embodiment of this application, constructing an automotive document and question-and-answer dataset includes: generating a simulated document based on a preset automotive crash test report template; generating questions with multiple expression forms and their corresponding field labels; and constructing an automotive document and question-and-answer dataset based on the simulated document and the questions and their corresponding field labels.

[0045] Specifically, the preset car crash test report template refers to a standardized template used to generate a realistic document by randomly filling in simulated data for 19 core fields such as vehicle model, test parameters, report number, and test date. It may include, but is not limited to, a cover page, vehicle details page, test conditions page, powertrain page, and random supplementary pages.

[0046] In actual implementation, the embodiments of this application can construct a highly realistic automotive document and question-and-answer dataset, and the specific steps are as follows.

[0047] (1) Realistic document generation: In this embodiment, the preset car crash test report template is first called, and a large number of highly realistic PDF documents are generated by randomly filling in the simulated data of 19 core fields such as car model, test parameters, report number, and test date.

[0048] (2) Diverse Question-Answer Pair Labeling: Based on the above 19 core fields, this application embodiment generates questions with multiple expressions and their corresponding field labels through a preset rule engine, forming a "question-field label" mapping pair. For example, for the "test date" field, questions such as "When was the test conducted? What is the report date?" can be generated.

[0049] (3) Automated cleaning and verification logic: During the construction of the dataset, the embodiments of this application have built-in automated cleaning and verification module. This module performs noise removal processing on the original extracted field values ​​(removing redundant spaces, illegal characters and specific prefixes) through regular expression matching and logical filtering algorithms, and performs consistency verification on the field format to ensure that the training data input to the model is accurate and standardized.

[0050] It should be noted that, regarding the dataset construction method, this embodiment of the application uses an automated question-and-answer pair generator to generate diverse question-and-answer pairs that combine professional terminology and colloquial expressions, based on the 19 core fields of the simulated document and focusing on multiple dimensions such as the existence, value, and unit of each core field.

[0051] For dataset quality control, this application embodiment performs systematic quality checks on the generated question-answer pairs, verifies the consistency between the question-answer pairs and the original field information, standardizes automotive domain terminology to eliminate semantic ambiguity, and finally outputs a high-quality automotive document and question-answer dataset suitable for model training.

[0052] like Figure 2 As shown, this application describes how to construct a highly realistic automotive documentation and question-and-answer dataset using a specific embodiment.

[0053] S201, Receive the generated parameters and select the vehicle power type.

[0054] In this embodiment, the user inputs generation parameters and generates relevant parameters according to the vehicle's power type: when the vehicle is a pure electric vehicle, only the motor model is generated; when the vehicle is a hybrid vehicle, the engine model, motor model, and fuel tank capacity are generated; when the vehicle is a gasoline vehicle, the engine model and fuel tank capacity are generated.

[0055] S202 utilizes the business data generation engine to generate core fields.

[0056] In this embodiment, core fields are generated based on relevant parameters, including basic information (vehicle model, report number, test date, etc.), size and mass information (length, width, height, wheelbase, weight, etc.), test condition information (speed, humidity, temperature, etc.), and unit and scoring information (production unit, test unit, test score, etc.).

[0057] S203 utilizes a dynamic PDF builder to generate realistic documents.

[0058] In this embodiment, the core fields are fitted into a preset car crash test report template to generate a realistic document. The preset car crash test report template includes a cover page (displaying the report number, test date, etc.), a vehicle details page (displaying length, width, height, wheelbase, weight, model, etc.), a test conditions page (displaying speed, humidity, temperature, etc.), a powertrain page (displaying engine, motor, fuel tank parameters, etc.), and random supplementary pages (used to randomly generate additional pages to simulate the diversity of real reports).

[0059] S204 uses a preset rule engine to generate questions with multiple forms of expression and their corresponding field labels.

[0060] In this embodiment, the core fields of the simulated document can be read, and a preset rule engine can be used to generate questions with multiple forms of expression and their corresponding field tags.

[0061] S205, outputs a dataset of automotive documents and question-and-answer questions.

[0062] In this embodiment, a structured dataset containing "document path, question, answer, and field labels" can be output.

