Artificial intelligence-based automobile information management method and system
By generating character-specific feature vectors for automotive news titles and summaries using artificial intelligence technology, and combining these with user behavior and request information, the problem of information organization and recommendation on existing automotive news platforms has been solved. This has enabled efficient personalized recommendations, improved user experience, and increased platform activity.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 智驭未来(广州)信息科技有限公司
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-29
Smart Images
Figure CN122113910A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of artificial intelligence and automotive technology, and in particular to artificial intelligence-based automotive information management methods and systems. Background Technology
[0002] With the rapid development of internet and mobile communication technologies, automotive information platforms have become an important channel for users to obtain automotive-related information. However, facing massive, multi-source, and heterogeneous automotive information content, how to efficiently organize, accurately understand, and personally recommend relevant content has become a key challenge to improve user experience and platform activity.
[0003] In recent years, the development of artificial intelligence, especially Natural Language Processing (NLP) technology, has brought new possibilities to information management and recommendation systems. Some existing technologies encode the entire news text into a high-dimensional vector and build a vector index, using the Approximate Nearest Neighbor (ANN) algorithm to achieve semantic similarity retrieval. However, the original news text is usually lengthy and contains a lot of redundant or irrelevant information. Directly vectorizing it can easily introduce noise, affecting the accuracy of semantic representation and the retrieval effect and efficiency. In addition, most existing systems only generate a single title, which cannot adapt to the significant differences in cognitive level, focus, and information needs of different user groups (such as professional engineers and ordinary consumers), resulting in insufficient title attractiveness, low click-through rate and reading completion rate, thus limiting the activity of automotive platforms due to uniform titles. Furthermore, in the early stage of new information launch (i.e., the "cold start" stage), due to the lack of user interaction data, traditional collaborative filtering or behavior-driven recommendation models are difficult to work effectively, resulting in insufficient exposure of high-quality content. Finally, existing systems often ignore the dynamic impact of user access information on information preferences, making it difficult to achieve truly real-time personalized recommendations. Summary of the Invention
[0004] To address the aforementioned technical problems, the purpose of this application is to provide an artificial intelligence-based automotive information management method and system, aiming to solve the technical problems mentioned above.
[0005] In a first aspect, embodiments of this application provide an artificial intelligence-based vehicle information management method, the method comprising:
[0006] Collect automotive information text;
[0007] The automotive information text is input into a trained multi-title generation model, which outputs automotive information titles corresponding to each preset role; wherein, the preset roles include automotive engineers and automotive consumers;
[0008] The automotive information text is input into the summary generation model, which outputs an automotive information summary.
[0009] Based on the automotive information summary, automotive information summary feature vectors are generated through a summary feature vector generation model, and each automotive information summary feature vector is indexed into a vector retrieval engine.
[0010] Based on the target user's reading behavior, target user's access request information, and target user's role, a user request feature vector is generated through a request feature vector generation model.
[0011] The vector retrieval engine is used to find the N automotive information summary feature vectors that are most similar to the user request feature vector, and the corresponding N automotive information items are determined based on the association between the automotive information summary feature vectors and their corresponding automotive information texts.
[0012] Based on the titles of the automotive news articles, the N automotive news articles are recommended to the target user.
[0013] Furthermore, the multi-title generation model includes an encoder, a decoder, and a role-content interaction gating module deployed between the encoder and the decoder;
[0014] The encoder is used to perform semantic encoding on the input automotive information text and output a document semantic sequence representation;
[0015] The role-content interaction gating module is used for:
[0016] Obtain static character feature vectors corresponding to multiple preset characters;
[0017] Generate a full-text semantic summary vector based on the document semantic sequence representation;
[0018] For each preset role, a dimensional activation weight vector is calculated using a lightweight gating network based on the static role feature vector and the full-text semantic summary vector. The dimensional activation weight vector is then used to modulate the static role feature vector element by element to generate a context-aware role guidance signal.
[0019] The decoder is used to fuse the guidance signals of each role with the semantic sequence representation of the document as generation conditions, and output K car information titles that match the corresponding preset roles in parallel via autoregression; where K is the number of preset roles.
[0020] Furthermore, the loss function of the multi-title generation model is:
[0021] ;
[0022] ;
[0023] ;
[0024] in, Main loss function, For role consistency regularization, T is the total number of terms in the reference title, T is the length of the terms in the reference title, and t is the time step index. For the true label of the t-th word, For all real words before time step t, To guide the character For document semantic sequence representation, Let r be the semantic anchor vector of character r. To balance the hyperparameters.
[0025] Furthermore, the step of generating a user request feature vector based on the target user's reading behavior, target user access request information, and target user role through a request feature vector generation model includes:
[0026] Find the summary feature vectors corresponding to all automotive information related to the target user's reading behavior, and calculate the average of all summary feature vectors of the target user's reading behavior to obtain the target user's preference feature vector;
[0027] The user request feature vector is generated by using the target user's preference feature vector, target user request information, and target user role through a request feature vector generation model.
