A method and apparatus for classifying standard automotive text.

Abstract

This application relates to the field of electronic digital processing technology, specifically to a method and apparatus for classifying automotive standard text. The method includes acquiring a deep learning-based model to be trained for classifying automotive text; training the model to be trained based on training samples to obtain a classification model; wherein the training samples include input text about automotive standards and labels, the labels including parent category labels and sub-category labels; during training, iterating the parameters in the classification model by minimizing a loss function; the loss function is constructed based on parent category bias, sub-category bias, and category hierarchy bias; inputting the user-inputted text about automotive standards to be classified into the classification model to obtain the parent category and sub-category to which the text to be classified belongs. This application achieves fast and accurate classification of automotive standard text using a neural network model.

Description

Technical Field

[0001] This application relates to the field of electronic digital processing technology, and more specifically, to a method and apparatus for classifying automotive standard text. Background Technology

[0002] Currently, companies in the automotive industry (such as OEMs, parts suppliers, and testing and certification bodies) frequently need to consult and apply standards in their daily R&D, production, certification, and compliance reviews. However, the existing standards management and application methods have the following significant drawbacks:

[0003] 1. Complex Standards System and Difficult Classification: The number of automotive standards is enormous, and their classification system is multidimensional and overlapping. A single standard may involve multiple technical fields. For example, a standard for "car seats" may belong to the category of "body and accessories," and also be related to "passive safety" and "in-vehicle environment (material emissions)." This complex cross-relationship makes it easy for even professionals to make oversights and misjudgments when manually classifying and searching, resulting in low efficiency.

[0004] 2. Inefficient reliance on manual retrieval and identification: Currently, users primarily search for standards through keyword searches in standard databases or by browsing manually pre-defined tree-like classification directories. When users input standard text and want to quickly determine its precise category, existing technology cannot provide a direct and intelligent solution. Users need to read the entire text themselves, make judgments based on their own knowledge, or repeatedly try different keywords, a cumbersome and time-consuming process.

[0005] 3. Insufficient understanding of natural language, hindering direct categorization guidance: When a user inputs a specific question about standards, such as "What are the standards for electric vehicle battery collision safety?" or "Which standard specifies the limit for formaldehyde in vehicle interiors?", existing systems typically only return a list of potentially relevant standards. They fail to directly understand the user's deeper intent and provide a clear, consistent categorization answer. Users must manually sift through massive amounts of search results, unable to obtain an instant, question-and-answer categorization guidance service.

[0006] In view of the above, this application is hereby submitted. Summary of the Invention

[0007] The purpose of this application is to provide a method and device for classifying automotive standard text, so as to quickly and accurately classify automotive standard text through a neural network model.

[0008] To achieve the above objectives, this application adopts the following technical solution:

[0009] Firstly, this application provides a method for classifying standard automotive texts, including:

[0010] Obtain a deep learning-based model to be trained for classifying car text;

[0011] The model to be trained is trained based on the training samples to obtain a classification model; wherein, the training samples include input text and labels about car standards, and the labels include parent category labels and sub-category labels;

[0012] During training, the parameters in the classification model are iteratively minimized by a loss function; the loss function is constructed based on parent class bias, child class bias, and class hierarchy bias.

[0013] The user-inputted text about automotive standards is fed into the classification model to obtain the parent and sub-categories of the text.

[0014] The classification model includes:

[0015] The feature vector extraction layer is used to extract features from the user-input text about automotive standards to obtain a sequence of feature vectors;

[0016] The feature enhancement layer is used to calculate the word vector similarity between the feature vector sequence and automotive domain knowledge, and to enhance the features of feature vectors that exceed the similarity threshold.

[0017] The parent category feature branch is used to input the enhanced feature vector sequence into the global attention pooling layer to obtain the parent category feature vector sequence.

[0018] The sub-category feature branch is used to input the enhanced feature vector sequence into the gated recurrent unit to obtain the sub-category feature vector sequence.

[0019] The parent category classification layer is used to classify the parent category feature vector sequence to obtain the predicted parent category;

[0020] The sub-category classification layer is used to classify the sub-category feature vector sequence to obtain the predicted sub-category.

