Intelligent retrieval method and device based on multi-modal file semantic naming and medium

Abstract

The application relates to the technical field of natural language processing, and discloses an intelligent retrieval method and device based on multi-modal file semantic naming and a medium, which comprises the following steps: receiving a natural language query sentence input by a user, and preprocessing the natural language query sentence; matching the preprocessed natural language query sentence with semantic file names of multi-modal files in a preset knowledge base, determining a target semantic file name matched with the preprocessed natural language query sentence as a retrieval result based on a matching result, wherein the preset knowledge base is composed of multi-modal files with semantic file names, and the semantic file names comprise a file type identification field and a content description field; parsing a target file type from the target semantic file name, and calling a rendering component corresponding to the target file type to load and show file content corresponding to the target semantic file name to a target user. The application can be applied to intelligent retrieval scenes in the fields of financial technology and medical health.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an intelligent retrieval method, apparatus and medium based on multimodal document semantic naming. Background Technology

[0002] Currently, intelligent question-answering and knowledge base retrieval systems are widely used in scenarios such as financial customer service, education, and personal health record inquiries.

[0003] Currently, enterprise knowledge bases or document management systems support attaching files to knowledge entries. The knowledge retrieval workflow involves matching user questions with the full-text content of files in the knowledge base. After the retrieval, a text knowledge entry is created for the user, and then a video file is manually uploaded and associated with it before being returned to the user. However, during the full-text content retrieval process, the large amount of file data leads to long retrieval times and consumes significant retrieval resources. Furthermore, the video file and text knowledge returned to the user are separate entities, making management complex. Summary of the Invention

[0004] This invention provides an intelligent retrieval method, device, and medium based on multimodal semantic naming of files, which solves the problems of large retrieval workload and complex file management in the prior art. The main advantage is that it can improve information retrieval efficiency and simplify the retrieval management process.

[0005] According to a first aspect of the present invention, an intelligent retrieval method based on multimodal document semantic naming is provided, comprising: Receive a natural language query statement input by a target user, and preprocess the natural language query statement, wherein the preprocessing includes information completion; The preprocessed natural language query statement is correlated with the file type identifier field and content description field of multiple semantic file names in the preset knowledge base. Corresponding file type correlation calculation results and content description correlation calculation results are obtained. Based on the file type correlation calculation results and the content description correlation calculation results, the target semantic file name that matches the preprocessed natural language query statement is determined, and the target semantic file name is used as the search result. The preset knowledge base is composed of multimodal files with semantic file names, and the semantic file name includes a file type identifier field and a content description field. The target file type is parsed from the target semantic file name, and the rendering component corresponding to the target file type is determined. The rendering component is then used to load and display the file content corresponding to the target semantic file name.

[0006] According to a second aspect of the present invention, an intelligent retrieval device based on multimodal document semantic naming is provided, comprising: The statement acquisition unit is used to receive a natural language query statement input by a target user and preprocess the natural language query statement, wherein the preprocessing includes information completion. The filename matching unit is used to perform relevance calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic filenames in a preset knowledge base, and obtain the corresponding file type relevance calculation results and content description relevance calculation results. Based on the file type relevance calculation results and the content description relevance calculation results, the unit determines the target semantic filename that matches the preprocessed natural language query statement and uses the target semantic filename as the retrieval result. The preset knowledge base is composed of multimodal files with semantic filenames, and the semantic filenames include a file type identifier field and a content description field. The file content display unit is used to parse the target file type from the target semantic file name, determine the rendering component corresponding to the target file type, and use the rendering component to load and display the file content corresponding to the target semantic file name.

[0007] According to a fourth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described intelligent retrieval method based on multimodal file semantic naming.

[0008] The present invention provides an intelligent retrieval method, device, and medium based on semantic naming of multimodal files. Compared with the current method of information retrieval by matching user questions with the full text content of files in a knowledge base, the present invention retrieves information by matching natural language query statements with semantic filenames of multimodal files in a preset knowledge base. Since the filename data volume is small, lightweight retrieval can be achieved, thereby improving retrieval efficiency and avoiding the consumption of retrieval resources. By utilizing the file type inherent in the semantic filename, the client rendering engine can directly infer the file type by parsing the string, eliminating the need for additional metadata query and format detection steps, improving the content presentation efficiency. Moreover, regardless of whether the underlying file is a video, PDF, or image, the front end can automatically match the most suitable player or reader or other rendering components based on the parsed type, without requiring the user to manually download or select an opening method, thus optimizing the user's reading experience. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This diagram illustrates an application environment provided by an embodiment of the present invention. Figure 2 The flowchart of an intelligent retrieval method based on multimodal file semantic naming provided by an embodiment of the present invention is shown. Figure 3 The flowchart of another intelligent retrieval method based on multimodal file semantic naming provided by an embodiment of the present invention is shown; Figure 4 This diagram illustrates the structure of an intelligent retrieval device based on multimodal document semantic naming, as provided in an embodiment of the present invention. Figure 5 This diagram illustrates the structure of another intelligent retrieval device based on multimodal document semantic naming, provided by an embodiment of the present invention. Figure 6 A schematic diagram of the physical structure of a computer device provided in an embodiment of the present invention is shown. Detailed Implementation

[0010] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0011] In this embodiment of the invention, the intelligent retrieval method, device, and medium based on multimodal document semantic naming can be applied to intelligent customer service platforms in fields such as finance, healthcare, and insurance. Intelligent customer service platforms can replace customer service personnel in providing customer service. Specifically, intelligent customer service is a system that uses artificial intelligence technology to simulate the thinking and behavior of customer service personnel, providing automated and intelligent services to customers.

