Enterprise information query method and device

Abstract

The present application provides a method and device for querying enterprise information; the method includes: combining a time series model (such as a hidden Markov model) with information foraging theory, and constructing a preference function (such as a user taste function) to identify the user's implicit query intention and preference for each type of enterprise information, and mapping this preference to the weight of the corresponding type, thereby achieving personalized re-ranking of query results. In addition, the embodiment of the present application also introduces a large language model as a real-time online adaptive adjustment layer to cope with temporary changes in user needs, thereby continuously improving the accuracy of queries and user satisfaction.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a method and device for querying enterprise information. Background Art

[0002] In the related technologies, enterprise information query platforms are mostly static presentations, and lack the ability to identify users' implicit query intentions and make personalized recommendations. For example, related technologies cannot fully utilize users' behavioral sequence data (such as access sequence, page dwell time, etc.) to accurately predict users' potential query intentions. At the same time, enterprise information query scenarios have multi-dimensional and vertical business needs of individual (i.e., C-end) users (such as job hunting, litigation rights protection, financial fraud prevention, etc.) and corporate (i.e., B-end) users (such as customer due diligence, anti-money laundering, supplier due diligence, compliance review, bidding, risk assessment, etc.), which requires the platform to be able to flexibly rearrange information types in different business scenarios. In addition, in the face of temporary changes in search preferences by specific high-value users (such as VIP users) or specific institutions (such as banks, insurance, funds, etc.), the query solutions based on offline model training provided by related technologies cannot respond quickly. Summary of the invention

[0003] The embodiments of the present application provide a method, device, electronic device, computer-readable storage medium and computer program product for querying enterprise information, which can identify users' potential query intentions and preferences for various types, thereby achieving personalized re-ranking of search results. At the same time, it can also respond to temporary changes in users' needs, thereby continuously improving the accuracy of queries and user satisfaction.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] The embodiment of the present application provides a method for querying enterprise information, which is applied to an enterprise information query platform, wherein an enterprise information library of the enterprise information query platform includes a plurality of types of collections, and one of the collections includes enterprise information of a plurality of enterprises of one type; the method comprises:

[0006] Receiving a query request sent by a terminal device, wherein the query request carries a keyword;

[0007] Acquire first operation data of the user of the terminal device on the enterprise information query platform, wherein the first operation data is generated before receiving the query request;

[0008] Based on the first operation data, a plurality of candidate states included in the state set are predicted respectively by a pre-trained time series model to obtain a selection probability of each candidate state, wherein each candidate state corresponds to a query intent, and the query intent represents a potential query target of a user when accessing the enterprise information query platform;

[0009] Sorting the multiple candidate states in descending order of the selection probability, and taking the query intention corresponding to at least one of the candidate states ranked first in the sorting result as the query intention of the user;

[0010] Building a preference function based on the multiple types and the state set, and determining the user's preference parameters for each type under the query intention through the preference function;

[0011] Mapping the plurality of preference parameters to weights of corresponding types respectively, and filling the plurality of weights and the keywords into a preset query data structure to obtain a query expression;

[0012] Querying the enterprise information database based on the query expression to obtain a plurality of enterprise information matching the keyword, and sorting the plurality of enterprise information based on a plurality of the weights;

[0013] When no preference adjustment instruction is received from the terminal device, sending the sorted plurality of enterprise information to the terminal device;

[0014] When the preference adjustment instruction sent by the terminal device is received, the sorted multiple enterprise information is sorted again through the large language model, and the re-sorted multiple enterprise information is returned to the terminal device.

[0015] The embodiment of the present application provides a query device for enterprise information, which is applied to an enterprise information query platform, wherein the enterprise information library of the enterprise information query platform includes a plurality of types of collections, and one of the collections includes enterprise information of a plurality of enterprises of one type; including:

[0016] A receiving module, configured to receive a query request sent by a terminal device, wherein the query request carries a keyword;

[0017] An acquisition module, configured to acquire first operation data of the user of the terminal device on the enterprise information query platform, wherein the first operation data is generated before receiving the query request;

[0018] A prediction module, configured to predict, based on the first operation data, a plurality of candidate states included in the state set by means of a pre-trained time series model, respectively, to obtain a selection probability of each candidate state, wherein each candidate state corresponds to a query intent, and the query intent represents a potential query target of a user when accessing the enterprise information query platform;

[0019] A sorting module, used to sort the multiple candidate states in descending order of the selection probability;

[0020] A determination module, configured to use the query intention corresponding to at least one of the candidate states ranked higher in the sorting result as the query intention of the user;

[0021] A construction module, configured to construct a preference function based on the plurality of types and the state set;

[0022] The determination module is further used to determine the preference parameters of the user for each of the types under the query intention through the preference function;

[0023] A mapping module, used for mapping the plurality of preference parameters into weights of corresponding types respectively;

[0024] A filling module, used to fill a plurality of the weights and the keywords into a preset query data structure to obtain a query expression;

[0025] A query module, used to query the enterprise information database based on the query expression to obtain multiple enterprise information matching the keyword;

[0026] The sorting module is further used to sort the plurality of enterprise information based on the plurality of weights;

[0027] a sending module, configured to send the sorted plurality of enterprise information to the terminal device when no preference adjustment instruction sent by the terminal device is received;

[0028] The sorting module is further configured to re-sort the sorted plurality of enterprise information by using a large language model when receiving a preference adjustment instruction sent by the terminal device;

[0029] The sending module is further used to return the re-sorted plurality of enterprise information to the terminal device.

[0030] The present application embodiment provides an enterprise information query platform, including:

[0031] A memory for storing computer executable instructions;

[0032] The processor is used to implement the enterprise information query method provided in the embodiment of the present application when executing the computer executable instructions stored in the memory.

[0033] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for implementing the enterprise information query method provided in the embodiment of the present application when executed by a processor.

[0034] An embodiment of the present application provides a computer program product, including a computer program or computer executable instructions, which are used to implement the enterprise information query method provided in the embodiment of the present application when executed by a processor.

[0035] The embodiments of the present application have the following beneficial effects:

[0036] Based on the historical operation data of users on the enterprise information query platform, the user's current query intention is identified through a pre-trained time series model, and by constructing a preference function, the user's preference parameters for different types of enterprise information under this query intention are determined, and this preference parameter is mapped to the weight of the corresponding type. In this way, multiple enterprise information matching keywords can be sorted based on the weight to meet the personalized needs of users. In addition, the embodiment of the present application also introduces a large language model as a real-time online adaptive adjustment layer to respond to temporary changes in user needs, further improving the user's query experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 1 is a schematic diagram of the architecture of a query system 100 for enterprise information provided in an embodiment of the present application;

[0038] Figure 2 is a schematic diagram of the structure of the enterprise information query platform 200 provided in the embodiment of the present application;

[0039] Figure 3 This is a first flow chart of the method for querying enterprise information provided by an embodiment of the present application;

[0040] Figure 4 It is a second flow chart of the method for querying enterprise information provided in an embodiment of the present application;

[0041] Figure 5 This is a third flow chart of the enterprise information query method provided in the embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0043] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0044] It is understandable that in the embodiments of the present application, related data such as user information (such as user operation data, log data, etc. on the enterprise information query platform) is involved. When the embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards.

[0045] In the following description, the terms "first\second\..." are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\..." can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0047] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0048] 1) In response to: used to indicate the conditions or states on which the executed operations depend. When the dependent conditions or states are met, one or more operations executed may be in real time or with a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations executed may be executed.

[0049] 2) Time series model: that is, time series model, is a theory and method to establish a mathematical model through curve fitting and parameter estimation based on the time series data obtained from system observation. It is generally carried out using curve fitting and parameter estimation methods. For example, it includes hidden Markov model (HMM), where HMM is a statistical model used to describe a Markov process with hidden unknown parameters. It generates unobservable sequence states through hidden Markov chains and generates observable sequences based on these hidden states.

[0050] 3) Information Foraging Theory (IFT): It is a theory about human information acquisition behavior provided by computer scientists and psychologists. It uses action foraging behavior as a model and believes that humans will also adopt strategies similar to action foraging in the process of information acquisition.

[0051] 4) Large Language Model (LLM): refers to a deep learning model trained with a large amount of text data, which enables the model to generate natural language text or understand the meaning of language text. These models can provide in-depth knowledge and language production on various topics by training on huge data sets. The core idea is to learn the patterns and structures of natural language through large-scale unsupervised training, and to simulate the human language cognition and generation process to a certain extent.

[0052] 5) Query Domain Specific Language (DSL): It is a language specifically used to build queries. It provides rich query syntax and functions, allowing users to flexibly define various complex queries. Queries in query DSL can be divided into two categories: query and filter. Query refers to calculating the relevance score of documents according to specified conditions and returning results sorted by relevance, while filtering refers to filtering out documents that meet the conditions according to specified conditions, but does not calculate the relevance score.

[0053] The applicant discovered during the implementation of the embodiments of the present application that the related technology has the following problems:

[0054] 1. Unable to effectively identify users’ hidden query intentions

[0055] Most of the solutions provided by related technologies rely on static keyword matching or pre-configured field weights, and lack dynamic identification of users' real needs at different times and in different contexts (such as C-end job hunting, B-end customer due diligence, supplier due diligence, etc.). In addition, even if a large amount of user behavior data (such as click order, dwell time, etc.) exists, it has not been fully utilized to infer users' deeper query intentions.

[0056] 2. Lack of adaptive personalized sorting capabilities

[0057] The solutions provided by related technologies can only score query results based on fixed weights, and cannot adjust the order of query results instantly according to temporary changes in user needs (such as sudden attention to "supplier information" or "anti-money laundering clues", etc.). In addition, for high-value users (such as VIP users or financial institution users), when searching for specific in-depth information, the system responds slowly and adjusts slowly, resulting in a poor user experience.