[0021] Secondly, this application provides an electronic device, comprising:

[0022] At least one processor, and a memory communicatively connected to at least one of the processors;

[0023] The memory stores instructions that can be executed by at least one of the processors, which enable the at least one processor to perform the above-described classification method for automotive standard text.

[0024] Compared with the prior art, the beneficial effects of this application are as follows:

[0025] This application employs a deep learning-based classification model to categorize text into parent and child categories. Deep learning, as an advanced artificial intelligence technology, has achieved significant results in various fields such as image recognition, natural language processing, and medical diagnosis. This application, by utilizing a deep learning-based classification model, can automatically extract and learn high-dimensional features of text, thereby achieving accurate text classification. Therefore, applying deep learning to automotive standard classification can not only reduce manual intervention and lower costs, but also achieve more efficient and accurate classification in diverse and complex automotive data, effectively improving classification efficiency and accuracy, and supporting the intelligentization and standardization of the automotive manufacturing industry. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a method for classifying standard automotive text provided in an embodiment of this application;

[0028] Figure 2 This is a schematic diagram of the structure of the classification model provided in the embodiments of this application;

[0029] Figure 3 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0030] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0031] The present application will be further described in detail below with reference to the embodiments.

[0032] Figure 1This is a flowchart illustrating a method for classifying automotive standard text according to an embodiment of this application. This method can be executed by a computer program and integrated into an electronic device. This application is applicable to scenarios where user-inputted text related to automotive standards undergoes dual classification using parent and subcategories.

[0033] like Figure 1 As shown, this embodiment provides a method for classifying standard automotive text, including the following steps:

[0034] S110. Obtain a deep learning-based model to be trained for classifying car text.

[0035] This embodiment does not limit the specific structure of the model to be trained. It can be a neural network model (such as the BERT model based on the Transformer architecture or its variants) or other deep learning-based models. The parameters in the model to be trained represent the initial training state. It has completed initial training on a massive general corpus and has gained a basic understanding of natural language, including the representation of vocabulary, grammar, and shallow semantics. It needs to continue iterating the parameters to enable the model to have the function of classifying standard automotive text.

[0036] S120. Train the model to be trained based on the training samples to obtain a classification model; wherein, the training samples include input text and labels about car standards, and the labels include parent category labels and sub-category labels.

[0037] Automotive standards need to be pre-defined into parent and subcategories, with a hierarchical relationship between them. Parent categories include automotive standards, non-automotive standards, and general standards. This embodiment only requires automotive standards to have subcategories. For example:

[0038] Parent category: Automotive safety standards; Subcategories include: a) Passive safety: involving vehicle collision safety, such as performance testing and regulations for body structure, seats, and airbag systems. b) Active safety: involving accident prevention technologies, such as anti-lock braking systems (ABS), electronic stability control (ESC), and automatic emergency braking (AEB).

[0039] Parent category: Emissions and environmental standards; Subcategories include: a) Exhaust emissions: including limits for harmful substances such as carbon monoxide, nitrogen oxides, and particulate matter in exhaust gases. b) Fuel evaporation control: regulating the sealing performance of fuel tanks and fuel systems to reduce volatile organic compound emissions.

[0040] Parent category: Automotive materials and parts standards; subcategories include: a) Body materials: Specifications for the strength and corrosion resistance of body materials (such as steel, aluminum alloys, etc.). b) Tires: Includes tire dimensions, wear resistance, anti-aging properties, etc. c) Glass: Standards related to the light transmittance and strength of windows and windshields.

[0041] Parent category: Electrical and electronic equipment standards; subcategories include: a) Electromagnetic compatibility: ensuring that in-vehicle electronic equipment is not affected by electromagnetic interference. b) Lighting systems: requirements for the brightness, angle, and chromaticity of headlights, turn signals, and brake lights. c) On-board diagnostics (OBD): standardizing OBD interfaces and data formats to facilitate vehicle self-diagnosis.

[0042] Parent category: Vehicle performance standards; Subcategories include: a) Fuel economy: Fuel consumption requirements for different vehicle models. b) Power performance: Performance indicators such as acceleration and braking. c) NVH: Control requirements for noise, vibration, and acoustic roughness.