[0028] Furthermore, the summary feature vector generation model includes a first BERT network, a second BERT network, a first embedding module, a second embedding module, and a first binary classifier; the first BERT network, the first embedding module, and the first binary classifier are connected in sequence; the second BERT network, the second embedding module, and the first binary classifier are connected in sequence; wherein, the first BERT network and the second BERT network have the same structure and share parameters.
[0029] Furthermore, the summary feature vector generation model is trained through the following steps:
[0030] From the user behavior logs, we count the users who interacted with each piece of automotive news, forming a set of interacting users corresponding to that piece of automotive news.
[0031] For any two automotive news articles a and b, obtain their corresponding sets of interacting users A and B respectively; when the following conditions are met... , and the size of the intersection When the condition is met, the car information pair (a, b) is marked as a positive sample; otherwise, it is marked as a negative sample. and The preset threshold;
[0032] The summary of car information 'a' from the positive or negative samples is input into the first BERT network of the summary feature vector generation model to be trained, and the summary of car information 'b' is input into the second BERT network of the summary feature vector generation model to be trained, and the corresponding feature vectors are output respectively. and ;
[0033] calculate and The similarity score is calculated, and based on the similarity score and sample labels, the parameters of the summary feature vector generation model to be trained are optimized using binary classification loss.
[0034] Furthermore, the request feature vector generation model includes a DNN network, a third embedding module, a first embedding module, a first BERT network, and a second binary classifier; the DNN network, the third embedding module, and the second binary classifier are connected in sequence; the first BERT network, the first embedding module, and the second binary classifier are connected in sequence.
[0035] Furthermore, the request feature vector generation model is trained through the following steps:
[0036] S81. Construct training samples from user behavior logs, where positive samples are user-information pairs consisting of the user and the car information in which the interaction occurred, and negative samples are user-information pairs consisting of the user and the car information in which the interaction did not occur.
[0037] S82: For each user in the training sample, perform the following sub-steps:
[0038] S821. Obtain a summary of all the car information in the user's historical interactions, and generate corresponding car information summary feature vectors respectively through the trained summary feature vector generation model.
[0039] S822. Calculate the mean of the feature vector of the car information summary to obtain the user's preference feature vector;
[0040] S823. Input the user's preference feature vector, user access request information and role information into the first DNN network, and output the user request feature vector by the third Embedding module;
[0041] S83: For the automotive information in each training sample, generate its corresponding automotive information summary feature vector through the summary feature vector generation model, and freeze the parameters of the summary feature vector generation model during the training process;
[0042] S84: Calculate the similarity score between the user request feature vector and the car information summary feature vector, and optimize the model parameters based on the similarity score and sample labels using a binary classification loss function.
[0043] Secondly, embodiments of this application provide an artificial intelligence-based vehicle information management system, characterized in that the system is used to execute the artificial intelligence-based vehicle information management method according to any one of claims 1-8, and the system includes:
[0044] The data acquisition module is used to collect automotive information text.
[0045] The first input module is used to input the automotive information text into a trained multi-title generation model and output automotive information titles corresponding to each preset role; wherein, the preset roles include automotive engineers and automotive consumers;
[0046] The second input module is used to input the automotive information text into the summary generation model and output the automotive information summary.
[0047] The first generation module is used to generate automotive information summary feature vectors based on the automotive information summary through a summary feature vector generation model, and index each automotive information summary feature vector into a vector retrieval engine.
[0048] The second generation module is used to generate user request feature vectors based on the target user's reading behavior, target user access request information, and target user role, through a request feature vector generation model.
[0049] The search module is used to use a vector retrieval engine to find the N automotive information summary feature vectors that are most similar to the user request feature vector, and to determine the corresponding N automotive information items based on the association between the automotive information summary feature vectors and their corresponding automotive information texts.
[0050] The recommendation module is used to recommend the N pieces of automotive news to the target user based on the automotive news titles.
[0051] Beneficial effects:
[0052] (1) The application embodiment obtains a multi-title generation model by pre-constructing and training, which can generate corresponding titles based on automotive information text according to different role characteristics. During recommendation, the generated titles are recommended to the corresponding target users. Since the titles match the user role characteristics, they are more attractive to users than traditional single titles, thereby increasing the reading volume of automotive information and thus increasing the activity of automotive information websites or platforms. In addition, the automotive information text is input into the summary generation model, and the automotive information summary is output; the summary feature vector generation model generates automotive information summary feature vectors, and each automotive information summary feature vector is indexed into the vector retrieval engine. In this way, any newly collected automotive information can quickly generate automotive information summary feature vectors and add the automotive information summary feature vectors to the vector retrieval engine. After being added, they are immediately retrieved. Therefore, this method can effectively solve the problem of real-time cold start of information. Since user access request information includes location and time information, a user request feature vector is generated based on user reading behavior, user access request information, and target user role. The N most similar automotive news summary feature vectors are then retrieved. Even without user reading behavior, suitable automotive news can still be recommended based on the user's location and time. Furthermore, this approach enables real-time personalized recommendations of diverse information for users. By generating automotive news summary feature vectors instead of automotive news feature vectors, compared to generating automotive news feature vectors, the summary length is much shorter than the original text, the vector semantics are more focused, there is less noise, and semantic matching accuracy is improved. On the other hand, it significantly reduces vector dimensional redundancy and computational overhead, improving response speed.
