Search method for content creation objects

Abstract

The embodiment of the present application provides a content creation object search method, which comprises: obtaining an input search request expression, the search request expression being expressed by using a natural language; based on internal part-of-speech codes assigned to content creation objects, the search request expression is split into single words or single phrases; based on the logical relationship between the single words or single phrases, a logical framework including multi-dimensional filtering conditions is constructed; the logical framework is converted into a structure expression, so as to match database table fields in a content creation object database based on the structure expression; based on the matched database table fields, the content creation object database is searched for content creation objects, and a plurality of content creation object description information is obtained; the search request expression is converted into a question vector, and each piece of content creation object description information is converted into a context sentence vector to calculate a matching degree evaluation score of the two, and the content creation object is returned according to the score.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of artificial intelligence, and in particular to a content creation object search method. BACKGROUND

[0002] With the rapid development of the Internet, a large amount of data related to net red or anchor has been accumulated in the long-term overseas operation process, and a huge database has been built for internal free use and external paid use. When searching information in the database, traditional search techniques are widely used to meet different search needs. Common conventional search techniques are mainly based on logical operators, such as using logical relationships such as "and", "or", and "not" for logical search, and filtering out the required data by accurately setting these logical conditions. In addition, full-text search technology also plays a role to some extent, which can search the text content more comprehensively and is more effective when searching for specific keywords. These traditional search techniques can indeed provide good support when dealing with some established and more explicit search needs, helping users to obtain the corresponding content creation object information from the database. However, in recent years, generative AI technology has developed rapidly and has shown great application potential in many fields. In the field of content creation object search, users expect to be able to express search intentions more naturally and conveniently, and the traditional search method relying on fixed logical operators and keyword settings gradually shows its limitations. Natural language is the most common expression method in people's daily communication, and if it can be applied to the search process of content creation objects such as net red or anchor, it will undoubtedly greatly improve the user's search experience, so that users no longer need to carefully design keywords and logical relationships that meet specific search syntax rules, but only need to input the content they want to find like daily conversation to initiate a search request. At present, although some search techniques have tried to develop in the direction of natural language search, there are still many problems to be solved. For example, how to accurately understand the complex logical relationship between each part of the search request expressed in natural language, how to reasonably convert this logical relationship into a filtering condition that can adapt to the database structure, and how to accurately find the most suitable object in the numerous content creation object description information from the natural language search request. SUMMARY

[0003] Therefore, the embodiment of the present application provides a content creation object search method to at least partially solve the above problems.

[0004] According to a first aspect of the embodiments of the present application, a search method for content creation objects is provided, which comprises: obtaining an input search request expression expressed in natural language; based on internal part-of-speech codes assigned to the content creation objects, splitting the search request expression into single words or single phrases; based on logical relationships between the single words or the single phrases, constructing a logical framework comprising multi-dimensional filtering conditions; converting the logical framework into a structured expression to perform matching of database table fields in a content creation object database based on the structured expression; based on the matched database table fields, searching for content creation objects in the content creation object database to obtain a plurality of content creation object description information; converting the search request expression into a question vector and converting each piece of content creation object description information into a context sentence vector; calculating a matching degree evaluation score of the question vector and the context sentence vector; and selecting and returning a content creation object corresponding to content creation object description information whose matching degree evaluation score is greater than a set matching degree evaluation score threshold.

[0005] In the scheme of the embodiments of the present application, the following technical benefits are achieved:

[0006] I. Improved user search experience

[0007] The method first obtains a search request expression expressed in natural language, which conforms to people's daily communication habits and is in contrast to the traditional search method relying on fixed logical operators and keywords. The user does not need to spend effort to construct keywords and logical relationships according to specific search syntax rules, but only needs to input the content to be searched as usual to initiate a search request, greatly reducing the user threshold and making the user operation more convenient and natural.

[0008] II. Accurate analysis of search intent and adaptation of database structure

[0009] After splitting the search request expression expressed in natural language into single words or single phrases based on the internal part-of-speech codes assigned to the content creation objects, the logical relationships between them are further analyzed to construct a logical framework comprising multi-dimensional filtering conditions. This conversion from natural language to a logical framework can deeply mine the complex logical relationships implied in the search request, and is no longer limited to simple keyword matching, making the understanding of the user's search intent more accurate and comprehensive.

[0010] The constructed logical framework is converted into a structural expression, which can be matched with the table fields of the content creation object database, so as to carry out search work in the database. This link realizes the connection from understanding the natural language intention of the user to adapting the database structure, and ensures that the required content creation object can be accurately found by using the data in the database, thereby overcoming the problem that the natural language logic is difficult to be well connected with the database structure in the past.

[0011] III. Precise matching of search results

[0012] The search request expression is converted into a question vector, each piece of content creation object description information is converted into a context sentence vector, and the matching degree evaluation score of the two is calculated. The vector representation can capture the deep semantic information of words and sentences, and the similarity between the two is measured by the matching degree evaluation score. Compared with the traditional key word matching method, the matching degree of the content creation object and the search request can be more accurately judged, and the object that truly meets the user's demand on the semantic level can be mined.

[0013] Finally, the content creation object corresponding to the content creation object description information with a matching degree evaluation score greater than the set matching degree evaluation score threshold is screened out and returned. By setting a reasonable threshold, the accuracy of the search results can be further controlled, and too many content creation objects with weak relevance can be avoided, so that the user is presented with highly matched and high-quality results. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0015] Figure 1 A search method flowchart of a content creation object provided by an embodiment of the present application.

[0016] Figure 2 A schematic diagram of a search device of a content creation object provided by an embodiment of the present application.

[0017] Figure 3 A structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] Figure 1 A search method flowchart of a content creation object provided by an embodiment of the present application. As shown in Figure 1 , it comprises:

[0019] obtaining an input search request expression expressed in natural language;

[0020] disassembling the search request expression into single words or single phrases based on internal part-of-speech codes assigned to the content creation objects;

[0021] constructing a logical framework including multi-dimensional filtering conditions based on logical relationships between the single words or the single phrases;

[0022] converting the logical framework into a structured expression to perform matching of database table fields in a content creation object database based on the structured expression;

[0023] performing a search of content creation objects in the content creation object database based on the matched database table fields to obtain a plurality of content creation object description information;

[0024] converting the search request expression into a question vector and converting each piece of content creation object description information into a context sentence vector;

[0025] calculating a matching degree evaluation score of the question vector and the context sentence vector;

[0026] filtering out a content creation object pointed to by the content creation object description information corresponding to the matching degree evaluation score greater than a set matching degree evaluation score threshold and returning the content creation object.

[0027] To this end, the search method for the content creation object provided in this embodiment optimizes and innovates from multiple links such as user input, intent understanding, database adaptation, and result matching, effectively overcomes many problems existing in the search of content creation objects such as web celebrities or anchors in the existing natural language search technology, and improves the accuracy, convenience, and practicality of the search.

[0028] Optionally, the obtaining of the input search request expression expressed in natural language includes: listening to an input event based on a front-end multi-modal event listener, capturing user input in real time, and converting the captured input into a character stream through a set conversion function, the input event including at least one of a keyboard input event, a voice input event, and a handwriting input event; adding the character stream to a configured buffer according to the order of input to form a character stream sequence; and performing preliminary syntax checking and automatic completion on the character stream sequence according to a set syntax rule to generate the search request expression.

[0029] To this end, the above-mentioned specific processing manner of obtaining the input search request expression has the following technical benefits:

[0030] I. Convenience and inclusiveness of multi-modal input

[0031] The front-end-based multi-modal event listener can listen to input events corresponding to various input modes, including keyboard input events, voice input events, handwriting input events, etc. Such a design fully considers the usage habits and actual needs of different users in different scenarios. Whether the user is accustomed to traditional keyboard typing input, prefers voice input in a mobile scenario, or inputs content through handwriting, the user can conveniently initiate a search request, greatly improving the applicability of the system to different user groups.

[0032] For example, a user can directly use voice input to initiate a search when walking or driving, etc. For some users who are not good at typing but are more convenient for writing, handwriting input can also meet their needs, broadening the ways for users to interact with the search system.

[0033] II. Accuracy and completeness of input processing

[0034] The conversion function converts the captured input of various forms into a character stream, unifying the format of the input data. The data collected by different input methods is in different forms, such as audio signals for voice input and graphical traces for handwriting input. After converting them into a character stream, subsequent processing can be carried out according to unified rules and processes, avoiding processing difficulties caused by mixed data formats, and laying a foundation for accurate analysis of user input.

[0035] For example, the content of voice input is converted into text by voice recognition technology and then further converted into a character stream, so that subsequent syntax checking or other processing steps can be operated like regular text input, ensuring the standardization of the entire processing flow.

[0036] The character stream is added to the configured buffer according to the order of input to form a character stream sequence. This approach can completely preserve the order and all content of user input. In the actual input process, the user may input content in multiple times and intermittently. The buffer mechanism can integrate these scattered inputs to ensure that the final character stream sequence accurately reflects the complete search request that the user wants to express, avoiding information loss or confusion, and improving the completeness and processability of the input content.

[0037] For example, the user first inputs "find", pauses for a few seconds, then inputs "will dance", and finally inputs "fitness anchor". Through the buffer, these sequentially input character streams can be combined into a complete character stream sequence "find will dance fitness anchor", which is convenient for subsequent generation of accurate search request expressions.

[0038] III. Syntax optimization to improve search efficiency and accuracy

[0039] According to the set syntax rules, the character stream sequence is preliminarily checked for syntax, and possible syntax errors in user input can be found in advance, such as missing necessary conjunctions, improper word collocation, and the like. Identifying these errors in a timely manner and making corresponding prompts or automatic corrections (within a certain range) can avoid the situation that the search request cannot be accurately parsed or the search result is greatly deviated due to syntax errors, and helps to improve the accuracy of the search.

[0040] For example, if the user inputs "dancing fitness anchor" (missing conjunctions and the like, and the syntax is not accurate), the syntax checking link can prompt the user to supplement "find fitness anchors who can dance", so that the search request expression is more in line with normal language logic, and the system can better understand the user's intention and perform subsequent processing.

[0041] The automatic completion function is based on syntax rules and common language expression habits to supplement and perfect the incomplete content input by the user. Not only does it facilitate user input and reduce the amount of input, but it also further standardizes the search request expression, making it more in line with the format that the system can accurately process. This helps to speed up user input and improve overall search efficiency, and also makes the generated search request expression more conducive to subsequent accurate conversion into effective filtering conditions and the like for search operations.

[0042] For example, after the user inputs "find funny", the system automatically completes it as "find funny pet bloggers" based on past popular searches and syntax rules. This not only conforms to common expression habits, but also makes the search request more specific and specific, making it easier for the system to accurately find the corresponding content creation object and improving the efficiency and accuracy of the entire search process.

[0043] Optionally, the front-end-based multi-modal event listener listens to input events and captures user input in real time. Before that, it includes:

[0044] For the keyboard input event, the multi-modal event listener is bound to the keydown, keyup, and keypress events of the search input box using the addEventListener component of JavaScript to trigger keyboard input event capture when the user inputs characters on the keyboard;

[0045] For the voice input event, the multi-modal event listener is bound to the click event of the voice input button or the voice wake-up event through the voice recognition interface provided by the browser to trigger voice input event capture when the user inputs characters by voice;

[0046] For the handwriting input event, the multimodal event listener is bound to the handwriting input area through the interface of the handwriting recognition library, so as to trigger the capture of the handwriting input event when the user inputs characters by handwriting.

[0047] To this end, by performing precise binding operations for different input events, the following technical benefits are achieved:

[0048] 1. Accurately capture keyboard input to ensure comprehensive monitoring of text input

[0049] 1. Comprehensive coverage of keyboard operations

[0050] Use JavaScript's addEventListener component to bind multimodal event listeners to the search input box's keydown, keyup, and keypress events. This allows you to capture all the input corresponding to different keyboard actions. The keydown event captures the moment a key is pressed, the keyup event corresponds to when a key is released, and the keypress event focuses on the input of printable characters. By monitoring these three key keyboard events, we ensure that no user input is missed, thereby fully and accurately capturing the user's text input.

[0051] For example, when a user quickly and continuously inputs a string of characters, the keydown event will respond to each key press in turn, the keypress event will process printable characters, and the keyup event will correspond to key releases. The system can record the characters corresponding to these operations in sequence. For example, when a user inputs "find funny anchors", regardless of the input speed, the complete input text can be accurately obtained through the joint monitoring of these three events, providing reliable raw data for the subsequent generation of accurate search request expressions.

[0052] 2. Closely integrate with the input box to improve input relevance

[0053] Binding a listener to the keyboard events corresponding to the search input box allows input capture to be closely tied to the specific search input scenario. Keyboard operations performed by users in this specific input box are intended to initiate search requests. This binding method directly focuses on the user's input content in search-related operations, avoiding interference with other unrelated keyboard operations. This improves the relevance and effectiveness of input capture, ensuring that only valid character information for search purposes is captured.

[0054] For example, there may be multiple input areas on a web page, but by accurately binding the listener to the relevant keyboard events of the search input box, only the keyboard input of the user in the input box dedicated for search will be treated as search request content, and text editing operations such as the user's text editing operation in other text boxes on the page will not be mistaken as search input, so that the input capture serves the search function more accurately.

[0055] II. Efficiently capture voice input and adapt to various voice interaction scenarios

[0056] 1. Simplify development and integration with browser interface

[0057] By using the voice recognition interface provided by the browser to bind the multimodal event listener to the click event of the voice input button or the voice wake-up event, and with the help of the browser's existing voice function module, developers do not need to build a complex underlying voice recognition architecture, reducing the development difficulty and cost, while ensuring the accuracy and stability of voice recognition. Because the voice recognition interface of the browser has usually undergone a lot of optimization and testing, it can better adapt to different devices, network environments, etc., and facilitate the integration of voice input function into the search system.

[0058] For example, in common modern browsers (such as Chrome, Firefox, etc.), there is a relatively mature voice recognition function built-in, and developers only need to bind the multimodal event listener according to the specification. When the user clicks the voice input button (such as the microphone icon next to the web search bar) or wakes up the voice input function through the voice wake-up word, the voice input event capture can be triggered smoothly, the user's voice can be quickly converted into text content that can be used for search, and the voice input function can be easily applied in the search system.

