Intelligent driving document data processing method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of data processing, and discloses an intelligent driving document data processing method and device, an electronic device and a storage medium. The intelligent driving document data processing method comprises the following steps: firstly, obtaining original document data to be processed; then, extracting text features from the original document data and converting the text features into continuous vector representations capable of representing the implied context semantic correlation information; on this basis, determining a target document from the original document data according to the matching degree between preset query features and the continuous vector representations; and finally, performing semantic analysis on the target document according to a preset information analysis dimension to obtain a corresponding analysis result. The application can effectively improve the accuracy of document semantic matching and the automation degree of document content analysis.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to methods, devices, electronic devices, and storage media for processing intelligent driving literature data. Background Technology

[0002] In document data processing scenarios, relevant technologies typically use explicit rules or literal keywords to determine document relevance, filtering target documents by detecting the presence of preset words. However, this approach relies solely on surface-level word matching and fails to effectively capture the implicit contextual semantic relationships within the text. This results in documents with similar semantics but different expressions being difficult to accurately identify, leading to low accuracy in relevance assessments. Furthermore, after document screening, existing technologies often lack automated semantic parsing capabilities, still requiring manual reading to obtain structured information, resulting in low processing efficiency.

[0003] Therefore, there is an urgent need for a method for processing literature data in intelligent driving to solve the problems of insufficient accuracy in determining the relevance of literature and low degree of automation in the analysis of literature content in related technologies. Summary of the Invention

[0004] This application provides a method, device, electronic device, and storage medium for processing intelligent driving literature data, in order to solve the problems of insufficient accuracy in determining the relevance of literature and low degree of automation in the analysis of literature content in related technologies.

[0005] Firstly, this application provides a method for processing intelligent driving literature data, the method comprising: Obtain the raw literature data to be processed; Extract text features from the original document data and convert the text features into corresponding continuous vector representations. The continuous vector representations are used to characterize the implicit contextual semantic association information of the text features. Based on the matching degree between preset query features and continuous vector representation, target documents are determined from the original document data; Based on the preset information analysis dimensions, semantic analysis is performed on the target document to obtain the analysis results.

[0006] In one alternative implementation, textual features of the original document data are extracted, including: Based on the joint determination of exclusion criteria for excluding non-target domain literature and retention criteria for retaining target domain literature, domain relevance filtering is performed on the original literature data to obtain a candidate literature set; Extract textual features from the candidate document set.

[0007] In one alternative implementation, the joint determination includes: The test examines whether the text to be judged in each document in the original literature data meets the exclusion criteria. If the exclusion criteria are met, then the literature that meets the exclusion criteria will be excluded. For documents that were not excluded, check whether the text to be judged in the documents that were not excluded meets the retention criteria; If the retention criteria are met, the documents that meet the retention criteria will be retained in the candidate document set.

[0008] In one optional implementation, the exclusion condition is that the text to be judged contains negative keywords, and the negative keywords include at least one of the characteristic words in the fields of medicine, biology, remote sensing, agriculture, indoor robots and robotic arm control. The retention condition is that the text to be judged contains positive keywords, which include at least one of the following: autonomous driving, object detection, BEV perception, point cloud, multi-sensor fusion, motion prediction, world model, and occupancy prediction related terms.

[0009] In one optional implementation, the preset query features are generated in the following manner: The preset query text is encoded using the same vector encoding model as that used in text feature transformation.

[0010] In one optional implementation, target documents are determined from the original document data based on the matching degree between preset query features and continuous vector representations, including: The candidate document set was sorted from highest to lowest matching degree; Select a predetermined number of documents as target documents corresponding to the target date; If the number of documents corresponding to the target date is less than the preset number, supplementary documents will be selected from the historical documents in the candidate document set until the preset number is reached.

[0011] In one alternative implementation, the analysis results include at least one of the following: paper abstract, research task, core innovation points, methodological structure, advantages analysis, limitations analysis, and recommendation index.

[0012] Secondly, this application provides an intelligent driving document data processing device, the device comprising: The acquisition module is used to acquire the raw literature data to be processed; The conversion module is used to extract text features from the original document data and convert the text features into corresponding continuous vector representations. The continuous vector representations are used to characterize the implicit contextual semantic association information of the text features. The matching module is used to determine target documents from the original document data based on the matching degree between preset query features and continuous vector representations; The parsing module is used to perform semantic parsing on the target document according to the preset information parsing dimensions, and obtain the analysis results of the target document.

[0013] Thirdly, this application provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the intelligent driving document data processing method of the first aspect or any corresponding embodiment described above.

[0014] Fourthly, this application provides a computer-readable storage medium storing computer instructions for causing a computer to execute the intelligent driving document data processing method described in the first aspect or any corresponding embodiment thereof.