[0063] This application embodiment provides the model with rich, automotive-domain-appropriate learning samples by constructing an automotive document and question-answer dataset, which is the foundation for improving the model's domain adaptability and semantic understanding accuracy. In addition, it can introduce data augmentation strategies, such as adding slang and colloquialisms, to further enhance the model's generalization ability.

[0064] Optionally, in one embodiment of this application, generating a lightweight Transformer semantic understanding model using automotive documents and question-answering datasets includes: acquiring automotive domain technical document corpus; using the automotive domain technical document corpus to further pre-train a preset BERT-base pre-trained model to update the weights of the preset BERT-base pre-trained model to obtain a pre-trained optimized model; and using the automotive documents and question-answering datasets to fine-tune the pre-trained optimized model to generate a lightweight Transformer semantic understanding model.

[0065] Specifically, automotive technical documentation corpus refers to dedicated corpus used for further pre-training of the pre-set BERT-base pre-trained model, which may include, but is not limited to, professional texts, technical specifications, and report excerpts related to powertrain systems and other related fields.

[0066] The pre-trained BERT-base model refers to the BERT-base-chinese model, which includes a multi-layer Transformer encoder and a fully connected layer. By pre-training on massive amounts of general text, it has preliminary text feature extraction and semantic matching capabilities. After further pre-training on automotive technical document corpora, it can be adapted to automotive semantic understanding tasks.

[0067] Pre-trained optimization model refers to the pre-training of the BERT-base pre-trained model through automotive technical document corpus, updating the model weights, learning automotive professional terminology and expression paradigms, and providing an optimized base model for subsequent fine-tuning based on automotive documents and question-answer datasets.

[0068] In actual implementation, the embodiments of this application can generate a lightweight Transformer semantic understanding model through model training.

[0069] First, this application describes the architecture of a lightweight Transformer semantic understanding model, which mainly includes three functional modules.

[0070] (1) Document semantic encoding module: The BERT-base-chinese model is used as the backbone network. Based on the original weights of the pre-trained model, a fully connected classification layer designed for 19 core fields is built on top of its output layer to map semantic vectors into specific field label probabilities.

[0071] (2) Domain Adaptation Fine-tuning Module: By introducing the GELU activation function and Dropout regularization in the fully connected layer, the feature representation capability of the model in the automotive domain is enhanced.

[0072] (3) Field classification generation module: Based on the semantic features optimized by the domain adaptation fine-tuning module, the corresponding document field categories and confidence scores are output through the multilayer perceptron classifier.

[0073] Next, the embodiments of this application describe the model training process.

[0074] (1) Domain-Adaptive Pre-training: In this embodiment, the BERT-base model is further pre-trained using automotive domain technical document corpus. The pre-training task is a masked language model task, enabling the model to learn automotive terminology and expression paradigms. The pre-training process adopts a linear learning rate decay strategy with warm-up, and the weight update formula can be, but is not limited to, the following: , in, For the first Step weights For the first Step weights For the loss function of the masked language model in the weights The loss value below, For gradient operators, For the first The dynamic learning rate of the step.

[0075] (2) Downstream Task Fine-tuning: In this embodiment, the model is fine-tuned using the constructed automotive document and question-answering dataset. A multi-task learning strategy is adopted, organizing the samples in the dataset into "document-question-field" triples, and labeling each field with a semantic category label. This application proposes a hybrid loss function based on BERT, the expression of which may be, but is not limited to: , in, For the total loss, Classification loss for the field, For semantic matching loss, For regularization loss, Classification loss for fields The weighting coefficients, semantic matching loss The weighting coefficients, For regularization loss The weighting coefficients.

[0076] It should be noted that the field classification loss Cross-entropy loss is used to measure the difference between the model's predicted field categories based on the [CLS] representation and the true labels; semantic matching loss is also employed. Cosine similarity loss is used to measure the semantic relevance between document content and user questions; regularization loss is also employed. L2 regularization is used to prevent the model from overfitting.

[0077] (3) Optimization strategy: The embodiments of this application adopt the AdamW optimizer, combined with linear learning rate decay with warm-up, gradient pruning and early stopping mechanism, and evaluate the model performance through five-fold cross-validation to ensure the generalization ability and stability of the model.