[0059] 2. Flexible triggering mechanism to meet diverse usage habits

[0060] Supporting binding to both the click event of the voice input button and the voice wake-up event fully considers different user scenarios and operation habits. Some users may be more accustomed to actively clicking the button to start voice input, while others prefer the more convenient and natural way of voice wake-up (similar to the wake-up method of smart speakers) to start voice input. Providing multiple triggering mechanisms can meet the individual needs of different user groups and improve the ease of use and user acceptance of the voice input function.

[0061] For example, when using the search function on a mobile device, the user can directly click the voice input button to perform a voice search when the handheld device is convenient to operate. When the device is placed aside or the user is not convenient to operate with both hands, the voice input is initiated by voice wake-up (such as saying "Hey, search assistant" or the like) to initiate a search request, and the diversified triggering modes provide the user with a more flexible voice interaction experience.

[0062] III. Reliable capture of handwriting input, expanding the diversity of input modes

[0063] 1. Function integration through handwriting recognition library interface

[0064] The multi-modal event listener is bound to the handwriting input area through the interface of the handwriting recognition library, and the powerful recognition capability of the professional handwriting recognition library is used to convert the graphical trajectory of handwriting into corresponding character information, thereby achieving effective capture of handwriting input. The handwriting recognition library is usually trained with a large number of handwriting samples and can accurately recognize handwriting content of different writing styles and fonts. By integrating the handwriting recognition library into the search system, a convenient input approach is provided for users who are accustomed to handwriting input or who are more suitable for handwriting in specific scenarios, and the input mode of the search system is enriched.

[0065] For example, on some devices with touch screens, such as tablet computers or notebook computers supporting a stylus, the user can directly write the content to be searched (such as "find drawing tutorial host") in the designated handwriting input area, the handwriting recognition library will convert the handwritten text into character information, and the multi-modal event listener can capture the handwriting input event and pass the corresponding character content to the subsequent link for processing, so that the search system can support handwriting, which is an intuitive and personalized input mode.

[0066] 2. Specific area binding enhances input targeting

[0067] The listener is bound to the handwriting input area, and the scope of handwriting input is clearly defined. Only handwriting operations within this specific area will be treated as search input content for processing, avoiding interference from other possible handwriting operations on the screen (such as handwriting annotations, drawings, and other non-search-related operations), ensuring the accuracy and targeting of handwriting input capture, and ensuring that the handwriting information received by the system is valid content for initiating a search request.

[0068] For example, in a web application with handwriting function, a dedicated handwriting input box is set up for search input. The user's handwriting in this box will be accurately captured as search request related content, and handwriting drawing operations in other parts of the page will not be misidentified as search input, improving the effectiveness of handwriting input for search.

[0069] Optionally, the preliminary syntax checking and automatic completion of the character stream sequence according to the set syntax rules to generate a search request expression comprises: parsing the syntax rules based on a set syntax rule parser to create keyword object word groups and conjunction object word groups, wherein the syntax rules include keyword rules, conjunction rules, and numerical range rules, the keyword rules list various attribute keywords related to the anchor and their possible variants; the conjunction rules define the usage of conjunctions such as "and", "or", "greater than", and "less than" in search requests; and the numerical range rules specify the expression of numerical search conditions; starting from the beginning of the character stream sequence, the syntax detection state transition mechanism defined by the set finite state machine is used to analyze the characters one by one, wherein for each character, first determine its category, whether it is a letter, a number, a punctuation mark, or a special character, if it is a letter, determine whether it is part of a keyword according to the context of the current character, and check whether there is a matching keyword by matching with the keyword object array, if it is a punctuation mark or a conjunction, check whether its use conforms to the syntax structure defined in the conjunction rules to generate a syntax detection result; based on the syntax detection result, determine the possible subsequent keywords or search conditions to generate automatic completion suggestions, and based on the user's historical search records, analyze the user's frequently used search patterns and keyword combinations, and integrate them into the automatic completion suggestions to generate a search request expression.

[0070] To this end, the above series of operations from syntax rule parsing, finite state machine-based syntax detection to automatic completion suggestion generation have the following technical benefits:

[0071] I. Syntax rule parsing and word group creation

[0072] The syntax rule parser performs parsing operations according to pre-set syntax rules. For keyword rules, it will sort out various attribute keywords related to the anchor and their possible variants, such as "number of fans", "number of fans", "number of followers", etc. These are sorted and classified into keyword object word groups to facilitate accurate identification of related content in the character stream sequence. For conjunction rules, it clearly specifies the specific usage of conjunctions such as "and", "or", "greater than", and "less than" in search requests, and parses and constructs conjunction object word groups to enable the system to clearly determine whether the use of conjunctions in the sentence conforms to the syntax specifications. The numerical range rules specify how numerical search conditions should be expressed, such as the combination of numbers and units, the format of range representation (e.g. "greater than 100,000", "between 5 and 10", etc.), and through the parsing of this complete set of syntax rules, a foundation is laid for subsequent syntax detection and completion work.

[0073] The created keyword object phrases and conjunction object phrases are equivalent to building a grammatical knowledge base. They clarify which words are valid and comply with the rules in the specific language environment of the search request, and what connection methods are reasonable for these words to be combined, providing a standardized reference basis for the system to accurately understand the user input content.

[0074] To this end, by clearly defining keyword and conjunction object phrases, the system can more accurately understand the semantics implied by the character stream sequence entered by the user. For example, when a user enters "more than 100,000 followers," the system can accurately determine that the user is referring to the number of followers of a live streamer based on the record of "number of followers" and its variations in the keyword object phrase. This avoids misunderstandings of user intent due to slight differences in wording, thereby improving the accuracy of semantic parsing of natural language search requests.

[0075] Furthermore, the connective object phrases and the grammatical framework constructed under each rule provide clear criteria for determining whether search request expressions conform to grammatical requirements. Based on these criteria, the system can accurately identify grammatical errors in user input, such as the proper use of connectives and the accuracy of numerical ranges. This helps to promptly detect and correct grammatical errors, ensure the standardization of search requests, and provide correct input content for subsequent effective search operations.

[0076] 2. Syntax Detection Based on Finite State Machine

[0077] The finite state machine defines a state transition mechanism for syntax checking, starting at the beginning of the character stream sequence and analyzing and processing each character one by one. During this process, each character is determined to belong to a category (such as a letter, number, punctuation mark, or special character), and then different analysis strategies are adopted based on the category and the context of the current character. For alphabetic characters, the context is combined to determine whether they are part of a keyword, and this is confirmed by matching them with the constructed array of keyword objects. For punctuation marks or conjunctions, their use is checked to see if they follow the grammatical structure defined in the conjunction rules. As each character is analyzed, the finite state machine transitions between different syntax checking states according to pre-set state transition rules. For example, from the starting state of identifying a keyword, to the intermediate state of gradually matching the complete keyword, to the associated state of determining whether the subsequent conjunctions are used correctly, and finally generates a comprehensive syntax checking result that reflects the grammatical compliance of the entire character stream sequence.

[0078] This character-by-character analysis and context-based judgment according to established rules can delve into the finest granularity level of the input content, not missing any possible grammatical issues, whether it is a single keyword spelling error, inappropriate use of conjunctions, or overall sentence structure that does not conform to grammatical norms. Through the rigorous state transition process of the finite state machine, it can ensure comprehensive and meticulous grammar checking of user input.

[0079] Therefore, since the finite state machine can analyze each character and its context in detail, it can discover many hidden and complex grammatical errors. For example, if the user inputs "find anchor with fan count greater than 100,000", the finite state machine can accurately detect the error of "greater than and" which does not conform to the definition of conjunction rules, and prompt the user to correct it in time, avoiding deviations in subsequent searches due to such grammatical errors, and improving the quality of search request expression.

[0080] No matter how complex the sequence of character streams input by the user is, or how irregular the format is, the finite state machine can effectively detect grammar based on its flexible state transition mechanism. It is not limited to fixed sentence templates and can handle search request inputs in various natural language expressions, enhancing the adaptability of the system to different user input styles and habits and improving the robustness of the overall system.

[0081] III. Automatic completion suggestion generation

[0082] Based on the grammar detection results, possible subsequent keywords or search conditions are determined to generate automatic completion suggestions. For example, if the grammar detection finds that the user has input "find funny", but has not yet input the complete description object, the system can determine that related keywords such as "pet anchor" may need to be supplemented based on keyword rules and common search sentence structures, and then give corresponding automatic completion suggestions. This completion is based on the analysis of the current sentence status by the previous grammar detection and the reasonable expression defined by the grammar rules, aiming to make the user's input content more consistent with the grammar specification and more complete and clear, facilitating accurate conversion into an effective search request.

[0083] At the same time, the system also analyzes the user's historical search records in depth to mine the user's frequently used search patterns and keyword combinations. For example, it is found that the user has searched for "funny pet anchor" and "funny food anchor" containing the "funny" keyword multiple times in the past. When detecting that the user has input a "funny" related expression again, these high-frequency related keyword combinations are integrated into the automatic completion suggestions, making the completion content more in line with the user's personalized search preferences and improving the relevance and practicality of automatic completion.

[0084] To this end, the automatic completion suggestion can help users reduce the workload of input and quickly generate complete and normative search request expressions. Users do not need to manually input complete long sentences, but only need to select or make a small amount of modification on the basis of reasonable completion suggestions given by the system, thereby saving input time, especially for some long and complex search requests, the convenience is more obvious, and the efficiency of user interaction with the search system is improved.

[0085] By incorporating user historical search records to optimize automatic completion suggestions, the system can better adapt to the unique search habits and preferences of each user and provide personalized input assistance for different users. The completion suggestions seen by each user are more likely to meet their past search needs and common expression methods, making the search process more in line with the user's personal expectations, enhancing the user's satisfaction and loyalty to the search system, and improving the overall user experience.

[0086] Optionally, based on the internal part-of-speech code assigned to the content creation object, the search request expression is divided into single words or single phrases, and the previous includes: collecting attribute features of the content creation object, defining different part-of-speech codes with different attributes and setting multiple matching modes for each part-of-speech code, the matching modes include synonyms, near synonyms and colloquial expression matching, wherein the attribute features include basic information, business-related information, style characteristics and social influence-related information; all part-of-speech codes and their corresponding matching modes are converted into a hash table for storage, and when a hash collision occurs, a free storage location is found according to a pre-defined probe sequence to store the conflicting elements, or the conflicting elements are stored in a linked list structure in the same hash bucket, so that the conflicting elements can be accessed through the traversal operation of the linked list.

[0087] To this end, the above processing of attribute features of content creation objects and the hash table storage and conflict processing mechanism have the following technical benefits for subsequent search request division and the entire search process in terms of semantic understanding, data management and retrieval efficiency.

[0088] I. Collecting attribute features and setting part-of-speech codes and matching modes

[0089] First, the attribute characteristics of content creation objects (such as net reds, anchors, etc.) are comprehensively collected and classified according to different attribute dimensions, covering basic information (such as name, gender, age, etc.), business-related information (such as broadcast category, broadcast time, joint company, etc.), style characteristics (such as humorous, serious, lively, cute, etc.), and social influence-related information (such as number of fans, number of likes, comment interaction, etc.). For these different categories of attributes, a unique segmentation code is defined for each attribute, and a mapping relationship from attribute to code is established. For example, for the attribute of "broadcast category", a specific segmentation code is assigned to facilitate subsequent unified reference and processing of information related to this attribute.

[0090] Multiple matching modes are set for each segmentation code, fully considering the diversity of natural language expression. Synonyms, near synonyms, and colloquial expressions are included, because when users actually input search requests, they may not use standard and unified vocabulary to describe relevant attributes. For example, for the attribute of "a large number of fans", users may use different expressions such as "a large number of fans" (colloquial expression) and "a large number of followers" (near synonym). By setting these matching modes, the system can associate these semantically similar but differently expressed words to the corresponding segmentation code, thereby more accurately understanding the specific attribute content involved in the user's input search request.

[0091] Therefore, by defining segmentation codes based on different attributes and setting rich matching modes, the system can accurately map the diverse expressions in natural language to the corresponding attribute concepts. No matter what vocabulary the user uses to describe a certain attribute of a content creation object, the system has a greater probability of accurately understanding its intent, avoiding missing or misunderstanding the key attribute information that the user wants to find due to expression differences, and improving the accuracy of semantic analysis of search requests. For example, when a user searches for "find a funny anchor", relying on the near synonym and colloquial expression matching modes set for the "humorous" style under the "style characteristics" attribute, the system can accurately identify that the user wants to find an anchor with a humorous style, and then accurately filter out the required objects.

[0092] The definition of segmentation codes realizes the unified management and standardized identification of different attribute characteristics. In the subsequent search process, whether it is the segmentation and processing of search request expressions or the matching with database fields, these standardized segmentation codes can be used for operation, avoiding the confusion caused by the complexity and diversity of attribute descriptions, making the connection between each link of the entire search system smoother, and improving the overall maintainability and processing efficiency of the system.

[0093] II. Hash table storage and conflict handling mechanism

[0094] All word segmentation codes and their corresponding matching patterns are converted and stored as a hash table. A hash function is used to calculate the word segmentation code (or its associated key information) to obtain a corresponding hash value, which determines the storage location of the element in the hash table. For example, for a word segmentation code representing the "number of fans" attribute, after calculation by the hash function, it will be mapped to a specific location in the hash table for storage, and its corresponding multiple matching patterns (such as the "fan super" and "follower numerous" vocabulary sets) are also stored in the location or associated structure. In this way, when searching for matching patterns related to a certain word segmentation code, the hash value can be quickly located to the corresponding storage location by recalculating the hash value, achieving efficient search operations with a time complexity of approximately O(1), greatly improving data retrieval speed.

[0095] However, different word segmentation codes may produce the same hash value after calculation by the hash function, resulting in hash collision. To solve this problem, a predefined probe sequence is used to find an open addressing method to find an idle storage location, that is, when a collision occurs, the subsequent storage locations are checked one by one according to the predetermined probe sequence (such as linear probe, quadratic probe, etc.), until an idle location is found to store the conflicting elements. Another common way to handle is to use a linked list structure to store conflicting elements in the same hash bucket (i.e., the storage location corresponding to the hash value that has collided), and then access these conflicting elements by traversing the linked list. In this way, even if there is a hash collision, all elements can be properly stored and accurately retrieved when needed.