[0015] The intelligent driving literature data processing method of this application continuously vectorizes the textual features of the original literature data, enabling the implicit contextual semantic association information of the text to be quantified and expressed. This allows subsequent relevance determination to be performed at the semantic level rather than at the lexical surface level, which is beneficial to improving the accuracy of target literature identification. After the target literature is identified, semantic analysis is further performed on it according to the preset information analysis dimensions, which can automatically output structured analysis results, avoiding the dependence on manual reading of each article, thereby improving the automation level and processing efficiency of literature content analysis. Attached Figure Description

[0016] 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.

[0017] Figure 1 This is a schematic flowchart of the first method for processing intelligent driving literature data according to an embodiment of this application; Figure 2 This is a schematic diagram of a second process for processing intelligent driving literature data according to an embodiment of this application; Figure 3 This is a structural block diagram of an intelligent driving document data processing device according to an embodiment of this application; Figure 4 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0020] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained.

[0021] arXiv: An open-access academic preprint platform where academic papers in fields such as computer science, physics, and mathematics can be published and accessed.

[0022] BEV perception: Bird's-eye view perception is a perception method in the field of autonomous driving that projects data from multiple sensors onto a top-down two-dimensional plane for environmental understanding, which helps to complete tasks such as target detection and map building more accurately.

[0023] Occupation network: a technique for understanding 3D scenes that enables refined modeling of complex driving environments by predicting whether each voxel (3D pixel unit) in space is occupied and its semantic category.

[0024] World Model: An AI model capable of internally modeling and predicting environmental states, which can be used in autonomous driving scenarios to predict future scene evolution and assist in planning and decision-making.

[0025] Text vector coding: a technique that converts natural language text into fixed-dimensional numerical vectors, enabling computers to understand and compare the semantic content of text in a numerical manner. Texts with similar semantics have vectors that are closer in distance in a high-dimensional space.

[0026] Semantic vector: A numerical vector generated by a text vector encoding model that can represent the implicit semantic information of a text. It is the basic data form for performing semantic similarity calculations.

[0027] Cosine similarity / dot product similarity: two commonly used methods for calculating vector similarity, used to measure the degree of directional similarity between two vectors in a high-dimensional space. The higher the value, the more similar the semantics of the corresponding texts.

[0028] SQLite: A lightweight, local relational database that can run without a separate server process, suitable for storing structured data in single-machine or small-scale systems.

[0029] Paper-QA: An intelligent analysis tool with the ability to understand academic literature. It can perform question-and-answer content parsing on the input paper text and output structured analysis results such as research tasks, core innovations, and methodological structure.

[0030] Positive keywords / negative keywords: Positive keywords are words that identify the core semantic direction of the target field and are used to retain relevant literature; negative keywords are words that identify non-target field and are used to exclude irrelevant literature. The two are used together to form a two-way filtering mechanism.

[0031] As an optional application scenario of this application embodiment, the intelligent driving literature data processing method can be deployed in a system that includes at least one terminal device and at least one server. The system may include the terminal device and the server, with the terminal device connected to the server via a network.

[0032] The terminal device can be a desktop computer, laptop, workstation, or other device with data interaction capabilities, or other electronic devices with network connectivity. The server can be a standalone physical server, a server cluster or distributed system, or a cloud server providing cloud computing services, used to host the core computing tasks of the intelligent driving literature data processing method. The network can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranets, local area networks (LANs), wide area networks (WANs), and combinations thereof.

[0033] In document data processing scenarios, related technologies typically rely on literal keyword matching to determine the relevance of documents, making it difficult to capture deep semantic connections within the text. This leads to documents that are semantically similar to the query topic but have different wording being easily missed. Furthermore, after document screening, there is a lack of ability to automatically and structurally analyze the document content, requiring manual reading of each document to obtain effective information. To address the aforementioned technical problems of inaccurate document relevance determination and low automation of content analysis, this application provides an intelligent driving document data processing method to improve the accuracy of document relevance determination and achieve automated and structural analysis of document content.

[0034] According to an embodiment of this application, an embodiment of a method for processing intelligent driving literature data is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment provides a method for processing intelligent driving literature data, which can be used with the aforementioned servers, such as independent physical servers, cloud servers, or distributed server clusters. Figure 1 This is a flowchart of an intelligent driving document data processing method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain the original literature data to be processed.

[0036] Raw literature data refers to the collection of original information from external data sources that has not yet been processed. The content format of the raw literature data can be diverse, and may include structured or semi-structured information such as the title, abstract, main text, publication date, and source links of the literature.

[0037] "Acquisition" refers to the process by which the server reads literature information from external data sources into the local processing flow through a preset data access method. Specifically, acquisition methods may include, but are not limited to: accessing the data interface of an external literature platform via a network request to retrieve literature information, reading stored literature records from a local or remote database, or loading batch literature data via file reading.