[0053] (2) Furthermore, the role-content interaction gating module proposed in this invention generates a dimension-level activation weight vector by calculating the interaction relationship between the global semantic representation of the automotive information text and the preset role vector, and dynamically weights the role vector to output a role guidance signal aligned with the current text topic. This mechanism enables the multi-headline generation model to generate highly specialized and context-consistent headlines for the differentiated information needs of the same role under different technical topics.
[0054] (3) Furthermore, the main loss-driven model learns to accurately reproduce manually written role-adapted titles given the main text and dynamic role guidance. By introducing a role consistency regularization term as an auxiliary loss, the semantic effectiveness of the role guidance signal is enhanced, thereby improving the title adaptability.
[0055] (4) Further, in this embodiment, the mean vector is obtained by averaging all the summary feature vectors of the target user's reading behavior, and this mean vector is used as the target user's preference feature vector. By averaging these vectors, the user's multi-dimensional interest signals can be effectively aggregated to form a low-dimensional, dense and semantically interpretable static preference representation, avoiding the computational overhead caused by complex sequence modeling.
[0056] (5) Furthermore, the summary feature vector generation model in this application adopts a dual BERT network architecture, and the first BERT network and the second BERT network have the same structure and share parameters. This design enables the model to extract features based on a unified parameter system and semantic understanding logic when processing two automotive information summaries, ensuring the consistency of the distribution of feature vectors for information summaries with similar semantics and high relevance. At the same time, the BERT network itself has a powerful contextual semantic understanding capability, and combined with a dedicated Embedding module to further optimize feature mapping, it can deeply mine professional semantic information (such as vehicle characteristics, technical parameters, industry dynamics, etc.) in automotive information summaries. The generated feature vectors can accurately represent the core connotation of the summary, greatly improving the accuracy of capturing the semantic association of automotive information, and providing a high-quality feature foundation for subsequent information similarity judgment.
[0057] (6) Further, in this embodiment of the application, training samples are constructed based on user interaction behavior logs. By statistically analyzing the set of interactive users for each piece of automotive information, the size of the set of interactive users and the size of their intersection are used as positive and negative labels for the samples. The summaries of automotive information a and automotive information b in the positive or negative samples are respectively input into the first Bert network and the second Bert network, which have the same structure and share parameters, and the corresponding feature vectors are output respectively. and ; and calculate and The similarity score is used to optimize the parameters of the summary feature vector generation model to be trained by using binary classification loss based on the similarity score and sample label. The summary feature vector generation model trained by this method can better express the user's interest features.
[0058] (7) The design of the request feature vector generation model includes a DNN network and a BERT network, and it is a two-branch model. The BERT network ensures fine-grained semantic understanding of the automotive information text, while the DNN branch efficiently processes high-dimensional sparse structured inputs. This request feature vector generation model design lays a structural foundation for the subsequent training of the request feature vector generation model to output user request feature vectors that are close to the feature vectors of information that users like, making the features extracted by the trained user request feature vector model more reflective of the features of information that users like.
[0059] (8) The training method of this application embodiment can make the user's request feature vector as close as possible to the feature vector of the information the user likes. Thus, the user request feature vector output by the trained request feature vector generation model is close to the feature vector of the information the user likes. In this way, the user request feature vector can be used to match the information the user likes, and the recall result is more accurate. Attached Figure Description
[0060] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0061] Figure 1 This is a schematic diagram of the process of artificial intelligence-based vehicle information management provided in an embodiment of this application;
[0062] Figure 2 This is a schematic diagram of the structure of the artificial intelligence-based vehicle information management system provided in the embodiments of this application. Detailed Implementation
[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0064] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.
[0065] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and unless as herein...
[0066] like Figure 1 As shown in the figure, this application provides an artificial intelligence-based vehicle information management method, the method comprising:
[0067] S1. Collect automotive information text;
[0068] S2. Input the automotive information text into the trained multi-title generation model and output the automotive information titles corresponding to each preset role; wherein, the preset roles include automotive engineers and automotive consumers;
[0069] S3. Input the automotive information text into the summary generation model and output the automotive information summary;
[0070] S4. Based on the automotive information summary, generate automotive information summary feature vectors through the summary feature vector generation model, and index each automotive information summary feature vector into the vector retrieval engine.