[0078] For example, such as Figure 3 As shown, a specific embodiment of this application illustrates how a lightweight Transformer semantic understanding model is generated through model training.

[0079] S301, Determine the automotive documentation and question-answering dataset and the preset BERT-base-chinese model.

[0080] S302, use automotive technical documentation corpus to continue pre-training the preset BERT-base pre-trained model to update the weights of the preset BERT-base pre-trained model in order to obtain the pre-trained optimized model.

[0081] S303 utilizes knowledge distillation to compress and pre-train the optimized model to obtain a pre-trained compressed model.

[0082] S304 utilizes automotive documents and question-answering datasets to perform multi-task joint fine-tuning on the pre-trained compressed model to obtain the pre-trained fine-tuned model.

[0083] S305 performs five-fold cross-validation on the pre-trained fine-tuned model to obtain the five-fold cross-validation results.

[0084] In this embodiment of the application, the average accuracy and standard deviation of the five-fold cross-validation can be used as evaluation metrics. The formula for calculating the average accuracy can be, but is not limited to, the following: , in, The average accuracy of the five-fold cross-validation. For the first Accuracy on the validation set; The formula for calculating the standard deviation of the mean accuracy can be, but is not limited to, as follows: , in, The standard deviation of the accuracy of the five-fold cross-validation is given.

[0085] S306 evaluates model performance based on the results of five-fold cross-validation and executes intelligent early stop.

[0086] In this embodiment, the model performance is evaluated based on the results of the five-fold cross-validation. If the model performance evaluation fails, the process returns to S304 and multi-task joint fine-tuning is performed again until the model performance evaluation passes, thus determining the optimal model.

[0087] S307 determined the best model to be the lightweight Transformer semantic understanding model.

[0088] The embodiments of this application adopt a lightweight model architecture design, which effectively reduces the model complexity. In addition, through domain-adaptive pre-training and targeted multi-task fine-tuning, the model fully learns the semantic features and question-field mapping relationship of the automotive domain, thereby achieving high-accuracy intent recognition and field classification.

[0089] Optionally, in one embodiment of this application, a car question-and-answer platform is constructed, comprising: a semantic understanding module, a document retrieval module, an answer generation module, and an interaction optimization module. The semantic understanding module receives user natural language questions and uses a lightweight Transformer semantic understanding model to output the field types and confidence levels corresponding to the user's natural language questions. The document retrieval module performs document retrieval based on the field types and confidence levels to obtain structured document information. The answer generation module generates structured answer content based on the structured document information. The interaction optimization module interacts with the user and stores the answer content. Based on the semantic understanding module, document retrieval module, answer generation module, and interaction optimization module, a car question-and-answer platform is established.

[0090] In practical implementation, the embodiments of this application can construct a high-precision OCR-based automotive question-answering platform. The automotive question-answering platform includes a semantic understanding module, a document retrieval module, an answer generation module, and an interaction optimization module, with the functions of each module as follows.

[0091] (1) Semantic understanding module: Receives user natural language questions, uses the pre-trained lightweight Transformer semantic understanding model to perform semantic parsing and intent recognition on the user natural language questions, and outputs the field type and confidence level corresponding to the question.

[0092] (2) Document retrieval module: includes an OCR preprocessing submodule, a table structure understanding submodule, and a cross-page association submodule. Among them, the OCR preprocessing submodule adopts a high-precision OCR engine to achieve nearly 100% character recognition for mixed Chinese and English text, complex tables, and special symbols; the table structure understanding submodule uses attention mechanism-based table detection and cell alignment technology to accurately parse complex structures such as merged cells and cross-page tables; the cross-page association submodule achieves automatic association and fusion of related information scattered across multiple pages through document structure analysis and semantic continuity detection.

[0093] (3) Answer generation module: Based on the structured document information retrieved by the document retrieval module, combined with the semantics of the user's natural language question, the module generates structured answer content and performs confidence calculation and anomaly handling.