[0038] In this step, the server reads literature information from one or more literature data sources according to a preset data acquisition strategy, obtaining a set of raw literature data to be processed. This raw literature data set may contain a varying number of literature entries, each corresponding to a single document and its related information. It should be noted that this step does not limit the specific form of the data source; any source that can provide literature information suitable for subsequent processing can be used as the source of the raw literature data.

[0039] Step S102: Extract the text features of the original document data and convert the text features into corresponding continuous vector representations. The continuous vector representations are used to characterize the implicit contextual semantic association information of the text features.

[0040] Textual features refer to the textual representations extracted from original document data that reflect the semantic information of the document's content. For example, for a document, the words, phrases, and their combinations contained in its title and abstract all belong to the document's textual features.

[0041] Continuous vector representation refers to the numerical vector obtained by mapping the above text features to a high-dimensional continuous real number space. Each document's text features correspond to a fixed-dimensional numerical vector, and the position and direction of this vector in space can reflect the semantic content of the document.

[0042] Implicit contextual semantic information refers to the deep semantic relationships implied in a text beyond its literal words. For example, although "vehicle perception" and "obstacle detection" are different in meaning, they are highly related in semantics. This relationship can be captured and quantified through continuous vector representation.

[0043] "Extraction" refers to identifying and selecting text content that is meaningful for subsequent processing from the original document data; "conversion" refers to using a text encoding model to map the extracted text features from character sequence form to numerical vector form. Specifically, the conversion process can be achieved through a text embedding model, which can encode the input text content into a fixed-dimensional numerical vector. Vectors corresponding to semantically similar texts are closer in high-dimensional space, while vectors corresponding to texts with significant semantic differences are farther apart.

[0044] In this step, the server extracts the textual features of each document record in the original document dataset and converts these features into corresponding continuous vector representations using a pre-defined text embedding model, forming a set of document semantic vectors. Through this conversion process, the natural language information in the original document data is transformed into semantic vectors that can be quantified by a computer, providing a numerical basis for subsequent semantic-level document matching.

[0045] Step S103: Based on the matching degree between the preset query features and the continuous vector representation, the target documents are determined from the original document data.

[0046] Query features refer to input information used to express the semantic direction of the retrieval target. They can be in the form of a natural language descriptive text, which, after the same encoding process as in step S102, yields the corresponding query vector.

[0047] Matching degree refers to the degree of similarity between the query vector and the continuous vector representation of each document in a high-dimensional semantic space. It can be quantified by vector similarity calculation methods such as dot product similarity and cosine similarity. The higher the matching degree value, the closer the semantic content of the corresponding document is to the query features.

[0048] In this step, the server first encodes the preset query features to obtain a query vector; then, it calculates the matching score between the query vector and the semantic vectors of each document generated in step S102; finally, it sorts the original document data from high to low according to the matching score and determines the target documents according to preset rules. It can be understood that the target documents can be determined by selecting documents with a matching score exceeding a preset threshold, or by selecting a preset number of documents with high matching scores. By performing similarity comparison in the semantic vector space, this step can identify documents that are highly semantically related to the query features, rather than being limited to documents that share the same words as the query text.

[0049] Step S104: Based on the preset information parsing dimensions, perform semantic parsing on the target document to obtain the analysis results of the target document.

[0050] Information parsing dimensions refer to pre-defined, structured analytical frameworks used to guide the semantic parsing process. These dimensions define the specific aspects to focus on when understanding the content of a target document. For example, information parsing dimensions may include the core research content, main methodological descriptions, and concluding information. Different information parsing dimension settings will guide the semantic parsing process to extract different levels of content information from the document, thereby outputting analytical results adapted to application needs.

[0051] In this step, the server performs semantic parsing on each target document identified in step S103 according to preset information parsing dimensions. That is, for each parsing dimension, the server identifies and extracts the corresponding semantic information from the text content of the target document, and finally integrates and outputs structured analysis results. Through this process, the scattered natural language content in the target document is transformed into structured information organized by dimensions, which facilitates subsequent storage, display, or further processing.

[0052] In summary, the intelligent driving literature data processing method provided in this application converts the textual features of the literature into continuous vector representations that can characterize implicit contextual semantic association information, enabling the determination of literature relevance to be matched at the semantic level rather than the literal level, which is beneficial to improving the accuracy of target literature identification. After identifying the target literature, it further performs semantic analysis based on preset information analysis dimensions and automatically outputs structured analysis results, which helps to reduce the reliance on manual reading and improve the processing efficiency of literature content analysis.

[0053] This embodiment provides a method for processing intelligent driving literature data, which can be used with the aforementioned servers, such as independent physical servers, cloud servers, or distributed server clusters. Figure 2 This is a flowchart of an intelligent driving document data processing method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain the original literature data to be processed.

[0054] In this step, the processing entity (such as a server) reads literature information from one or more literature data sources according to a preset data collection strategy to obtain a set of raw literature data to be processed. This set may contain a varying number of literature entries, each corresponding to a literature document and its related information.