[0071] S5. Based on the target user's reading behavior, target user's access request information, and target user's role, generate a user request feature vector through a request feature vector generation model.
[0072] S6. Use a vector retrieval engine to find the N automotive information summary feature vectors that are most similar to the user request feature vector, and determine the corresponding N automotive information items based on the association between the automotive information summary feature vectors and their corresponding automotive information texts.
[0073] S7. Based on the title of the automotive news, recommend the N pieces of automotive news to the target user.
[0074] In this embodiment of the application, the automotive information text is input into a trained multi-title generation model, which outputs automotive information titles corresponding to each preset role. For example, a short automotive information text is:
[0075] The all-new electric SUV, the 'X7', has officially launched, with a starting price of 239,800 yuan. This model features an 800V high-voltage fast-charging platform, capable of replenishing 400 kilometers of range in just 10 minutes; it is equipped with a Level 3 autonomous driving assistance system and utilizes unibody die-casting technology, reducing the overall vehicle weight by 15%. Furthermore, its intelligent cockpit supports multimodal voice interaction and an AR-HUD head-up display.
[0076] The preset characters and their corresponding generated titles are shown below:
[0077] Role: Car consumer
[0078] Headline: "Starting at 230,000 yuan! Jichi X7 electric SUV launched, 400 kilometers on 10-minute charge."
[0079] Role: Automotive Engineer
[0080] Title: "Jichi X7 debuts with 800V high-voltage platform and integrated die-cast body, reducing weight by 15%, and bringing L3 intelligent driving to fruition."
[0081] This application embodiment pre-builds and trains a multi-title generation model, which can generate corresponding titles based on automotive information text and the characteristics of different user roles. During recommendation, the generated titles are recommended to the corresponding target users. Because the titles match the user's role characteristics, they are more attractive to users than traditional single titles, thereby increasing the readership of automotive information and thus improving the activity of automotive information websites or platforms. Furthermore, the automotive information text is input into a summary generation model, which outputs the automotive information summary. A summary feature vector generation model generates automotive information summary feature vectors, and each automotive information summary feature vector is indexed into a vector retrieval engine. This allows any newly collected automotive information to quickly generate automotive information summary feature vectors, which are then added to the vector retrieval engine and immediately retrieved. Therefore, this method effectively solves the real-time cold start problem for information. Since user access request information includes location information, time information, etc., a user request feature vector is generated based on user reading behavior, user access request information, and the target user role. The N automotive information summary feature vectors most similar to the user request feature vector are retrieved. Even without user reading behavior, suitable automotive information can still be recommended based on the user's location information, time, etc. Furthermore, this approach enables real-time personalized recommendations of diverse information for users. By generating automotive information summary feature vectors instead of automotive information feature vectors, compared to generating automotive information feature vectors, on the one hand, the summary length is much shorter than the original text, the vector semantics are more focused, there is less noise, and the semantic matching accuracy is improved. On the other hand, it significantly reduces vector dimensional redundancy and computational overhead, improving response speed.
[0082] In one embodiment, the multi-title generation model includes an encoder, a decoder, and a role-content interaction gating module deployed between the encoder and the decoder;
[0083] The encoder is used to perform semantic encoding on the input automotive information text and output a document semantic sequence representation;
[0084] The role-content interaction gating module is used for:
[0085] Obtain static character feature vectors corresponding to multiple preset characters;
[0086] Generate a full-text semantic summary vector based on the document semantic sequence representation;
[0087] For each preset role, a dimensional activation weight vector is calculated using a lightweight gating network based on the static role feature vector and the full-text semantic summary vector. The dimensional activation weight vector is then used to modulate the static role feature vector element by element to generate a context-aware role guidance signal.
[0088] The decoder is used to fuse the guidance signals of each role with the semantic sequence representation of the document as generation conditions, and output K car information titles that match the corresponding preset roles in parallel via autoregression; where K is the number of preset roles.
[0089] In a preferred embodiment, the encoder may employ the BERT model, which is a Transformer-based bidirectional encoder that outputs a sequence of context-sensitive vectors. The document semantic sequence output by the encoder is represented as follows: Where L is the sequence length, d is the hidden layer dimension, and R is the set of real numbers. The static character feature vector The semantic description text is obtained as follows: For each preset role *r*, a semantic description text is predefined, summarizing the role's typical concerns and language style in the automotive field. The description text is then encoded using a pre-trained sentence vector model to generate a corresponding fixed-dimensional semantic vector, which is stored as a role feature library before model deployment. The full-text semantic summary vector *s* can be generated based on the document's semantic sequence through attention pooling. The role feature vector is then... and full-text semantic summary vector After concatenation, input a lightweight gating network, which can be an MLP+Sigmoid network:
[0090]
[0091] Where g is the dimensional activation weight vector, It is the weight matrix of the gated network. It is the bias vector of the gating network. , For the Sigmoid function, each dimension of g This represents the activation strength of the i-th dimension of the role vector in the current context. MLP (Multilayer Perceptron) is a multilayer perceptron.