[0043] Parent category: Standards for electric vehicles and new energy vehicles; subcategories include: a) Battery safety: including thermal management, fire prevention, and explosion protection of battery systems. b) Charging interface: ensuring the universality and compatibility of charging interfaces. c) Range and energy consumption: standards for the range and energy consumption of electric vehicles.

[0044] Parent category: In-vehicle environmental standards; subcategories include: a) In-vehicle air quality: regulations on the content of harmful gases such as volatile organic compounds (VOCs) in the vehicle interior. b) Interior materials: regulations on the flame retardancy and toxicity of interior materials.

[0045] Training samples are collected in advance. For example, historical user input of automotive standard-related text (i.e., input text) is collected through the human-computer interface of the automotive standard retrieval system, such as "standards related to active safety," "what standard requires a low-speed warning sound during low-speed vehicle operation," and "which standard specifies the control requirements for noise, vibration, and acoustic roughness." Simultaneously, the standard parent category labels (i.e., the actual parent category) and sub-category labels (i.e., the actual sub-categories) corresponding to the input text are collected. The parent and sub-category labels can be obtained through retrieval or determined by experts.

[0046] S130. During training, the parameters in the classification model are iterated by minimizing the loss function; the loss function is constructed based on parent class bias, child class bias, and class hierarchy bias.

[0047] The training process is essentially an iterative optimization process. The model to be trained receives input text and, through its complex internal parameter iterations, outputs the predicted probability distributions for the parent and child categories.

[0048] Parent category bias quantifies the difference between the parent category predicted by the model (called the predicted parent category) and the parent category label. Subcategory bias quantifies the difference between the subcategory predicted by the model (called the predicted subcategory) and the subcategory label. Category hierarchy bias measures the logical reasonableness of the combination of parent and subcategories predicted by the model. For example, if the model predicts "environmental protection" as the parent category but predicts "steering system" as the subcategory, it will output a large hierarchy bias, used to constrain the correlation between the predicted subcategories and the predicted parent categories.

[0049] The parent class bias, child class bias, and class hierarchy bias are combined with certain weights to form the loss function. Stochastic gradient descent or its variant Adam is used to calculate the gradient of the loss function with respect to the model parameters, and then these parameters are fine-tuned along the direction of gradient descent. After several iterations, the model's internal parameters are gradually adjusted to a state that minimizes the loss function. At this point, the model training is complete, and it becomes a usable "classification model".

[0050] S140. Input the text to be classified about automobile standards input by the user into the classification model to obtain the parent category and sub-category to which the text to be classified belongs.

[0051] For example, the text to be classified regarding automotive standards is "When a vehicle brakes at an initial speed of 100 km / h on a dry asphalt road, the braking distance should not exceed 42 meters." This text is input into a pre-trained classification model. Based on the model's internal forward propagation mechanism, after feature extraction and classification calculations, the model ultimately outputs a predicted parent category (e.g., "vehicle safety standards") and a predicted subcategory (e.g., "braking performance"), thus obtaining the parent and subcategories stated in the text to be classified.

[0052] To achieve accurate text classification, embodiments of this application provide a novel classification model. Figure 2 This is a schematic diagram of the structure of the classification model provided in an embodiment of this application. The classification model includes:

[0053] The feature vector extraction layer extracts features from the user-input text about automotive standards to be classified, resulting in a sequence of feature vectors. Specifically, in the feature vector extraction layer, the text to be classified is first segmented into words and converted into a sequence of lexical identifiers. Then, the sequence of lexical identifiers is fed into an embedding layer, which converts each discrete word into a dense, low-dimensional word vector. These word vectors initially capture the semantic information of the words. Next, the word vectors are processed by an encoder (e.g., a bidirectional Transformer encoder). The contextual information of each word in the sequence is considered, thereby generating a context-rich representation for each word vector. Finally, the output of this feature vector extraction layer is a sequence of feature vectors for the text to be classified, where each vector represents the meaning of the corresponding word in the specific context of the entire text.