[0055] Step S202: Extract the text features of the original document data and convert the text features into corresponding continuous vector representations. The continuous vector representations are used to characterize the implicit contextual semantic association information of the text features.

[0056] Before formally performing text feature extraction and vectorization transformation, the original literature data can be filtered for domain relevance to improve the efficiency and accuracy of subsequent processing. Specifically, this includes: Step S2021: Based on the joint determination of the exclusion conditions for excluding non-target domain literature and the retention conditions for retaining target domain literature, domain relevance filtering is performed on the original literature data to obtain a candidate literature set.

[0057] The term "joint judgment" refers to a two-stage sequential judgment of literature by simultaneously introducing exclusion and retention criteria, rather than relying solely on a single criterion. Specifically, the joint judgment process is as follows: First, the text to be judged in each document in the original literature data is checked to see if it meets the exclusion criteria. The text to be judged can be a complete text paragraph formed by combining the title and abstract of the document, so as to comprehensively utilize the topic direction reflected in the title and the methods, tasks and application background supplemented by the abstract, thereby improving the accuracy of the judgment.

[0058] The exclusion criterion is that the text to be judged contains negative keywords. Negative keywords include at least one of the characteristic words in the fields of medicine, biology, remote sensing, agriculture, indoor robotics, and robotic arm control. Negative keywords are used to mark domain feature words that are semantically similar to the target domain but not actually related, such as "tumor," "gene sequence," "satellite image," and "crop detection." If any of the above negative keywords are detected in the text to be judged, the corresponding document will be excluded and will not proceed to the next processing step. It should be noted that the specific content of negative keywords can be flexibly configured according to the actual application scenario and is not limited to the examples above. As long as the words can effectively distinguish between the target domain and non-target domain, they can be used as negative keywords.

[0059] Secondly, for the documents that were not excluded, their texts to be judged are checked to see if they meet the retention criteria. The retention criteria require that the text contains positive keywords, which include at least one of the following: autonomous driving, object detection, BEV perception, point cloud, multi-sensor fusion, motion prediction, world model, and occupancy prediction. Positive keywords are used to mark the core semantic direction of the target domain, such as "autonomous driving," "BEV," "LiDAR point cloud," and "multimodal fusion." If the text to be judged meets the retention criteria, the corresponding document is retained in the candidate document set as input for subsequent processing.

[0060] By employing the two-stage joint decision-making mechanism of first performing exclusion condition detection and then performing retention condition detection, the interference of documents that are semantically similar to the target domain but actually unrelated to the candidate results can be effectively reduced, thereby improving the domain purity of the candidate document set and providing more accurate input for subsequent vectorization processing and relevance ranking.

[0061] Step S2022: Extract the text features of each document in the candidate document set and convert them into corresponding continuous vector representations using a preset text embedding model.

[0062] Text embedding models can integrate contextual information of words during the encoding process, so that the vectors corresponding to semantically similar texts are closer in high-dimensional space, thus providing a numerical basis for subsequent semantic relevance matching.

[0063] Step S203: Based on the matching degree between the preset query features and the continuous vector representation, the target documents are determined from the original document data.

[0064] To ensure that query features and document continuous vector representations reside in the same semantic space, thus making the matching degree calculation meaningful, query features are obtained by encoding a pre-defined query text using the same vector encoding model employed in text feature transformation. For example, if a specific sentence vector encoding model is used in the document vectorization stage, the query text is also encoded using this model, resulting in a query vector with the same dimension and unified semantic space as the document vector. The pre-defined query text can be flexibly set according to the actual retrieval objective, such as "autonomous driving," or, depending on a specific sub-direction, "BEV perception," "sensor fusion," or "world model," to meet the retrieval needs of different application scenarios. Encoding the query text using the same vector encoding model ensures the consistency of semantic measurement between the query vector and the document vector, thereby making the matching degree calculation results more accurate and reliable.

[0065] Matching degree refers to the degree of similarity between the query feature vector and the continuous vector representation of each document in a high-dimensional semantic space. It can be quantified by dot product similarity, cosine similarity, or other vector similarity calculation methods. After calculating the matching degree, the candidate document set is sorted from high to low according to the matching degree, and the target document is determined from it.

[0066] Regarding the selection strategy for target literature, a preset number of articles are selected from the literature corresponding to the target date. The target date can be understood as the time range of literature corresponding to a specific time range of interest to the user (e.g., the current day or the previous day). The preset number can be configured according to actual needs, for example, set to 10 articles. When the number of articles corresponding to the target date is less than the preset number, supplementary selections are made from historical articles in the candidate literature set until the preset number is reached. The above selection strategy takes into account both the timeliness and completeness of the recommendation results: prioritizing the selection of literature within the target date helps ensure that the recommended content can reflect the latest research developments in a timely manner; introducing historical articles as a supplement when the number of articles within the target date is insufficient ensures the stability of the number of recommendations each time, avoiding the situation where there are too few recommendations due to a small number of articles on a certain date.