[0092] The dimensional activation weight vector is a control signal with the same dimension as the character vector. Each element independently determines the activation level of the corresponding dimension in the character's semantics, thereby achieving fine-grained, context-aware character adaptation. In other words, the dimensional activation weight vector is used to modulate the character feature vector of a preset character dimension by dimension, so that the character feature dimension related to the semantics of the current automotive information text is enhanced, while irrelevant dimensions are suppressed.
[0093] The static character feature vector is modulated element-wise using the dimensional activation weight vector to generate a context-aware character guidance signal, as shown in the following formula:
[0094]
[0095] in, It is a context-aware role guidance signal. For Hadamard products, it means multiplying corresponding positions.
[0096] Context-aware role guidance signals refer to conditional vectors obtained by dynamically adjusting statically preset role feature vectors given the semantic content (context) of the input document, which are used to guide text generation.
[0097] The role guidance signal and the document semantic sequence representation are fused as generation conditions to autoregressively output a car information title matching the preset role. Specifically, the decoder adopts a Transformer-based decoder structure, including multiple decoding layers connected in sequence. Each decoding layer includes a masked self-attention sublayer, an encoder-decoder cross-attention sublayer, and a feedforward neural network sublayer. The cross-attention sublayer converts the document semantic sequence representation output by the encoder... Context-aware role guidance signals The sequences are concatenated along the sequence dimension to form an extended (Key) and (Value) matrix, which, together with the decoder query, participates in attention calculation to achieve deep integration of role semantics and document content.
[0098] The role-content interaction gating module proposed in this invention generates a dimensional activation weight vector by calculating the interaction relationship between the global semantic representation of automotive information text and preset role vectors. This vector is then dynamically weighted to output a role guidance signal aligned with the current text topic. This mechanism enables the multi-headline generation model to generate highly specialized and context-consistent headlines for the differentiated information needs of the same role under different technical topics.
[0099] In one embodiment, the loss function of the multi-title generation model is:
[0100] ;
[0101] ;
[0102] ;
[0103] in, Main loss function, For role consistency regularization, T is the total number of terms in the reference title, T is the length of the terms in the reference title, and t is the time step index. For the true label of the t-th word, For all real words before time step t, To guide the character For document semantic sequence representation, Let r be the semantic anchor vector of character r. To balance the hyperparameters.
[0104] In this embodiment, to enable the multi-title generation model to generate adapted titles for different preset roles, the present invention adopts an end-to-end joint training strategy, uniformly optimizing all components, including the encoder, decoder, and role-content interaction gating module. The specific training process is as follows:
[0105] 1. Training data construction
[0106] (1) Construct a large-scale automotive information triplet training set, where each sample includes:
[0107] Automotive news articles: Authentic automotive news, technical reviews, or industry analysis articles;
[0108] Preset role tags: such as "Automotive Engineer", "Automotive Salesperson", "Automotive Procurement Personnel", "Consumer", "Investor", etc.
[0109] Manually annotated reference titles: High-quality titles written by professional editors for the main text and for a specific role.
[0110] (2) End-to-end joint optimization
[0111] Input the above triples into the multi-heading generation model and perform the following steps:
[0112] The encoder encodes the information text and outputs a semantic sequence representation of the document. ;
[0113] The character encoder maps character tags to static character feature vectors. ;
[0114] The role-content interaction gating module is based on document semantic sequence representation. Combined state role feature vector Generate context-aware role guidance signals ;
[0115] decoder with Given H as a condition, an autoregressive method generates predicted titles.
[0116] All trainable parameters of the model, including encoder parameters, decoder parameters, character encoder parameters, and gating network parameters in the character-content interaction gating module. and —All updates are synchronized through backpropagation, achieving end-to-end optimization of the overall architecture.
[0117] In this embodiment, the primary optimization objective is defined as maximizing the probability of the model generating a reference title. Specifically, this is achieved through the aforementioned primary loss function, which is actually a negative log-likelihood loss. This loss drives the model to learn how to accurately reproduce manually written, role-appropriate titles given the main text and dynamic role guidance. To further enhance the role guidance signal... To assess semantic effectiveness, a role consistency regularization term is introduced as an auxiliary loss. First, semantic anchors for each predefined role need to be defined in the role semantic space. These anchors are semantic center vectors obtained by vectorizing a large amount of related text for that role. It should be understood that a semantic anchor is a typical representative vector of a role in the semantic space vector; it should summarize the topics, terminology, and expression style that the role is most concerned with.