[0094] (4) Interaction Optimization Module: Supports quick answers to preset questions, complete question-and-answer history (including timestamps, questions, identification fields, confidence levels, answers, etc.), and local storage and export of history (TXT format) for easy archiving and auditing. As shown in Figure 5, the effect of the car question-and-answer platform is demonstrated. Figure 6 As shown, this displays the historical results of questions and answers on the automotive Q&A platform.

[0095] For example, such as Figure 4 As shown below, a specific embodiment of this application will be used to illustrate how a high-precision OCR automotive question-and-answer platform can be used for question answering.

[0096] S401, Input user's natural language question.

[0097] Among them, the user's natural language questions cover both technical terms and colloquial questions related to automotive documents.

[0098] S402, the semantic understanding module is used to perform semantic parsing and intent recognition on the user's natural language question in order to obtain the field type and confidence level corresponding to the question.

[0099] In this embodiment, the semantic understanding module is used to call the pre-trained lightweight Transformer semantic understanding model to perform semantic parsing and intent recognition on the user's natural language question, and output the field type and confidence level corresponding to the question.

[0100] S403 utilizes the document retrieval module to perform document retrieval based on field type and confidence level to obtain structured document information.

[0101] In this embodiment, the document retrieval module performs a search in a preset automotive document database based on the field types and confidence levels output by the semantic understanding module. The retrieval process combines OCR preprocessing, table structure understanding, and cross-page association to extract structured document information.

[0102] S404, Use the answer generation module to generate structured answer content based on structured document information.

[0103] In this embodiment, the answer generation module generates structured answer content based on the structured document information output by the document retrieval module and the semantics of the user's natural language question, while simultaneously performing confidence calculation and anomaly handling.

[0104] S405, output structured response content.

[0105] In this embodiment of the application, after outputting structured answer content, the interaction optimization module can be used to store the history of this question and answer (including timestamp, question, identification field, confidence level, answer, etc.).

[0106] This application embodiment integrates a high-precision OCR engine through a document retrieval module, effectively ensuring the accurate extraction of original information from complex automotive documents. In addition, through a semantic understanding module, it can combine a pre-trained lightweight Transformer semantic understanding model to perform semantic parsing of users' natural language questions. Furthermore, through an answer generation module, it effectively ensures that the answers are both accurate and comprehensive. And through the interaction optimization function of the interaction optimization module, it improves ease of use and practicality.

[0107] Furthermore, embodiments of this application can perform system performance evaluation on the constructed automotive question-and-answer platform.

[0108] Specifically, this application embodiment constructs a pre-defined hierarchical quantitative evaluation system, with field extraction accuracy and question-answering accuracy as the core indicators.

[0109] Firstly, at the field extraction evaluation level, this application's embodiments introduce field extraction accuracy.

[0110] Optionally, in one embodiment of this application, the formula for calculating the field extraction accuracy may be, but is not limited to, the following: , in, Extract accuracy for the field. The number of fields correctly extracted by the model. This represents the total number of fields in all documents within the test set.

[0111] In actual implementation, this application embodiment calculates the field extraction accuracy by measuring the ratio of the number of fields correctly extracted by the model in the test set to the total number of fields in all documents.

[0112] The embodiments of this application, by calculating the field extraction accuracy, can directly reflect the accuracy of the model in locating and extracting target fields in automotive documents, and objectively measure the consistency of the model's parsing and classification of text semantics, table structure and field boundaries.

[0113] Secondly, in terms of question-and-answer performance evaluation, the embodiments of this application construct a multi-level evaluation index system based on the accuracy of question-and-answer.

[0114] In one embodiment of this application, the formula for calculating the accuracy of a single question answer may be, but is not limited to, the following: , in, For the accuracy of answering a single question, The number of questions the model answered correctly. This represents the total number of issues in the test set.

[0115] In actual implementation, this application embodiment calculates the question-answering accuracy by calculating the ratio of the number of questions correctly answered by the model in the test set to the total number of questions.

[0116] This application embodiment can directly quantify the overall correctness of the model's response to users' natural language questions by calculating the question-answering accuracy rate, and comprehensively reflect the overall execution accuracy of the model in understanding user intent, locating fields, and generating answers.

[0117] Furthermore, the formula for calculating field type accuracy can be, but is not limited to, as follows: , in, For field type accuracy, The total number of field types, Index for field type, For the first The number of correct answers in the category field. For the first Total number of issues for class fields.