[0067] Step S204: Based on the preset information parsing dimensions, perform semantic parsing on the target document to obtain the analysis results of the target document.

[0068] In this embodiment, the analysis results include at least one of the following: paper abstract, research task, core innovation points, methodological structure, advantages analysis, limitations analysis, and recommendation index.

[0069] Specifically, the abstract is a concise summary of the core content of the paper; the research task refers to the specific problem or goal addressed by the paper; the core innovation points refer to the novelty of the paper compared to existing work in terms of methods, models, or applications; the methodological structure refers to the technical route and module composition adopted by the paper; the advantages analysis refers to the superiority of the paper's solution in terms of performance, efficiency, or applicability compared to similar work; the limitations analysis refers to the shortcomings of the paper's solution in terms of applicable scenarios, assumptions, or performance indicators; and the recommendation index is a score of the paper's reference value output based on a comprehensive evaluation of the above dimensions. These information analysis dimensions can be flexibly selected and combined according to actual application needs, and are not limited to the dimension types listed above.

[0070] Semantic parsing refers to the use of analytical tools with text understanding capabilities to identify, extract, and integrate corresponding semantic information from the text content of a target document for each preset information parsing dimension, ultimately outputting structured analysis results organized by dimension. By performing semantic parsing on target documents according to preset information parsing dimensions, the scattered natural language content in the target document can be automatically transformed into structured information, allowing users to quickly obtain the core content of each dimension without having to read the entire document, thereby effectively improving the automation level of document content analysis and information acquisition efficiency.

[0071] In summary, the intelligent driving literature data processing method provided in this application introduces a joint judgment mechanism based on exclusion and retention conditions to perform two-stage domain relevance filtering on the original literature data. This effectively eliminates interference from non-target domain literature before formal vectorization processing, improving the domain purity of the candidate literature set. Furthermore, by converting the textual features of candidate literature into continuous vector representations that characterize implicit contextual semantic relationships, and using the same vector encoding model as the literature encoding to encode the query text, it ensures that the query features and the literature vectors are in the same semantic space. This allows relevance matching to be performed at the semantic level rather than the literal level. This effectively improves the accuracy of target document identification, enabling the accurate identification of documents with similar semantics to the query topic but different wording. Furthermore, by prioritizing selection from documents corresponding to the target date and supplementing from historical documents when the number is insufficient, the timeliness of the recommendation results is ensured while maintaining quantity stability. Finally, by performing semantic analysis on the target documents based on preset information analysis dimensions, the system automatically outputs structured analysis results covering dimensions such as paper abstracts, research tasks, core innovations, methodological structure, advantages analysis, limitations analysis, and recommendation index. This effectively reduces the reliance on manual reading of each document and improves the automation level and overall processing efficiency of document content analysis.

[0072] To better illustrate the intelligent driving literature data processing method of this application, a preferred embodiment will be provided below. This embodiment is intended to describe the implementation process of this application in detail, but is not intended to limit the scope of protection of this application.

[0073] This embodiment uses the automatic daily paper scraping, semantic ranking, analysis, and Lark group push based on the arXiv platform for the field of intelligent driving as a specific preferred example for illustration. This example is chosen because the field of intelligent driving generates a large number of new papers on the arXiv platform daily, making the need for automated paper screening, ranking, and analysis typical and urgent, and fully demonstrating the practical value of each technical aspect of this application embodiment. It should be noted that this application is not limited to this specific example; the paper source platform, target field, and push channel can be flexibly replaced and expanded according to actual needs. Furthermore, this application can be implemented as an intelligent driving literature data processing method, or correspondingly as an electronic device or computer-readable storage medium carrying the above method. In this embodiment, the method flow will be the main focus for comprehensive explanation.

[0074] Against the backdrop of rapid development in areas such as autonomous driving perception, prediction and planning, multi-sensor fusion, BEV perception, occupancy networks, world models, and data closed loops, the arXiv platform publishes a large number of new papers related to or closely related to intelligent driving every day. For autonomous driving algorithm engineers and research teams, timely discovery of highly relevant papers and rapid understanding of their core value are of great significance for technology pre-research and R&D decisions. However, existing methods of obtaining papers mainly rely on manual browsing, subscription push, or simple keyword matching, which is not only time-consuming and labor-intensive, but also suffers from problems such as mistakenly including papers from non-target fields and missing papers with different wording but semantically related ones. Furthermore, they lack the ability to perform automated in-depth analysis of paper content, failing to meet the actual needs of teams for continuous and efficient tracking of cutting-edge papers. To address this, this application provides a paper processing method for the field of intelligent driving, which includes the following steps: The first step is to obtain information about the paper.