[0012] An intelligent customer service platform, as a type of intelligent question answering system (IQAS), is an artificial intelligence application that can understand questions posed by users in natural language and automatically retrieve, extract, and generate accurate answers from knowledge bases or data sources. For example, an intelligent customer service platform can integrate technologies such as Natural Language Processing (NLP), machine learning, knowledge graphs, and speech recognition. By understanding the questions entered by customers and automatically matching answers from the knowledge base, it can complete business processing, provide business consultations, and even handle emotion perception.

[0013] like Figure 1As shown, an intelligent customer service platform can include a client and a server. The client refers to the terminal through which users directly interact, responsible for data collection, display, and initial verification, and for sending user requests to the server. The client can also include applications, web pages, and operating systems deployed on the terminal device.

[0014] Depending on the business scenario corresponding to the application environment, clients can be of different types. For example, in a financial insurance claims scenario, clients can be divided into user clients and operation and maintenance management clients based on the user's target audience. User clients can be accessed on the user's terminal device through applications, mini-programs, web pages, etc. Operation and maintenance management clients can be accessed on the terminal devices used by operation and maintenance personnel through a backend management system (web) or a monitoring dashboard for operation and maintenance interactions.

[0015] When the application's business scenario is a healthcare service scenario, the client can be categorized based on the user: patient-side, doctor-side, operations and maintenance management-side, and integration-side. The patient-side allows user interaction via applications, mini-programs, or web pages on the terminal devices used by patients or related users (such as family members). The doctor-side allows user interaction via desktop applications, mobile workstation apps, or dedicated tablet interfaces on the terminal devices used by doctors or other healthcare-related users. The operations and maintenance management-side allows interaction via a backend management system (web) or monitoring dashboard on the terminal devices used by operations and maintenance personnel. The integration-side generally refers to terminal devices such as Hospital Information System (HIS) interfaces, medical insurance settlement terminals, and wearable device gateways, used for interactions related to performing integration functions.

[0016] Servers can be deployed in the cloud or data centers, responsible for core business processing, persistent data storage, security management, and integration with external systems. In some embodiments, the server may include one or more combinations of servers such as web servers, application servers, database servers, and storage servers.

[0017] The client and server can establish a communication connection via a network, thereby building an intelligent customer service platform. For example, when the intelligent customer service platform is applied to business scenarios such as finance and insurance, it can perform intelligent question-and-answer interaction based on the user's natural language text in the business scenario. By combining the user's question information with relevant knowledge retrieval based on business data generated in the business scenario, it can generate and provide feedback on financial and insurance business-related answers to meet business needs such as business consultation, business processing, and business recommendations.

[0018] Currently, the method of matching user questions with the full text of documents in the knowledge base for information retrieval results in long retrieval times and high resource consumption due to the large amount of data in the documents.

[0019] To address the aforementioned problems, embodiments of the present invention provide an intelligent retrieval method based on multimodal file semantic naming, such as... Figure 2 As shown, the method includes: 101. Receive the natural language query statement input by the target user, and preprocess the natural language query statement, including information completion.

[0020] In this embodiment of the invention, the natural language query statement input by the target user is first received through the input interface. The query statement can be in text form or in text form after speech transcription. To improve the quality of natural language queries, preprocessing can be performed first, such as information completion. Information completion can include specifying the required file type. This completion process utilizes a predefined keyword library for file types, covering common multimodal file types such as "video," "PDF," "image," and "audio," along with their synonyms or aliases. Upon receiving the query, word segmentation and entity recognition are performed, and the extracted words are matched against the keyword library to determine if the query explicitly contains keywords specifying the required file type. If the result is yes (containing keywords related to file type), the query is directly used as input for subsequent searches. If the result is no (not containing keywords related to file type, for example, if the user only entered "how to configure the server environment" without specifying the file type), a content completion interaction mechanism is triggered. In this case, instead of blindly performing a full database search, a clarifying question or option suggestion is generated, such as: "Would you like to view a video tutorial, a text document, or a code example?" The system then provides feedback to the target user through the user interface. Upon receiving feedback from the target user regarding this interaction, natural language understanding technology is used to analyze the user's intent. The parsed file type qualifiers are dynamically inserted or integrated into the original query statement, thereby structurally modifying the original statement. This ultimately generates a natural language query statement that includes both core search content and clearly defines the file type, providing a standardized input basis for subsequent accurate retrieval. This embodiment of the invention, by interactively modifying the user's input natural language query statement to specify the required file type, ensures that the retrieved files meet the user's file type requirements.

[0021] 102. Perform relevance calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic file names in the preset knowledge base, and obtain the corresponding file type relevance calculation results and content description relevance calculation results. Based on the file type relevance calculation results and content description relevance calculation results, determine the target semantic file name that matches the preprocessed natural language query statement, and use the target semantic file name as the search result. The preset knowledge base consists of multimodal files with semantic file names, and the semantic file name contains a file type identifier field and a content description field.