[0096] Therefore, the hash table storage structure itself provides extremely fast search speed, which is crucial for operations that frequently need to search for matching patterns based on word segmentation codes during the search process. Whether it is to quickly determine the matching pattern corresponding to the word when the search request is expressed, or to verify whether the attribute-related vocabulary meets the set matching range in other aspects, the hash table can quickly complete the operation, significantly reducing data search time and improving the execution efficiency of the entire search process, especially when dealing with large-scale word segmentation codes and matching patterns data, the advantage is more obvious.

[0097] By reasonable hash conflict processing mechanism, even if the different segmentation code mapping to the same position, all the data can be stored and accessed. The open addressing method of detection sequence can effectively use the free space of hash table, and the linked list structure solves the conflict in a simple and intuitive way, so that the system has good adaptability and stability when facing complex and diverse data storage requirements and inevitable hash conflicts, and will not lose or misprocess data due to conflict problems, ensuring the reliability of the entire data storage and retrieval mechanism.

[0098] Optionally, the splitting the search request expression into single words or single phrases based on the internal segmentation code allocated to the content creation object comprises: normalizing and removing stop words from the search request expression to obtain a to-be-cut search request expression;

[0099] Starting from the beginning of the to-be-cut search request expression, the longest phrase matching mode defined in the internal segmentation code library is tried to match words or phrases in sequence, and the words or phrases are identified according to the context relationship, which is obtained by analyzing the co-occurrence probability of the previous and subsequent words or phrases through a set context window; it is judged whether the identified words or phrases exist in the internal segmentation code library, if yes, the words or phrases are determined as split words or phrases; if not, the similarity between the identified words or phrases and the segmentation codes of the internal segmentation code library is calculated, and if the similarity is greater than a set similarity threshold, the words or phrases are supplemented as split words or phrases.

[0100] Therefore, the above series of operations from the preprocessing of the search request expression to the splitting and judging of the words or phrases based on the segmentation code library have the following technical benefits:

[0101] I. Normalization and stop word removal of search request expression

[0102] The normalization of the search request expression aims to unify the format, encoding and some common text specifications of the input text, such as converting full-width characters to half-width characters, unifying case forms, etc. This can reduce the complexity of subsequent processing caused by inconsistent text formats, so that the system can treat the user's input search request expression in a relatively unified and standardized format, avoiding recognition errors or processing difficulties caused by format differences.

[0103] Stop words refer to words that frequently appear in natural language but do not help express the substantive content of the semantics, such as "of", "is", "in", etc. By removing these stop words from the search request expression through a specific stop word list or rules based on word frequency, part of speech, etc., the text content can be effectively simplified, the key word information can be highlighted, the subsequent word and phrase recognition can be focused on more semantically valuable parts, unnecessary processing burden can be reduced, and the accuracy of capturing core semantics can be improved.

[0104] To this end, the standardization operation enables various formats of search request expressions input by different users to be converted into a unified format that meets the system processing requirements, facilitating subsequent processing with uniform rules and algorithms, reducing the probability of errors caused by format differences, and improving the consistency and accuracy of the entire system processing search requests. For example, whether the user inputs "FIND FUNNY HOST" in uppercase or "find funny host" in lowercase, after standardization, they can all be in a unified format (such as lowercase) for subsequent analysis, avoiding recognition problems caused by case differences.

[0105] Removing stop words can quickly focus the system on the core words and phrases that truly reflect the search intent, avoiding interference from a large number of words with no substantive semantics. For example, after removing the stop words "I", "want", "a", and "very" from the user input "I want to find a very funny host", the remaining "find funny host" more clearly highlights the key search content, which helps subsequent more accurate splitting and recognition of words and phrases related to the search intent, improving the efficiency and effectiveness of semantic understanding.

[0106] II. Longest phrase matching mode and context relationship analysis

[0107] Matching according to the longest phrase matching mode defined in the internal word segmentation code library is based on the idea that in natural language, there are many phrase structures composed of multiple words with specific semantics, such as "food sharing host" and "outdoor exploration blogger". Starting from the beginning of the search request expression to be cut, the longest possible phrase is preferred for matching, rather than simply splitting by single words, which can better preserve semantic integrity, avoid incorrect splitting of words with inherent semantic relationships, and improve the accuracy of semantic understanding of user input content. Because often, the meaning expressed by a complete phrase is not simply the sum of its constituent words, but has a unique overall semantics, the longest phrase matching mode can better capture this semantics.

[0108] The co-occurrence probability of the front and rear words and phrases is analyzed through the set context window to determine the context relationship. The context window is like an observation range. For example, if the window size is set to 3, the frequency of the co-occurrence of 3 words or phrases before and after a certain middle word or phrase is checked. Based on a large amount of text data statistics or pre-trained language models, the co-occurrence probability distribution of different words and phrases in a specific context can be obtained. According to the probability information, the relevance and semantic role of the currently recognized words and phrases in the specific context can be determined to assist more accurate recognition of words and phrases and further improve the semantic understanding of the entire search request expression.

[0109] Therefore, the longest phrase matching mode can maximize the semantic integrity of phrases in natural language and avoid losing semantic information due to excessive splitting. For example, for a search request expression such as “find anchors who like to share travel stories”, if it is split according to individual words, the complete semantic structure of “share travel stories” may be destroyed. However, the longest phrase matching mode can recognize it as a whole, which is more consistent with the actual meaning that the user wants to express, thereby providing a more accurate semantic basis for subsequent accurate matching of content creation objects.

[0110] The context relationship analysis considers the co-occurrence probability of words and phrases and other context information, so that the system has a more accurate understanding of the semantic role of each recognized part in the entire search request expression. For example, in “find interesting outdoor food anchors”, by analyzing the context, the association between “outdoor” and “food” in the subject-related attribute of “anchor” can be determined, avoiding isolated consideration of each word, which helps to accurately split words and phrases that meet the actual intent according to the complete semantic context, and improves the accuracy and effectiveness of subsequent searches based on these split results.

[0111] III. Matching and similarity judgment based on word segmentation code library

[0112] In the internal word segmentation code library, it is checked whether there is a word segmentation code that matches the recognized words and phrases. This is based on the standard mapping relationship established by the previously constructed word segmentation code library. The word segmentation code library organizes and encodes various attribute-related words and phrases of content creation objects. When a word or phrase is recognized, the corresponding code is directly searched in this library. If a matching code is found, it means that the word or phrase meets the system's pre-set standard expression related to the attribute of the content creation object, so it can be determined as an effective word or phrase obtained by splitting and used for subsequent search process operations.

[0113] When there is no direct matching in the word segmentation code library, further judgment is made by calculating the similarity of the recognized words and phrases with the word segmentation codes in the library. Similarity calculation can use various methods, such as cosine similarity based on word vectors, edit distance, etc. Taking cosine similarity as an example, the words or phrases are converted into vector representation (usually using a pre-trained word vector model), and then the cosine value of the angle between the vectors corresponding to the word segmentation codes is calculated. The closer the value is to 1, the higher the similarity. If the calculated similarity is greater than a set similarity threshold, it means that although the expressions are not completely consistent, the semantics are relatively close, and the split words or phrases can be supplemented to include some reasonable variant expressions in natural language, thereby expanding the effective understanding range of user search requests.

[0114] Therefore, through matching judgment with the word segmentation code library, effective words and phrases can be filtered according to the system preset standard expressions related to the content creation object attributes, ensuring that the elements entering the subsequent search process are consistent with the established specifications and can accurately correspond to the corresponding attribute concepts, ensuring the accuracy of the search operation and avoiding the problem of large search result deviation caused by randomly accepting non-standard words and phrases.

[0115] The similarity judgment mechanism fully considers the flexibility and diversity of natural language, and recognizes words and phrases that have similar semantics although there is no completely consistent match in the word segmentation code library, so that they can participate in the subsequent search process. This enables the system to better adapt to the diverse expression habits and expression methods of different users, improves the compatibility of various natural language search requests, further improves the user search experience and the recall rate of search results, and finds content creation objects that meet the user's intent as comprehensively as possible.

[0116] Optionally, the logical framework including multi-dimensional filtering conditions is constructed based on the logical relationship between the single words or the single phrases, including: performing syntax structure analysis on the single words or the single phrases to obtain the hierarchical structure and logical relationship between the single words or the single phrases; performing semantic relationship recognition on the single words or the single phrases to obtain the semantic association relationship between the single words or the single phrases; constructing multi-dimensional filtering conditions according to the hierarchical structure and logical relationship, semantic association relationship between the single words or the single phrases; marking the importance weight of each filtering condition in grammar and semantics to construct a logical framework including multi-dimensional filtering conditions.

[0117] Optionally, when performing syntax structure analysis on individual words or individual phrases, a syntax analysis tool or algorithm in natural language processing can be used, such as a rule-based syntax analyzer, a statistical syntax analysis model (e.g., a probabilistic context-free grammar, PCFG), etc. Taking the rule-based syntax analyzer as an example, it analyzes how individual words and phrases in the input are combined according to a predefined set of syntax rules, such as word class collocation rules (e.g., a noun and an adjective can form a attributive structure, such as “funny anchor”) and phrase formation rules (e.g., a verb-object structure such as “share interesting stories”), to determine the hierarchical structure between them. For example, for “find anchors with a large number of fans and who often share food”, through syntax analysis, it can be identified that “a large number of fans” and “often share food” are two components that modify “anchor” in parallel, and each of them has different internal syntax structures, which clarifies the hierarchical relationship between words and phrases.

[0118] The analysis result based on the syntax structure further deduces the logical relationship. For example, by identifying the conjunction “and”, it can be clear that there is a logical “and” relationship between the two phrases “a large number of fans” and “often share food”, indicating that anchors that meet both conditions are to be found. For example, if the conjunction “or” appears, it indicates that there is a logical “or” relationship between different components. At the same time, there is a logical relationship between the modifier and the center (e.g., an adjective modifying a noun), which can be analyzed by analyzing the structural characteristics presented at the syntax level, and the complex logical relationship between the components can be completely clarified.

[0119] Therefore, syntax structure analysis can convert the search request expressed in natural language input by the user into a structure with clear hierarchy and logical relationship, so that the system can accurately grasp the query intent expressed by the user. For example, for the complex sentence “find anchors with more than 100,000 fans, good at dancing or singing, from Europe and the United States”, after analyzing the logical relationship between the parts such as “and” and “or” and the modifier and the center, the system clearly knows that anchors that meet the fan number condition, the specific “or” relationship in terms of talent, and the regional condition are to be filtered, avoiding the problem of inaccurate search results due to confusion in understanding the logic of the sentence.

[0120] The clear syntax hierarchy and logical relationship provide a basic framework for subsequent construction of filtering conditions and the entire search process. Subsequent operations can be based on these well-organized relationships to develop database query strategies, set filtering conditions, etc., ensuring the orderliness and accuracy of the entire search process, and making a reasonable transition between natural language requests and actual database operations.

[0121] Optionally, semantic relation recognition often relies on semantic role labeling (SRL) techniques and knowledge graphs. Semantic role labeling can determine the semantic role played by each word or phrase in a sentence, such as the performer, recipient, or time and location of an event. For example, in "the host shares food in the live room", "host" is the performer of the action "share", and "food" is the object being shared. Through such labeling, the semantic relationship between words can be clearly presented. Knowledge graphs contain a wealth of entity and semantic relationship knowledge. By matching the words and phrases in the search request with the nodes and edges in the knowledge graph, deeper semantic associations can be uncovered. For example, if the knowledge graph indicates that "food" is closely related to "cooking", when the user searches for "host sharing food", the system can use the knowledge from the knowledge graph to identify concepts such as "cooking" that are semantically related to "food", expanding the understanding of the semantic association relationship of the search request.

[0122] Semantic similarity calculation methods (such as cosine similarity based on word vectors) are also used to determine the degree of semantic similarity between different words and phrases, and to uncover semantic associations by analyzing the hierarchical structure of concepts (for example, in a classification system, "pets" is a subclass of "animals", and they have a hierarchical semantic relationship). For example, for "find cute animal host", through semantic similarity calculation, it is found that "small animals" and "pets" have high similarity, and are related in concept hierarchy, so the semantic association relationship between them can be identified, and it is known that the user may be interested in hosts who show various small animals, and may also have a search demand for hosts related to pets.

[0123] Therefore, semantic relation recognition can break through the surface understanding of grammar and deeply uncover the semantic connection behind words and phrases, more comprehensively and accurately grasping the true meaning of the user's search request. For example, from a grammatical perspective, "find host who likes outdoor sports" may simply understand the relationship between each part based on structure, but through semantic relation recognition, it can be linked to "outdoor sports" which includes activities such as hiking, trekking, and cycling, and it can be understood that the user may have a search intention for hosts related to these specific outdoor sports, enriching the understanding dimension of the search request and improving the depth and accuracy of semantic understanding.

[0124] Based on semantic association, on the one hand, the search range can be appropriately expanded to include content creation objects that are semantically related although they are expressed differently, improving recall rate; on the other hand, by accurately grasping the semantic association, the objects that truly meet the user's intention can be more accurately filtered out, avoiding the omission of hosts that meet the requirements from a semantic point of view due to only relying on grammatical structure, and improving the quality and accuracy of the search results.

[0125] Optionally, in the specific construction of multi-dimensional screening conditions and label importance weights, according to the hierarchical structure, logical relationship and semantic association relationship between the single words or phrases obtained from the previous analysis, screening conditions are constructed from different dimensions. For example, from the perspective of grammatical structure, if there is an adjective phrase modifying a noun, a screening condition of one dimension can be constructed for the attribute represented by the adjective (such as "funny" corresponding to the screening condition of the "style feature" dimension); from the perspective of logical relationship, for multiple phrases connected by "and", the attributes corresponding to each phrase become independent and need to be satisfied at the same time (such as "many fans" and "often share food" corresponding to the screening conditions of "social influence" and "business-related information" two different dimensions, respectively); from the perspective of semantic association relationship, if the semantic related extension concept is identified, it will also be included in the screening condition of the corresponding dimension (such as "cooking" associated with "food" into the "business-related information" dimension). The screening conditions constructed in this way from multiple dimensions can comprehensively cover all aspects involved in the user search request, forming a three-dimensional screening framework.