[0058] 3. Model update and long-term evolution mechanism without a closed loop

[0059] Related technologies lack closed-loop collection and model retraining for user feedback (such as satisfaction, dissatisfaction, browsing time, etc.), and cannot continuously evolve at the system level. At the same time, they are also unable to automatically capture user preference shifts and update indexing strategies or weight configurations, resulting in query results that are gradually out of touch with users' real needs during long-term use.

[0060] 4. Lack of online intelligent scheduling capabilities

[0061] The solutions provided by related technologies do not involve secondary semantic analysis or contextual understanding of the results through large language models, and can only return query results based on string matching, lacking deep semantic recommendations in complex retrieval scenarios. At the same time, they are unable to respond efficiently and intelligently to instantaneous preference changes.

[0062] In view of this, the embodiments of the present application identify the hidden query intentions of users by integrating HMM and information foraging theory. For example, the hidden query intentions of users can be dynamically inferred by constructing HMM and combining the click sequence, residence time, etc. of users on the enterprise information query platform. At the same time, the concept of "information smell" in information foraging theory can be combined to quantify the user's "taste" (i.e., degree of preference) for different types of enterprise information, and perform real-time preference weighting during the retrieval process. In this way, multi-scenario intentions such as C-end and B-end can be automatically detected and distinguished, avoiding simple reliance on static weights or manual selection. In addition, in specific circumstances (such as a temporary sudden demand shift by the user), the embodiments of the present application can introduce a large language model to perform secondary sorting on the preliminary sorting results. The large language model can capture more sophisticated semantic clues and preference transfers, and form an efficient matching result that is presented instantly, thereby making up for the disadvantage of the long HMM training cycle, quickly responding to preference mutations in a short period of time, and improving the user experience. In addition, the embodiments of the present application can also collect user feedback on query results (for example, including clicks, dwell time, dissatisfaction annotations, etc.), and the preference transfer events can be recorded by a large language model. At the same time, the HMM can be retrained and the search strategy can be adjusted regularly to achieve long-term adaptive evolution, thereby solving the limitations of "static model + no feedback update" in related technologies, allowing the enterprise information query platform to understand users more and more.

[0063] That is to say, the embodiments of the present application provide a method, device, enterprise information query platform, computer-readable storage medium and computer program product for querying enterprise information, which can identify the user's potential query intention and preference for each type, thereby realizing personalized re-ranking of search results, and at the same time, can also respond to temporary changes in user needs, thereby continuously improving the accuracy of queries and user satisfaction. Figure 1 The method for querying enterprise information provided in the embodiment of the present application is described in detail.

[0064] For example, see Figure 1 , Figure 1 1 is a schematic diagram of the structure of the enterprise information query system 100 provided in the embodiment of the present application. Figure 1 As shown, the enterprise information query system 100 provided in the embodiment of the present application includes: an enterprise information query platform 200, a network 300, a terminal device 400 and an enterprise information library 500, wherein the network 300 can be a local area network or a wide area network, or a combination of the two, the terminal device 400 is a user-associated terminal device, and a client 410 (such as a client or browser of the enterprise information query platform, etc.) runs on the terminal device 400, and the enterprise information library 500 is connected to the enterprise information query platform 200, and the enterprise information library 500 can include multiple types of collections, wherein each collection can include enterprise information of multiple enterprises of one type.

[0065] In some embodiments, a search box may be displayed in the human-computer interaction interface provided by the client 410, and the user may enter the keyword (such as the name of a certain enterprise) that he wants to query in the search box, and then the terminal device 400 may send a query request to the enterprise information query platform 200 through the network 300, wherein the query request may carry at least one keyword input by the user. After receiving the query request sent by the terminal device 400, the enterprise information query platform 200 may first obtain the first operation data (i.e., historical operation data) of the user associated with the terminal device 400 on the enterprise information query platform 200, wherein the first operation data is generated before receiving the query request, and then the enterprise information query platform 200 may predict the multiple candidate states included in the state set based on the first operation data through a pre-trained time series model (such as HMM) to obtain the selection probability of each candidate state, wherein each candidate state may correspond to a search intention, and the query intention may be used to characterize the potential query target or task type of the user when accessing the enterprise information query platform 200. After obtaining the selection probability corresponding to each candidate state, the enterprise information query platform 200 can sort the multiple candidate states in order of selection probability from high to low, and use the query intent corresponding to at least one candidate state ranked at the top in the sorting result as the user's current query intent; then the enterprise information query platform 200 can construct a preference function (such as a user taste function) based on multiple types and state sets, and determine the user's preference parameters (such as preference degree or preference value) for each type under this query intent through the preference function. After obtaining the preference parameters, the enterprise information query platform 200 can map the multiple preference parameters to the weights of the corresponding types, and fill the multiple weights and keywords into the preset query data structure to obtain a query expression (such as a query DSL). Then the enterprise information query platform 200 can query the enterprise information library 500 based on the generated query expression to obtain multiple enterprise information matching the keywords from the enterprise information library 500, and sort the multiple enterprise information based on multiple weights.Finally, the enterprise information query platform 200 can also determine whether it has received a preference adjustment instruction sent by the terminal device 400 (i.e., determine whether the user has temporarily adjusted the preference). If it has received a preference adjustment instruction sent by the terminal device 400 (i.e., the user has temporarily adjusted the preference, for example, the user manually checked a type in the type filtering panel), the enterprise information query platform 200 can re-sort the sorted multiple enterprise information through the large language model, and return the re-sorted multiple enterprise information to the terminal device 400 through the network 300, so that the terminal device 400 calls the client 410 for presentation; if it has not received a preference adjustment instruction sent by the terminal device 400 (i.e., the user has not temporarily adjusted the preference), the enterprise information query platform 200 can directly send the sorted multiple enterprise information to the terminal device 400 through the network 300.

[0066] It should be noted that Figure 1 The enterprise information query platform 200 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content distribution networks (CDN, Content Delivery Network), and big data and artificial intelligence platforms. The terminal device 400 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, a car terminal, etc., but is not limited to this. The terminal device 400 and the enterprise information query platform 200 can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiments of the present application.

[0067] The enterprise information query method provided in the embodiment of the present application has a wide range of application scenarios. For example, it can be oriented to C-end users, such as providing accurate and personalized enterprise information search results for personal demand scenarios such as job applications, litigation and rights protection, and financial fraud prevention; it can also be oriented to B-end users, such as meeting the vertical needs of professional users such as banks, securities, insurance, funds, and Internet financial institutions (for example, including customer due diligence, supplier chain analysis, risk control, and anti-money laundering investigations, etc.), significantly improving the information acquisition efficiency and decision-making quality of professional users; in addition, it can also be oriented to member users (such as VIP users), for example, it can achieve a customized and highly flexible search experience, respond to preference mutations instantly, and provide data support for subsequent long-term strategy optimization, thereby enhancing product stickiness and competitive advantage.

[0068] The following is a description of the structure of the enterprise information query platform provided in the embodiment of the present application. Figure 2 , Figure 2is a schematic diagram of the structure of the enterprise information query platform 200 provided in the embodiment of the present application, Figure 2 The enterprise information query platform 200 shown includes: at least one processor 210, a memory 240, and at least one network interface 220. The various components in the enterprise information query platform 200 are coupled together through a bus system 230. It can be understood that the bus system 230 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 230 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 230 is not described in detail. Figure 2 Various buses are labeled as bus system 230 .

[0069] The processor 210 may be an integrated circuit chip having signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., wherein the general-purpose processor may be a microprocessor or any conventional processor, etc.

[0070] The memory 240 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 240 may optionally include one or more storage devices that are physically remote from the processor 210.

[0071] The memory 240 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 240 described in the embodiment of the present application is intended to include any suitable type of memory.

[0072] In some embodiments, memory 240 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0073] Operating system 241, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0074] A network communication module 242, used to reach other computing devices via one or more (wired or wireless) network interfaces 220, exemplary network interfaces 220 include: Bluetooth, Wireless Compatibility Certification (WiFi), and Universal Serial Bus (USB), etc.;

[0075] In some embodiments, the device provided in the embodiments of the present application can be implemented in software. Figure 2 The query device 243 of enterprise information stored in the memory 240 is shown, which can be software in the form of programs and plug-ins, including the following software modules: receiving module 2431, acquisition module 2432, prediction module 2433, sorting module 2434, determination module 2435, construction module 2436, mapping module 2437, filling module 2438, query module 2439, sending module 24310, input module 24311, training module 24312 and weighting model 24313. These modules are logical, so they can be arbitrarily combined or further split according to the functions implemented. It should be pointed out that in Figure 2 For the sake of convenience, all the above modules are shown at once, but it should not be regarded as that the enterprise information query device 243 excludes the implementation of only including the receiving module 2431, the acquisition module 2432, the prediction module 2433, the sorting module 2434, the determination module 2435, the construction module 2436, the mapping module 2437, the filling module 2438, the query module 2439 and the sending module 24310. The functions of each module will be explained below.

[0076] The following will be combined Figure 3 The method for querying enterprise information provided in the embodiment of the present application is described in detail.

[0077] It should be noted that the enterprise information query method provided in the embodiment of the present application can be applied to an enterprise information query platform. The enterprise information library of the enterprise information query platform may include multiple types of collections, wherein each collection may include enterprise information of multiple enterprises of one type.

[0078] For example, see Figure 3 , Figure 3 This is a first flow chart of the method for querying enterprise information provided by the embodiment of the present application, which will be combined with Figure 3 The steps shown are explained.

[0079] In step 101, a query request sent by a terminal device is received.

[0080] Here, the query request may carry at least one keyword, for example, the keyword may be the name of an enterprise, the name of a legal person, etc.