[0022] In this embodiment of the invention, a preset knowledge base is first constructed and maintained. For each multimodal file, such as video, audio, image, or PDF document, its core semantic information is extracted, and a unique semantic filename is generated according to a predefined naming convention. This semantic filename uses a structured natural language format, explicitly including key elements such as the file's content theme, applicable scenario (purpose or business behavior), and file format type. For example, a filename named "This is a video explaining the order payment steps.mp4" carries key semantics: content subject (order payment steps), file type (format, i.e., video), and purpose (explanation). These multimodal files with semantic filenames are indexed and stored in the preset knowledge base, forming a lightweight index library with semantic filenames as search keys. In another embodiment, semantic filenames and their corresponding files can also be stored in the preset knowledge base, meaning the files in the preset knowledge base are named with semantic filenames. When a natural language query is received, the query is used as a search vector or keyword combination and compared with the semantic filenames of all multimodal files in the preset knowledge base for similarity calculation or exact matching. Since the query statement already contains explicit file type qualifiers, the matching process prioritizes weighted matching or filtering of the file type identifier field in the semantic filename to ensure that the search results meet the user's specified format requirements. Then, the content description field is matched. Based on matching algorithms, such as the matching score output by cosine similarity, one or more semantic filenames with a matching score higher than a preset threshold are selected from the knowledge base and identified as target semantic filenames. These target semantic filenames are directly returned to the front-end presentation layer as the final results of this search. Because the search process only involves comparing this short file (the semantic filename), there is no need for real-time parsing or deep feature extraction of the actual content of the multimodal file. Therefore, it can significantly reduce search latency, thus improving search efficiency, while avoiding false recall problems caused by redundant information within the file, ensuring that the returned search results match the user's expectations in both content and format. Figure 1 To meet users' search needs.

[0023] 103. Parse the target file type from the target semantic file name, determine the rendering component corresponding to the target file type, and use the rendering component to load and display the file content corresponding to the target semantic file name.

[0024] In this embodiment of the invention, after obtaining the target semantic filename as the retrieval result, the file type is first determined by extracting the suffix identifier or type descriptor from the target semantic filename using preset naming parsing rules or regular expressions, such as "MP4", "PDF", "JPG", etc. A rendering component mapping table is maintained, which predefines the correspondence between different file types and specific front-end rendering components. For example, "MP4" corresponds to a video player component, "PDF" corresponds to a document preview component, and "JPG / PNG" corresponds to an image viewer component. After determining the target file type, the matching rendering component is obtained and instantiated according to the rendering component mapping table. Then, the file represented by the target semantic filename is obtained. During this process, if the file is stored in a preset knowledge base, it can be directly obtained from the preset knowledge base. If the file is stored on a server, the actual multimodal file on the storage server can be located using the file path information or unique resource identifier carried in the target semantic filename, and the access address or data stream of the file is passed to the instantiated rendering component. After receiving file data, the rendering component automatically loads and renders the content of the multimodal file on the user's terminal display interface. For example, if the target file type is video, a playback control is directly embedded in the page and buffering begins; if it is a document, a page-turning preview interface is loaded. In this way, a seamless connection is achieved between the display of search results and the preview of file content. Users do not need to perform additional operations such as downloading, jumping, or manually selecting an opening method; they can directly view the multimodal file content that matches their search intent, greatly improving the efficiency of information acquisition and the smoothness of the interactive experience.

[0025] This invention provides an intelligent retrieval method based on semantic naming of multimodal files. Compared with the current method of matching user questions with the full text of files in a knowledge base for information retrieval, this invention retrieves information by matching natural language query statements with semantic filenames of multimodal files in a preset knowledge base. Due to the smaller amount of filename data, lightweight retrieval can be achieved, thereby improving retrieval efficiency and avoiding the consumption of retrieval resources. By utilizing the file type inherent in the semantic filename, the client rendering engine can directly infer the file type by parsing the string, eliminating the need for additional metadata query and format detection steps, improving the efficiency of content presentation. Moreover, regardless of whether the underlying file is a video, PDF, or image, the front end can automatically match the most suitable player or reader or other rendering components based on the parsed type, eliminating the need for users to manually download or select an opening method, thus optimizing the user's reading experience.

[0026] Furthermore, to better illustrate the intelligent retrieval process based on multimodal document semantic naming, as a refinement and extension of the above embodiments, this invention provides another intelligent retrieval method based on multimodal document semantic naming, such as... Figure 3 As shown, the method includes: 201. Obtain multimodal files, which include at least video files, image files, and text files.

[0027] 202. Perform multi-dimensional information identification on multimodal files. The multi-dimensional information includes a file type identifier field and a content description field. The file type identifier field is used to represent the format category of the multimodal file, and the content description field is used to represent the entity information and business information of the multimodal file.

[0028] Among them, business information can be the purpose of the document, and entity information is information that can characterize the main content of the document.

[0029] Specifically, the process begins by acquiring a collection of multimodal files to be processed through file upload interfaces, web crawlers, or local storage scanning. These multimodal files encompass various data formats, including at least video files, image files, and text files such as PDF, DOCX, and TXT formats. The first step is to identify the content entity information of the multimodal files: for video and image files, computer vision technology is used for keyframe extraction, object detection, and OCR text recognition to extract visual semantic elements such as objects, scenes, people, and embedded text; for text files, natural language processing technology is used for word segmentation and named entity recognition to extract keywords, proper nouns, and core summaries. The second step is to identify business information: by analyzing the file's metadata attributes or associated business logs, the file's creation time, author attribution, and applicable business process nodes, such as onboarding training and troubleshooting, are obtained. This information reflects the file's functional attributes and purpose in a specific business scenario. Finally, the file type is identified: by reading the file's binary header information or extension, the physical format of the file is accurately determined, providing a basis for subsequent matching of rendering components. Through the above multi-dimensional identification, the originally unstructured and difficult-to-understand multimodal files are transformed into a structured feature dataset containing "key content, business behavior and type", laying the data foundation for the next step of generating semantic filenames that can accurately describe the full picture of the file.