[0054] The feature enhancement layer calculates the word vector similarity between the feature vector sequence and automotive domain knowledge, and enhances the feature vectors that exceed a similarity threshold. Specifically, an automotive domain knowledge base is pre-built, containing a dictionary of standard automotive terminology (such as "ESC," "thermal management," "NVH," "BMS," and "Crashworthiness"), with each term represented by an accurate word vector. In the feature enhancement layer, for each word's feature vector in the feature vector sequence, the similarity is calculated with all word vectors in the automotive domain knowledge base (e.g., using cosine similarity). A similarity threshold is set. If the similarity between a word's feature vector and a word vector in the knowledge base exceeds this threshold, the input word is considered a key entity. The feature vectors of these identified key entities are then "enhanced." Enhancement methods can include weighted fusion of the word's feature vector with the corresponding term's word vector in the knowledge base, or directly scaling up the word's feature vector. The feature vectors of other non-key entities remain unchanged.

[0055] The parent category feature branch is used to input the enhanced feature vector sequence into the global attention pooling layer to obtain the parent category feature vector sequence. The global attention pooling layer automatically learns which parts of the sequence are most important for determining the parent category through a learnable attention scoring mechanism, and assigns them higher weights. Finally, by performing weighted global pooling on the entire enhanced feature vector sequence, a fixed-dimensional vector that can represent the overall theme of the text to be classified is obtained, i.e., the parent category feature vector.

[0056] The sub-category feature branch is used to input the enhanced feature vector sequence into the gated recurrent unit (GRU) to obtain the sub-category feature vector sequence. This sub-category feature branch runs in parallel with the parent branch. The GRU is a type of recurrent neural network that excels at processing sequential data. Unlike global attention pooling layers that focus on the overall picture, the GRU processes each word in the feature vector sequence progressively, utilizing its internal gating mechanism (update and reset gates) to retain important historical information and determine how to incorporate new information into the current state. This makes the GRU adept at understanding the technical processes, conditional relationships, and detailed descriptions within the text. The output sub-category feature vector of this branch is a comprehensive representation of all the fine-grained technical information in the text.

[0057] Optionally, the classification model also includes: a sub-category constraint branch, used to generate a logical gating signal based on the parent category feature vector. The logical gating signal is then multiplied element-wise with the sub-category feature vector to obtain the constrained sub-category feature vector; this sub-category feature vector is then passed to the sub-category classification layer. Specifically, the parent category feature vector is input into a fully connected neural network layer to map it into a vector with the same dimension as the sub-category feature vector, i.e., the "logical gating signal." The range of each numerical element in this logical gating signal is constrained to between 0 and 1 by a sigmoid activation function. The value of each element in the logical gating signal is multiplied by the value of the corresponding element in the sub-category feature vector. For example, when the value of an element in the logical gating signal is close to 1, it indicates that the parent category feature vector considers the information carried by the corresponding dimension in the sub-category feature vector to be "relevant and permissible." This element of the sub-category feature vector is preserved as completely as possible. When the value of an element in the logical gating signal is close to 0, it indicates that the parent category feature vector judges that the element of the sub-category feature vector is "irrelevant or contradictory" to the parent category. This element of the feature vector for that subclass is suppressed, or even set to zero.

[0058] The parent category classification layer (e.g., implemented using softmax) is used to classify the parent category feature vector sequence to obtain the predicted parent category; the sub-category classification layer (e.g., implemented using softmax) is used to classify the sub-category feature vector sequence to obtain the predicted sub-category.

[0059] The technical effect of the classification model provided in this embodiment is as follows: Through the feature enhancement layer, the model can recognize the specific meaning of these texts within the automotive standard domain, rather than their general meaning, thereby making classification judgments consistent with domain common sense. By using sub-category constraint branches, elements of feature vectors in sub-categories that contradict the parent category are suppressed, thus guiding the final sub-category prediction in the correct direction, consistent with the parent category logic. Simultaneous classification of parent and sub-categories is achieved through parallel parent and sub-category classification layers.