[0081] For example, a client may be running on a terminal device, and a search box may be displayed in the human-computer interaction interface of the client. The user may enter the keywords he or she wants to search in the search box. For example, the user may enter the name of the company he or she wants to search or a natural language description (for example, "check the risk information and supplier status of XX company") in the search box. The terminal device may then send a query request carrying the company name entered by the user to the enterprise information query platform.

[0082] In some embodiments, before executing step 101, the enterprise information query platform may pre-classify multiple enterprise information (e.g., including full enterprise names, abbreviations, trademarks, former names, social unified credit codes, etc.) of multiple enterprises stored in the enterprise information database into a plurality of business types (i.e., the multiple enterprise information of multiple enterprises stored in the enterprise information database may be pre-classified into a plurality of types of sets). Assume that the plurality of types of sets obtained by classification can be recorded as ,For example Can be a collection of basic information types (i.e. Includes basic information of multiple enterprises), Can be a collection of financial information types (i.e. Includes financial information of multiple companies), It can be a collection of customer due diligence types ( Includes customer due diligence information of multiple companies), It can be a collection of risk warning types (i.e. Includes risk warning information for multiple companies), Can be a collection of supplier information types (i.e. Includes supplier information of multiple companies), Can be a collection of job posting types (i.e. In the process of searching and browsing, users will click or visit certain types of enterprise information according to their needs. t =1,2,… T In the , the user's observable behavior can be recorded ,For example Indicates that the user accessed the enterprise information of the basic information type in the first step.

[0083] In step 102, first operation data of a user of a terminal device on an enterprise information query platform is obtained.

[0084] Here, the first operation data is generated before the query request is received, that is, the first operation data is the user's historical operation data.

[0085] In some embodiments, after receiving a query request sent by a terminal device, the enterprise information query platform can obtain operational data such as keywords entered by the user of the terminal device, the sequence of enterprise information clicked, dwell time, bounce rate, historical access data, etc. from the system log.

[0086] It should be noted that if it is impossible to obtain sufficient historical behavior data of the user in certain scenarios (for example, the current user is a new user), the large language model can be directly used as the main query scheduler to intelligently sort the query results. The embodiments of the present application do not make specific limitations on this.

[0087] In addition, it should be noted that when the first operation data cannot be obtained, a preset default value can also be used, that is, the default value can be used to infer the user's potential query intention, wherein the default value is equivalent to the initial value in mathematical modeling. For example, for a new user who cannot obtain the first operation data, it can be assumed that the new user's potential goal when accessing the enterprise information query platform is to query the basic information of the enterprise.

[0088] In step 103, based on the first operation data, a plurality of candidate states included in the state set are predicted respectively by a pre-trained time series model to obtain a selection probability of each candidate state.

[0089] Here, each candidate state can correspond to a query intent, wherein each query intent represents a potential query target or task type (such as job hunting, company due diligence, or supplier due diligence, etc.) of the user when accessing the enterprise information query platform.

[0090] In some embodiments, the first operation data may include: the user's click sequence, input keywords and stay time on the enterprise information query platform, etc., and the above-mentioned step 103 can be implemented in the following manner: based on the user's click sequence, input keywords and stay time on the enterprise information query platform, determine the user's historical query intention; input the historical query intention into a pre-trained time series model, so that the pre-trained time series model determines the posterior probability of the user being in the query intention corresponding to each candidate state included in the state set at the current moment; and use each posterior probability as the selection probability of the corresponding candidate type.

[0091] For example, taking the time series model as HMM as an example, in the technical solution provided in the embodiment of the present application, a set of hidden state sets can be created in advance, assuming that it is recorded as , where each candidate state Corresponding to a specific query intent, such as It can correspond to the job application scenario of C-end users. It can correspond to the C-end user litigation and rights protection scenario, It can correspond to the financial fraud prevention scenario of C-end users. It can correspond to the B-side user customer due diligence scenario, It can correspond to the anti-money laundering review scenario of B-side users. It can correspond to the B-side user supplier due diligence scenario. It should be noted that different candidate states point to potential decision goals or task types of users when accessing the enterprise information query platform. These states are hidden, that is, the actual query intention of the user is not directly visible and needs to be inferred by observing user behavior. Then, the user's most recent query keywords, the sequence of types of clicked enterprise information, the length of stay and other behavioral data can be obtained from the cache or log, and the user's previous query intention (that is, the user's query intention before making this query) can be determined based on the obtained behavioral data. The user's previous query intention can then be input into the pre-trained HMM, so that the pre-trained HMM calculates the posterior probability of the user's query intention corresponding to each candidate state, and uses each posterior probability as the selection probability of the corresponding candidate state.

[0092] In some embodiments, see Figure 4 , Figure 4 is a second flow chart of the method for querying enterprise information provided by the embodiment of the present application, such as Figure 4 As shown, in the execution Figure 3 Before step 103 shown, you can also perform Figure 4 Steps 111 to 115 shown in the figure are combined with Figure 4 The steps shown are explained.

[0093] In step 111, an initialized timing model is constructed.

[0094] Here, the parameters of the initialized time series model (eg, HMM) may include: a state set (ie, ), the initial state probability (assuming it is recorded as ), state transition probability matrix (assuming it is recorded as A ) and the observed probability distribution (assuming it is recorded as B ).

[0095] In step 112, the log data of the user on the enterprise information query platform is obtained, and an observation sequence is created based on the log data.

[0096] It should be noted that the log data and the first operation data may be the user's behavior data on the enterprise information query platform at different periods. For example, the log data may be the user's behavior data on the enterprise information query platform in the past month, and the first operation data may be the user's behavior data on the enterprise information query platform in the past week.

[0097] In step 113, a first probability of the observation sequence existing in each candidate state included in the state set at each time point is determined, and a second probability of the observation sequence appearing in each candidate state from the current time point to the end of the sequence is determined.

[0098] Here, the first probability is the forward probability and the second probability is the backward probability.

[0099] In step 114, based on the first probability and the second probability, the expected number of transitions between each candidate state and the expected number of observations in each candidate state are determined.

[0100] In step 115, the initial state distribution, the state transition probability matrix, and the observation probability distribution are updated based on the expected number of transitions and the expected number of observations.

[0101] In some embodiments, taking the time series model as HMM as an example, the embodiments of the present application can use log data from users (such as access type sequence, stay duration, click behavior, etc.) as training samples of the model, supplemented by query intentions manually or with rule annotations. At the same time, in the process of training the HMM, the log-likelihood function can be used as the loss function to maximize the probability of the observation sequence under the model. In addition, the training process can use the forward-backward algorithm + expectation maximization (EM, Expectation Maximization) iteration to update the initial state probability, state transition probability matrix and observation probability distribution included in the HMM.

[0102] In some other embodiments, when defining a hidden state set Afterwards, the embodiment of the present application may also introduce a preference function (such as a user taste function, assuming that it is recorded as ), used to measure the state Corresponding query intent type The attractiveness to users (similar to the information smell in IFT). Then, the initial value can be estimated based on the user's access frequency, stay time, domain expert experience and other data under the marked query intent. The access type under the corresponding query intent The probability of enterprise information is defined as:

[0103] (1)

[0104] in, To control the parameters, represents exponential operation, N is the total number of different types obtained by classification.

[0105] That is, the embodiment of the present application can train the initial state probability, state transition probability matrix and observation probability distribution using the user's historical behavior data through the forward-backward algorithm of HMM (such as Baum-Welch algorithm). Through iterative optimization, HMM can infer the user's potential query intention based on the user's current behavior at any time, and thereby calculate the user's preference parameters (such as preference values) for various types of corporate information under this query intention.

[0106] For example, in the technical solution provided in the embodiment of the present application, the HMM may be composed of the following parameters:

[0107] Hidden state collection: S

[0108] Initial state probability: ,in,

[0109] State transition probability matrix: ,in,

[0110] ,and

[0111] Observation probability distribution: ,in,

[0112] In the embodiments of the present application, Indicates that the user is at time step t Access Type , HMM assumes that the state transition between time steps satisfies the Markov property and the observation depends only on the current state.

[0113] For example, the embodiment of the present application can be derived from the user taste function Set out and get in shape User access type under the corresponding query intent Probability For example, suppose the user is in state The corresponding query intent is for type The probability of selection depends on their relative taste values. For this reason, the embodiment of the present application can introduce a normalization function (such as a Softmax function) to Normalize, that is:

[0114] (2)

[0115] in, is a control parameter and is greater than 0, used to adjust the sensitivity to taste differences. . The larger it is, the sharper the user preference distribution is. The smaller it is, the smoother the user preference distribution is.

[0116] In HMM, the observation probability , considering that the embodiment of the present application has converted the user's preference for type into a probability distribution, we can directly obtain:

[0117] (3)

[0118] In addition, if there is a stay time in the actual business Or other features, it can be further expanded to:

[0119] (4)

[0120] For example, when training HMM, you can first preset , and The initial value of The initial value of can be set to uniform distribution or based on historical data statistics of users' initial access preferences; The initial value of can be based on prior business logic (e.g., the probability of a user switching from a basic query state to a risk assessment state) or statistical transition frequency from data; The initial value of can come from statistical data, for example, In a labeled data set (such as user query intent manually or with rules), the user is in the state The corresponding query intent access type When there is no domain expert knowledge, the embodiment of the present application can also use the logarithmic frequency ratio as the initial value, that is:

[0121] (5)

[0122] in, and is the smoothing parameter, It represents logarithmic operation. If there is expert experience, linear weighted adjustment can also be performed on this basis.

[0123] The following will continue to explain the process of updating the parameters of the HMM using the Baum-Welch algorithm.