[0030] 203. Determine the file function of the multimodal file, determine the preset naming template based on the file function, use the preset naming template to integrate multi-dimensional information into semantic file names of the multimodal file, and construct a preset knowledge base from the semantic file names of the multimodal file.

[0031] In this embodiment of the invention, the file function of the multimodal file is first determined based on the business behavior information and content entity information extracted from the multimodal file. For example, if the file content contains a large number of operation step diagrams and the business attribute is marked as after-sales support, its file function is determined to be operation guidance; if the file content is meeting minutes and contains decision items, its function is determined to be meeting minutes; if the file is a product promotion video and contains marketing keywords, its function is determined to be marketing promotion. This determination process maps unstructured content features to standardized function category labels. After determining the file function, a preset naming template library is accessed. This template library stores pre-designed structured naming templates for different file function categories. Each template defines the components, arrangement order, and connector specifications of the semantic file name. For example, for the operation guidance function, the selected preset naming template may be "{core topic}{applicable object}{operation step overview}{file type}"; while for the meeting minutes function, the selected template is "{meeting name}{date}{resolution item}{file type}". Based on the determined file function tags, a corresponding preset naming template is accurately retrieved from the library as the skeleton structure for generating semantic filenames. This ensures that the generated filenames not only contain content information but also intuitively reflect the file's business purpose. Furthermore, template-based naming guarantees naming standardization and improves subsequent information retrieval efficiency. Further, after selecting a suitable preset naming template, the semantic filenames for multimodal files need to be generated based on this model. The method includes: obtaining historical natural language query statements; identifying historically high-frequency query terms from the historical natural language query statements; correcting the information representation of the multi-dimensional information based on the historically high-frequency query terms; and integrating the corrected multi-dimensional information into the semantic filename of the multimodal file using the preset naming template.

[0032] Specifically, the system retrieves a set of historical natural language queries generated by a user group within a preset time period, such as the past three months, from log databases or search engine backends. Text mining techniques are then used to segment, remove stop words, and perform word frequency statistics on this set to identify historically high-frequency query terms. These terms represent the most commonly used terms, abbreviations, or specific expressions by users in the current business context; for example, users often search for "how to reinstall the system" rather than "operating system reinstallation guide." The identified multi-dimensional information is then semantically compared with the extracted historically high-frequency query terms. If certain professional terms or formal expressions in the multi-dimensional information are found to be synonymous or near-synonymous with the high-frequency query terms, the high-frequency query terms are used to replace or supplement the multi-dimensional information. For example, if the file content in the multi-dimensional information is identified as "how to find personal health record information," while the historically high-frequency term is "personal health record information query guide," the multi-dimensional information is corrected to "personal health record information query guide." This process narrows the semantic gap between file naming and user search habits, improving information search efficiency and accuracy. Finally, the corrected multi-dimensional information is used as filler elements and injected into a preset naming template. Following the field order and connector rules defined in the template, the corrected content entities, business information, and file types are concatenated and integrated to generate the final semantic filename. Filenames generated in this way retain the accuracy of the file content while incorporating the user's search term features, thus significantly improving the accuracy of subsequent search matching and the user experience.

[0033] Ultimately, a pre-defined knowledge base is formed from multimodal files with semantically named filenames.

[0034] 204. Receive the natural language query statement input by the target user, and preprocess the natural language query statement, including information completion.

[0035] In this embodiment of the invention, after receiving a natural language query statement input by a user, in order to ensure the efficiency and accuracy of subsequent information search, it is necessary to pre-identify and correct defects in the natural language query statement. Based on this, the method includes: identifying defects in the natural language query statement, determining a target defect correction method that matches the natural language query statement based on the defect identification result, and correcting defects in the natural language query statement using the target defect correction method.

[0036] Specifically, the method identifies defects in natural language queries, such as whether they contain ambiguous words or lack information necessary for accurate information retrieval, to obtain defect identification results. Then, based on the defect identification results, a target defect correction method needs to be determined to correct the defects in the natural language query. Therefore, the method includes: determining the defect type and severity based on the defect identification results; determining the defect compensation capability of the preceding historical natural language queries input by the target user; and determining the target user's interaction preference based on the preceding historical natural language queries; and determining a target defect correction method matching the natural language query based on the defect type, the defect severity, the defect compensation capability, and the interaction preference.