[0075] Following a pre-defined scheduling strategy, data requests are periodically sent to the arXiv platform to automatically retrieve newly added paper information under the target category. arXiv is a globally accessible academic preprint platform where researchers publish their latest papers daily. This embodiment selects three categories—Computer Vision (CV), Robotics, and Artificial Intelligence (AI)—as the retrieval sources to cover core research directions in the field of autonomous driving. After obtaining the paper information, further filtering is performed based on the submission date, prioritizing papers added within the target date range (e.g., the previous day) to ensure the timeliness and novelty of subsequent recommendation results. The retrieved information for each paper includes at least the paper title, abstract, source link, and submission date; this information serves as the basic input for subsequent filtering, vectorization, and analysis.

[0076] The second step is domain relevance filtering.

[0077] Due to the massive number of new papers added daily on the arXiv platform, and the fact that categories such as computer vision and robotics contain a large number of papers that are semantically similar to but practically unrelated to autonomous driving (such as medical image processing, agricultural remote sensing, and indoor service robots), directly performing vectorization on all papers would introduce a significant amount of noise, reducing the accuracy of subsequent recommendations. Therefore, before vectorization, domain relevance filtering is performed on the original paper data, specifically employing a joint judgment mechanism combining positive and negative keywords.

[0078] First, the title and abstract of each paper are combined to form the text to be judged. The reason for using the title and abstract together as the judgment criteria is that the title can reflect the topic direction of the paper, and the abstract can supplement the paper's methods, tasks and application background. The combination of the two helps to improve the accuracy of the filtering judgment.

[0079] The joint determination is carried out in two phases: The first stage is negative keyword detection (excluding documents from non-target fields): The system detects whether the text to be judged contains preset negative keywords. Negative keywords are representative words that can identify non-target fields, including words representing fields such as medicine, biology, physics, remote sensing, agriculture, indoor robotics and robotic arm control, such as "tumor," "gene," "satellite image," and "crops." If any negative keyword is detected, the paper is directly excluded and will not proceed to the next processing step.

[0080] The second stage, positive keyword matching (retaining literature in the target domain): For papers not excluded by negative keywords, the system further checks whether the text to be judged contains at least one positive keyword. Positive keywords refer to words that can identify the core semantic directions in the field of intelligent driving, including related representation words in directions such as autonomous driving, object detection, semantic segmentation, depth estimation, motion prediction, lane detection, BEV (bird's-eye view perception), occupancy prediction, point cloud, multi-sensor fusion, world model, neural rendering, active learning, and domain adaptation. If the text to be judged contains at least one positive keyword, the paper is retained as a candidate paper related to intelligent driving.

[0081] By combining the above two stages of judgment, papers irrelevant to the target domain can be effectively filtered out before the formal semantic vectorization is performed, thereby improving the domain purity of the candidate paper set and providing more accurate input for subsequent processing.

[0082] The third step is text vector encoding and database storage.

[0083] After filtering by domain relevance, text vector encoding is performed on the titles and abstracts of candidate papers to obtain a semantic vector for each paper. Text vector encoding refers to using a text embedding model to transform natural language text into a fixed-dimensional continuous numerical vector. This vector can represent the implicit contextual semantic relationships of the text in a high-dimensional semantic space, and the vectors corresponding to semantically similar texts are closer in space. The text embedding model can be a sentence vector model or other embedding models suitable for semantic retrieval.

[0084] After vectorization, the title, abstract, source link, submission date, and corresponding semantic vector of each paper are written as a complete record to the database. The database can be a lightweight local database such as SQLite, or other relational or vector databases. A unique identifier record is created for each paper. If a unique record corresponding to the current paper already exists in the database, the duplicate writing is skipped to avoid duplicate recommendations in the future.

[0085] In addition, to improve the readability of the content pushed to the group, the English titles of newly added papers are translated into Chinese, and the translation results are updated to the corresponding records in the database. The translation task is executed asynchronously, that is, after the papers are crawled and stored in the database, an independent task is started to translate the English titles that have not yet been translated, avoiding the translation process from blocking the main crawling process, thereby improving the overall task execution efficiency.

[0086] The fourth step is query vector generation and semantic similarity ranking.

[0087] After the candidate papers are vectorized and stored, the preset query text is encoded to obtain the query vector. To ensure that the query vector and the semantic vector of the candidate papers are in the same semantic space and that their similarity calculation results are actually comparable, the query text is encoded using the same text embedding model as the paper vectorization stage. The preset query text can be flexibly set according to the team's focus. For example, it can be set to "autonomous driving," or refined to specific sub-directions such as "BEV perception," "sensor fusion," "world model," or "occupancy prediction," to meet the recommendation needs of different teams in different application scenarios.