[0060] Optionally, the following loss function can be constructed before the model training process:

[0061] ;

[0062] in, The cross-entropy loss represents the difference between the parent category label and the predicted parent category. The cross-entropy loss represents the difference between the sub-category label and the predicted sub-category. This refers to the category hierarchy bias. The calculation process for category hierarchy bias includes: calculating the hierarchy bias between the predicted parent category and the predicted sub-category; calculating the semantic bias between the text to be classified and the description of the predicted sub-category; and weighted summing the hierarchy bias and semantic bias to obtain the category hierarchy bias.

[0063] ;

[0064] Where N is the total number of samples. The predicted parent category for the i-th sample. Let be the parent category label of the i-th sample.

[0065] If the predicted parent category belongs to the automotive category, the cross-entropy loss between the sub-category label and the predicted sub-category is non-zero; if the predicted parent category belongs to a non-automotive category or a general category, the cross-entropy loss between the sub-category label and the predicted sub-category is zero. The following equation holds:

[0066] ;

[0067] in, This represents the number of samples whose parent category belongs to the "Automotive Standard". For the indicator function, when the parent category label The value is 1 if it belongs to the "Automotive Standard", otherwise it is 0. For the predicted subclass of the i-th sample, Let be the subclass label of the i-th sample.

[0068] Because there is a clear hierarchical relationship among the standard categories of automobiles, this application introduces a hierarchical structure loss into the total loss to optimize model training. Specifically, to encourage the model to better learn the relationship between parent and subcategories, a hierarchical structure loss term is added. This loss term quantifies the dependency between parent and subcategories, and its loss is:

[0069] ;

[0070] in, For hierarchical structure loss, The weights for the hierarchical loss are used to control the contribution of this term to the total loss. Used to measure the parent category and subcategories This application considers the hierarchical differences between categories and the semantic distance between subcategory descriptions and input text. To more accurately measure the hierarchical relationships between categories and combine them with contextual semantic analysis, this application designs a special distance function, Dist(), which not only considers the hierarchical differences between categories but also incorporates the user-input text to be classified, updating the distance under the technical scenario or user intent to improve model accuracy and adaptability.

[0071] ;

[0072] in, It is the weight of the hierarchical bias, with a value range of [0,1], used to balance the contributions of hierarchical bias and semantic bias. This indicates the hierarchical deviation between the predicted parent category and the predicted child category. This indicates the semantic deviation between the predicted subclass description and the user-input text to be classified.

[0073] The calculation process of the hierarchical deviation between the predicted parent category and the predicted child category includes: a tree structure consisting of multiple parent categories and multiple child categories; and obtaining the hierarchical deviation based on the path length of the predicted parent category and the predicted child category on the tree structure.

[0074] Specifically, a tree structure is constructed based on the classification system of automotive standards. Parent and child categories are nodes in the tree structure, with the parent category being the parent node of the child category. The distance between a parent and child category is measured by the path length in the tree. For example, in automotive standards, the distance from "Automotive Safety Standards" to "Passive Safety" is 1, and the distance from "Passive Safety" to "Body Structure" is 2. The longer the distance between nodes, the greater the difference between categories. The path length between parent category c1 and child category c2 in the tree structure is also considered. Defined as:

[0075] ;

[0076] in, It is the shortest path from c1 to c2, and length is the length of the path.

[0077] The process of calculating the semantic bias between input and output includes: performing semantic extraction on the input text and the description text of the predicted sub-category to obtain feature vectors; and calculating the semantic bias based on the distance between the feature vectors of the input text and the feature vectors of the description text.

[0078] Specifically, a pre-trained language model is introduced. The input text and the description text of the predicted subclass are respectively input into the language model to obtain feature vectors (representing semantics). Cosine similarity is used to measure the similarity between feature vectors. High similarity means closer proximity, and vice versa. This can effectively capture contextual semantic changes and accurately find the best match given user information. Assuming the input text... If the feature vectors of the predicted subclass description text are v1 and v2, then the semantic bias is defined as:

[0079] ;

[0080] Based on the aforementioned loss function, this application uses stochastic gradient descent (SGD) as the optimization method to train the model. The basic idea of ​​SGD is to minimize the loss function by iteratively updating the model parameters. The specific parameter update formula is as follows:

[0081] ;

[0082] in, These are the parameters in the classification model obtained in the (t+1)th iteration. These are the parameters in the classification model obtained in the t-th iteration. It's the learning rate. It is a loss function Regarding parameters The gradient.