[0126] The importance weight of each screening condition in grammar and semantics is usually considered from multiple factors. In terms of grammar, the screening conditions corresponding to the core subject, predicate and other key components are often given higher weights, because they directly determine the key elements such as the subject and core behavior of the search object. For example, in "anchor shares interesting stories", "anchor" as the subject will have a higher weight, and "share" as the core behavior will also have a corresponding weight. In terms of semantics, the screening conditions corresponding to the words or phrases that better reflect the user's core intent and play a key role in distinguishing search results will have higher weights. For example, the user wants to find anchors with "many fans", so this condition is more important than other secondary description conditions in semantics, so it is given a higher weight. By comprehensively considering grammar and semantic factors to reasonably allocate weights, the subsequent screening process can have more emphasis, and important screening requirements can be prioritized according to the weight size.

[0127] Therefore, the construction of multi-dimensional screening conditions can fully consider various information from grammar to semantics in the user search request, avoid missing the content creation object that meets the requirements by only screening from a single angle, ensure that the screening process is comprehensive and detailed, and maximize the user's real search intent, improve the completeness and accuracy of the search results. For example, for a comprehensive description of the search request, multi-dimensional screening can accurately find the corresponding anchor from different attribute directions, whether from basic information, business content or style features, etc.

[0128] The importance weight label enables the system to flexibly adjust the screening focus when facing search requests with different complexity and emphasis. For some requests that emphasize a certain key attribute, the system can prioritize ensuring that the screening condition corresponding to the important attribute is met. For relatively broad and balanced requests, the system can also reasonably allocate screening resources based on weights, improving the system's adaptability and processing effect for various types of search requests, and making the search results better meet the user's diverse expectations.

[0129] Optionally, the method further comprises: vectorizing the single word or the single phrase to obtain a vector sequence; inputting the vector sequence to a syntax attention head and a semantic attention head based on a set time step; performing syntax feature extraction on the vector sequence based on learned part-of-speech information, syntax patterns, and syntax rules of the words by invoking the syntax attention head, to perform syntax structure analysis on the single word or the single phrase, and obtain hierarchical structure and logical relationships between the single words or the single phrases; and performing semantic feature extraction on the vector sequence by invoking the semantic attention head, to perform semantic relationship identification on the single word or the single phrase, and obtain semantic association relationships between the single words or the single phrases.

[0130] Optionally, when implementing single word or phrase vectorization, a pre-trained word vector model (such as Word2Vec, GloVe, BERT, etc.) is used to perform vectorization operations on single words or phrases. These models are trained based on large-scale corpora and can learn rich information about words at the semantic and syntactic levels and map them to low-dimensional vector spaces. Taking Word2Vec as an example, it learns word vectors through word context information, and words with similar contexts will be close in the vector space. For example, “happy” and “joy” will have very close vector representations because they often appear in similar contexts. For phrase vectorization, some models treat phrases as a whole for learning, or obtain phrase vectors by combining the vectors of individual words in the phrase (such as simple weighted averaging, concatenation, etc.). In this way, each word or phrase is converted into a vector with a specific numerical representation, and together they form a vector sequence that can be used to mine the syntactic and semantic information contained in the sequence through vector operations in subsequent processing.

[0131] To this end, vectorization converts words and phrases in natural language into numerical vector form that computers can easily handle, so that different words and phrases can be analyzed and compared in a unified vector space. No matter how complex and diverse the original text form is, subsequent calculations and operations can be based on vectors, providing a standardized data basis for syntax and semantic analysis using deep learning and other technologies, facilitating subsequent model processing, and improving the operability and compatibility of the entire process.

[0132] The vector representation learned by the pre-trained word vector model itself contains rich semantic and syntactic information, which enables subsequent analysis based on vectors to better exploit the inherent relationship between words and phrases with the help of these latent information. For example, the semantic similarity of words can be preliminarily judged by the distance relationship in the vector space, providing valuable clues for more in-depth semantic relationship recognition and other operations, which helps to improve the accuracy and effectiveness of the overall analysis.

[0133] Optionally, when the vector sequence is input to the syntax attention head and the semantic attention head based on the time step, considering that when processing natural language, the role and association of words or phrases in different positions in syntax and semantics are different due to the sequential nature of the text, the time step is a mechanism for processing such sequential information. In recurrent neural networks (RNN) and its variants (such as long short-term memory network, LSTM; gated recurrent unit, GRU) or models based on the Transformer architecture, the corresponding vectors are input to the network in different time steps according to the order of words or phrases in the sentence. For example, for the sentence "anchor shares interesting stories", the vector corresponding to "anchor" is first input in the first time step, the vector corresponding to "share" is input in the second time step, and so on. In this way, the model can capture the sequential dependency between words and phrases according to the order of the time step, and then better analyze their syntactic structure and semantic association, which is crucial for understanding the dynamic changes and contextual relevance of natural language.

[0134] The setting of the syntax attention head and the semantic attention head is an application of a multi-head attention mechanism. Different attention heads can focus on different aspects of feature extraction. The syntax attention head focuses on information related to syntax, such as the part of speech of a word, the syntactic structure pattern of a sentence, etc. The semantic attention head focuses on mining semantic associations and features, such as semantic similarity between words, semantic hierarchical relationship between concepts, etc. Through this way of division of labor and cooperation, the input vector sequence can be analyzed from different angles at the same time, avoiding the problem of insufficient or confused feature extraction caused by mixing syntax and semantic information together. Each attention head can extract the most relevant features, and then integrate these features to fully grasp the various relationships between words and phrases.

[0135] To this end, the time step-based input method enables the model to effectively capture the sequential features of natural language, fully considering the sequential dependency of words or phrases in the entire sentence, which is crucial for accurately analyzing syntactic structures (such as the order relationship of subject-predicate-object, the position relationship between modifiers and center words, etc.) and semantic associations (such as the logical order of events, cause-and-effect relationships, etc. which are reflected through word order). For example, "eat first and walk later" and "walk first and eat later" have the same words but different orders, resulting in differences in semantics and syntax. Through the time step mechanism, the differences can be accurately distinguished, improving the accuracy of natural language understanding.

[0136] Using the syntax attention head and the semantic attention head for feature extraction respectively can more accurately extract the key features of words and phrases in terms of syntax and semantics, achieving more detailed and in-depth structural analysis and relationship recognition. The syntax attention head can accurately sort out the syntactic framework of a sentence, such as determining the hierarchical nesting relationship of components in a complex sentence. The semantic attention head can deeply mine semantic connections, such as discovering metaphors, metonymies, and other deep semantic associations, providing strong support for comprehensively and accurately understanding user search requests and building reasonable filtering conditions, which helps to improve the processing capability of the entire search system for natural language and the quality of search results.

[0137] Optionally, when the syntax attention head performs syntax feature extraction and analysis, the syntax attention head performs syntax feature extraction according to learned part-of-speech information, syntax patterns, and syntax rules. Part-of-speech information (such as nouns, verbs, adjectives, etc.) is the basis for understanding sentence structure. Words with different parts of speech play different grammatical roles in a sentence and have specific combination rules. For example, nouns often serve as subjects or objects, and adjectives are often used to modify nouns. Syntax patterns are summaries of common sentence structure combinations, such as basic structures like “subject-verb-object” and “subject-predicate-attribute” and more complex nested structure patterns (such as “noun modified by a relative clause”). These patterns are learned from a large amount of corpus data and their feature representations in vector form. Syntax rules further specify the legality of word collocation and sentence construction, such as semantic and grammatical collocation requirements between verbs and objects. The syntax attention head focuses on and extracts these information from each vector in the input vector sequence, dynamically focuses on key parts related to syntax using the attention mechanism, calculates the degree of association between different words or phrases in the syntax layer, and then analyzes their hierarchical structure and logical relationship.

[0138] In analyzing the hierarchical structure, the syntax attention head determines which words are in a modifying and modified, limiting and limited relationship based on the grammatical association between words and the matching of syntax patterns, thereby determining the hierarchical nested structure of the sentence from the whole to the part. For example, for “young anchor who likes outdoor sports”, by focusing on the grammatical associations that “young” modifies “anchor” and “outdoor sports” is dominated by “likes” and modifies “anchor” as a whole, the hierarchical structure from outer modification to inner core is teased out. In terms of logical relationship analysis, the logical connection between different parts of the sentence is determined based on the corresponding vector features of conjunctions (such as “and”, “or”, “because”, “so” etc.) and the grammatical parallelism and subordination of different phrases, such as the logical “and” relationship between “likes outdoor sports” and “is good at sharing stories” in “anchor who likes outdoor sports and is good at sharing stories”.

[0139] Therefore, the syntax attention head can accurately parse the complex syntax structure in the natural language search request based on professional syntax-related criteria for feature extraction and analysis, whether it is a simple sentence or a long sentence containing multiple layers of modification and multiple clauses. It can tease out clear hierarchical and logical relationships. This avoids the subjectivity and inaccuracy that may exist in manual syntax analysis, and accurately captures the syntax features of the user input sentence in an automated and intelligent way, providing a reliable syntax basis for subsequent conversion of natural language into executable filtering conditions and other operations, ensuring the accuracy of the search process.

[0140] Since it works based on learned extensive grammar patterns and rules, it can adapt to various grammar expressions of different users, effectively processes sentences that are novel and complex in form but conform to grammar specifications, improves the system's inclusiveness and adaptability to the diversity of natural language grammar, and enables different users to express search requests in their own grammar, while the system accurately understands the grammar connotation, improving user experience.

[0141] Optionally, when the semantic attention head performs semantic feature extraction and analysis, the semantic attention head mainly focuses on the essential connection of words and phrases in the semantic layer when performing semantic feature extraction on the vector sequence. On the one hand, the semantic association is mined through the distance and similarity relationship in the vector space. Because the pre-trained word vector model makes the vectors of semantically similar words close in distance, vector operations such as cosine similarity calculation can be used to find which words or phrases have semantic similarity, and then infer the possible semantic association between them. For example, the vector similarity of "delicacy" and "delicacy" is high, indicating that they have a close semantic association and may represent similar concept categories in the search request. On the other hand, external knowledge and technical means such as knowledge graph and semantic role labeling are used to further enrich semantic analysis. Entities and relationships in the knowledge graph can help expand the scope of semantic understanding. For example, knowing that "delicacy" and "cooking" and "restaurant" are related entities in the graph can help identify broader semantic connections. Semantic role labeling can clearly define the specific semantic roles of words in events or situations, helping to understand how different words interact semantically. For example, the roles of the words in "anchor shares delicacy in live room" clearly show their semantic associations such as action and object. The semantic attention head integrates these information to dynamically focus on and extract key features related to semantics, and analyzes the semantic association between words and phrases.

[0142] When mining semantic association relationships, the semantic attention head not only limits to superficial synonym and near-synonym relationships, but also goes deep into concept hierarchical relationships, whole-part relationships, cause-and-effect relationships and other deep semantic connections. For example, for "find cute animal anchor", not only can it identify the semantic similarity between "small animal" and "pet", but also can mine the upper concept of "animal" through knowledge graph or concept hierarchy analysis, thereby expanding the understanding of the semantic category of the search request and understanding that the user may have the intention to find anchors related to various animals, more comprehensively and deeply grasping the semantic intention behind the user's search request, and providing rich semantic basis for more accurate selection of content creation objects.

[0143] To this end, the semantic attention head can deeply mine semantic features and relationships in multiple ways, breaking through the surface understanding of the text and touching the deeper semantic connotations behind natural language, more comprehensively and accurately grasping the user's true search intent. This is very helpful for processing search requests that are semantically ambiguous, contain metaphors, or require background knowledge to understand, avoiding situations where search results do not meet user expectations due to literal understanding, and improving the depth and accuracy of natural language semantic understanding.

[0144] With various semantic analysis means, the semantic attention head can expand the scope of semantic association, consider more concepts related to the search request semantics but expressed differently, expand the potential scope of the search, and improve recall. For example, when searching for anchors related to "healthy life", multiple related concepts such as "fitness", "nutrition diet", and "mental health" are mined through semantic association, making the search results more comprehensive and rich, and more likely to meet the user's diverse needs, further optimizing the performance and user experience of the search system.

[0145] Optionally, the converting the logical framework into a structural expression to perform database table field matching in the content creation object database based on the structural expression comprises: calling a conversion mode of content creation object attribute features between the logical framework and the structural expression, so that each filtering condition in the logical framework is combined according to a predetermined template to convert the logical framework into a structural expression; a query optimizer generates a query plan according to the structural expression index information to determine the access path and connection order of the data; and the database table field matching in the content creation object database is performed according to the access path and connection order of the data.

[0146] Optionally, during the conversion of the logical framework to the structural expression, first, for various attribute features of the content creation object, an explicit correspondence and conversion mode between the logical framework and the structural expression is established during the system design phase. For example, if the content creation object has attribute features such as "anchor name", "broadcast category", and "fan count", the filtering conditions described in natural language in the logical framework (such as "find food anchors with more than 100,000 fans") are presented, while in the structural expression, they need to be mapped to specific table structures and field representations in the database. For example, the "fan count" attribute may be stored in a field named "fan_count" in the database, and "food anchors" correspond to records with a value of "food" in the "category" field. This mapping relationship is determined based on the design patterns of the database (such as table structure, field definition, etc.) and the standardized storage method of the content creation object attributes.

[0147] Each filtering condition is combined according to a predetermined template, which usually follows the specification requirements of a database query language (such as the syntax structure of SQL statements). For example, for multiple filtering conditions with an "and" relationship in the logical framework, when converted into a structured expression, the template for connecting multiple conditions with "AND" in the "WHERE" clause of SQL is used for combination, and the field conditions corresponding to each attribute are spliced together. Taking "find anchors with fan count greater than 100,000 and category as food" as an example, the structured expression may be similar to "SELECT * FROM anchor table WHERE fan_count > 100,000 AND category = 'food'". Through such combination conversion according to the predetermined template, the logical framework can be accurately converted into a structured expression that can be recognized and executed by the database, realizing the transition from natural language intent to database operation statement.

[0148] To this end, through explicit attribute feature mapping and predetermined template combination, the user search intent embodied in the logical framework can be accurately translated into an expression that meets the requirements of the database structure, so that the system can directly use the query function of the database for subsequent operations. Avoid the problem of being unable to effectively execute the search due to the difference between natural language and database structure, ensure smooth connection from user input to database query, and improve the reliability of accurate execution of search requests at the database level.