[0118] It should be understood that the main loss in this application focuses on generating quality, directly aligning with the manual title, and the role consistency regularization term enhances the semantic consistency of the role, preventing the gating module from outputting noise signals that deviate from the essence of the role, thereby improving the matching degree of the title generation.
[0119] The end-to-end training described above ensures collaborative optimization among modules, avoiding information loss during phased training. Through this training method, the role-content interaction gating module of this invention can learn to dynamically activate relevant dimensions of the role vector based on the text content. The multi-headline generation model can establish a mapping capability of "text + role preference → headline style and terminology." Ultimately, this enables the generation of highly differentiated and professionally credible headlines for the same text but different roles.
[0120] In one embodiment, the step of generating a user request feature vector based on the target user's reading behavior, target user access request information, and target user role through a request feature vector generation model includes:
[0121] Find the summary feature vectors corresponding to all automotive information related to the target user's reading behavior, and calculate the average of all summary feature vectors of the target user's reading behavior to obtain the target user's preference feature vector;
[0122] The user request feature vector is generated by using the target user's preference feature vector, target user request information, and target user role through a request feature vector generation model.
[0123] In this embodiment, the mean vector is obtained by averaging all the summary feature vectors of the target user's reading behavior, and this mean vector is used as the target user's preference feature vector. By averaging these vectors, multi-dimensional interest signals of users can be effectively aggregated to form a low-dimensional, dense, and semantically interpretable static preference representation, avoiding the computational overhead of complex sequence modeling.
[0124] In one embodiment, the summary feature vector generation model includes a first BERT network, a second BERT network, a first embedding module, a second embedding module, and a first binary classifier; the first BERT network, the first embedding module, and the first binary classifier are connected in sequence; the second BERT network, the second embedding module, and the first binary classifier are connected in sequence; wherein the first BERT network and the second BERT network have the same structure and share parameters.
[0125] The feature vector generation model in this application adopts a dual BERT network architecture, with the first and second BERT networks having identical structures and sharing parameters. This design enables the model to extract features based on a unified parameter system and semantic understanding logic when processing two automotive news summaries, ensuring the consistency of feature vector distribution for summaries with similar semantics and high relevance. Simultaneously, the BERT network itself possesses powerful contextual semantic understanding capabilities. Combined with a dedicated embedding module, it further optimizes feature mapping, enabling in-depth mining of professional semantic information (such as vehicle characteristics, technical parameters, industry dynamics, etc.) within automotive news summaries. The generated feature vectors accurately represent the core connotations of the summaries, significantly improving the accuracy of capturing semantic relationships in automotive news and providing a high-quality feature foundation for subsequent similarity judgment.
[0126] In one embodiment, the summary feature vector generation model is trained through the following steps:
[0127] From the user behavior logs, we count the users who interacted with each piece of automotive news, forming a set of interacting users corresponding to that piece of automotive news.
[0128] For any two automotive news articles a and b, obtain their corresponding sets of interacting users A and B respectively; when the following conditions are met... , and the size of the intersection When the condition is met, the car information pair (a, b) is marked as a positive sample; otherwise, it is marked as a negative sample. and The preset threshold;
[0129] The summary of car information 'a' from the positive or negative samples is input into the first BERT network of the summary feature vector generation model to be trained, and the summary of car information 'b' is input into the second BERT network of the summary feature vector generation model to be trained, and the corresponding feature vectors are output respectively. and ;
[0130] calculate and The similarity score is calculated, and based on the similarity score and sample labels, the parameters of the summary feature vector generation model to be trained are optimized using binary classification loss.
[0131] In this embodiment, the interactive behavior includes one or more of clicking, liking, collecting, or sharing; the similarity score is dot product similarity or cosine similarity; and the binary classification loss function is the binary cross-entropy loss function. In this embodiment, after training, the trained summary feature vector generation model uses a first BERT network and a first embedding module, or a second BERT network and a second embedding module, to generate automotive information summary feature vectors.
[0132] This application embodiment constructs training samples based on user interaction behavior logs. By statistically analyzing the set of interacting users for each piece of automotive information, the size of the user set and the intersection of the user sets are used as positive and negative labels for the samples. The summaries of automotive information 'a' and automotive information 'b' from the positive or negative samples are respectively input into a first Bert network and a second Bert network with identical structures and shared parameters, and the corresponding feature vectors are output respectively. and ; and calculate and The similarity score is used to optimize the parameters of the summary feature vector generation model to be trained by using binary classification loss based on the similarity score and sample label. The summary feature vector generation model trained by this method can better express the user's interest features.
[0133] In one embodiment, the request feature vector generation model includes a DNN network, a third embedding module, a first embedding module, a first BERT network, and a second binary classifier; the DNN network, the third embedding module, and the second binary classifier are connected in sequence; the first BERT network, the first embedding module, and the second binary classifier are connected in sequence.