[0118] Confidence distribution analysis assesses the degree of model calibration by calculating the correlation between the confidence level of the model output and the true accuracy. The calculation formula can be, but is not limited to, the following: , in, To calibrate the model confidence level, Number of buckets for confidence level. Bucketed index for confidence level For the first The actual accuracy of the bucket For the first The average confidence level of the bucket.

[0119] It should be noted that, according to testing, the built automotive question-answering platform can accurately extract all the necessary core field information from automotive documents, and the success rate of parsing complex tables and document structures exceeds 98%. The model achieves an average accuracy of over 95% in question-answering tasks, demonstrating extremely strong classification consistency and domain adaptability.

[0120] like Figure 7 As shown below, the semantic extraction method for automotive documents proposed in this application is illustrated with a specific embodiment.

[0121] S701, Construct a car documentation and question-and-answer dataset.

[0122] In this embodiment, a simulated document is generated based on a preset car crash test report template, and questions in various forms and their corresponding field tags are generated. The car document and question-and-answer dataset is constructed by combining the document path, questions, answers and field tags.

[0123] S702 generates a lightweight Transformer semantic understanding model.

[0124] In this embodiment, technical document corpus in the automotive field is obtained, and the corpus is used to further pre-train a preset BERT-base pre-trained model to obtain a pre-trained optimized model. Then, the pre-trained optimized model is fine-tuned using automotive documents and question-answering datasets to generate a lightweight Transformer semantic understanding model.

[0125] S703, establish a car Q&A platform.

[0126] In this embodiment, a semantic understanding module is constructed using a lightweight Transformer semantic understanding model. Combined with a document retrieval module, an answer generation module, and an interaction optimization module, an automotive question-and-answer platform is established. The semantic understanding module receives users' natural language questions and outputs the field types and confidence levels corresponding to the users' natural language questions using the lightweight Transformer semantic understanding model. The document retrieval module performs document retrieval based on the field types and confidence levels to obtain structured document information. The answer generation module generates structured answer content based on the structured document information. The interaction optimization module interacts with users and stores the answer content.

[0127] S704 uses a pre-defined, tiered quantitative evaluation system to assess the field extraction accuracy and question-answering accuracy of an automotive question-and-answer platform.

[0128] In this embodiment, a pre-defined hierarchical quantitative evaluation system can be used to first quantitatively evaluate the field extraction accuracy and single-question answer accuracy of the automotive Q&A platform.

[0129] S705, when both the field extraction accuracy and question-answering accuracy meet the corresponding indicator conditions, uses an automotive question-answering platform to extract core field information from the target automotive document.

[0130] In this embodiment, the accuracy of field extraction and the accuracy of answering a single question are compared with preset indicator conditions to determine whether the platform performance meets the standards. Under the premise that both indicators meet the corresponding requirements, the core field information of the target car document is extracted by the car question and answer platform.

[0131] The automotive document semantic extraction method proposed in this application constructs an automotive document and question-and-answer dataset and generates a lightweight Transformer semantic understanding model. This model enables the vehicle to understand automotive terminology and complex query intents, significantly reducing model complexity and computational resource consumption while maintaining high accuracy and improving inference speed. By utilizing an automotive question-and-answer platform to extract core field information from target automotive documents, it supports functions such as natural language questioning, quick answers to preset questions, question-and-answer history recording and export, enhancing practicality and achieving one-stop, implementable intelligent processing of automotive documents. This effectively supports R&D, testing, and after-sales processes in the automotive industry. Therefore, it solves the problem that related technologies rely on historical question-and-answer text, making it difficult to cover the full-scenario semantic needs of automotive documents, resulting in low data annotation and utilization efficiency.

[0132] Next, the semantic extraction apparatus for automotive documents according to embodiments of this application is described with reference to the accompanying drawings.

[0133] Figure 8 This is a block diagram of a semantic extraction device for automotive documents provided according to an embodiment of this application.

[0134] like Figure 8 As shown, the automotive document semantic extraction device 80 includes: a first construction module 100, a second construction module 200, and an extraction module 300.