[0149] Using a predetermined template for conversion follows a unified rule, making the entire conversion process standardized. No matter how complex and diverse the filtering conditions in the logical framework are, they can be converted into structured expressions in a predetermined manner, facilitating system management and maintenance. If the database structure changes later (such as adding new fields, modifying table names, etc.), only the corresponding mapping relationship and template part need to be adjusted, without the need for large-scale modification of the entire search logic, reducing system maintenance cost and improving system adaptability and expandability.

[0150] Optionally, when the query optimizer generates the query plan, one of the key factors is the structural expression index information. Index in a database is a data structure that is similar to a book's table of contents, which can help quickly locate and access data. For example, an index is established for the "fan_count" field in the "anchor table" that is frequently used for query filtering. When the structural expression involves a query on the "fan_count" condition (such as "fan_count > 100000"), the query optimizer will identify this index and analyze how to find the data records that meet the condition most quickly based on the storage structure and data distribution of the index. Different types of indexes (such as B-tree index, hash index, etc.) are suitable for different data characteristics and query scenarios. The query optimizer will select the most effective utilization method based on the characteristics of the index and the specific conditions in the structural expression to determine the data access path, i.e., the order of operations such as which tables to access first, which index to use to find data, etc.

[0151] In a multi-table join query (such as finding a condition that satisfies both the anchor's own attributes and its joint brand attributes, which may involve a join query of "anchor table" and "joint brand table"), the query optimizer also needs to determine the join order between tables. It will consider the size of the table, the data distribution of the associated fields, the index situation, and the impact of different join orders on query performance, etc. By analyzing the statistical information (such as the number of records in each table, the distribution frequency of different values, etc.) and the association relationship reflected in the structural expression, the optimal join order is calculated to enable the multi-table join operation to consume the least resources and take the shortest time to obtain the required data, improving query efficiency.

[0152] Therefore, using index information to generate a query plan can quickly locate data that meets the filtering conditions, avoid time-consuming operations such as full table scanning, greatly shorten the query time, and especially in the case of large-scale data content creation object databases, the efficiency improvement effect is more significant. Reasonably determining the join order can also reduce unnecessary data association and calculation, so that multi-table queries can be efficiently executed, the overall system response speed is improved, and the user's search experience is enhanced, enabling users to quickly obtain search results.

[0153] By scientifically planning the data access path and join order, the database system can more reasonably allocate system resources such as memory and disk I / O, avoiding wasting resources on unnecessary query operations. For example, preferentially using index to find data can reduce the amount of data read from the disk, reducing the disk I / O pressure, so that the system can run more smoothly and efficiently under limited resources, while also being able to handle more concurrent search requests, enhancing the overall performance and stability of the system.

[0154] Optionally, when performing database table field matching based on the access path and connection order, the database system will access the relevant tables in sequence according to the determined access path, and perform matching operations on the field conditions involved in the structural expression when accessing each table. For example, when accessing the "anchor table", it checks whether the value of the "fan_count" field meets the condition of "greater than 100,000" and whether the value of the "category" field is "food", etc. Through the database's own query execution engine, it compares the field values in the records with the conditions in the structural expression row by row. For multi-table join query cases, the relevant records in different tables are associated and matched according to the connection order, and the association fields (such as "anchor ID" in the "anchor table" and "cooperation anchor ID" in the "cooperation brand table") are used to establish the connection between the data. Then, the matching of other field conditions is continued in the connected result set to ensure that the records that fully meet all the filtering conditions set by the structural expression are found, that is, the database records corresponding to the content creation objects that meet the user's search request are returned.

[0155] During the entire matching process, the database system will follow its internal data consistency mechanism to ensure the integrity and accuracy of the data. For example, in a relational database, through transaction processing, foreign key constraints, and other mechanisms, it ensures that in the case of multi-table association operations and field value updates, deletions, etc., the data will not be inconsistent, lost, or incorrectly associated. In this way, no matter how complex the query operation is, the final matching result returned is based on accurate and reliable data, which can truly reflect the content creation object information that meets the user's search intent.

[0156] Therefore, strictly following the determined access path and connection order for database table field matching can accurately filter out the database records corresponding to the content creation objects that meet all the filtering conditions, avoiding missing or incorrectly matching data, ensuring the accuracy and completeness of the search results, so that the object information returned to the user is truly in line with their search request, improving the search quality and meeting the user's demand for finding specific content creation objects.

[0157] Relying on the database's own data consistency mechanism ensures the quality and reliability of the data during the entire matching process, so that users can rely on the data source of the search results, avoid questioning the search results due to data problems, enhance the user's trust in the search system, and improve the overall reputation and usability of the system.

[0158] Optionally, based on the matched database table fields, a search for content creation objects is performed in the content creation object database to obtain a number of content creation object description information, including: dynamically constructing an SQL query statement based on the matched database table fields; creating a search task according to the SQL query statement, the search task including multiple subtasks; and performing paginated query and batch reading of the content creation object database based on multiple subtasks.

[0159] Optionally, when dynamically constructing an SQL query statement based on the matched database table fields, after obtaining the matched database table fields, it is necessary to dynamically construct an SQL query statement based on these field information and the query syntax rules of the database. First, determine the target table to be queried (usually the main table that stores information related to the content creation object and other related tables that may involve associated queries). For example, when looking for information related to the anchor, the corresponding main table may be the "anchor information table". Then, the matched fields are combined according to the grammatical structure of the SQL statement, generally including a "SELECT" clause to specify the list of fields to be returned (which may be all fields or some key fields, such as the anchor's name, number of fans, broadcast categories, etc.), a "FROM" clause to specify which table to query data from, and a "WHERE" clause to logically combine the matched field conditions, such as "WHERE fan_count>100000AND category='美食'" to splice them together according to specific filtering conditions. If a query involves multiple tables, the "FROM" clause will use appropriate join methods (such as "JOIN" statements) and related fields to connect different tables to ensure that complete information that meets the conditions is obtained. In addition, an "ORDER BY" clause may be added to sort the results (for example, in descending order by the number of followers) and a "LIMIT" clause may be added to limit the number of results returned. Through the appropriate combination of these components, a complete SQL query statement that meets the current search requirements is dynamically generated.

[0160] To enhance the flexibility and versatility of query statements, parameterized SQL query statements are often constructed. For example, specific numerical values ​​(such as a specific threshold for the number of followers) and string values ​​(such as the specific category of the broadcast category) in the filtering conditions are passed in as parameters rather than being hard-coded directly in the statement. This way, under different search requests, simply changing the corresponding parameter values ​​can easily generate different query statements to adapt to various changing filtering conditions. This can also effectively prevent SQL injection attacks and improve system security.

[0161] To this end, by constructing a SQL query statement according to the matched database table fields, it can be ensured that the query statement accurately reflects the filtering conditions corresponding to the user search request, so that the query operation performed by the database completely matches the feature requirements of the content creation object to be found, thereby accurately obtaining the data records meeting the conditions, improving the matching degree of the search results and the user's intention, and avoiding querying irrelevant or inaccurate information.

[0162] The parameterized construction and dynamic combination in accordance with the SQL syntax rules enable the system to easily cope with various search requests. Whether the user is searching for content creation objects with a specific fan number range, different live broadcast categories, specific regions, or any other condition combination, the system can quickly generate the corresponding query statement for querying, thereby enhancing the adaptability of the system to different search scenarios and requirements, and improving the versatility and scalability of the system.

[0163] Optionally, when creating a search task and dividing sub-tasks according to the SQL query statement, a search task is created according to the constructed SQL query statement. This search task is a logical whole operation and represents the complete process of finding content creation objects meeting the requirements from the database. In actual execution, in order to improve efficiency and facilitate management, the whole search task is divided into multiple sub-tasks. For example, according to the storage structure and query characteristics of the database, the sub-tasks can be divided according to the partitioning of data (if the database adopts partitioned storage, such as time partitioning, regional partitioning, etc.) or logical grouping of data (such as grouping according to different live broadcast categories, etc.). Each sub-task is responsible for querying a part of the data range, such as a sub-task responsible for querying anchor information meeting the conditions in a certain regional partition, another sub-task responsible for querying related anchor data under another live broadcast category group, etc. Through reasonable division, the whole search task can be executed in parallel or distributed, improving the overall speed of querying.

[0164] There will be a corresponding task scheduler in the system to manage these sub-tasks. It will arrange the execution order and allocate computing resources of the sub-tasks according to the resource status of the database server (such as CPU usage, memory free situation, etc.), the priority of the sub-tasks (which can be set according to the importance, urgency, etc. of the search request), and the dependency relationship between the sub-tasks (some sub-tasks may need to wait for other sub-tasks to complete part of the pre-operation before starting). For example, when the server CPU resources are sufficient, the scheduler will start multiple sub-tasks that can be executed in parallel at the same time, speeding up the search process; when the resources are tight, the sub-tasks will be executed in order according to the priority, ensuring that important search requests can be processed first, guaranteeing the efficient operation and rational use of resources of the system as a whole.

[0165] To this end, the search task is divided into multiple sub-tasks and reasonably scheduled, which can fully utilize the multi-core CPU, distributed storage and other resource advantages of the database server, and realize parallel or distributed query. For example, on a multi-core database server, multiple sub-tasks can be executed simultaneously on different cores, greatly shortening the overall query time, especially when dealing with large-scale data content creation object database, the advantage of parallel processing is more obvious, which can quickly respond to user search requests,

[0166] Through the task scheduler, the execution of sub-tasks is coordinated according to factors such as resource status and task priority, so that the system can flexibly cope with different load situations, avoid excessive use or waste of resources, and ensure that each search task can be processed within a reasonable time, improving the utilization of database server resources, enhancing the stability and reliability of the system in high-concurrency scenarios, and enabling the system to better serve multiple users simultaneously initiated search requirements.

[0167] Optionally, in the content creation object search based on sub-tasks for paging query and batch reading, the paging query method is used when executing sub-tasks, considering that when the query result set may be very large, if all the data is read at once, it will occupy a large amount of memory resources, and even may cause memory overflow and other problems, and also increase the burden of network transmission. Paging query is based on the set number of records per page (such as displaying 100 records per page) and page number (such as querying page 1, page 2, etc.), and is realized in the SQL query statement by combining the "LIMIT" clause with the "OFFSET" clause (used to specify the offset, that is, from which record in the result set to start taking). For example, to query the second page of content creation object information with 100 records per page, the corresponding SQL statement may be "SELECT * FROM anchor information table WHERE... LIMIT 100 OFFSET 100", so that only a small amount of data of the specified page is obtained each time, and the entire result set that meets the conditions is traversed gradually, which can meet the query requirements and effectively control the memory usage and network transmission volume.

[0168] Batch reading is to read a batch of data (such as reading 10 pages of data at a time) at a time based on each paging query, rather than getting data record by record page by page, which can reduce the overhead of frequent database connection and query operation, and improve reading efficiency. The database system internally caches and manages the batch read data to optimize data transmission and storage. When the sub-task performs a query operation, it first obtains the corresponding data block from the database according to the batch reading strategy, and then further filters, processes and other operations (such as sorting and deduplicating the read content creation object description information according to specific search requirements) on these data in memory to obtain the final content creation object description information that meets the requirements, and complete the entire search process.

[0169] To this end, the pagination query effectively avoids the memory and network resource pressure problems caused by large-scale data reading at one time, so that the system can stably process large-scale query result sets, is suitable for content creation object databases of various sizes, and improves the robustness of the system. Batch reading reduces database connection and query overhead while improving data acquisition efficiency, optimizing the performance of the entire search process, enabling the system to quickly acquire the required content creation object description information, and improving the overall efficiency and response speed of the search.

[0170] The pagination query method also facilitates user viewing and management of large search results, and users can choose to view different page contents according to their own needs, in line with common user browsing habits. Moreover, through batch reading and further processing of data in memory, the system can optimize the results according to specific circumstances, such as removing duplicate information, arranging in a more reasonable order, etc., improving the quality of search results and providing users with a better and more convenient search experience.

[0171] Optionally, the conversion of the search request expression into a question vector includes: after text preprocessing of the search request expression, using a word segmentation tool to segment the search request expression into individual words or phrase sequences; assigning a pre-trained word vector to each segmented word or phrase sequence, and mapping it to a low-dimensional vector space; assigning a semantic weight according to the semantic importance of each segmented word or phrase sequence; based on the semantic weight, performing fusion processing on all word vectors to obtain the question vector.

[0172] Optionally, during text preprocessing and word segmentation operations, during text preprocessing of the search request expression, normalization operations are first performed, which include unifying the case form of the text (such as converting all characters to lowercase, because in many natural language processing tasks, case does not affect semantics but increases processing complexity), removing unnecessary spaces and special characters (such as some meaningless punctuation marks, tab characters, etc., which may interfere with subsequent word segmentation and semantic analysis), and standardizing some special format texts (such as converting full-width characters to half-width characters, etc.). Through these normalization operations, the input search request expression is more uniform and regular in format, facilitating subsequent processing procedures.

[0173] Stop words are those words in natural language that appear frequently but do not help much in expressing core semantics, such as "of", "is", "in", etc. The system will identify and remove these stop words according to a pre-constructed stop word list (which is usually a collection of common stop words based on statistical analysis of a large corpus) or through some statistical methods (such as according to word frequency, set a frequency threshold, and determine the stop words as the words with a frequency higher than the threshold and with a more virtual semantics). Removing stop words can simplify the text content, focus the subsequent word segmentation operation on more substantial semantic words and phrases, reduce unnecessary computational load, and highlight the key information in the search request expression.

[0174] The pre-processed search request expression is segmented into individual words or phrase sequences using a word segmentation tool. Common word segmentation tools such as Jieba (for Chinese), NLTK (for English and other languages), etc. generally use statistical machine learning methods or rule matching methods for word segmentation. Taking the statistical method as an example, statistical information such as the frequency of word appearance in a large corpus and the adjacency relationship between words is used to determine where to segment to be the most reasonable. For example, for the Chinese sentence "I like to watch funny pet videos", by statistically analyzing the probability of words such as "like" and "funny" appearing as independent words and the probability of them forming other words with the words before and after them, the sentence can be accurately segmented into individual word sequences such as "I", "like", "watch", "funny", "pet", "video", etc. For some fixed collocation phrases (such as "pet video"), they can also be identified as a phrase unit based on factors such as their frequency of appearance as a whole in the corpus, so that the segmented result is more consistent with the actual usage habits and semantic structure of the language.