[0083] In summary, the technical effects of this application include:

[0084] 1. Improved Classification Accuracy: By introducing a dual-label system of parent and child categories, the classification model can further refine the subcategories based on the parent category classification, improving the accuracy and robustness of the classification. Especially in complex and diverse automotive standard data, this method can more accurately capture the subtle differences between different standards.

[0085] 2. Adaptable to multi-category scenarios: Traditional single-label classification methods are weak when dealing with multi-category and multi-level label tasks. However, this method, through the hierarchical classification structure of parent and child categories, can effectively adapt to the complex classification requirements of automotive standards, make detailed distinctions between different dimensions of standards, and is suitable for multi-level classification scenarios.

[0086] 3. Enhance the generalization ability of the model: Through hierarchical classification and feature extraction, this method can better adapt to the differences between different categories, improve the generalization ability of the model, and enable the model to maintain high classification accuracy and stability when facing new standard data or slightly different classification systems.

[0087] 4. Reduce misclassification risk: In the design of the loss function, the sub-category is only considered when the parent category is predicted as "car category standard". This helps to reduce the risk of misclassification in the sub-category and avoid sub-category errors caused by misclassification, thereby further improving the overall performance of the model.

[0088] 5. Easy to expand and maintain: The design of this dual-label system and hierarchical loss function facilitates the expansion of standard classification systems. When it is necessary to add new categories or refine existing classifications, new labels can be added to the existing parent and sub-categories without large-scale changes to the model structure, thereby reducing the maintenance cost of the system.

[0089] like Figure 3 As shown, this embodiment provides an electronic device, including:

[0090] At least one processor; and

[0091] A memory that is communicatively connected to at least one processor; wherein,

[0092] The memory stores instructions executable by at least one processor, which, when executed, enables the processor to perform the described method. Since at least one processor in this electronic device is capable of performing the described method, it possesses at least the same advantages as the described method.

[0093] Optionally, the electronic device also includes interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The components are interconnected using different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI (Graphical User Interface) on an external input / output device (such as a display device coupled to the interface). In other embodiments, multiple processors can be used with multiple memories, and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple electronic devices (e.g., as a server array, a group of blade servers, or a multiprocessor system) can be connected, each providing some of the necessary operations. Figure 3 Take processor 301 as an example.

[0094] The memory 302, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the classification method for automotive standard text in this embodiment. The processor 301 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 302, thereby implementing the aforementioned classification method for automotive standard text.

[0095] The memory 302 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on terminal usage. Furthermore, the memory 302 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 302 may further include memory remotely located relative to the processor 301, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0096] The electronic device may further include an input device 303 and an output device 304. The processor 301, memory 302, input device 303, and output device 304 can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0097] Input device 303 can receive input digital or character information, and output device 304 may include a display device, an auxiliary lighting device (e.g., an LED), and a haptic feedback device (e.g., a vibration motor). The display device may include, but is not limited to, a liquid crystal display (LCD), a light-emitting diode (LED) display, and a plasma display. In some embodiments, the display device may be a touchscreen.

[0098] This embodiment provides a computer-readable storage medium storing computer instructions for causing a computer to perform the methods described above. The computer instructions on this computer-readable storage medium, used to cause a computer to perform the methods described above, thus have at least the same advantages as the methods described above.

[0099] The medium in this application may be any combination of one or more computer-readable media. The medium may be a computer-readable signal medium or a computer-readable storage medium. The medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of the medium (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, the medium may be any tangible medium containing or storing a program that may be used by or in connection with an instruction execution system, apparatus, or device.

[0100] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0101] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF (Radio Frequency), or any suitable combination thereof.

[0102] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0103] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means, such as coaxial cable, optical fiber, digital subscriber line (DSL), or wireless means, such as infrared, wireless, microwave, etc. The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium, etc. It is worth noting that the computer-readable storage medium mentioned in the embodiments of this application can be a non-volatile storage medium, in other words, it can be a non-transient storage medium.