[0135] The first building module 100 is used to build a car document and question-and-answer dataset.

[0136] The second building module 200 is used to generate a lightweight Transformer semantic understanding model using automotive documents and question-answering datasets, in order to build an automotive question-answering platform based on the lightweight Transformer semantic understanding model.

[0137] The extraction module 300 is used to evaluate the field extraction accuracy and question-answering accuracy of the automotive question-answering platform based on a preset hierarchical quantitative evaluation system. When both the field extraction accuracy and question-answering accuracy meet the corresponding indicator conditions, the core field information of the target automotive document can be extracted using the automotive question-answering platform.

[0138] Optionally, in one embodiment of this application, the first building module 100 includes: a first generation unit, a second generation unit, and a first building unit.

[0139] The first generation unit is used to generate a simulated document based on a preset car crash test report template.

[0140] The second generation unit is used to generate questions with multiple forms of expression and their corresponding field labels.

[0141] The first building unit is used to construct a car document and question-and-answer dataset based on simulated documents and questions and their corresponding field labels.

[0142] Optionally, in one embodiment of this application, the second construction module 200 includes: an acquisition unit, a first training unit, and a second training unit.

[0143] The acquisition unit is used to acquire technical document corpus in the automotive field.

[0144] The first training unit is used to further pre-train the preset BERT-base pre-trained model using automotive technical documentation corpus, in order to update the weights of the preset BERT-base pre-trained model and obtain a pre-trained optimized model.

[0145] The second training unit is used to fine-tune the pre-trained optimization model using automotive documents and question-answering datasets to generate a lightweight Transformer semantic understanding model.

[0146] Optionally, in one embodiment of this application, the second building module 200 includes a second building unit and a third building unit.

[0147] The second building unit is used to construct a semantic understanding module, a document retrieval module, an answer generation module, and an interaction optimization module. The semantic understanding module receives user natural language questions and uses a lightweight Transformer semantic understanding model to output the field types and confidence levels corresponding to the user's natural language questions. The document retrieval module performs document retrieval based on field types and confidence levels to obtain structured document information. The answer generation module generates structured answer content based on the structured document information. The interaction optimization module interacts with the user and stores the answer content.

[0148] The third building block is used to establish an automotive question-and-answer platform based on a semantic understanding module, a document retrieval module, an answer generation module, and an interaction optimization module.

[0149] Optionally, in one embodiment of this application, the formula for calculating the field extraction accuracy may be, but is not limited to, the following: , in, Extract accuracy for the field. The number of fields correctly extracted by the model. This represents the total number of fields in all documents within the test set.

[0150] Optionally, in one embodiment of this application, the formula for calculating the question-answering accuracy may be, but is not limited to, the following: , in, For question-and-answer accuracy, The number of questions the model answered correctly. This represents the total number of issues in the test set.

[0151] It should be noted that the foregoing explanation of the embodiment of the automotive document semantic extraction method also applies to the automotive document semantic extraction device of this embodiment, and will not be repeated here.

[0152] The automotive document semantic extraction device proposed in this application constructs an automotive document and question-and-answer dataset and generates a lightweight Transformer semantic understanding model. This model enables the vehicle to understand automotive terminology and complex query intents, significantly reducing model complexity and computational resource consumption while maintaining high accuracy and improving inference speed. By utilizing an automotive question-and-answer platform to extract core field information from target automotive documents, it supports functions such as natural language questioning, rapid answering of preset questions, question-and-answer history recording and export, enhancing practicality. This achieves one-stop, implementable intelligent processing of automotive documents, effectively supporting R&D, testing, and after-sales processes in the automotive industry. Therefore, it solves the problem that related technologies rely on historical question-and-answer text, making it difficult to cover the full-scenario semantic needs of automotive documents, resulting in low data annotation and utilization efficiency.

[0153] Figure 9 This is a schematic diagram of a vehicle structure according to an embodiment of this application. The vehicle may include: The memory 901, the processor 902, and the computer program stored on the memory 901 and capable of running on the processor 902.

[0154] When processor 902 executes the program, it implements the automotive document semantic extraction method provided in the above embodiments.