[0124] When training HMM, we can first define an observation sequence, assuming that it is recorded as , and you can also define forward variables for:

[0125] (6)

[0126] And you can define backward variables for:

[0127] (7)

[0128] Thus, the forward-backward algorithm can be used to find , A , B Calculation in case and .

[0129] In getting and After that, we can calculate the time t Belong to the status The posterior probability (assuming it is ( i )),Right now:

[0130] (8)

[0131] Similarly, the posterior probability of state transition can be calculated (assuming it is recorded as ),Right now:

[0132] (9)

[0133] In getting ( i )and After that, you can use ( i )and Update the parameters of the HMM, such as the updated initial state probability (assuming it is recorded as ) can be:

[0134] ( i ) (10)

[0135] The updated state transition probability (assuming it is recorded as ) can be:

[0136] (11)

[0137] It should be noted that for the update of the observation probability distribution part, it is necessary to update parameters, thus affecting .because Existence Item, the log-likelihood needs to be maximized when updating, and it may be necessary to use iterative optimization (such as gradient ascent or EM nested process), where the log-likelihood function is as follows:

[0138] (12)

[0139] in, .

[0140] In the M step, to update , need to consider ( i ) for a specific For a given state , observed as The probability and Closely related, namely:

[0141] (13)

[0142] make:

[0143] (14)

[0144] but:

[0145] (15)

[0146] Statistics of all ( t , i ), the user is in state When observed The expected frequency (assuming it is )for:

[0147] (16)

[0148] in, is the indicator function, when updating When , we hope that under the maximum likelihood estimation:

[0149] (17)

[0150] in, It means to find partial derivatives.

[0151] Ignore not dependent on When the term is , the log-likelihood is about The partial derivative of is:

[0152] (18)

[0153] Thus, by iterative methods (such as gradient ascent) The correction is continued until convergence. In practical applications, fixed point iteration or Newton and other numerical methods can be used.

[0154] After training the HMM using the above training method, we can obtain a set of updated initial state probabilities, state transition probability matrices, and user taste functions. Thus, given the user's current and past observation sequences (e.g., the set of types visited, the most recent click behavior, etc.), the forward-backward algorithm can be used to quickly calculate the posterior probability of the user's query intent corresponding to each candidate state, and then the most likely candidate state (e.g., the candidate state with the largest posterior probability) can be selected and the user taste function can be used to calculate the posterior probability of the user's query intent corresponding to each candidate state. Calculate the user's preference parameters for each type of enterprise information under the query intent corresponding to this candidate state, such as . Later, you can also Determine the user's preference strength for each type at this moment (i.e., weight, which is equivalent to the quantitative measurement in information smell), and provide basic data for subsequent ES (ElasticSearch, an open source, distributed, highly scalable real-time search and analysis engine) query weighting and result sorting. In other words, the embodiment of the present application can combine user behavior, business scenario requirements, and preferences for each type into a computable and trainable probability model, thereby providing a solid theoretical basis for subsequent personalized result sorting, dynamic tuning, and large language model auxiliary strategies.

[0155] It should be noted that in addition to using HMM to infer the hidden query intention of users, if there is enough data volume and computing power, recurrent neural networks (RNN, Recurrent Neural Network), long short-term memory networks (LSTM, Long Short-Term Memory), etc. can also be used for sequence prediction to learn the user's access pattern on the enterprise information query platform, so as to identify the user's potential query intention. In addition, Bayesian networks or other graph models can be used to replace HMM, and the user's potential query intention (or business intention) can also be probabilistically inferred. For example, multi-dimensional user feature nodes (such as access type frequency, stay time, context category, etc.) can be defined, and then these user feature nodes jointly point to the "query intention" node. For more complex scenarios or multi-factor combinations, Bayesian networks may have more flexible scalability than HMM. For example, when the sources of user behavior clues are diverse and the structure is relatively dispersed, you can consider using graph models to infer the user's potential query intention.

[0156] In step 104, the plurality of candidate states are sorted in descending order of selection probability, and the query intention corresponding to at least one candidate state ranked high in the sorting result is used as the query intention of the user.

[0157] In some embodiments, after obtaining the selection probabilities corresponding to the multiple candidate states, the multiple candidate states can be sorted in descending order of the selection probabilities, and the query intent corresponding to at least one candidate state ranked at the top in the sorting result is used as the user's current query intent. For example, taking the multiple candidate states as 10 candidate states, assuming that they are states To status , assuming the state The probability of being selected is 89%, status The probability of being selected is 86%, and the status The probability of being selected is 91%, and the status The probability of being selected is 94%, and the status The probability of selection is 80%, state The probability of being selected is 84%, status The probability of selection is 96%, state The probability of being selected is 81%, status The probability of selection is 95%, state The probability of being selected is 85%. Then the 10 candidate states are sorted in descending order of probability of being selected as follows: ,state ,state ,state , , , , , , , then the status The corresponding query intention (for example, assuming a job application scenario) is used as the user's current query intention.

[0158] In step 105, a preference function is constructed based on multiple types and state sets, and the preference parameters of the user for each type under the query intention are determined through the preference function.

[0159] In some embodiments, the above-mentioned preference function (e.g., user taste function) based on multiple types and state sets can be implemented in the following manner: ): Obtain the marked query intents from the multiple query intents corresponding to the multiple candidate states included in the state set; obtain the type of enterprise information that the user interacts with (such as clicks or browses) under the marked query intent, and the interaction parameters for the enterprise information (such as click frequency or browsing time, etc.); take the marked query intent and type as independent variables, and the interaction parameters as dependent variables, and construct a preference function from the independent variable to the dependent variable.

[0160] For example, the present application embodiment can use the information foraging theory, which regards users as "information predators" who search for valuable information in the information environment. In the present application embodiment, each type For users in state The corresponding query intent has a certain appeal, similar to "Information Scent". In view of this, the embodiment of the present application can expand this concept and define it as "User Taste Function" (ie preference function):

[0161] (19)

[0162] in, Represents the number domain, user taste function The higher the function value, the better the Under the corresponding query intent, users have different The stronger the potential preference of The more interested they are in the company's information.

[0163] It should be noted that the user taste function The construction process of has been described above, and the embodiments of the present application will not be repeated here. In addition, collaborative filtering or matrix decomposition methods can be directly used to regard "user multiplied by type" as a scoring matrix, and the user's preference for each type of corporate information can be inferred by analyzing the user's historical clicks or length of stay. In addition, if each type of corporate information has a specific topic label at the text level (such as risk, finance, recruitment, etc.), the "type attractiveness" score can also be obtained through topic modeling or simple keyword statistics, thereby replacing the user taste function in IFT.

[0164] In some embodiments, the above-mentioned determination of the user's preference parameters for each type under the query intent through the preference function can be achieved in the following manner: the following processing is performed for each type: the query intent and the type are input into the preference function to obtain the function value output by the preference function; the function value is normalized through a normalization function (such as a Softmax function), and the normalized function value is used as the user's preference parameter (such as preference degree) for the type under the query intent.

[0165] For example, the preference function is the user taste function For example, when constructing the user taste function After that, you can combine the query intent (such as the job application scenario of C-end users) and the type Enter user taste function , and The output function value is used as the user's type in the job application scenario potential preference for corporate information, The larger the output function value, the more likely the user is to The higher the potential preference for corporate information, the higher the potential preference for corporate information. Then, a normalization function (such as the Softmax function) can be used to The output function value is normalized and the normalized result is used as the user's job application scenario for the type The preference parameter of the enterprise information, that is, the value range of the preference parameter is between 0 and 1.

[0166] In other words, the embodiment of the present application lays a rigorous foundation for personalized sorting strategies through the HMM+IFT fusion model, mapping complex user behaviors and business needs into a set of trainable and optimizable probability models and parameter update mechanisms.

[0167] In some embodiments, a type filter panel may also be displayed in the human-computer interaction interface provided by the client, and multiple candidate types may be displayed in the type filter panel for the user to select. If the user manually checks or excludes certain types in the type filter panel during the query process, the preference parameters calculated by the preference function may be modified according to the user's check result or exclusion result in the type filter panel. For example, assuming that the user manually checks the "job search and recruitment category" in the type filter panel, the user's preference parameters for the "job search and recruitment category" calculated by the preference function may be increased, so that in the subsequent search process, a greater weight may be assigned to the corporate information of the "job search and recruitment category"; assuming that the user manually excludes the "financial information category" in the type filter panel, the user's preference parameters for the "financial information category" calculated by the preference function may be reduced, that is, a smaller weight may be assigned to the corporate information of the "financial information category" in the subsequent search process, thereby satisfying the user's temporary preference adjustment.

[0168] In other embodiments, when there is sufficient behavioral data (i.e., the amount of the first operation data is greater than the data amount threshold), the embodiments of the present application can further combine collaborative filtering (CF, Collaborative Filtering), graph neural network (GNN, Graph Neural Network) and other algorithms to perform multi-dimensional modeling of user interests, and then integrate it with the HMM results to improve the accuracy of type preference prediction.

[0169] In step 106, the plurality of preference parameters are respectively mapped to weights of corresponding types, and the plurality of weights and keywords are filled into a preset query data structure to obtain a query expression.

[0170] In some embodiments, the above-mentioned mapping of multiple preference parameters to corresponding types of weights can be achieved in the following ways: determining the weight difference amplitude that matches the business needs; selecting a target mapping method corresponding to the weight difference amplitude from multiple mapping methods (such as mapping formulas), wherein the multiple mapping methods may include: linear mapping, logarithmic mapping, exponential mapping; for each preference parameter, mapping the preference parameter through the target mapping method to obtain a corresponding type of weight.