[0175] Therefore, text preprocessing effectively reduces noise and redundant information in the text data through normalization and stop word removal, making the subsequent processed data cleaner and more concise. This not only reduces the computational complexity of subsequent word segmentation and vector mapping operations, improves processing efficiency, but also avoids word segmentation errors or semantic understanding biases caused by irrelevant information, improving the data quality and accuracy of the entire processing flow.

[0176] Using professional word segmentation tools for segmentation can accurately divide the search request expression into individual words or reasonable phrase sequences based on the internal rules and statistical characteristics of the language, providing a precise semantic unit basis for subsequent vector mapping and semantic weight allocation based on word and phrase levels, ensuring that each unit can better carry and reflect a part of the semantic information, which helps to more accurately represent and analyze the entire search request expression, and thus improves the accuracy of the search.

[0177] Optionally, when assigning pre-trained word vectors

[0178] Pre-trained word vector models (such as Word2Vec, GloVe, BERT, etc.) are trained based on large-scale corpora. These models learn the features of words at the semantic and syntactic levels through different training methods and represent them as low-dimensional vectors. Taking Word2Vec as an example, it has two common training architectures, CBOW (Continuous Bag of Words) and Skip-gram. CBOW predicts a word from its context, for example, given a few words before and after the word "like" (such as "I" "watch" "funny" etc.), it predicts that the middle word is "like"; Skip-gram is the opposite, it predicts the context words from a word. Through such a large amount of corpus training, the model can capture the semantic similarity between words, so that semantically similar words are closer in vector space. For example, "happy" and "happy" are often used in similar contexts, so their vector representations are very close, and similarly, for phrases, some models can also generate corresponding vector representations through specific methods (such as weighted average, concatenation, etc.).

[0179] When assigning pre-trained word vectors to each segmented word or phrase sequence, the segmented language units are matched with the vocabulary in the pre-trained word vector model, and the corresponding pre-trained vector representation is found, so that the word or phrase is mapped from text form to low-dimensional vector space. For example, for the segmented word "funny", find its corresponding vector in the Word2Vec model vocabulary (assuming a 300-dimensional vector, the specific dimension depends on the model settings), then use this vector to represent the numerical representation of the "funny" word in subsequent semantic analysis and other operations, so that the text information that is difficult to perform mathematical calculations and analysis can be converted into a vector form that can be easily processed by the computer, laying the foundation for subsequent vector-based semantic weight assignment and fusion processing operations.

[0180] Therefore, the pre-trained word vector model contains a large amount of semantic and syntactic information, which is learned from massive corpora. Mapping segmented words or phrases into such a vector space is equivalent to indirectly using this rich prior knowledge, which can better capture the semantic similarity and relevance between words and phrases, providing strong support for understanding the semantics of search request expressions. For example, when judging the similarity of the two search request expressions "funny pet host" and "funny animal host", based on the similarity of the word vectors "funny" and "funny", "pet" and "animal" in the word vector space, the semantic similarity of the two search request expressions can be more accurately evaluated, and the accuracy of the search result matching can be improved.

[0181] The different words and phrases are all converted into fixed-dimensional vector form, realizing the standardization and numerical representation of the text data, so that subsequent mathematical operations and algorithms can be used to process these vectors, such as vector addition, multiplication, and distance and similarity calculation between vectors. This unified data representation greatly simplifies the processing flow, facilitates semantic weight allocation and fusion processing at the vector level, improves the operability and efficiency of the entire search request vector representation process, and also provides convenience for connection with other vector-based processing modules (such as subsequent vector fusion and vector matching with content creation object description information).

[0182] Optionally, when assigning semantic weights, the semantic weights are assigned according to the semantic importance of each segmented word or phrase sequence. This evaluation process usually considers multiple factors. On the one hand, from the perspective of grammatical structure, the words or phrases corresponding to the core subject, predicate, and other key components often have high semantic importance. For example, in "anchor shares interesting stories", "anchor" is the subject of the action, and "shares" is the core action, which plays a key role in expressing the core semantics of the entire sentence, so they are assigned relatively high weights. On the other hand, from the semantic level, those words or phrases that better reflect the user's core search intent and have an important influence on distinguishing search results also receive higher weights. For example, if the user wants to find an anchor with a large number of fans, the phrase "a large number of fans" has high semantic importance in the search request "find anchors with a large number of fans", and accordingly is assigned a large semantic weight. In addition, statistical information of words or phrases in a large-scale corpus (such as term frequency, inverse document frequency, etc., which measures their importance through TF-IDF algorithms) and their association degree in a specific domain knowledge graph can be used to assist in determining the semantic weight, and multiple factors are considered to more accurately measure the importance of each semantic unit.

[0183] In actual weight assignment, various methods can be used, such as manually setting the weight range (e.g., setting a range of 0 to 1, and assigning different words or phrases corresponding values according to importance), rule-based weight assignment (assigning fixed weight values to words or phrases that meet specific rules according to pre-set grammar and semantic rules), and learning weights through machine learning models (using labeled training data to let the model automatically learn appropriate weights for different words or phrases in different search request expressions). For example, for the search request expression "find anchors who are humorous and have a large number of fans", it is analyzed that both "humorous" and "a large number of fans" are important, and the weights are manually assigned as 0.5 (assuming the total weight is 1, and both are equally important and share the weight), and then this weight is used to affect their role in subsequent operations such as vector fusion processing.

[0184] To this end, by reasonably allocating semantic weights, those words and phrases in the search request expression that are most critical to embodying the user's intent can be highlighted, so that in the subsequent process of fusing multiple word vectors into a question vector, these core semantic information can be fully embodied and avoid being overwhelmed by some secondary, relatively weak semantic words and phrases, so that the question vector is more focused on the content creation object features that the user really wants to find, improving the representativeness and accuracy of the question vector to the search request semantics, which helps to more accurately match the content creation object that meets the user's intent.

[0185] Different search request expressions have different key points and core semantics, and the semantic weight allocation mechanism can flexibly adjust the weight of each word or phrase according to the specific request content. Whether the search request focuses on describing the anchor style characteristics, business-related information, or other attributes, appropriate weight allocation can accurately reflect its semantic importance, enhancing the adaptability and processing capacity of the system to various types of search requests. No matter how the user emphasizes the search focus, the system can effectively capture and reflect it in the vector representation, improving the user experience.

[0186] Optionally, when performing vector fusion processing based on semantic weights, all word vectors are fused based on the allocated semantic weights. Common fusion methods include weighted average method, weighted summation method, etc. Taking the weighted average method as an example, assuming that after tokenization and word vector allocation, the word vector corresponding to the word "anchor" is with a weight w1, and the word vector corresponding to "funny" is with a weight w2, then the fused question vector can be calculated by the following formula: (Here, only two vectors are used as an example, and in actual situations, all word vectors corresponding to the tokenized words or phrases will be included). By such weighted average, each word vector is integrated according to its semantic weight, so that the fused question vector integrates the semantic information of different words and phrases, and highlights the important semantic part according to the weight. The weighted summation method is similar, except that it does not need to divide by the sum of the weights, i.e. No matter which method, the core is to integrate each word vector according to the weight, so that the final question vector can better reflect the overall semantics of the entire search request expression.

[0187] In the process of vector fusion, it is necessary to ensure that the dimensions of each word vector are consistent (which is usually guaranteed when assigning pre-trained word vectors, which generally output fixed-dimensional vectors) so that corresponding mathematical operations can be performed. At the same time, sometimes in order to ensure that the numerical range and properties of the fused vector meet the requirements of subsequent processing (such as in some similarity calculation and other operations, it is hoped that the length and other properties of the vector are in a suitable range), the fused vector is also normalized. Common normalization methods include dividing the vector by its length to make it a unit vector (i.e., a vector with a length of 1), or mapping the values of the elements of the vector to a specific interval (such as the interval from 0 to 1), through these processing, the question vector is more conducive to subsequent matching and other operations with the content creation object description information vector in terms of numerical value and properties.

[0188] To this end, the vector fusion process integrates the word vectors corresponding to each segmented word or phrase according to semantic weights, generating a question vector that can represent the entire search request expression. This question vector integrates the information of each semantic unit in the search request, realizing the conversion from dispersed words and phrases to overall semantic representation, providing a unified and comprehensive semantic representation basis for subsequent vector-level matching and comparison with content creation object description information, facilitating accurate measurement of the matching degree between the search request and the content creation object.

[0189] Through dimension processing and normalization of the fused vector, the numerical properties of the question vector are optimized, making it more reasonable and comparable when performing similarity calculation, matching degree evaluation and other operations with other vectors. For example, after normalization, the range of the result of the cosine similarity calculation is more in line with the conventional semantic similarity judgment habit (the cosine similarity is between -1 and 1, and normalization helps to make the result more accurately reflect the actual semantic matching situation), improving the accuracy and reliability of subsequent search result screening based on vector matching, and thus improving the performance and search result quality of the entire search system.

[0190] Optionally, the converting each piece of content creation object description information into a context sentence vector comprises:

[0191] For each piece of content creation object description information, the key attributes therein are taken as nodes, and the relationships between the attributes are taken as edges to construct an attribute relationship graph;

[0192] So that the nodes propagate information through the edges, wherein in each layer of propagation, the nodes update their vector representation according to the information of their neighbor nodes, and the propagation formula can be expressed as: Wherein is the vector representation of node i at the l+1 layer, N(i) is the neighbor node set of node i, and a ijis the attention weight of the edge between node i and node j (used to determine the importance of neighbor node information), W (l) is the weight matrix of the l-th layer, and σ is the activation function, and α ij The correlation score between the node itself and the vector representation of the neighbor node is calculated, and then normalized by the softmax function to obtain the weight;

[0193] Based on the created virtual center node instance, it is connected with all other nodes in the attribute relationship graph, and then the center node learns the comprehensive information of the whole graph through information propagation, to obtain the vector of the center node and take it as the context sentence vector.

[0194] To this end, the technical processing process of converting the content creation object description information into the context sentence vector has the following technical benefits in terms of mining semantic association, enriching vector semantics, highlighting important relationships, and generating unified representation, etc. through the construction of attribute relationship graph, information propagation to update node vector, and the use of virtual center node to integrate information, etc.

[0195] I. The benefits of constructing an attribute relationship graph

[0196] By taking the key attributes in the content creation object description information as nodes and the relationships between the attributes as edges to construct an attribute relationship graph, the originally relatively scattered and disordered description information can be converted into a structured representation form. For example, for the description information related to the anchor, attributes such as "anchor name", "broadcasting category", "number of fans", "style characteristics", etc. become nodes in the graph, and relationships such as "the anchor is good at the live content category" between "anchor name" and "broadcasting category" exist as edges to connect them. This graph structure can clearly present the internal logic and association between each attribute, no longer treating each attribute in isolation, but grasping how different attributes of the content creation object are influenced and associated with each other as a whole, which helps to deeply mine the deep semantic and feature information.

[0197] Therefore, this structured representation helps to more accurately capture the overall picture of the content creation object in the subsequent process, avoiding missing the synergistic effect and implicit semantics between attributes due to only focusing on a single attribute. For example, when searching for anchors related to "food sharing and large number of fans", through the attribute relationship graph, it can be found that there may be a positive correlation between the "food sharing" broadcasting category attribute and the "number of fans" (because high-quality food sharing content often attracts more fans), so that this correlation can be better reflected in the vector representation, improving the recognition accuracy of the content creation object that meets the requirements, and making the search results more in line with the user's search intent.

[0198] II. The benefits of updating the node vector representation based on information propagation

[0199] In the information propagation process, the node updates its vector representation according to the information of its neighbor nodes, which means that each node is no longer limited to the attribute information it initially carries, but can absorb the information of other attributes (neighbor nodes) associated with it. Through the propagation formula, the vector representation of the neighbor nodes, the attention weight of the edge, and the corresponding weight matrix and activation function are comprehensively considered. For example, for the "style characteristics" node describing the anchor, its neighbor nodes may include "launch category" and "fan interaction situation". During information propagation, according to the relevance between them and their importance (reflected by the attention weight), part of the information of these neighbor nodes is integrated into the vector representation of the "style characteristics" node, making the semantics contained in its vector more rich and reflecting more characteristics of the anchor in different aspects.

[0200] Therefore, the complex characteristics of the content creation object can be more comprehensively described, and the vector representation corresponding to each attribute can contain information of other attributes related to it, reflecting the specific situation of the attribute in the entire content creation object from multiple dimensions. This is very crucial for accurately understanding and representing the content creation object, especially when facing complex situations with multiple attributes influencing each other. It can avoid the one-sidedness of single attribute vector representation, improve the accuracy and completeness of vector description of the content creation object, and further improve the precision of subsequent search based on vector matching.

[0201] The attention weight used in the propagation process is used to determine the importance of the neighbor node information. It calculates the correlation score between nodes and then normalizes it through the softmax function to get the weight value. This mechanism allows different importance weights to be assigned to neighbor nodes according to their actual association with the current node when integrating neighbor node information. For example, for the "fan number" node, the "launch category" node may have a stronger correlation in some cases (such as specific popular categories that are more likely to attract fans), so the attention weight corresponding to the "launch category" node will be relatively high, and its information will have a larger proportion in updating the vector representation of the "fan number" node, highlighting more critical association information and filtering out relatively minor association influences.

[0202] To this end, the attention mechanism can focus on the truly important relationships between the attributes of content creation objects, avoiding interference from irrelevant or weakly relevant information. This allows the vector representation to more accurately reflect the factors that significantly influence the characteristics of the content creation object and its match with the search request. During the search process, this helps to more accurately distinguish the differences in the association of key attributes between different content creation objects, improving the ability to select content creation objects that meet the user's search intent, making search results more targeted and reliable.

[0203] 3. Benefits of Learning Comprehensive Graph Information as Context Sentence Vectors through Virtual Central Nodes

[0204] Create a virtual central node instance and connect it to all other nodes in the attribute relationship graph, and then let it learn the comprehensive information of the entire graph through information propagation. This virtual central node is like a convergence point, which can collect and integrate all the information contained in each attribute node and their mutual relationships. After multiple rounds of information propagation, the information it learns covers the full picture of the content creation object represented by the entire attribute relationship graph. By expressing this comprehensive information in the form of a vector, we can obtain a context sentence vector that can represent the descriptive information of the content creation object. For example, for a piece of information that describes all aspects of the anchor in detail, through the information aggregation and integration of the central node, its final vector can fully reflect the comprehensive characteristics of the anchor from basic information to business content, style characteristics, and social influence, forming a unified and holistic semantic representation.