[0104] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0105] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method of classifying a car standard text, characterized by, The method comprises the following steps: obtaining a deep learning-based to-be-trained model for classifying automobile texts; training the to-be-trained model according to training samples to obtain a classification model; wherein the training samples comprise input texts about automobile standards and labels, and the labels comprise parent category labels and subcategory labels; in the training process, the parameters in the classification model are iterated by minimizing a loss function; the loss function is constructed according to a parent category bias, a subcategory bias and a category hierarchy bias; the category hierarchy bias is used to measure the logical rationality of the combination of the parent category and the subcategory predicted by the model; inputting a to-be-classified text about automobile standards input by a user into the classification model to obtain a parent category and a subcategory to which the to-be-classified text belongs. The classification model comprises: a feature vector extraction layer, which is used to extract features of the to-be-classified text about automobile standards input by the user to obtain a feature vector sequence; a feature enhancement layer, which is used to calculate a word vector similarity between the feature vector sequence and automobile domain knowledge, and to enhance the feature vector that exceeds a similarity threshold; a parent category feature branch, which is used to input the enhanced feature vector sequence into a global attention pooling layer to obtain a parent category feature vector sequence; a subcategory feature branch, which is used to input the enhanced feature vector sequence into a gated recurrent unit to obtain a subcategory feature vector sequence; a parent category classification layer, which is used to classify the parent category feature vector sequence to obtain a predicted parent category; a subcategory classification layer, which is used to classify the subcategory feature vector sequence to obtain a predicted subcategory.

2. The classification method for automotive standard text according to claim 1, characterized in that, The classification model further comprises: a subcategory constraint branch, which is used to generate a logical gating signal according to the parent category feature vector, multiply the logical gating signal with the subcategory feature vector element by element to obtain a constrained subcategory feature vector, and pass the subcategory feature vector to the subcategory classification layer.

3. The classification method for automotive standard text according to claim 1, characterized in that, The loss function is: ； wherein, a cross-entropy loss between the parent class label and the predicted parent class, a cross-entropy loss between the child class label and the predicted child class, is a class hierarchy bias; The calculation process of the category hierarchy bias comprises: calculating a hierarchy bias between the predicted parent category and the predicted subcategory; calculating a semantic bias between the to-be-classified text and a description of the predicted subcategory; and performing weighted summation on the hierarchy bias and the semantic bias to obtain the category hierarchy bias.

4. The method of claim 3, wherein the automobile standard text is classified into a category of a manual transmission automobile or a category of an automatic transmission automobile. The calculation process of the hierarchy bias between the predicted parent category and the predicted subcategory comprises: a tree structure is formed by a plurality of parent categories and a plurality of subcategories; the hierarchy bias is obtained according to the path length of the predicted parent category and the predicted subcategory on the tree structure.

5. The method of claim 3, wherein the automobile standard text is classified into a category of a manual transmission vehicle or an automatic transmission vehicle. The calculation process of the semantic bias between the input and the output comprises: performing semantic extraction on the input text and the description text of the predicted subcategory respectively to obtain feature vectors; calculating the semantic bias according to the distance between the input text feature vector and the description text feature vector.

6. The method of claim 3, wherein the automobile standard text is classified into a category of a manual transmission vehicle or an automatic transmission vehicle. The parent category comprises automobile standard, non-automobile standard and general standard; if the predicted parent category belongs to the automobile standard, the cross-entropy loss between the subcategory label and the predicted subcategory is not 0; if the predicted parent category belongs to the non-automobile standard or the general standard, the cross-entropy loss between the subcategory label and the predicted subcategory is 0.

7. The method of classifying automotive standard texts according to any one of claims 1 to 6, characterized in that, The parameters in the classification model are iterated by minimizing the loss function, comprising: the parameters in the classification model are iterated by using the following formula: ； wherein, is a parameter in the classification model obtained at the t+1th iteration, is a parameter in the classification model obtained at the tth iteration, is a learning rate, is a loss function with respect to the parameter .

8. An electronic device, comprising: ​ at least one processor, and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method for classifying automobile standard texts according to any one of claims 1-7.

Images