[0171] For example, taking the time series model as HMM, by modeling HMM and preference functions (such as user taste functions), the user's preference for each type of enterprise information under the query intent corresponding to the most likely state (such as the candidate state with the highest probability of selection) can be obtained during real-time query, that is, The output function value is then normalized by the Softmax function to obtain , that is, the user's most likely state at the moment Under the corresponding query intent, preference parameters for each type of enterprise information, such as the user's most likely current state The relative preference probability of accessing each type of enterprise information under the corresponding query intent. In order to facilitate the use of weight or boost parameters in ES (a parameter used to modify the relevance of documents, the default value is 1), you can The value of (i.e., preference parameter) is mapped to a positive weight suitable for ES treatment.

[0172] It should be noted that in ES, boost or weight parameters are usually multiplied or added to the original relevance score. If the interval is too large, the "preference" may have too strong an impact on the sorting result, while if the interval is too small, the original relevance score may not be effectively distinguished. Therefore, the embodiment of the present application can adjust the upper and lower limits of the interval appropriately according to the test results or experimental feedback to achieve the ideal sorting effect. The value of is between 0 and 1. In order to make the weight have a significant difference in ES, the embodiment of the present application can define a linear mapping, for example, it can be converted into a range, such as between 1 and 5, that is, if The larger the value of is, the greater the weight of the mapping (assuming it is recorded as ) is closer to 5; if The smaller the value of The closer to 1. In addition, the embodiment of the present application can also adjust the mapping formula according to business needs, for example, logarithmic mapping or exponential mapping can be used to increase the difference between different weights. For example, when the difference between different weights needs to be more significant, exponential mapping can be used to The value of .

[0173] In other embodiments, the enterprise information query platform can divide the enterprise information in the enterprise information database into multiple types according to business needs, for example, including: basic information type ( ), including basic enterprise registration information, industrial and commercial registration, equity structure, financial information (i.e. ), including the company's financial statements, annual operating conditions, credit rating data, etc., risk information (i.e. ), including litigation, execution, dishonest persons subject to execution, administrative penalties, risk warnings for affiliated companies, etc., recruitment information (i.e. ), including company recruitment positions, employee evaluation, team structure information, supplier information (i.e. ), including supplier chain, upstream and downstream enterprise relationships, compliance records, etc., customer due diligence (i.e. ), including due diligence reports, related relationships, cooperation records, etc., anti-money laundering (assuming that it is recorded as ), including suspected money laundering records, abnormal data of cross-border fund transactions, etc. Of course, other types can be expanded according to other business situations. For example, the user's preference parameters for basic information categories under the query intention predicted based on HMM can be mapped to obtain the weight of the basic information category (assuming it is recorded as ); Map the user's preference parameters for risk information under the query intention predicted based on HMM to obtain the weight of the risk information class (assuming it is recorded as ). Thus, the weights corresponding to each type can be obtained. After obtaining the weights corresponding to each type, the keywords input by the user and the weights corresponding to each type can be filled into the preset query data structure to obtain the corresponding query expression (such as query DSL).

[0174] For example, ES queries generally include three parts: basic queries that match user search keywords, such as match, bool, filter, must conditions, etc.; adding custom scoring strategies (such as scoring function function_score) to the results of basic matching; controlling the way scores are merged (such as score_mode and boost_mode). For example, in the query DSL, assume that the basic query condition can be the company name (i.e., keyword) entered by the user, and you can also add a must condition to the bool query to match the name, and you can also fill in the weights corresponding to each type in the functions array of the scoring function function_score. Taking the score mode (score_mode) as an example, after obtaining multiple company documents that match the company name, you can first calculate the basic relevance score, then obtain the types of each company document, and weight the relevance score according to the weight corresponding to the type (for example, multiply it). The more the type meets the user's taste, the higher the final score of the company document.

[0175] In step 107, the enterprise information database is queried based on the query expression to obtain multiple enterprise information matching the keyword, and the multiple enterprise information is sorted based on multiple weights.

[0176] In some embodiments, the enterprise information library may include multiple sub-libraries, wherein each sub-library can be used to store enterprise information of multiple enterprises of one type. The above-mentioned querying the enterprise information library based on the query expression to obtain multiple enterprise information matching the keywords can be implemented in the following ways: querying multiple sub-libraries based on the query expression to obtain enterprise information matching the keywords from the multiple sub-libraries respectively; merging the enterprise information obtained from the multiple sub-libraries to obtain multiple enterprise information matching the keywords.

[0177] For example, after the enterprise information query platform divides the multiple enterprise information in the enterprise information database into several business types according to business needs, it can also store different types of enterprise information in different sub-databases. For example, basic information enterprise documents can be stored in sub-database 1; financial information enterprise documents can be stored in sub-database 2; risk information enterprise documents can be stored in sub-database 3, and so on. Each sub-database only stores one type of enterprise document. In addition, corresponding indexes can be constructed for multiple sub-databases respectively. In this way, multiple indexes can be queried simultaneously during the search (i.e., multi-index strategy), and weighted processing can be performed in the query result merging stage.

[0178] It should be noted that if different types of enterprise information are not stored separately, a single index multi-type tagging strategy can be adopted. For example, the same index can be used but fields can be added to enterprise documents to mark the type of enterprise documents. In this index, the type field can be basic information, risk information, financial information, supplier information, etc., to distinguish the types of enterprise documents. In practical applications, the above two methods can be flexibly selected. One recommended solution is to use a single index and distinguish through the type field to simplify the search request structure.

[0179] In other embodiments, the above-mentioned sorting of multiple enterprise information based on multiple weights can be achieved in the following manner: for each enterprise information in the multiple enterprise information, determine the type to which the enterprise information belongs; based on the weight corresponding to the type, weight the relevance between the enterprise information and the keyword to obtain the weighted relevance; sort the multiple enterprise information in descending order of the weighted relevance. That is, by mapping the user's preference parameters (such as preference values) for each type under the query intent predicted by the time series model (such as HMM) to the ES indexing process, it can be ensured that the corresponding weights can be assigned to different types of enterprise information in the search stage, and the query results returned by ES are more in line with the user's hidden query intent and taste preferences after the above-mentioned weight adjustment.

[0180] It should be noted that when calculating the relevance between enterprise information and keywords, if the content of the enterprise information is large (for example, the number of words in the text included in the enterprise information is greater than the word count threshold) and a more accurate semantic search is required, the enterprise information can be vectorized first, and then the relevance can be calculated through the vector engine.

[0181] That is to say, after the query request sent by the user reaches the enterprise information query platform, the enterprise information query platform can first infer the user's session information and the latest click behavior (i.e., the first operation data) through the pre-trained HMM, calculate the user's posterior probability in each candidate state, and select the candidate state corresponding to the largest posterior probability, and the query intent corresponding to the candidate state as the user's current query intent. Then, the user's preference parameters for each type of enterprise information under the query intent can be calculated through the user taste function, and the preference parameters can be mapped to the weights of the corresponding types. Then, the ES query DSL (i.e., query expression) can be dynamically constructed. For example, the functions array of function_score can be filled with the above weights, and the generated query DSL can be submitted to ES for execution. Finally, ES returns the query results and returns them to the user after sorting according to multiple weights.

[0182] It should be noted that, in order to improve performance, in the technical solution provided in the embodiment of the present application, the enterprise information query platform can also cache repeated requests of users in a short period of time. For example, if the user queries the same keyword multiple times in a short period of time, the calculated query intent and weight can be reused. In addition, the parameters of the HMM and the user taste function can be updated regularly to cope with the long-term changes in user preferences.

[0183] In step 108 , it is determined whether a preference adjustment instruction sent by the terminal device is received, if so, step 109 is executed; if not, step 110 is executed.

[0184] In some embodiments, the preference adjustment instruction may be sent by the terminal device after receiving a selection operation of the user for multiple candidate types, wherein the preference adjustment instruction may carry the preference type selected by the user from the multiple candidate types.

[0185] For example, a type option panel may also be displayed in the human-computer interaction interface of the client running on the terminal device. Multiple candidate types (for example, basic information, financial information, risk information, recruitment information, supplier due diligence, customer due diligence, anti-money laundering investigation, etc.) may be displayed in the type option panel for the user to choose from. When the user makes a query, the user may directly make the query without making a selection, or may select the type of his or her preference in the type option panel. In other words, the user may switch preferences for certain types by himself or herself. For example, assuming that the user selects "risk information" in the type option panel, the terminal device may send a signal of temporary preference adjustment of the user to the enterprise information query platform, that is, the terminal device may send a preference adjustment instruction to the enterprise information query background, and the preference adjustment instruction may carry the preference type (for example, "risk information") selected by the user in the search option panel.

[0186] In step 109, the sorted plurality of enterprise information is sorted again by using the large language model, and the sorted plurality of enterprise information is returned to the terminal device.

[0187] In some embodiments, when a preference adjustment instruction is received from a terminal device, the received preference adjustment instruction can be first parsed to obtain the preference type carried by the preference adjustment instruction. Then, for each enterprise information among multiple enterprise information, the similarity between the summary of the enterprise information and the preference type can be determined through a large language model. Finally, the sorted multiple enterprise information can be sorted again in order of similarity from high to low.

[0188] For example, the embodiments of the present application can introduce a large language model to instantly respond to temporary changes in the needs of VIP users or users of specific institutions. For example, after receiving a preference adjustment instruction sent by a terminal device, the enterprise information query platform can input the preference type carried by the preference adjustment instruction into the large language model, so that the large language model can perform additional sorting fine-tuning based on the user's temporary preferences. That is, the natural language understanding ability of the large language model can be used to perform secondary sorting on the sorting results based on the user's current temporary intention changes, so that a more matching intelligent sorting result can be presented to the user.