[0205] This provides a complete and representative vector representation for subsequent matching and comparison with the question vector, avoiding the difficulty of comprehensively measuring the degree of match with the search request due to the use of multiple, scattered attribute vectors. This contextual sentence vector can be directly used to perform similarity calculations with the question vector, making it more intuitive and effective to determine whether the content creation object meets the user's search intent. This improves the convenience and accuracy of the vector matching process and helps to quickly and accurately screen suitable content creation objects.

[0206] No matter how complex the structure of the content creation object description information itself is, the corresponding attribute relationship graph structure can be very different, but the way of learning comprehensive information through a virtual center node has good adaptability. Because no matter how the association relationship between attributes is simple or complex, sparse or dense, the virtual center node can obtain and integrate the corresponding information through the connection with each node and the information propagation process, and finally generate the corresponding context sentence vector. For example, for a content creation object information that only contains a few key attributes, or a piece of information that contains a large number of detailed attributes and complex association relationships, it can be converted into a suitable vector representation through this mechanism, so that the system can perform subsequent search matching operations in the form of a unified vector when processing various description information with different levels of detail and structure.

[0207] Therefore, the processing capability of the system for content creation object description information in different formats and different levels of detail is enhanced, the versatility and compatibility of the system are improved, and the system can better cope with the diversified data that may occur in actual applications, ensuring that the description information can be accurately converted into a vector for matching and screening in various situations, and ensuring the stable operation of the search system and the output of effective search results.

[0208] Optionally, the calculation of the matching degree evaluation score of the question vector and the context sentence vector comprises:

[0209] The question vector and the context sentence vector are spliced to obtain an input vector, and the input vector is input into a hidden layer;

[0210] The question vector and the context sentence vector are spliced to obtain an input vector Input vector is the question vector and the context sentence vector are spliced.

[0211] Calculate the output of the hidden layer First calculate For each neuron j in the hidden layer, calculate Then apply the ReLU function, that is W 1： is the weight matrix from the input layer to the hidden layer, b 1： is the bias vector of the hidden layer, W 1ji is the element in the jth row and ith column of the weight matrix W 1： ;

[0212] Calculate the output of the output layer as the matching degree score y: first calculate That is Then apply the Sigmoid function to obtain the matching degree evaluation score, W2： is a weight matrix from the hidden layer to the output layer, W 21j represents the weight matrix W 2： the element in the first row and jth column of the matrix.

[0213] The following are the technical benefits of the above technical process of calculating the matching degree evaluation score of the question vector and the context sentence vector:

[0214] I. The benefits of vector splicing

[0215] Splicing the question vector and the context sentence vector to obtain the input vector can integrate the information represented by the question vector and the information contained in the context sentence vector that embodies the description information of the content creation object into a unified vector representation. For example, the question vector carries key semantic information such as "funny style, anchor with large number of fans" that the user wants to find, and the context sentence vector covers the actual attribute situation of a specific anchor in terms of style characteristics, number of fans, etc. Through splicing, the subsequent neural network can consider both aspects of information at the same time, avoid the information fragmentation problem that may occur when processing separately, and lay the foundation for measuring the matching degree of the two as a whole.

[0216] Therefore, this integration method provides convenience for the comprehensive analysis and processing of information by the subsequent neural network, so that the system can judge the fit degree of the content creation object and the search request based on complete information in a unified framework, more comprehensively and accurately evaluate the matching degree, avoid the matching result deviation caused by only focusing on part of the information, improve the accuracy and reliability of the matching degree evaluation, and thus ensure that the search result can better meet the user's search intent.

[0217] II. The benefits of calculating through the hidden layer

[0218] When calculating the output of the hidden layer, linear transformation is first performed (calculate ), where the weight matrix W1 and the bias vector b1 play a role in linear mapping of the input vector. Different weight values determine the scaling, translation, etc. of the input vector in the mapping process, which helps to preliminarily extract features and integrate information of the spliced input vector. Then apply the ReLU function ReLU as a nonlinear activation function can introduce nonlinear factors, so that the neural network can learn the complex nonlinear relationship between the dimensions of the input vector. For example, the dimensions corresponding to different attributes in the question vector and the context sentence vector may not be simply linearly related (for example, there may be a complex, nonlinear relationship between "humorous style" and "fan number", not a simple proportional or inverse proportional relationship). Through the nonlinear transformation of the ReLU function, the neural network can mine and capture the complex relationships hidden in the data, further enrich and deepen the feature representation of the input information.

[0219] Therefore, it can better adapt to the complex semantic relationship commonly existing in natural language and content creation object related data, mine deep features and associations that are difficult to discover by linear models, make the matching relationship analysis of the question vector and the context sentence vector more accurate and detailed, help improve the effectiveness of the matching degree evaluation score, and enable the system to more accurately determine which content creation objects are truly matched with the user search request in terms of semantics and features, thereby improving the quality of the search results.

[0220] The multiple neurons in the hidden layer perform feature extraction and abstraction on the input vector from different angles through the above calculation process. The elements in the weight matrix \(W_1\) of each neuron are different, meaning that they focus on different combinations of dimensions of the input vector and different processing methods for these dimension information, thereby extracting diversified features. For example, some neurons may focus on extracting features related to style characteristics, and other neurons may focus on features related to fan number, category of debut, and other attributes. With the increase in the number of hidden layer neurons and their mutual cooperation, the original, specific information in the input vector can be gradually abstracted into higher-level, more representative feature representations, and the matching related features between the question vector and the context sentence vector can be described from multiple dimensions, facilitating further comprehensive judgment of the matching degree.

[0221] Therefore, the system's processing capability for complex information contained in the input vector is enhanced, and through the feature extraction and abstraction of multiple neurons, complex, high-dimensional vector information can be converted into a more concise representation that reflects key matching features, avoiding problems such as dimension disaster when directly processing the original vector. At the same time, the matching degree evaluation process can focus on core features that truly affect the matching result, improving the efficiency and accuracy of the matching degree evaluation, and providing a strong intermediate feature basis for outputting a reasonable matching degree evaluation score.

[0222] III. Benefits of calculating the matching degree score through the output layer

[0223] When calculating the output of the output layer as the matching degree score y, a linear transformation is first performed Similarly, the weight matrix W2 and the bias vector b2 are used to further linearly map and integrate the feature vector output by the hidden layer, and then the Sigmoid function is applied. The Sigmoid function compresses the output value into the range (0, 1). This way, regardless of the range of the eigenvector output by the hidden layer after linear transformation, the resulting match evaluation score falls within the easily understood and interpretable range (0, 1), where 0 represents a complete mismatch, 1 represents a complete match, and values ​​in between represent varying degrees of match. For example, a calculated match score of 0.8 can be intuitively understood as a high degree of match between the content creation object and the search request, making it easier for users and the system to intuitively judge and subsequently process the matching results.

[0224] To this end, an intuitive and standardized matching measurement method is provided to help users clearly understand the degree of fit between the content creation object and their search request. At the same time, it also provides the system with a clear numerical basis for subsequent operations such as screening content creation objects based on matching, enhancing the interpretability and practicality of the matching evaluation results, which is conducive to improving the user experience and operational convenience of the entire search system.

[0225] In the output layer, the features extracted by multiple neurons in the hidden layer are integrated, and different weights of importance are assigned to different features by combining the weight matrix W2 to comprehensively judge the matching degree between the question vector and the context sentence vector. 21j Determines the different features of the hidden layer (output h by the hidden layer neurons j The weight of each feature in the final matching evaluation enables the system to reasonably weigh the contribution of each feature to the overall matching according to the learned pattern, and determine the matching score from a holistic and comprehensive perspective. For example, if the feature "style matching" extracted in the hidden layer is found in previous training to be very important for determining whether the content creation object meets the user's needs, then the corresponding weight W in the output layer 21j The impact of this feature will be more significant when calculating the matching score, thus achieving accurate matching evaluation based on multi-dimensional features.

[0226] To this end, we fully utilize the rich feature information extracted from the hidden layer to avoid the one-sidedness of judging the matching degree by a single feature. By comprehensively considering multiple features and their relationships, we can more scientifically and accurately measure the matching between the question vector and the context statement vector, so that the matching evaluation score can truly reflect the degree of fit between the content creation object and the search request in terms of semantics, features, etc., thereby improving the accuracy and reliability of the matching evaluation, and providing a more powerful basis for screening out content creation objects that meet user intentions, thereby improving the overall performance of the search system and the quality of search results.

[0227] Optionally, the filtering out and returning the content creation object pointed to by the corresponding content creation object description information when the matching evaluation score is greater than the set matching evaluation score threshold includes: creating a matching_scores array and a content_objects array, the matching_scores array storing the matching score, and the content_objects array storing the corresponding content creation object description information; assigning the set matching evaluation score threshold to a threshold variable; using a loop to traverse each matching evaluation score in the matching_scores array; performing a comparison operation with the threshold variable for each matching evaluation score to determine whether each matching evaluation score is greater than the set matching evaluation score threshold; when finding a case where the matching evaluation score is greater than the set matching evaluation score threshold, recording the index index of the matching evaluation score in the matching_scores array; and obtaining the content creation object pointed to by the corresponding content creation object description information from the content_objects array according to the index index.

[0228] Optionally, during the array creation and initialization, when creating the `matching_scores` array and the `content_objects` array, the size of the array should be reasonably allocated according to the estimated number of matching score evaluation scores and the number of content creation object description information to avoid subsequent out-of-bound errors and other error conditions. Some initialization operations can be used to ensure that the array elements have appropriate initial values, such as initializing the elements of the `matching_scores` array to 0 and initializing the elements of the `content_objects` array to null pointers (if using a programming language that supports pointers) or specific default values (indicating no valid content creation object description information). The creation of these two arrays is usually implemented using the corresponding array declaration syntax according to the specific programming language environment, such as using a list (List) in Python to simulate an array, creating empty lists using `matching_scores = []` and `content_objects = []`, and then adding elements according to the actual calculated matching scores and corresponding description information. In Java, use statements such as `double[] matching_scores = new double[size];` (assuming `size` is the estimated number of elements, and here we use a double-precision matching score as an example) and `ContentObject[] content_objects = new ContentObject[size];` (assuming `ContentObject` is a custom class for storing content creation object description information) to create arrays and allocate memory space.

[0229] The estimation of the array size should take into account the number of possible search results and system resource conditions. If the estimation is too small, it may result in frequent array expansion operations (in some dynamic array implementation languages) during element addition, affecting performance; while an overestimation may cause waste of memory space. Therefore, the appropriate array size can be reasonably estimated according to the past search data statistics, system load conditions, and the approximate size of the content creation object database, etc. to ensure that all possible matching results can be accommodated and memory resources can be efficiently utilized.

[0230] Optionally, when assigning a value to the threshold variable, the matching score threshold set is assigned to the `threshold` variable. The determination of this threshold is a key step. The value of the threshold usually needs to be determined through multiple experiments and data analysis, considering the accuracy of search results, recall rate, and user expectations. First, test different thresholds on a small test dataset to observe whether the returned content creation objects meet the expected search intent. For example, for a search request for "funny style anchors", set different thresholds to see at which threshold the returned anchors indeed have obvious funny style, while not missing too many potential qualified anchors. In this way, a suitable threshold range is determined, and then further fine-tuned according to actual application scenarios and user feedback. In the code, according to the type of programming language, the assignment operation may be a simple direct assignment statement, for example, in C++ it can be `double threshold = 0.8;` (assuming the threshold is set to 0.8). The threshold type (such as `double` representing double-precision floating-point number) should match the type of matching score stored in the `matching_scores` array to ensure the correctness of subsequent comparison operations.

[0231] The threshold is not fixed and may need to be re-evaluated and updated periodically as the content creation object database updates, user search habits change, and system search accuracy requirements adjust. In addition, in a multi-user, multi-scenario application environment, personalized thresholds can be set for different types of search requests or different user groups, and user portraits, search history analysis, and other means can be used to dynamically adjust the threshold for each user to better meet diverse search needs.

[0232] Optionally, for the loop iteration and comparison operation, when iterating through each matching score in the `matching_scores` array, common loop structures include `for` loops, `while` loops, etc. The choice depends on the characteristics of the programming language and the need for logical clarity of the code. Taking the `for` loop as an example, in Python it can be written as follows:

[0233]

[0234] In the comparison operation, make sure the logic of the comparison meets the expected semantics, that is, determine whether the match degree evaluation score is strictly greater than the set threshold (here it is the `>` operator), and take into account the precision problem that may occur when comparing floating-point numbers (due to the storage method of floating-point numbers in computers, some small errors may occur). In order to more accurately compare floating-point numbers, some appropriate methods can be used, such as setting an allowed error range (such as determining whether the absolute value of the difference between two floating-point numbers is less than a very small value, such as `abs(score-threshold)<0.00001`, the specific value is determined according to the actual precision requirement), when the difference is within this range, they are considered equal in value or meet the comparison requirements, avoiding false comparison results due to floating-point precision problems. In the loop traversal process, if the number of array elements is very large, performance problems may occur, especially when the operation in each loop is complex (such as in addition to the comparison operation, there are other complex calculations or database access operations, etc.). In order to improve performance, you can consider optimizing the loop, such as using the compiler's optimization options (in compiled languages), using more efficient algorithm logic to replace some complex operations, or by parallel processing technology (if supported) to distribute the comparison operation in the loop to multiple threads or processes simultaneously, speeding up the traversal and comparison speed. In addition, for floating-point precision problems, in addition to the above-mentioned method of setting an error range, you can also perform appropriate numerical processing on the match degree evaluation score according to the specific application scenario, such as rounding, etc., so that it is more in line with the actual semantic requirements when compared.