[0037] Specifically, based on the sentence defect identification results, the defect type is determined, such as ambiguous reference, missing keywords, grammatical errors, unclear intent, and ambiguous words. The severity of the defect is quantified by considering its potential impact on the accuracy of search results, for example, categorized as high, medium, and low. Then, the target user's previous natural language queries entered in the current session window are reviewed to assess the defect compensation capability. This is done by analyzing whether the preceding statements contain contextual elements that can complete the missing information in the current statement, such as the subject, time range, or specific business scenario mentioned earlier. If the preceding statements provide sufficient background information, the defect compensation capability is considered strong; otherwise, it is considered weak. Simultaneously, based on the target user's historical interaction data, including historical query frequency, click-through rate, acceptance of correction feedback, and dwell time, the user's interaction preference score is determined. This score reflects the user's acceptance of frequent interactions. Finally, a comprehensive analysis of the defect type, defect severity, defect compensation capability, and interaction preference determines the target defect correction method. For example, when the defect severity is low and the preceding historical statements have strong defect compensation capabilities, non-user-interactive correction methods are prioritized. This means that information can be directly supplemented using the context and the search can be executed, avoiding interruption of the user's workflow. When the defect severity is high but the user's interaction preference indicates a tendency towards self-control, or the preceding statements cannot provide effective compensation, user-interactive correction methods are adopted. These methods guide the user to confirm or supplement information through pop-up clarification options, recommended candidate words, or highlighted prompts. When the defect severity is high, the compensation capability is weak, and the user's interaction preference indicates a preference for efficient automated services, confidence-based automatic correction can be attempted, with a feedback entry point at the edge of the interface as a fallback. Through this dynamic adaptation mechanism, user habits can be satisfied to the greatest extent and the interactive experience can be optimized while ensuring search accuracy.

[0038] Furthermore, after determining the target defect correction method, if the target defect correction method is a non-user-interactive correction method, i.e., an automatic system correction method, then the defect in the natural language query statement needs to be corrected using the non-user-interactive correction method. Based on this, the method includes: if the target defect correction method is a non-user-interactive correction method, in the case of an ambiguous word defect in the natural language query statement, obtaining the target user's previous historical natural language query statements; concatenating the natural language query statement and the previous historical natural language query statements to form an extended context; performing syntactic dependency analysis on the extended context; extracting context words with syntactic dependency relationships with the ambiguous word as disambiguation feature words based on the syntactic dependency analysis results in the extended context; determining multiple candidate meanings of the ambiguous word, and determining the domain label for each candidate meaning; determining the target user's domain preference based on the previous historical natural language query statements; determining the disambiguation feature domain based on the domain label and the domain preference; and correcting the ambiguous word defect based on the disambiguation feature words and the disambiguation feature domain.

[0039] Specifically, the preceding historical natural language queries entered by the target user within the current dialog window are retrieved from the session log or memory cache. These historical queries form the key contextual basis for understanding the current intent. The current natural language query is then concatenated with the preceding historical natural language queries in chronological order to construct an extended context containing the complete dialogue flow. Natural language processing techniques are then used to perform syntactic dependency analysis on the generated extended context. By constructing a dependency syntax tree, the modification, dominance, and association relationships between words in the sentence can be accurately identified. Based on this, words with direct syntactic dependency relationships with the current ambiguous word, such as subject-verb, verb-object, and attributive-head relationships, are located in the extended context and extracted as disambiguation feature words. For example, if the ambiguous word is "apple," and dependency analysis reveals that it is dominated by "eat" or modified by "red," then "eat" and "red" are the key disambiguation feature words. Simultaneously, a knowledge base is searched for this ambiguous term to determine all possible candidate meanings. For example, "apple" can refer to either fruit or a technology company. Each candidate meaning is pre-labeled with a corresponding domain tag, such as "catering / fresh food" or "technology / digital." To further narrow down the scope, the keyword distribution in previous historical natural language queries is analyzed, and the frequency or weight of each domain is statistically analyzed to determine the target user's current domain preference. For example, if the user has previously searched for "phone repair" and "system update" multiple times, their preference is determined to be in the "technology field." Then, the semantic similarity between the extracted disambiguation feature words and each candidate meaning is used for initial screening. In this process, the disambiguation feature domain obtained by matching the domain tag with the user's domain preference is used as a strong constraint, prioritizing candidate meanings that both conform to the syntactic context and fall within the user's current preference domain as the final correction result. For example, under the technology domain preference, combined with the dependency word "system," the system will automatically correct the "apple" in the query to a semantic index related to "Apple Inc." instead of an index for fruit, thus achieving high-precision query intent understanding without disturbing the user. When the method is determined to be a non-user interaction correction method, the present invention constructs an extended context by splicing previous historical sentences and uses syntactic dependency analysis to accurately extract disambiguation feature words that are grammatically related to the ambiguous words. This grammatical constraint based on local context can automatically lock the true meaning of the ambiguous words in the current dialogue without disturbing the user, thereby achieving silent and imperceptible accurate retrieval and improving the user experience.

[0040] In another embodiment of the present invention, if the target defect correction method is a user interaction correction method, then the natural language query statement needs to be corrected using the user interaction correction method. Based on this, the method includes: if the target defect correction method is a user interaction correction method, generating a defect correction interaction request based on the statement defect identification result, sending the defect correction interaction request to the target user terminal, and receiving the defect correction interaction result fed back by the target user terminal; and correcting the natural language query statement based on the defect correction interaction result.

[0041] Specifically, based on the results of statement defect identification, targeted defect correction interaction requests are dynamically generated. These requests are not generic error messages, but rather structured data containing specific correction suggestions. For example, if an ambiguous referent is identified in the query, the interaction request will include several possible referents inferred from the context as options; if a keyword is missing, the request will include a list of words recommended based on semantic association. This defect correction interaction request is sent to the target user terminal in real time. The target user terminal then receives the request and renders it on a graphical user interface. The display format can be a pop-up, drop-down menu, speech bubble, or embedded card, designed to guide the user to quickly make a selection or input. After the user performs actions on the interface, such as clicking to confirm an option, selecting a recommended word from the list, or manually entering the corrected content, the terminal encapsulates these actions into defect correction interaction results and sends them back to the server. Finally, the server receives and parses the defect correction interaction results, accurately replacing or filling the corresponding defect location in the original natural language query with the parsed user confirmation information or new input content. For example, if a user selects "referring to object A", the vague pronoun in the original sentence is replaced with the specific name of "referring to object A"; if the user manually adds missing qualifiers, they are inserted into the correct node of the syntactic structure. Through this correction process, the original natural language query is transformed into a semantically complete and clearly defined standardized query, which is then sent to the subsequent retrieval process. This explicit interaction mechanism ensures the absolute accuracy of retrieval results in critical business scenarios.