[0189] It should be noted that in the technical solution provided in the embodiment of the present application, before using the large language model, instruction fine-tuning or domain fine-tuning can be performed for the "enterprise information query scenario" so that the large language model can be more familiar with the semantics of keywords such as finance, suppliers, and risks. In addition, in addition to using the large language model to re-sort the sorted multiple enterprise information, a lightweight classifier or sentence vector model can also be used to identify temporary changes in user needs. For example, in a resource-constrained environment or a scenario that is sensitive to latency, the use of a lightweight small model can be more flexible.

[0190] In addition, it should be noted that when the terminal device calls the client to present multiple enterprise information after re-sorting, it can also guide the user to view other types of enterprise information that may be of interest through guided questions and answers. For example, a real-time feedback component can also be presented in the client's human-computer interaction interface, and the user can give "satisfied / unsatisfied" or "like / dislike" feedback on the current sorting results. The enterprise information query platform can input the user's feedback data into the large language model or HMM to trigger short-term or long-term optimization. In addition, when the user needs to go deeper or dynamically adjust the query intention, the large language model can also give intelligent question and answer prompts such as "Do you need to view the supplier due diligence?" to further guide the user.

[0191] In step 110, the sorted information of multiple enterprises is returned to the terminal device.

[0192] In some embodiments, continuing with the above example, when the user does not select multiple candidate types in the search options panel when querying, that is, the user does not temporarily adjust his or her preferences, the enterprise information query platform can directly return the sorted multiple enterprise information to the user's terminal device.

[0193] In other embodiments, in some extremely vertical B-side scenarios, such as only facing users of specific financial institutions, and with stable rules, an expert system or rule engine can also be used to re-sort the sorted multiple enterprise information. For example, when it is detected that the user has searched for the keyword "supplier risk" multiple times in this round of conversation, the priority of enterprise information of the "supplier due diligence" category can be increased in the sorting.

[0194] In some embodiments, see Figure 5 , Figure 5 is a third flow chart of the method for querying enterprise information provided in the embodiment of the present application, such as Figure 5 As shown, after executing Figure 3 After step 109 or step 110 shown, you can also perform Figure 5 Steps 116 to 118 shown will combine Figure 5 The steps shown are explained.

[0195] In step 116, second operation data of the user on the enterprise information query platform is obtained.

[0196] Here, the second operation data is generated after the enterprise information query platform returns the sorted multiple enterprise information to the terminal device, or returns the re-sorted multiple enterprise information. For example, the second operation data may include: the keywords that the user has queried recently, the type of results that have been clicked, the length of time that the page stays, etc. In addition, if the user is a VIP user or a user of a specific institution, the user's previous interaction history and long-term query preference characteristics can also be provided.

[0197] In step 117, the second operation data is input into the large language model, so that the large language model determines whether the user's preference has shifted based on the second operation data.

[0198] In some embodiments, the second operation data may include operation data of multiple observation cycles, and step 117 may be implemented in the following manner: the second operation data is input into the large language model so that the large language model performs the following processing: the following processing is performed for the operation data of each observation cycle: the type of enterprise information whose interaction frequency (for example, click frequency) is greater than the interaction frequency threshold is obtained from the operation data of the observation cycle (that is, the user's preference type in the observation cycle); the type is compared with the user's preference type determined based on the first operation data, or the type corresponding to the preference parameter greater than the parameter threshold (that is, the user's preference type predicted based on the timing model and the preference function); when the comparison result is that the two are inconsistent, a preference transfer signal is generated, wherein the preference transfer signal can be used to characterize that the user's preference in the observation cycle has changed.

[0199] For example, the embodiment of the present application can introduce a large language model as an online adapter for specific users (such as VIP users or specific institutions). When the user deviates from the query intention predicted by the HMM (for example, from focusing on risk information to focusing on supplier information), the large language model can detect the user's preference changes in real time by understanding the context of the user's recent query keywords, click behavior, HMM state and ES results. In addition, the large language model can also store the detected preference transfer information (for example, the user shifts from focusing on "risk assessment" to "supplier due diligence") in the user portrait warehouse. If the change shows a certain degree of continuity (for example, the number of times the user's preference changes reaches a threshold), the HMM can be trained again.

[0200] Specifically, taking the time series model as HMM as an example, in most cases, the HMM and ES weight adjustment strategies can meet the personalized needs of users. However, for specific users (such as VIP users, key cooperative banks, insurance institutions or fund companies, etc.), their query behavior may present more complex and changeable characteristics. For example, such users may focus on a certain category (such as risk information) for a period of time, and suddenly shift to another category (such as supplier information) for in-depth query after a period of time. At this time, the state prediction of HMM and the ES solution based on offline strategy optimization may not be able to capture such rapid turns in time. In view of this, the embodiment of the present application uses the large language model as an online, intelligent enhancement layer to provide more timely and delicate personalized responses for special or high-value users. Among them, the task of the large language model is to analyze the user's recent behavioral data and determine whether the current user has generated a preference shift. For example, assuming that the user frequently clicked on risk-related corporate documents before, but after this query, he continuously clicked and paid attention to supplier information-related corporate documents, then the large language model can conclude that the user's temporary demand has shifted from the risk category to the supplier due diligence category; or, assuming that the type predicted by the HMM is "risk assessment", but the user's actual recent behavior is more inclined to another type (such as customer due diligence), the large language model can understand the user's input keywords, click sequences, document summaries and user portrait data at the semantic level to obtain preference transfer signals.

[0201] It should be noted that when the large language model concludes that the user's preference has shifted, it can perform semantic understanding and matching evaluation on each enterprise document based on the current temporary environment. For example, assuming that the user is more concerned about supplier-related enterprise information in the short term, the large language model can give a higher sorting priority to the enterprise information belonging to the supplier category among the sorted multiple enterprise information; or, the large language model can also combine natural language reasoning to optimize and improve the summary of the enterprise information based on whether it contains text features that are highly relevant to the user's latest preferences. After processing, the large language model can output a new sorted list or a weight distribution for scoring each enterprise information. The enterprise information query platform can re-sort the sorted multiple enterprise information based on this, and return the final sorting result to the user, thereby achieving instant, high-precision, personalized response.

[0202] In other embodiments, the large language model can also record the user's preference transfer information. For example, when the large language model determines that the user's preference has changed significantly, it can output a preference transfer signal. After receiving the signal, the enterprise information query platform can record the preference transfer information in the user portrait warehouse. The warehouse can be an independent data storage, for example, it can use a key-value pair or a graph database structure to store the user's search history, preference transfer trajectory, demand shift at a specific time point and corresponding context information.

[0203] In step 118, when the user's preference shifts and the number of shifts to the same type is greater than a threshold number, the time series model is retrained based on the second operation data.

[0204] In some embodiments, taking 10 observation cycles as an example, assuming that in these 10 observation cycles, the large language model generates a total of 8 preference transfer signals, that is, the user's preference has changed within 8 observation cycles, and assuming that the preset number threshold is 7 times, it can be determined that the user's current preference has changed significantly (that is, the user's preference has undergone multiple similar transfers in subsequent time), and the timing model can be trained again.

[0205] For example, for temporary preference shifts of users, the large language model can perform instant secondary sorting of the current query request, thereby presenting users with results that better meet temporary needs. The next time a user queries, they can enjoy customized results based on the latest preference shifts without waiting for offline retraining of the HMM or ES strategy. In addition, when the user's preference shift records accumulate to a certain level (for example, the number of times the user's preference changes is greater than the number threshold), the enterprise information query platform can conduct regular evaluations. For example, when the large language model repeatedly points out that the user has shifted from the "HMM predicted state" to other types, and this situation continues or occurs frequently (for example, the user has repeatedly encountered situations that do not match the HMM prediction results in the past two weeks), the retraining process can be initiated. For example, the enterprise information query platform can import the actual user access records and preference shift events during this period into the HMM training pipeline, and optimize the model parameters again through the Baum-Welch algorithm, thereby updating the state transition probability and the user taste function, so that the HMM can better adapt to the user's new long-term demand characteristics.

[0206] In addition, when users have shown strong preferences for certain types for a long time, the enterprise information query platform can also update the ES word segmenter, adjust the dictionary, and enhance the weight initialization parameters of related fields. It is also possible to perform preset optimizations in the scoring function (function_score) or the basic weight strategy of the corresponding type to meet the user's new basic preferences more quickly, without relying on the intervention of the large language model every time. For example, if long-term statistics show that the frequency of requests for anti-money laundering and supplier due diligence by a certain bank user has increased significantly, a higher initial ES weight can be set for the user in advance, so that the baseline ranking of ES is closer to the determined demand characteristics of such users.

[0207] In other words, through the integration of the above long-term and short-term strategies, an adaptive evolutionary closed loop can be formed: 1. Preferences change in the short term → large language model instantly re-sorts → users immediately enjoy a better experience; 2. Multiple preference change records are summarized → HMM is retrained and ES strategy is adjusted → optimization results are solidified in the long term; 3. The optimized HMM and ES strategies are more efficient in future queries, reducing dependence on the immediate intervention of the large language model; 4. If the user has a new preference change again, the large language model will capture and correct it again, forming a positive feedback loop of continuous iteration and evolution.

[0208] It can be seen that this multi-level optimization can not only meet the temporary and tricky needs of individual users, but also maintain the optimal adaptation for the entire user group or specific important users in the long term. For VIP users, the real-time understanding ability of the large language model can be used to quickly adjust the results to avoid users being disappointed because of dissatisfaction with the search results. At the same time, by continuously using user behavior feedback for offline retraining of HMM and ES strategies, the basic capabilities of the enterprise information query platform are constantly evolving. In the future, when facing similar changes in demand, there is no need to rely too much on the frequent intervention of the large language model, which reduces complexity and computing costs, and improves overall query efficiency and user satisfaction. In the long-term operation, the accumulated behavioral data of the user portrait warehouse, the continuous update and iteration of HMM, the refined weight allocation of ES strategy, combined with the advantages of the large language model in natural language and context understanding, can make the enterprise information query platform have the characteristics of self-adaptation and self-evolution, which is in line with the development trend and application value of future intelligent search and recommendation systems. That is, the short-term preference transfer of users can be instantly adapted by the large language model, and the long-term trend can be solidified by recording preference changes and then training HMM and updating ES strategies. In addition, the temporary adjustment data accumulated from multiple user sessions can be used as a reference for subsequent HMM training and ES optimization, allowing the enterprise information query platform to show higher matching and intelligence levels for specific users and scenarios in the long-term evolution.