[0088] Subsequently, the semantic vectors of each candidate paper in the database are read, and the similarity score between the query vector and the semantic vector of each paper is calculated. The similarity can be calculated using cosine similarity, dot product similarity, or other vector similarity methods. A higher value indicates a stronger semantic association between the corresponding paper and the query topic.

[0089] Candidate papers are sorted from highest to lowest based on similarity scores. A predetermined number of papers (e.g., 10) with the highest similarity are selected from the subset of candidate papers corresponding to the target date as the target paper set. If the number of candidate papers corresponding to the target date is less than the predetermined number, supplementary selections are made from historical candidate papers over a larger time range until the predetermined number is reached, ensuring the completeness and stability of the quantity of results pushed each time.

[0090] Step 5: Automated paper analysis.

[0091] Automated content analysis is performed on each paper in the target paper collection. Specifically, paper titles, abstracts, and other information are input into a paper analysis tool (such as Paper-QA), which then performs literature understanding and structured information extraction. Paper-QA is an intelligent tool with question-and-answer analysis capabilities for academic literature, capable of identifying and extracting corresponding semantic information from the paper text for preset analysis dimensions.

[0092] In this embodiment, the following structured analysis results are extracted from each target paper: paper abstract (a concise summary of the core content of the paper), research task (the specific problem or goal addressed by the paper), core innovation points (the novelty of the paper in terms of method or application), methodological architecture (the technical route and module composition adopted by the paper), advantage analysis (the superiority over similar works), limitation analysis (the shortcomings of the solution in terms of applicable scenarios or performance indicators), and recommendation index (a reference value score output based on a comprehensive evaluation of each dimension).

[0093] Through automated paper analysis, the core reading conclusions of each paper can be provided to users simultaneously during the recommendation stage, avoiding users judging the value of a paper based solely on its title and significantly improving the efficiency of paper tracking.

[0094] Step 6: Structured content organization and group push.

[0095] The target set of papers and their corresponding automated analysis results are organized into group message text. The group message content may include the paper's recommendation number, Chinese title, original English title, source link, abstract summary, description of core innovations, and recommendation index, enabling team members to quickly assess the reference value of each paper upon receiving the message. Subsequently, this structured message is automatically pushed to the target group via a group robot interface, allowing team members to receive the latest research updates daily without manual intervention.

[0096] Step 7: Log recording, failure retry, and exception alerts.

[0097] The entire process of crawling, filtering, translating, analyzing, and pushing is logged. The log information includes task start and end time, target date, total number of papers crawled, number of papers filtered, number of new papers added to the database, number of papers translated, and abnormal information at each stage, so as to support subsequent operation status review and problem investigation.

[0098] If any processing step experiences interface rate limiting, network anomaly, database anomaly, or service anomaly, automatic retries will be performed according to the preset number of retries and retry interval. If the maximum number of retries is reached and the problem is still not resolved, an error log will be generated and a failure message will be sent.

[0099] Step 8: External query interface service.

[0100] In some implementations, interface services can also be provided externally. For example, a web interface can be used to receive user-input query text, invoke a paper retrieval module to perform semantic retrieval on candidate papers in the database, further perform automated analysis on the retrieved papers, and finally return the analysis results to the caller or push them to the target group. In this way, this embodiment not only supports daily automatic scheduled pushes but also responds to on-demand interactive query requests, further enhancing flexibility and practicality.

[0101] In summary, this application integrates timed automatic paper retrieval, combined positive and negative keyword filtering, semantic vectorization encoding, similarity ranking, automated structured analysis, and group robot push into an end-to-end automated processing flow. This reduces the cost of manual retrieval while effectively improving the accuracy of paper selection in the target field, the semantic relevance of the recommendation results, and the automation level of paper content analysis. This enables team members to obtain filtered, ranked, and deeply analyzed cutting-edge paper abstracts daily at low cost and high efficiency.

[0102] This embodiment also provides an intelligent driving document data processing device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0103] This embodiment provides an intelligent driving document data processing device, such as... Figure 3 As shown, it includes: The acquisition module 301 is used to acquire the original literature data to be processed; The conversion module 302 is used to extract the text features of the original document data and convert the text features into corresponding continuous vector representations. The continuous vector representations are used to characterize the implicit contextual semantic association information of the text features. The matching module 303 is used to determine the target document from the original document data based on the matching degree between the preset query features and the continuous vector representation; The parsing module 304 is used to perform semantic parsing on the target document according to the preset information parsing dimensions, and obtain the analysis results of the target document.

[0104] In one alternative implementation, the conversion module 302 is configured to: Based on the joint determination of exclusion criteria for excluding non-target domain literature and retention criteria for retaining target domain literature, domain relevance filtering is performed on the original literature data to obtain a candidate literature set. Extract textual features from the candidate document set.