[0042] Furthermore, after correcting the defects in the natural language query statement, if the corrected query statement does not specify the required file type, it is necessary to determine the required file type through user interaction and insert the file type into the appropriate position in the natural language query statement to obtain a natural language query statement with specified required file type.

[0043] 205. Perform relevance calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic file names in the preset knowledge base, and obtain the corresponding file type relevance calculation results and content description relevance calculation results. Based on the file type relevance calculation results and content description relevance calculation results, determine the target semantic file name that matches the preprocessed natural language query statement, and use the target semantic file name as the search result. The preset knowledge base consists of multimodal files with semantic file names, and the semantic file name contains a file type identifier field and a content description field.

[0044] Specifically, the corrected or completed natural language query is parsed to extract the user-specified file types, such as "PDF," "video," and "code file," as well as core search keywords. Then, the combined query conditions containing these type restrictions and keywords are matched against semantic filenames of multimodal files already generated in a pre-defined knowledge base. The matching process is based not only on literal similarity but also on a comprehensive comparison of structured tags embedded in the semantic filenames, such as business scenarios and functional attributes. Finally, target semantic filenames that both meet the specified file type requirements and semantically align with the query intent are selected as precise search results, thus achieving targeted searching of multimodal resources.

[0045] 206. Parse the target file type from the target semantic file name, determine the rendering component corresponding to the target file type, and use the rendering component to load and display the file content corresponding to the target semantic file name.

[0046] Specifically, after obtaining the target semantic filename, the system first accurately identifies the target file type by parsing the structured identifiers or metadata tags in the filename. Then, based on the identified file type, it dynamically calls a matching dedicated rendering engine or playback control from a pre-built component library. The rendering component then loads the actual file pointed to by the target semantic filename and displays it adaptively on the user's terminal interface. For text files, the rendering component performs typesetting and content parsing to achieve a preview of the original text; for audio and video files, the player control is launched for streaming loading and playback. This file type-based differentiated rendering mechanism ensures that multimodal file content is correctly and efficiently presented to the user in the search results.

[0047] According to another intelligent retrieval method based on semantic naming of multimodal files provided by the present invention, compared with the current method of information retrieval by matching user questions with the full text content of files in the knowledge base, the present invention retrieves information by matching natural language query statements with semantic filenames of multimodal files in the preset knowledge base. Since the filename data volume is small, lightweight retrieval can be achieved, thereby improving retrieval efficiency and avoiding the consumption of retrieval resources. By utilizing the file type inherent in the semantic filename, the client rendering engine can directly infer the file type by parsing the string, eliminating the need for additional metadata query and format detection steps, improving the content presentation efficiency. Moreover, regardless of whether the underlying file is a video, PDF, or image, the front end can automatically match the most suitable player or reader or other rendering components based on the parsed type, without requiring the user to manually download or select an opening method, thus optimizing the user's reading experience.

[0048] Furthermore, as Figure 1 In specific implementation, embodiments of the present invention provide an intelligent retrieval device based on multimodal document semantic naming, such as... Figure 4 As shown, the device includes: a statement acquisition unit 31, a file name matching unit 32, and a file content display unit 33.

[0049] The statement acquisition unit 31 can be used to receive a natural language query statement input by a target user and preprocess the natural language query statement, wherein the preprocessing includes information completion. The filename matching unit 32 can be used to perform correlation calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic filenames in the preset knowledge base, and obtain the corresponding file type correlation calculation results and content description correlation calculation results. Based on the file type correlation calculation results and the content description correlation calculation results, the target semantic filename that matches the preprocessed natural language query statement is determined, and the target semantic filename is used as the retrieval result. The preset knowledge base is composed of multimodal files with semantic filenames, and the semantic filenames include a file type identifier field and a content description field. The file content display unit 33 can be used to parse the target file type from the target semantic file name, determine the rendering component corresponding to the target file type, and use the rendering component to load and display the file content corresponding to the target semantic file name.

[0050] In specific application scenarios, in order to build a pre-defined knowledge base, such as Figure 5 As shown, the device also includes a knowledge base construction unit 34.

[0051] The knowledge base construction unit 34 can be used to acquire multimodal files, wherein the multimodal files include at least video files, image files, and text files; to perform multi-dimensional information recognition on the multimodal files, wherein the multi-dimensional information includes a file type identifier field and a content description field, wherein the file type identifier field is used to characterize the format category of the multimodal file, and the content description field is used to characterize the entity information and business information of the multimodal file; to determine the file function of the multimodal file, to determine a preset naming template based on the file function, to integrate the multi-dimensional information into a semantic filename of the multimodal file using the preset naming template, and to construct the preset knowledge base from the semantic filename of the multimodal file.

[0052] In specific application scenarios, in order to determine the semantic file name of multimodal files, the knowledge base construction unit 34 includes a first statement acquisition module 341 and a file name determination module 342.