[0209] In other embodiments, in addition to retraining the timing model after multiple similar preference shifts occur in the user, online learning or reinforcement learning solutions may also be used to perform small-step iterations after each receipt of user feedback data, that is, user feedback can immediately affect model parameters and converge more quickly.

[0210] In summary, the enterprise information query method provided in the embodiment of the present application has the following beneficial effects:

[0211] 1. Intelligent intention prediction: The embodiment of the present application combines the hidden Markov model with the information foraging theory to accurately characterize the user's implicit query intention in the process of enterprise information query and the preference for various types of enterprise information, and significantly improves the prediction accuracy of the user's preferred type in the user's diversified search scenarios;

[0212] 2. Personalized re-ranking of search results: After calculating the user's preference parameters for different types of enterprise information under the most likely query intent at the moment through the preference function, the embodiment of the present application maps multiple preference parameters to weights of corresponding types, thereby achieving refined re-ranking of search results at the type level, helping users to efficiently find the required content in the complex amount of enterprise credit data;

[0213] 3. Dynamic adaptive adjustment: The embodiment of the present application introduces a large language model as an online enhancement layer, which can respond immediately when the preferences of specific users (such as VIP users or financial institutions) suddenly change, and perform secondary intelligent sorting on the preliminary sorting results to avoid damage to user experience due to untimely updates of traditional strategies;

[0214] 4. Closed-loop self-evolution: In the technical solution provided in the embodiment of the present application, the large language model can continuously feed back dynamic change information such as user preference transfer to the user portrait warehouse and the underlying HMM, ES strategy optimization module, etc. In subsequent iterations, the system automatically retrains and updates the core model and index strategy, forming a closed-loop optimization process from short-term remediation to long-term evolution, and continuously improving overall intelligence and matching.

[0215] In other words, the embodiment of the present application constructs a user taste function model by integrating HMM and information foraging theory, and uses this as a basis to achieve type-level personalized reordering in ES search. Assisted by the real-time adaptive adjustment of specific users and long-term strategy closed-loop optimization of the large language model, the enterprise information query platform has the characteristics of rapid response, multi-level evolution and continuous improvement of user experience. The technical solution provided by the embodiment of the present application can effectively meet the diversified and dynamic user needs in the field of enterprise information query, and provide higher-value intelligent information services for business decision-making and risk control.

[0216] The following is a description of an exemplary structure of the enterprise information query device 243 provided in the embodiment of the present application implemented as a software module. In some embodiments, for example Figure 2 As shown, the software modules in the query device 243 for enterprise information stored in the memory 240 may include: a receiving module 2431, an acquisition module 2432, a prediction module 2433, a sorting module 2434, a determination module 2435, a construction module 2436, a mapping module 2437, a filling module 2438, a query module 2439 and a sending module 24310.

[0217] A receiving module 2431 is used to receive a query request sent by a terminal device, wherein the query request carries keywords; an acquiring module 2432 is used to acquire first operation data of a user of the terminal device on the enterprise information query platform, wherein the first operation data is generated before receiving the query request; a predicting module 2433 is used to predict, based on the first operation data, a plurality of candidate states included in a state set by means of a pre-trained time series model, to obtain a selection probability of each candidate state, wherein each candidate state corresponds to a query intent, and the query intent represents a potential query target of a user when accessing the enterprise information query platform; a sorting module 2434 is used to sort the plurality of candidate states in descending order of the selection probability; a determining module 2435 is used to use a query intent corresponding to at least one of the candidate states ranked higher in the sorting result as the query intent of the user; a constructing module 2436 is used to construct a preference function based on the plurality of types and the state set; a determining module 2435 is also used to determine the user's preference parameters for each type under the query intention through the preference function; the mapping module 2437 is used to map multiple preference parameters to corresponding types of weights; the filling module 2438 is used to fill multiple weights and keywords into a preset query data structure to obtain a query expression; the query module 2439 is used to query the enterprise information database based on the query expression to obtain multiple enterprise information matching the keyword; the sorting module 2434 is also used to sort the multiple enterprise information based on the multiple weights; the sending module 24310 is used to return the sorted multiple enterprise information to the terminal device when the preference adjustment instruction sent by the terminal device is not received; the sorting module 2434 is also used to re-sort the sorted multiple enterprise information through a large language model when the preference adjustment instruction sent by the terminal device is received; the sending module 24310 is also used to return the re-sorted multiple enterprise information to the terminal device.

[0218] In some embodiments, the preference adjustment instruction is sent by the terminal device after receiving the user's selection operation for multiple candidate types, wherein the preference adjustment instruction carries the preference type selected by the user from the multiple candidate types; the sorting module 2434 is also used to perform the following processing: parsing the preference adjustment instruction to obtain the preference type carried by the preference adjustment instruction; for each enterprise information among the multiple enterprise information, determining the similarity between the summary of the enterprise information and the preference type through a large language model; and re-sorting the sorted multiple enterprise information in order of the similarity from high to low.

[0219] In some embodiments, after the sending module 24310 returns the sorted multiple enterprise information to the terminal device, or returns the multiple enterprise information after re-sorting, the acquisition module 2432 is also used to obtain the second operation data of the user on the enterprise information query platform, wherein the second operation data is generated after returning the sorted multiple enterprise information, or returning the multiple enterprise information after re-sorting; the enterprise information query device 243 also includes an input module 24311 and a training module 24312, wherein the input module 24311 is used to input the second operation data into the large language model so that the large language model determines whether the user's preference has shifted based on the second operation data; the training module 24312 is used to re-train the time series model based on the second operation data when the user's preference has shifted and the number of transfers to the same type is greater than a number threshold.

[0220] In some embodiments, the second operation data includes operation data of multiple observation cycles; the input module 24311 is also used to input the second operation data into the large language model so that the large language model performs the following processing: the following processing is performed for the operation data of each observation cycle: the type of enterprise information whose interaction frequency is greater than an interaction frequency threshold is obtained from the operation data of the observation cycle; the type is compared with the user's preference type determined based on the first operation data, or the type is compared with the type corresponding to the preference parameter greater than a parameter threshold; when the comparison result is inconsistent, a preference transfer signal is generated, wherein the preference transfer signal indicates that the user's preference in the observation cycle has shifted.

[0221] In some embodiments, the mapping module 2437 is also used to determine the weight difference amplitude that matches the business needs; select a target mapping method corresponding to the weight difference amplitude from multiple mapping methods, wherein the multiple mapping methods include: linear mapping, logarithmic mapping, exponential mapping; for each of the preference parameters, map the preference parameters through the target mapping method to obtain a weight of the corresponding type; the determination module 2435 is also used to determine the type of each enterprise information in the multiple enterprise information; the enterprise information query device 243 also includes a weighting module 24313, which is used to weight the correlation between the enterprise information and the keyword based on the weight corresponding to the type to obtain a weighted correlation; the sorting module 2434 is also used to sort the multiple enterprise information in order from high to low according to the weighted correlation.

[0222] In some embodiments, the acquisition module 2432 is further used to obtain the marked query intents among the multiple query intents corresponding to the multiple candidate states included in the state set; and to obtain the type of enterprise information interacted by the user under the marked query intent, and the interaction parameters for the enterprise information; the construction module 2436 is further used to construct a preference function from the independent variable to the dependent variable using the marked query intent and the type as independent variables, and the interaction parameters as dependent variables.

[0223] In some embodiments, the determination module 2435 is also used to perform the following processing for each of the types: input the query intent and the type into the preference function to obtain a function value output by the preference function; normalize the function value through a normalization function, and use the normalized function value as the user's preference parameter for the type under the query intent.

[0224] In some embodiments, the first operation data includes: the click sequence, input keywords and stay time of the user on the enterprise information query platform; the determination module 2435 is also used to determine the user's historical query intention based on the user's click sequence, input keywords and stay time on the enterprise information query platform; the prediction module 2433 is also used to input the historical query intention into a pre-trained time series model so that the pre-trained time series model determines the posterior probability of the user being in the query intention corresponding to each candidate state included in the state set at the current moment; each of the posterior probabilities is used as the selection probability of the corresponding candidate state; the training module 24312 is also used to perform the following processing before the prediction module 2433 predicts the multiple candidate states included in the state set through the pre-trained time series model : Construct an initialized time series model, wherein the parameters of the time series model include: the state set, the initial state distribution, the state transition probability matrix and the observation probability distribution; obtain the log data of the user on the enterprise information query platform, and create an observation sequence based on the log data; determine the first probability of the existence of each candidate state included in the state set at each time point of the observation sequence, and determine the second probability of the observation sequence appearing in each candidate state from the current time point to the end of the sequence; based on the first probability and the second probability, determine the expected number of transitions between each candidate state and the expected number of observations under each candidate state; based on the expected number of transitions and the expected number of observations, update the initial state distribution, the state transition probability matrix and the observation probability distribution.

[0225] In some embodiments, the enterprise information library includes multiple sub-libraries, wherein each of the sub-libraries is used to store enterprise information of multiple enterprises of one type; the query module 2439 is also used to query the multiple sub-libraries based on the query expression to obtain enterprise information matching the keyword from the multiple sub-libraries respectively; the enterprise information obtained from the multiple sub-libraries are merged to obtain multiple enterprise information matching the keyword.