[0155] Furthermore, automobiles also include: Communication interface 903 is used for communication between memory 901 and processor 902.

[0156] The memory 901 is used to store computer programs that can run on the processor 902.

[0157] The memory 901 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0158] If the memory 901, processor 902, and communication interface 903 are implemented independently, then the communication interface 903, memory 901, and processor 902 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0159] Optionally, in a specific implementation, if the memory 901, processor 902, and communication interface 903 are integrated on a single chip, then the memory 901, processor 902, and communication interface 903 can communicate with each other through an internal interface.

[0160] The processor 902 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0161] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described automotive document semantic extraction method.

[0162] This application also provides a computer program product, including a computer program that, when executed, implements the above-described automotive document semantic extraction method.

[0163] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0164] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0165] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0166] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0167] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0168] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0169] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0170] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for semantic extraction from automotive documents, characterized in that, Includes the following steps: Constructing a dataset of automotive documentation and question-and-answer materials; A lightweight Transformer semantic understanding model is generated using the aforementioned automotive documents and question-answering dataset, and an automotive question-answering platform is constructed based on the lightweight Transformer semantic understanding model. Based on a pre-defined hierarchical quantitative evaluation system, the field extraction accuracy and question-answering accuracy of the automotive question-answering platform are evaluated. Under the condition that both the field extraction accuracy and the question-answering accuracy meet the corresponding indicator conditions, the core field information of the target automotive document is extracted using the automotive question-answering platform.

2. The method according to claim 1, characterized in that, The construction of the automotive document and question-answer dataset includes: Generate a realistic document based on a preset car crash test report template; Generate questions with multiple expression formats and their corresponding field labels; Based on the simulated document and the question and its corresponding field labels, the car document and question-and-answer dataset is constructed.

3. The method according to claim 1, characterized in that, The process of generating a lightweight Transformer semantic understanding model using the aforementioned automotive documents and question-answering dataset includes: Obtain technical documentation corpus in the automotive field; The preset BERT-base pre-trained model is further pre-trained using the automotive technical documentation corpus to update the weights of the preset BERT-base pre-trained model, thereby obtaining a pre-trained optimized model. The pre-trained optimization model is fine-tuned using the automotive documents and question-answering dataset to generate the lightweight Transformer semantic understanding model.

4. The method according to claim 1, characterized in that, The construction of the automotive Q&A platform includes: The system comprises a semantic understanding module, a document retrieval module, an answer generation module, and an interaction optimization module. The semantic understanding module receives user natural language questions and uses the lightweight Transformer semantic understanding model to output the field types and confidence levels corresponding to the user's natural language questions. The document retrieval module performs document retrieval based on the field types and confidence levels to obtain structured document information. The answer generation module generates structured answer content based on the structured document information. The interaction optimization module interacts with the user and stores the answer content. The automotive Q&A platform is established based on the semantic understanding module, the document retrieval module, the answer generation module, and the interaction optimization module.

5. The method according to claim 1, characterized in that, The formula for calculating the accuracy of field extraction is as follows: ， in, Extract the accuracy for the field. The number of fields correctly extracted by the model. This represents the total number of fields in all documents within the test set.

6. The method according to claim 1, characterized in that, The formula for calculating the accuracy of the question-and-answer session is as follows: ， in, The accuracy rate of the question and answer is... The number of questions the model answered correctly. This represents the total number of issues in the test set.

7. A semantic extraction device for automotive documents, characterized in that, include: The first building module is used to construct the automotive documentation and question-and-answer dataset; The second building module is used to generate a lightweight Transformer semantic understanding model using the car documents and question-and-answer dataset, so as to build a car question-and-answer platform based on the lightweight Transformer semantic understanding model; The extraction module is used to evaluate the field extraction accuracy and question-answering accuracy of the automotive question-answering platform based on a preset hierarchical quantitative evaluation system, so as to extract the core field information from the target automotive document using the automotive question-answering platform when both the field extraction accuracy and the question-answering accuracy meet the corresponding indicator conditions.

8. A car, characterized in that, include: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, the processor executing the program to implement the automotive document semantic extraction method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the automotive document semantic extraction method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement the automotive document semantic extraction method as described in any one of claims 1-6.

Images