[0053] The first statement acquisition module 341 can be used to acquire historical natural language query statements and identify historical high-frequency query words in the historical natural language query statements.

[0054] The file name determination module 342 can be used to correct the information representation of the multi-dimensional information based on the historical high-frequency query words, and to integrate the corrected multi-dimensional information into the semantic file name of the multimodal file using the preset naming template.

[0055] In specific application scenarios, in order to preprocess natural language query statements, the device further includes a defect correction unit 35.

[0056] The defect correction unit 35 can be used to identify defects in the natural language query statement, determine a target defect correction method that matches the natural language query statement based on the defect identification result, and correct the defects in the natural language query statement using the target defect correction method.

[0057] In specific application scenarios, in order to correct defects in natural language query statements, the defect correction unit 35 includes a second statement acquisition module 351, a disambiguation feature word extraction module 352, a disambiguation feature domain determination module 353, and a defect correction module 354.

[0058] The second statement acquisition module 351 can be used to acquire the target user's previous historical natural language query statements when the target defect correction method is a non-user interaction correction method and the natural language query statement has an ambiguous word defect.

[0059] The disambiguation feature word extraction module 352 can be used to concatenate the natural language query statement with the preceding historical natural language query statement to form an extended context, perform syntactic dependency analysis on the extended context, and extract context words that have a syntactic dependency relationship with the ambiguous word as disambiguation feature words based on the syntactic dependency analysis results.

[0060] The disambiguation feature domain determination module 353 can be used to determine multiple candidate meanings of the ambiguous word, and determine the domain label of each candidate meaning, determine the domain preference of the target user based on the preceding historical natural language query statement, and determine the disambiguation feature domain based on the domain label and the domain preference.

[0061] The defect correction module 354 can be used to correct the defects of the ambiguous words based on the disambiguation feature words and the disambiguation feature domain.

[0062] In specific application scenarios, in order to correct defects in natural language query statements, the defect correction unit 35 further includes a request generation module 355.

[0063] The request generation module 355 can be used to generate a defect correction interaction request based on the statement defect identification result if the target defect correction method is a user interaction correction method, send the defect correction interaction request to the target user terminal, and receive the defect correction interaction result fed back by the target user terminal.

[0064] The defect correction module 354 can also be used to correct defects in the natural language query statement based on the defect correction interaction results.

[0065] In specific application scenarios, in order to determine the target defect correction method that matches the natural language query statement, the defect correction unit 35 further includes: a correction method determination module 356.

[0066] The second statement acquisition module 351 can also be used to determine the defect type and severity based on the statement defect identification result, determine the defect compensation capability of the preceding historical natural language query statement input by the target user, and determine the interaction preference of the target user based on the preceding historical natural language query statement.

[0067] The correction method determination module 356 can also be used to determine a target defect correction method that matches the natural language query statement based on the defect type, the defect severity, the defect compensation capability, and the interaction preference.

[0068] In specific application scenarios, in order to match the preprocessed natural language query statement with the semantic file name of multimodal files in the preset knowledge base, the file name matching unit 32 can be used to correlate the natural language query statement with the file type identifier field and content description field of multiple semantic file names in the preset knowledge base, and determine the target semantic file name that matches the preprocessed natural language query statement as the retrieval result based on the correlation calculation results of the file type identifier field and the correlation calculation results of the content description field.

[0069] It should be noted that other corresponding descriptions of the functional modules involved in the intelligent retrieval device based on multimodal document semantic naming provided in this embodiment of the invention can be found in the following references. Figure 1 The corresponding description of the method shown will not be repeated here.

[0070] Based on the above, Figure 1 Correspondingly, this embodiment of the invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the program performs the following steps: receiving a natural language query statement input by a target user; preprocessing the natural language query statement, wherein the preprocessing includes information completion; performing correlation calculations between the preprocessed natural language query statement and file type identifier fields and content description fields of multiple semantic file names in a preset knowledge base, obtaining corresponding file type correlation calculation results and content description correlation calculation results; determining a target semantic file name that matches the preprocessed natural language query statement based on the file type correlation calculation results and the content description correlation calculation results; using the target semantic file name as a search result; wherein the preset knowledge base is composed of multimodal files with semantic file names, and the semantic file name includes a file type identifier field and a content description field; parsing the target file type from the target semantic file name, determining a rendering component corresponding to the target file type, and using the rendering component to load and display the file content corresponding to the target semantic file name.

[0071] Based on the above, Figure 1 The method shown and as Figure 4 The embodiment of the device shown in the invention also provides a physical structure diagram of a computer device, such as... Figure 6As shown, the computer device includes: a processor 41, a memory 42, and a computer program stored on the memory 42 and executable on the processor. Both the memory 42 and the processor 41 are mounted on a bus 43. When the processor 41 executes the program, it performs the following steps: receiving a natural language query statement input by a target user; preprocessing the natural language query statement, including information completion; performing correlation calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic filenames in a preset knowledge base, obtaining corresponding file type correlation calculation results and content description correlation calculation results; determining a target semantic filename that matches the preprocessed natural language query statement based on the file type correlation calculation results and the content description correlation calculation results; using the target semantic filename as the search result; wherein the preset knowledge base consists of multimodal files with semantic filenames, and the semantic filenames include a file type identifier field and a content description field; parsing the target file type from the target semantic filename, determining a rendering component corresponding to the target file type, and using the rendering component to load and display the file content corresponding to the target semantic filename.