[0226] It should be noted that the description of the device in the embodiment of the present application is similar to the description of the method embodiment above, and has similar beneficial effects as the method embodiment, so it will not be repeated. Figure 3 , Figure 4 ,or Figure 5 The present invention should be understood by referring to the description of any one of the accompanying drawings.

[0227] The embodiment of the present application provides a computer program product, which includes a computer program or a computer executable instruction, and the computer program or the computer executable instruction is stored in a computer-readable storage medium. The processor of the computer device reads the computer executable instruction from the computer-readable storage medium, and the processor executes the computer executable instruction, so that the computer device executes the enterprise information query method described in the embodiment of the present application.

[0228] The present application embodiment provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, the processor will execute the enterprise information query method provided by the present application embodiment, for example, Figure 3 , Figure 4 ,or Figure 5 A method for querying enterprise information is shown.

[0229] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various devices including one or any combination of the above memories.

[0230] In some embodiments, executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0231] As an example, the executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed at multiple sites and interconnected by a communication network.

[0232] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for querying enterprise information, characterized in that: Applied to an enterprise information query platform, the enterprise information database of the enterprise information query platform includes a plurality of types of collections, and one of the collections includes enterprise information of a plurality of enterprises of one type; The method comprises: Receiving a query request sent by a terminal device, wherein the query request carries a keyword; Acquire first operation data of the user of the terminal device on the enterprise information query platform, wherein the first operation data is generated before receiving the query request; Based on the first operation data, a plurality of candidate states included in the state set are predicted respectively by a pre-trained time series model to obtain a selection probability of each candidate state, wherein each candidate state corresponds to a query intent, and the query intent represents a potential query target of a user when accessing the enterprise information query platform; Sorting the multiple candidate states in descending order of the selection probability, and taking the query intention corresponding to at least one of the candidate states ranked first in the sorting result as the query intention of the user; Building a preference function based on the multiple types and the state set, and determining the user's preference parameters for each type under the query intention through the preference function; Mapping the plurality of preference parameters to weights of corresponding types respectively, and filling the plurality of weights and the keywords into a preset query data structure to obtain a query expression; Querying the enterprise information database based on the query expression to obtain a plurality of enterprise information matching the keyword, and sorting the plurality of enterprise information based on a plurality of the weights; When no preference adjustment instruction sent by the terminal device is received, returning the sorted plurality of enterprise information to the terminal device; When the preference adjustment instruction sent by the terminal device is received, the sorted multiple enterprise information is sorted again using a large language model, and the re-sorted multiple enterprise information is returned to the terminal device.

2. The method according to claim 1, characterized in that The preference adjustment instruction is sent by the terminal device after receiving the user's selection operation for multiple candidate types, wherein the preference adjustment instruction carries the preference type selected by the user from the multiple candidate types; The step of re-sorting the sorted plurality of enterprise information by using the large language model comprises: Parsing the preference adjustment instruction to obtain the preference type carried by the preference adjustment instruction; For each piece of enterprise information among the plurality of enterprise information, determining the similarity between a summary of the enterprise information and the preference type by using a large language model; The sorted plurality of enterprise information are sorted again in the order of the similarity from high to low.

3. The method according to claim 1, characterized in that After returning the sorted plurality of enterprise information to the terminal device, or returning the re-sorted plurality of enterprise information, the method further includes: Acquire second operation data of the user on the enterprise information query platform, wherein the second operation data is generated after returning the sorted plurality of enterprise information or returning the re-sorted plurality of enterprise information; inputting the second operation data into the large language model, so that the large language model determines whether the user's preference has shifted based on the second operation data; When the user's preference shifts and the number of shifts to the same type is greater than a number threshold, the time series model is retrained based on the second operation data.

4. The method according to claim 3, characterized in that The second operation data includes operation data of a plurality of observation periods; The step of inputting the second operation data into the large language model so that the large language model determines whether the user's preference has shifted based on the second operation data, includes: The second operation data is input into the large language model so that the large language model performs the following processing: For the operation data of each observation period, the following processing is performed: Obtaining the type of enterprise information whose interaction frequency is greater than an interaction frequency threshold from the operation data of the observation period; Comparing the type with a preference type of the user determined based on the first operation data, or comparing the type with a type corresponding to the preference parameter that is greater than a parameter threshold; When the comparison result is inconsistent, a preference transfer signal is generated, wherein the preference transfer signal indicates that the user's preference has shifted during the observation period.

5. The method according to claim 1, characterized in that The mapping of the plurality of preference parameters to weights of corresponding types respectively includes: Determine the weight difference range that matches business needs; Selecting a target mapping method corresponding to the weight difference amplitude from a plurality of mapping methods, wherein the plurality of mapping methods include: linear mapping, logarithmic mapping, and exponential mapping; For each of the preference parameters, the preference parameters are mapped by the target mapping method to obtain a weight of a corresponding type; The sorting of the plurality of enterprise information based on the plurality of weights comprises: For each piece of enterprise information among the plurality of enterprise information, determining the type to which the enterprise information belongs; Based on the weight corresponding to the type, weighting the relevance between the enterprise information and the keyword to obtain a weighted relevance; The plurality of enterprise information are sorted in descending order of the weighted relevance.

6. The method according to claim 1, characterized in that The constructing a preference function based on the multiple types and the state set includes: Obtaining the marked query intents among the multiple query intents respectively corresponding to the multiple candidate states included in the state set; Acquire the type of enterprise information that the user interacts with under the marked query intent, and interaction parameters for the enterprise information; The marked query intent and the type are taken as independent variables, and the interaction parameter is taken as a dependent variable, and a preference function from the independent variables to the dependent variable is constructed.

7. The method according to claim 1, characterized in that Determining the user's preference parameters for each type under the query intention by using the preference function includes: For each of these types the following processing is performed: Inputting the query intent and the type into the preference function to obtain a function value output by the preference function; The function value is normalized by a normalization function, and the normalized function value is used as the user's preference parameter for the type under the query intent.

8. The method according to claim 1, characterized in that: The first operation data includes: the click sequence, input keywords and stay time of the user on the enterprise information query platform; The step of predicting a plurality of candidate states included in the state set based on the first operation data by using a pre-trained time series model to obtain a selection probability of each candidate state includes: Determine the user's historical query intention based on the user's click sequence, input keywords and stay time on the enterprise information query platform; Inputting the historical query intent into a pre-trained time series model, so that the pre-trained time series model determines the posterior probability that the user is in the query intent corresponding to each candidate state included in the state set at the current moment; Taking each of the posterior probabilities as the probability of selection of the corresponding candidate state; Before predicting the multiple candidate states included in the state set respectively by using the pre-trained time series model, the method further includes: Constructing an initialized time series model, wherein the parameters of the time series model include: the state set, the initial state distribution, the state transition probability matrix, and the observation probability distribution; Obtaining log data of the user on the enterprise information query platform, and creating an observation sequence based on the log data; Determine a first probability that each time point of the observation sequence exists in each candidate state included in the state set, and determine a second probability that the observation sequence appears in each candidate state from the current time point to the end of the sequence; Determine an expected number of transitions between each of the candidate states and an expected number of observations in each of the candidate states based on the first probability and the second probability; Based on the expected number of transitions and the expected number of observations, the initial state distribution, the state transition probability matrix, and the observation probability distribution are updated.

9. The method according to any one of claims 1 to 8, characterized in that: The enterprise information database includes multiple sub-databases, wherein each of the sub-databases is used to store enterprise information of multiple enterprises of one type; The querying the enterprise information database based on the query expression to obtain a plurality of enterprise information matching the keyword includes: querying the multiple sub-databases based on the query expression to respectively obtain enterprise information matching the keyword from the multiple sub-databases; The enterprise information respectively obtained from the plurality of sub-databases is merged to obtain a plurality of enterprise information matching the keyword.

10. A device for querying enterprise information, characterized in that: Applied to an enterprise information query platform, the enterprise information database of the enterprise information query platform includes a plurality of types of collections, and one of the collections includes enterprise information of a plurality of enterprises of one type; The device comprises: A receiving module, configured to receive a query request sent by a terminal device, wherein the query request carries a keyword; An acquisition module, configured to acquire first operation data of the user of the terminal device on the enterprise information query platform, wherein the first operation data is generated before receiving the query request; A prediction module, configured to predict, based on the first operation data, a plurality of candidate states included in the state set by means of a pre-trained time series model, respectively, to obtain a selection probability of each candidate state, wherein each candidate state corresponds to a query intent, and the query intent represents a potential query target of a user when accessing the enterprise information query platform; A sorting module, used to sort the multiple candidate states in descending order of the selection probability; A determination module, configured to use the query intention corresponding to at least one of the candidate states ranked higher in the sorting result as the query intention of the user; A construction module, configured to construct a preference function based on the plurality of types and the state set; The determination module is further used to determine the preference parameters of the user for each of the types under the query intention through the preference function; A mapping module, used for mapping the plurality of preference parameters into weights of corresponding types respectively; A filling module, used to fill a plurality of the weights and the keywords into a preset query data structure to obtain a query expression; A query module, used to query the enterprise information database based on the query expression to obtain multiple enterprise information matching the keyword; The sorting module is further used to sort the plurality of enterprise information based on the plurality of weights; a sending module, configured to return the sorted plurality of enterprise information to the terminal device when no preference adjustment instruction sent by the terminal device is received; The sorting module is further configured to re-sort the sorted plurality of enterprise information by using a large language model when receiving a preference adjustment instruction sent by the terminal device; The sending module is further used to return the re-sorted plurality of enterprise information to the terminal device.

Images