[0134] This application embodiment designs a request feature vector generation model comprising a DNN network and a BERT network, with two branches. The BERT network ensures fine-grained semantic understanding of the automotive information text, while the DNN branch efficiently processes high-dimensional, sparse, structured input. This request feature vector generation model design lays a structural foundation for the subsequent training of the model to produce user request feature vectors that closely resemble the feature vectors of user-preferred information, making the features extracted by the trained user request feature vector model more reflective of user-preferred information characteristics.
[0135] In one embodiment, the request feature vector generation model is trained through the following steps:
[0136] S81. Construct training samples from user behavior logs, where positive samples are user-information pairs consisting of the user and the car information in which the interaction occurred, and negative samples are user-information pairs consisting of the user and the car information in which the interaction did not occur.
[0137] S82: For each user in the training sample, perform the following sub-steps:
[0138] S821. Obtain a summary of all the car information in the user's historical interactions, and generate corresponding car information summary feature vectors respectively through the trained summary feature vector generation model.
[0139] S822. Calculate the mean of the feature vector of the car information summary to obtain the user's preference feature vector;
[0140] S823. Input the user's preference feature vector, user access request information and role information into the first DNN network, and output the user request feature vector by the third Embedding module;
[0141] S83: For the automotive information in each training sample, generate its corresponding automotive information summary feature vector through the summary feature vector generation model, and freeze the parameters of the summary feature vector generation model during the training process;
[0142] S84: Calculate the similarity score between the user request feature vector and the car information summary feature vector, and optimize the model parameters based on the similarity score and sample labels using a binary classification loss function.
[0143] The training method in this application embodiment can make the user's request feature vector as close as possible to the feature vector of the information the user likes. As a result, the user request feature vector output by the trained request feature vector generation model is close to the feature vector of the information the user likes. Therefore, the user request feature vector can be used to match the information the user likes, and the recall result is more accurate.
[0144] like Figure 2 As shown in the illustration, this application provides an artificial intelligence-based vehicle information management system. The system is used to execute the artificial intelligence-based vehicle information management method described in any of the preceding claims. The system includes:
[0145] Data collection module 1 is used to collect automotive information text.
[0146] The first input module 2 is used to input the automotive information text into the trained multi-title generation model and output the automotive information titles corresponding to each preset role; wherein, the preset roles include automotive engineers and automotive consumers;
[0147] The second input module 3 is used to input the automotive information text into the summary generation model and output the automotive information summary.
[0148] The first generation module 4 is used to generate automotive information summary feature vectors based on the automotive information summary through a summary feature vector generation model, and index each automotive information summary feature vector into a vector retrieval engine.
[0149] The second generation module 5 is used to generate user request feature vectors based on the target user's reading behavior, target user access request information, and target user role, through a request feature vector generation model.
[0150] The search module 6 is used to use a vector retrieval engine to find the N automotive information summary feature vectors that are most similar to the user request feature vector, and to determine the corresponding N automotive information items based on the association between the automotive information summary feature vectors and their corresponding automotive information texts.
[0151] Recommendation module 7 is used to recommend the N pieces of automotive news to the target user based on the automotive news titles.
[0152] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0153] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.
[0154] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
Claims
1. A vehicle information management method based on artificial intelligence, characterized in that, The method includes: Collect automotive information text; The automotive information text is input into a trained multi-title generation model, which outputs automotive information titles corresponding to each preset role; wherein, the preset roles include automotive engineers and automotive consumers; The automotive information text is input into the summary generation model, which outputs an automotive information summary. Based on the automotive information summary, automotive information summary feature vectors are generated through a summary feature vector generation model, and each automotive information summary feature vector is indexed into a vector retrieval engine. Based on the target user's reading behavior, target user's access request information, and target user's role, a user request feature vector is generated through a request feature vector generation model. The vector retrieval engine is used to find the N automotive information summary feature vectors that are most similar to the user request feature vector, and the corresponding N automotive information items are determined based on the association between the automotive information summary feature vectors and their corresponding automotive information texts. Based on the titles of the automotive news articles, the N automotive news articles are recommended to the target user.
2. The vehicle information management method based on artificial intelligence according to claim 1, characterized in that, The multi-title generation model includes an encoder, a decoder, and a role-content interaction gating module deployed between the encoder and the decoder; The encoder is used to perform semantic encoding on the input automotive information text and output a document semantic sequence representation; The role-content interaction gating module is used for: Obtain static character feature vectors corresponding to multiple preset characters; Generate a full-text semantic summary vector based on the document semantic sequence representation; For each preset role, a dimensional activation weight vector is calculated using a lightweight gating network based on the static role feature vector and the full-text semantic summary vector. The dimensional activation weight vector is then used to modulate the static role feature vector element by element to generate a context-aware role guidance signal. The decoder is used to fuse the guidance signals of each role with the semantic sequence representation of the document as generation conditions, and output K car information titles that match the corresponding preset roles in parallel via autoregression; where K is the number of preset roles.