[0072] Through the technical solution of this invention, information is retrieved by matching natural language query statements with semantic filenames of multimodal files in a preset knowledge base. Since the filename data volume is small, lightweight retrieval can be achieved, thereby improving retrieval efficiency and avoiding the consumption of retrieval resources. By utilizing the file type inherent in the semantic filename, the client rendering engine can directly infer the file type by parsing the string, eliminating the need for additional metadata query and format detection steps, improving the content presentation efficiency. Moreover, regardless of whether the underlying file is a video, PDF, or image, the front end can automatically match the most suitable player or reader or other rendering components based on the parsed type, without requiring the user to manually download or select an opening method, thus optimizing the user's reading experience.

[0073] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0074] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0075] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent retrieval method based on multimodal document semantic naming, characterized in that, include: Receive a natural language query statement input by a target user, and preprocess the natural language query statement, wherein the preprocessing includes information completion; The preprocessed natural language query statement is correlated with the file type identifier field and content description field of multiple semantic file names in the preset knowledge base. Corresponding file type correlation calculation results and content description correlation calculation results are obtained. Based on the file type correlation calculation results and the content description correlation calculation results, the target semantic file name that matches the preprocessed natural language query statement is determined, and the target semantic file name is used as the search result. The preset knowledge base is composed of multimodal files with semantic file names, and the semantic file name includes a file type identifier field and a content description field. The target file type is parsed from the target semantic file name, and the rendering component corresponding to the target file type is determined. The rendering component is then used to load and display the file content corresponding to the target semantic file name.

2. The method according to claim 1, characterized in that, Before performing relevance calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic filenames in the preset knowledge base, the method further includes: Obtain multimodal files, which include at least video files, image files, and text files; The multimodal file is subjected to multi-dimensional information identification, wherein the multi-dimensional information includes a file type identifier field and a content description field. The file type identifier field is used to characterize the format category of the multimodal file, and the content description field is used to characterize the entity information and business information of the multimodal file. The file function of the multimodal file is determined, a preset naming template is determined based on the file function, the multidimensional information is integrated into the semantic file name of the multimodal file using the preset naming template, and the preset knowledge base is constructed from the semantic file name of the multimodal file.

3. The method according to claim 2, characterized in that, The multi-dimensional information is integrated into a semantic filename for the multimodal file using the preset naming template, including: Obtain historical natural language query statements, identify historical high-frequency query words in the historical natural language query statements; modify the information representation of the multi-dimensional information based on the historical high-frequency query words, and integrate the modified multi-dimensional information into the semantic file name of the multimodal file using the preset naming template.

4. The method according to claim 1, characterized in that, Preprocessing the natural language query statement includes: The natural language query statement is subjected to statement defect identification, and a target defect correction method matching the natural language query statement is determined based on the statement defect identification results; The natural language query statement is modified using the target defect correction method.

5. The method according to claim 4, characterized in that, The natural language query statement is corrected using the target defect correction method, including: If the target defect correction method is a non-user interaction correction method, in the case that the natural language query statement has an ambiguous word defect, the previous historical natural language query statements of the target user are obtained; The natural language query statement is concatenated with the preceding historical natural language query statement to form an extended context. Syntactic dependency analysis is performed on the extended context. Based on the syntactic dependency analysis results, context words that have a syntactic dependency relationship with the ambiguous word are extracted from the extended context as disambiguation feature words. Multiple candidate meanings of the ambiguous word are determined, and a domain label for each candidate meaning is determined. The domain preference of the target user is determined based on the preceding historical natural language query statement. The disambiguation feature domain is determined based on the domain label and the domain preference. Based on the disambiguation feature words and the disambiguation feature domain, the defects of the ambiguous words are corrected.

6. The method according to claim 4, characterized in that, The natural language query statement is corrected using the target defect correction method, including: If the target defect correction method is a user interaction correction method, then a defect correction interaction request is generated based on the statement defect identification result, the defect correction interaction request is sent to the target user terminal, and the defect correction interaction result fed back by the target user terminal is received. Based on the defect correction interaction results, the natural language query statement is corrected for defects.

7. The method according to claim 4, characterized in that, Based on the statement defect identification results, the target defect correction method matching the natural language query statement is determined, including: Based on the defect identification results, the defect type and severity are determined, the defect compensation capability of the preceding historical natural language query statement input by the target user is determined, and the interaction preference of the target user is determined based on the preceding historical natural language query statement. Based on the defect type, the defect severity, the defect compensation capability, and the interaction preference, a target defect correction method matching the natural language query statement is determined.

8. An intelligent retrieval device based on multimodal document semantic naming, characterized in that, include: The statement acquisition unit is used to receive a natural language query statement input by a target user and preprocess the natural language query statement, wherein the preprocessing includes information completion. The filename matching unit is used to perform relevance calculations between the preprocessed natural language query statement and the file type identifier field and content description field of multiple semantic filenames in a preset knowledge base, and obtain the corresponding file type relevance calculation results and content description relevance calculation results. Based on the file type relevance calculation results and the content description relevance calculation results, the unit determines the target semantic filename that matches the preprocessed natural language query statement and uses the target semantic filename as the retrieval result. The preset knowledge base is composed of multimodal files with semantic filenames, and the semantic filenames include a file type identifier field and a content description field. The file content display unit is used to parse the target file type from the target semantic file name, determine the rendering component corresponding to the target file type, and use the rendering component to load and display the file content corresponding to the target semantic file name.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images