[0105] In one alternative implementation, the conversion module 302 is configured to: The test examines whether the text to be judged in each document in the original literature data meets the exclusion criteria. If the exclusion criteria are met, then the literature that meets the exclusion criteria will be excluded. For documents that were not excluded, check whether the text to be judged in the documents that were not excluded meets the retention criteria; If the retention criteria are met, the documents that meet the retention criteria will be retained in the candidate document set.

[0106] In one optional implementation, the exclusion condition is that the text to be judged contains negative keywords, and the negative keywords include at least one of the characteristic words in the fields of medicine, biology, remote sensing, agriculture, indoor robots and robotic arm control. The retention condition is that the text to be judged contains positive keywords, which include at least one of the following: autonomous driving, object detection, BEV perception, point cloud, multi-sensor fusion, motion prediction, world model, and occupancy prediction related terms.

[0107] In one optional implementation, the preset query features are generated in the following manner: The preset query text is encoded using the same vector encoding model as that used in text feature transformation.

[0108] In one alternative implementation, the matching module 303 is configured to: The candidate document set was sorted from highest to lowest matching degree; Select a predetermined number of documents as target documents corresponding to the target date; If the number of documents corresponding to the target date is less than the preset number, supplementary documents will be selected from the historical documents in the candidate document set until the preset number is reached.

[0109] In one alternative implementation, the analysis results include at least one of the following: paper abstract, research task, core innovation points, methodological structure, advantages analysis, limitations analysis, and recommendation index.

[0110] The intelligent driving document data processing device provided in this application embodiment can execute the intelligent driving document data processing method provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0111] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0112] The following is a detailed reference. Figure 4 The diagram illustrates a structural schematic suitable for implementing the electronic device described in the embodiments of this application. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from memory 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device. The processor 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0113] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 408 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0114] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 409, or installed from a memory 408, or installed from a ROM 402. When the computer program is executed by the processor 401, it performs the functions defined in the intelligent driving document data processing method of embodiments of this application.

[0115] Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0116] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the intelligent driving document data processing method shown in the above embodiments is implemented.

[0117] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0118] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for processing intelligent driving literature data, characterized in that, The method includes: Obtain the raw literature data to be processed; The text features of the original document data are extracted and converted into corresponding continuous vector representations, which are used to characterize the implicit contextual semantic association information of the text features; Based on the matching degree between the preset query features and the continuous vector representation, the target documents are determined from the original document data; Based on the preset information parsing dimensions, the target document is semantically parsed to obtain the analysis results of the target document.

2. The method according to claim 1, characterized in that, The extraction of text features from the original document data includes: Based on the joint determination of exclusion criteria for excluding documents from non-target domains and retention criteria for retaining documents from target domains, domain relevance filtering is performed on the original document data to obtain a candidate document set. Extract textual features from the candidate document set.

3. The method according to claim 2, characterized in that, The joint determination includes: Detect whether the text to be judged in each document in the original document data meets the exclusion conditions; If the exclusion criteria are met, the literature that meets the exclusion criteria will be excluded. For documents that have not been excluded, check whether the text to be judged in the documents that have not been excluded meets the retention conditions; If the retention conditions are met, the documents that meet the retention conditions will be retained in the candidate document set.

4. The method according to claim 3, characterized in that, The exclusion condition is that the text to be judged contains negative keywords, and the negative keywords include at least one of the characteristic words in the fields of medicine, biology, remote sensing, agriculture, indoor robots and robotic arm control. The retention condition is that the text to be judged contains positive keywords, and the positive keywords include at least one of the following: autonomous driving, object detection, BEV perception, point cloud, multi-sensor fusion, motion prediction, world model, and occupancy prediction related characterization words.

5. The method according to claim 2, characterized in that, The preset query features are generated in the following way: The preset query text is encoded using the same vector encoding model as that used in the text feature transformation.

6. The method according to claim 5, characterized in that, The determination of target documents from the original document data based on the matching degree between preset query features and the continuous vector representation includes: The candidate document set is sorted from high to low according to the matching degree; Select a predetermined number of documents as target documents corresponding to the target date; If the number of documents corresponding to the target date is less than the preset number, supplementary documents are selected from the historical documents in the candidate document set until the preset number is reached.

7. The method according to claim 1, characterized in that, The analysis results include at least one of the following: paper abstract, research task, core innovation points, methodological structure, advantages analysis, limitations analysis, and recommendation index.

8. A smart driving document data processing device, characterized in that, The device includes: The acquisition module is used to acquire the raw literature data to be processed; The conversion module is used to extract the text features of the original document data and convert the text features into corresponding continuous vector representations, wherein the continuous vector representations are used to characterize the implicit contextual semantic association information of the text features; The matching module is used to determine target documents from the original document data based on the matching degree between preset query features and the continuous vector representation; The parsing module is used to perform semantic parsing on the target document according to a preset information parsing dimension, and obtain the analysis results of the target document.

9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the intelligent driving document data processing method according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the intelligent driving document data processing method according to any one of claims 1 to 7.

Images