[0235] Optionally, for index records and object acquisition, when a matching evaluation score is found that is greater than the set matching evaluation score threshold, the index `index` of the matching evaluation score in the `matching_scores` array is recorded. A variable can be used to store this index value. For example, in most programming languages, an integer variable `index` can be defined and assigned within a code block that meets the conditions, such as `index=i;` (assuming `i` is a loop variable used to traverse the index of the array). When obtaining the content creation object pointed to by the corresponding content creation object description information from the `content_objects` array according to the index `index`, the legitimacy of the index must also be ensured, that is, it cannot exceed the range of the `content_objects` array, so it is best to perform an index range verification before the acquisition operation (such as `if(index>=0&&index <content_objects.length)`这样的条件判断,这里以类似Java的语法为例)。获取对象的具体操作取决于存储内容创作对象描述信息以及对象本身的数据结构和编程语言特性,例如如果`content_objects`数组存储的是对象的引用(指针),那么可以通过这个引用直接访问对应的对象内容,在Python中如果是存储的对象实例或者字典等数据结构表示对象信息,就可以通过相应的索引或键值来获取具体的对象属性和内容等。

[0236] In order to ensure the accuracy and completeness of the obtained content creation objects, it is necessary to ensure that the elements in the `content_objects` array are logically one-to-one corresponding to the elements in the `matching_scores` array, that is, each matching evaluation score corresponds to the matching degree of the object pointed to by the description information in the `content_objects` array at the same index position. During the entire search and matching process, it is necessary to ensure that operations on the two arrays (such as adding elements, modifying elements, etc.) follow the same logic and order to avoid data inconsistencies. For example, when adding new matching scores and corresponding object description information to two arrays at the same time, it is necessary to ensure their data synchronization and consistency through synchronization mechanisms (in a multi-threaded or multi-process environment) or strictly follow the order of operations. Otherwise, it may result in obtaining the wrong object or missing the object that should be returned.

[0237] To this end, the relevant data is orderly organized and classified by creating the `matching_scores` array and the `content_objects` array to store the matching degree scores and the corresponding content creation object description information respectively. The design of this data structure enables clear positioning and processing of the corresponding elements according to the array index in subsequent filtering, comparison, and object retrieval operations, just like using a table, where the two elements (matching degree score and object description information) in each row (corresponding to the same index) are associated with each other, facilitating unified management and operation. For example, when traversing and comparing matching degree scores, the corresponding object description information can be directly accessed through the index, eliminating the need for complex lookup or association operations to determine the correspondence between the two, improving code readability and maintainability, and also facilitating batch processing and analysis of data, such as counting the number of objects that meet different threshold conditions, analyzing the distribution of matching degree scores, etc.

[0238] In the actual development and maintenance of search systems, a clear data structure helps developers quickly understand the code logic, extend functions, and debug. If new attributes need to be added or the matching degree evaluation mechanism needs to be improved later, it is convenient to modify the code and update the data based on this array structure, such as adding another array next to the `matching_scores` array to store other related evaluation indicators (such as different dimensions of matching weights), and easily associating these new data with existing object description information, ensuring system scalability and flexibility, and reducing development and maintenance costs. In addition, precise screening is achieved by comparing the threshold value, setting the matching degree evaluation score threshold, and comparing it with each matching degree evaluation score, which can accurately screen content creation objects according to the user's desired matching degree standard. The threshold value is like a screening "filter net", only content creation objects with a certain level of matching degree (greater than the threshold value) will be selected, which can avoid returning too many objects with weak relevance or not meeting the user's search intent, improving the quality and relevance of search results. For example, when searching for "professional knowledge of technology anchor", if no threshold is set, many anchors who only occasionally involve technology content or have not deep professional knowledge may be returned, but by setting a reasonable threshold (such as 0.8), anchors with high matching degree in the field of professional knowledge and technology can be screened out, allowing users to quickly obtain the most suitable content creation objects, improving user experience. In different search scenarios and user needs, the adjustability of the threshold further enhances the flexibility and accuracy of screening. For users who require high accuracy of search results and want to obtain the most suitable content creation objects, a higher threshold can be set; for users who want to obtain more extensive related information and focus on recall, the threshold can be appropriately lowered. The system can well adapt to diverse user needs through this threshold comparison mechanism, whether professional users are looking for accurate information or ordinary users are exploring a wide range of information, they can get relatively satisfactory search results, improving the universality and applicability of the search system. The index mechanism ensures accurate association and acquisition, using indexes to record and associate matching degree evaluation scores with content creation object description information and the final acquisition object, ensuring accurate correspondence of data during the entire screening process. By finding the index of the score that meets the threshold condition in the `matching_scores` array, and then obtaining the object pointed to by the object description information from the `content_objects` array according to this index, the matching degree score can be accurately associated with the correct object, avoiding errors caused by data confusion or incorrect association when acquiring objects.For example, no matter how the elements in the `matching_scores` array and the `content_objects` array dynamically change (such as constantly adding new search results), as long as the indexing mechanism is correctly maintained, it can always accurately filter the corresponding objects according to the matching degree, ensuring the reliability and accuracy of the entire filtering process. When dealing with large-scale search results and complex data structures, the accuracy of the indexing mechanism is particularly important. It enables the system to efficiently and accurately obtain the required objects even when faced with a large amount of matching degree evaluation data and numerous content creation object description information, without errors or omissions due to the large amount of data. This indexing-based accurate association and acquisition method ensures that the search system can run stably under different data sizes and output reliable search results, enhancing the robustness and stability of the system and meeting the high requirements for search accuracy and reliability in practical applications.

[0239] Figure 2 A schematic diagram of a content creation object search device is provided for the embodiments of the present application. As shown in Figure 2 , it includes: a first program unit for obtaining an input search request expression, the search request expression being expressed in natural language; a second program unit for splitting the search request expression into individual words or individual phrases based on the internal part-of-speech codes assigned to content creation objects; a third program unit for constructing a logical framework including multi-dimensional filtering conditions based on the logical relationships between the individual words or individual phrases; a fourth program unit for converting the logical framework into a structured expression to match database table fields in a content creation object database based on the structured expression; a fifth program unit for searching content creation objects in the content creation object database based on the matched database table fields, obtaining several pieces of content creation object description information; a sixth program unit for converting the search request expression into a question vector and converting each piece of content creation object description information into a context sentence vector; a seventh program unit for calculating the matching degree evaluation score of the question vector and the context sentence vector; and an eighth program unit for filtering out the content creation objects pointed to by the content creation object description information corresponding to the matching degree evaluation score greater than the set matching degree evaluation score threshold and returning them. The above Figure 2 In the embodiments, exemplary explanations of the technical processing procedures performed by each program unit can be referred to the descriptions in the above Figure 1 .

[0240] Figure 3 A structural schematic diagram of an electronic device is provided for the embodiments of the present application. As shown in Figure 3As shown, the electronic device includes a memory and a processor, the memory has stored thereon a computer executable program, and the processor executes the computer executable program to implement the following steps: obtaining an input search request expression, the search request expression being expressed in natural language; based on internal part-of-speech codes assigned to content creation objects, splitting the search request expression into individual words or individual phrases; based on logical relationships between the individual words or the individual phrases, constructing a logical framework including multi-dimensional filtering conditions; converting the logical framework into a structured expression to perform matching of database table fields in a content creation object database based on the structured expression; based on the matched database table fields, searching for content creation objects in the content creation object database to obtain a plurality of pieces of content creation object description information; converting the search request expression into a question vector and converting each piece of content creation object description information into a context sentence vector; calculating a matching degree evaluation score of the question vector and the context sentence vector; and selecting and returning a content creation object corresponding to content creation object description information whose matching degree evaluation score is greater than a set matching degree evaluation score threshold. Figure 3 In the embodiments, exemplary explanations of technical processing procedures of each step can refer to the descriptions of the above Figure 1 .

[0241] The above embodiments are only used for describing the embodiments of the present application, but not for limiting the embodiments of the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present application. Therefore, all equivalent technical solutions belong to the scope of the embodiments of the present application, and the patent protection scope of the embodiments of the present application should be defined by the claims. The systems, devices, modules or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions.

Claims

1. A method for searching content creation objects, characterized in that: include: Obtaining an input search request expression, where the search request expression is expressed in natural language; Normalizing the search request expression and removing stop words to obtain a search request expression to be segmented; Starting from the beginning of the search request expression to be segmented, the longest phrase matching pattern defined in the internal segmentation code library is used to sequentially try to match words or phrases and identify words or phrases based on their contextual relationships. The contextual relationships are determined by analyzing the co-occurrence probabilities of the preceding and following words or phrases within a set context window. Determine whether the identified words or phrases have matching word segmentation codes in the internal word segmentation code library; if so, determine the words or phrases as the words or phrases obtained by splitting; If it does not exist, the similarity between the identified word or phrase and the word segmentation code of the internal word segmentation code library is calculated. If the similarity is greater than the set similarity threshold, the word or phrase is supplemented with the word or phrase obtained by splitting; Performing grammatical structure analysis on the single words or the single phrases to obtain hierarchical structures and logical relationships between the single words or the single phrases; Performing semantic relationship recognition on the single words or the single phrases to obtain semantic association relationships between the single words or the single phrases; Constructing multi-dimensional screening conditions based on the hierarchical structure, logical relationship, and semantic association relationship between the individual words or the individual phrases; Mark each filter condition with its importance weight in syntax and semantics to build a logical framework including multi-dimensional filter conditions; Converting the logical framework into a structural expression to match database table fields in a content creation object database based on the structural expression; Based on the matched database table fields, searching for content creation objects in the content creation object database to obtain a plurality of content creation object description information; Convert the search request expression into a question vector, and convert each content creation object description information into a context sentence vector; Calculating a matching evaluation score between the question vector and the context sentence vector; The content creation object pointed to by the corresponding content creation object description information when the matching evaluation score is greater than the set matching evaluation score threshold is filtered out and returned.

2. A method for searching content creation objects according to claim 1, characterized in that: The input search request expression is obtained, and the search request expression is expressed in natural language: A front-end-based multimodal event listener monitors input events, captures user input in real time, and converts the captured input into a character stream through a set conversion function, wherein the input event includes at least one of a keyboard input event, a voice input event, and a handwriting input event; Add the character stream to the configured buffer in the order of input to form a character stream sequence; According to the set grammatical rules, preliminary grammar checking and automatic completion are performed on the character stream sequence to generate a search request expression.

3. A method for searching content creation objects according to claim 2, characterized in that: The front-end-based multimodal event listener listens for input events and captures user input in real time. Previously, it included: For the keyboard input event, use the addEventListener component of JavaScript to bind the multimodal event listener to the keydown, keyup and keypress events of the search input box to trigger keyboard input event capture when the user enters characters on the keyboard; For the voice input event, the multimodal event listener is bound to the click event or voice wake-up event of the voice input button through the voice recognition interface provided by the browser, so as to trigger the voice input event capture when the user inputs characters by voice; For the handwriting input event, the multimodal event listener is bound to the handwriting input area through the interface of the handwriting recognition library, so as to trigger the capture of the handwriting input event when the user inputs characters by handwriting.

4. A method for searching content creation objects according to claim 2, characterized in that: The process of performing preliminary grammar checking and automatic completion on the character stream sequence according to the set grammar rules to generate a search request expression includes: The grammatical rules are parsed based on a set grammatical rule parser to create keyword object phrases and connective object phrases, wherein the grammatical rules include keyword rules, connective rule, and numerical range rules. The keyword rules list various attribute keywords related to the anchor and their possible variations; the connective rule defines how the connectives "and" or "greater than" and "less than" are used in search requests; the numerical range rule specifies how to express numerical search conditions; Based on the grammar detection state transition mechanism defined by the set finite state machine, starting from the beginning of the character stream sequence, each character is analyzed character by character, wherein for each character, the category to which it belongs is first determined, whether it is a letter, a number, a punctuation mark or a special character; if it is a letter, based on the context of the current character, it is determined whether it may be part of a keyword, and by matching it with the keyword object array, it is checked whether there is a matching keyword; if it is a punctuation mark or a conjunction, its use is checked to see if it conforms to the grammatical structure defined in the conjunction rule, so as to generate a grammar detection result; Based on the grammar detection results, possible subsequent keywords or search conditions are determined to generate automatic completion suggestions, and based on the user's historical search records, the search patterns and keyword combinations frequently used by the user are analyzed and integrated into the automatic completion suggestions to generate a search request expression.

5. The method for searching content creation objects according to claim 1, wherein: The step of splitting the search request expression into individual words or phrases based on the internal word segmentation code assigned to the content creation object includes: Collecting attribute features of the content creation object, defining different word segmentation codes based on different attributes, and setting multiple matching modes for each word segmentation code, wherein the matching modes include synonyms, near-synonyms, and colloquial expression matching, wherein the attribute features include: basic information, business-related information, style characteristics, and social influence-related information; All word segmentation codes and their corresponding matching patterns are converted into a hash table for storage. When a hash conflict occurs, a free storage location is found according to a predefined detection sequence to store the conflicting elements. Alternatively, the conflicting elements are stored in a linked list structure in the same hash bucket so that the conflicting elements can be accessed through a linked list traversal operation.

6. A method for searching content creation objects according to claim 1, characterized in that: The method further comprises: Vectorizing the single word or the single phrase to obtain a vector sequence; Based on the set time step, the vector sequence is input into the syntactic attention head and the semantic attention head; By calling the grammatical attention head, based on the learned part-of-speech information, grammatical patterns, and grammatical rules of the words, grammatical feature extraction is performed on the vector sequence to perform grammatical structure analysis on the single words or the single phrases to obtain the hierarchical structure and logical relationship between the single words or the single phrases; By calling the semantic attention head to extract semantic features from the vector sequence, and to identify semantic relationships between the single words or the single phrases, the semantic association relationships between the single words or the single phrases are obtained.

7. A method for searching content creation objects according to claim 1, characterized in that: The converting the logical framework into a structural expression to match database table fields in a content creation object database based on the structural expression includes: Invoking a method for converting content creation object attribute features between a logical framework and a structural expression, combining various screening conditions in the logical framework according to a predetermined template to convert the logical framework into a structural expression; The query optimizer generates a query plan based on the structure expression index information to determine the access path and connection order of the data; According to the access path and connection sequence of the data, database table fields are matched in the content creation object database.

8. The method for searching content creation objects according to claim 1, wherein: Based on the matched database table fields, a search for content creation objects is performed in the content creation object database to obtain a plurality of content creation object description information, including: Dynamically build SQL query statements based on the matched database table fields; Creating a search task according to the SQL query statement, wherein the search task includes multiple subtasks; The content creation object database is searched for content creation objects by performing paged queries and batch reading based on the plurality of subtasks.

Images