3. The vehicle information management method based on artificial intelligence according to claim 2, characterized in that, The loss function of the multi-title generation model is: ； ; ； in, Main loss function, For role consistency regularization, T is the total number of terms in the reference title, T is the term length of the reference title, and t is the time step index. For the true label of the t-th word, For all real words before time step t, To guide the character For document semantic sequence representation, Let r be the semantic anchor vector of character r. To balance the hyperparameters.
4. The vehicle information management method based on artificial intelligence according to claim 1, characterized in that, The step of generating a user request feature vector based on the target user's reading behavior, target user access request information, and target user role through a request feature vector generation model includes: Find the summary feature vectors corresponding to all automotive information related to the target user's reading behavior, and calculate the average of all summary feature vectors of the target user's reading behavior to obtain the target user's preference feature vector; The user request feature vector is generated by using the target user's preference feature vector, target user request information, and target user role through a request feature vector generation model.
5. The vehicle information management method based on artificial intelligence according to claim 1, characterized in that, The summary feature vector generation model includes a first BERT network, a second BERT network, a first embedding module, a second embedding module, and a first binary classifier; the first BERT network, the first embedding module, and the first binary classifier are connected in sequence; the second BERT network, the second embedding module, and the first binary classifier are connected in sequence; wherein, the first BERT network and the second BERT network have the same structure and share parameters.
6. The artificial intelligence-based vehicle information management method according to claim 5, characterized in that, The abstract feature vector generation model is trained through the following steps: From the user behavior logs, we count the users who interacted with each piece of automotive news, forming a set of interacting users corresponding to that piece of automotive news. For any two automotive news articles a and b, obtain their corresponding sets of interacting users A and B respectively; when the condition is met... , and the size of the intersection When the condition is met, the car information pair (a, b) is labeled as a positive sample; otherwise, it is labeled as a negative sample. and The preset threshold; The summary of car information 'a' from the positive or negative samples is input into the first BERT network of the summary feature vector generation model to be trained, and the summary of car information 'b' is input into the second BERT network of the summary feature vector generation model to be trained, and the corresponding feature vectors are output respectively. and ; calculate and The similarity score is calculated, and based on the similarity score and sample label, the parameters of the summary feature vector generation model to be trained are optimized by binary classification loss.
7. The vehicle information management method based on artificial intelligence according to claim 6, characterized in that, The requested feature vector generation model includes a DNN network, a third Embedding module, a first Embedding module, a first BERT network, and a second binary classifier; the DNN network, the third Embedding module, and the second binary classifier are connected in sequence; the first BERT network, the first Embedding module, and the second binary classifier are connected in sequence.
8. The vehicle information management method based on artificial intelligence according to claim 7, characterized in that, The request feature vector generation model is trained through the following steps: S81. Construct training samples from user behavior logs, where positive samples are user-information pairs consisting of the user and the car information in which the interaction occurred, and negative samples are user-information pairs consisting of the user and the car information in which the interaction did not occur. S82: For each user in the training sample, perform the following sub-steps: S821. Obtain a summary of all the car information in the user's historical interactions, and generate corresponding car information summary feature vectors respectively through the trained summary feature vector generation model. S822. Calculate the mean of the feature vector of the car information summary to obtain the user's preference feature vector; S823. Input the user's preference feature vector, user access request information and role information into the first DNN network, and output the user request feature vector by the third Embedding module; S83: For the automotive information in each training sample, generate its corresponding automotive information summary feature vector through the summary feature vector generation model, and freeze the parameters of the summary feature vector generation model during the training process; S84: Calculate the similarity score between the user request feature vector and the car information summary feature vector, and optimize the model parameters based on the similarity score and sample labels using a binary classification loss function.
9. An artificial intelligence-based vehicle information management system, characterized in that, The method is used to execute the artificial intelligence-based vehicle information management method according to any one of claims 1-8, wherein the system includes: The data acquisition module is used to collect automotive information text. The first input module is used to input the automotive information text into a trained multi-title generation model and output automotive information titles corresponding to each preset role; wherein, the preset roles include automotive engineers and automotive consumers; The second input module is used to input the automotive information text into the summary generation model and output the automotive information summary. The first generation module is used to generate automotive information summary feature vectors based on the automotive information summary through a summary feature vector generation model, and index each automotive information summary feature vector into a vector retrieval engine. The second generation module is used to generate user request feature vectors based on the target user's reading behavior, target user access request information, and target user role, through a request feature vector generation model. The search module is used to use a vector retrieval engine to find the N automotive information summary feature vectors that are most similar to the user request feature vector, and to determine the corresponding N automotive information items based on the association between the automotive information summary feature vectors and their corresponding automotive information texts. The recommendation module is used to recommend the N pieces of automotive news to the target user based on the automotive news titles.