Information retrieval method and device, electronic equipment, storage medium and computer program product

Abstract

Embodiments of the invention provide an information retrieval method and apparatus, an electronic device, a storage medium and a computer program product. The method comprises the steps of obtaining a query statement of a target object and historical retrieval data of the target object; performing intention recognition on the query statement based on the historical retrieval data to obtain an intention recognition result; generating at least one retrieval condition and retrieval result sorting words and sentences based on the intention recognition result; wherein each retrieval condition corresponds to one retrieval type, and the retrieval conditions under the same retrieval type comprise at least one retrieval dimension; the retrieval result sorting words and sentences are used for appointing a sorting mode of multiple pieces of retrieved information during information retrieval; generating a dynamic retrieval statement according to the at least one retrieval condition and the retrieval result sorting word and sentence; and performing information retrieval by executing the dynamic retrieval statement to obtain an information retrieval result matched with the query statement. According to the method, intention recognition is performed on the prompt information, and the dynamic retrieval statement is generated according to the recognition result, so that the retrieval precision is improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence, and relate to but are not limited to an information retrieval method, device, electronic device, storage medium, and computer program product. Background Art

[0002] With the rapid development of information technology, information retrieval has become an indispensable part of business operations, scientific research and daily life. In related technologies, traditional information retrieval methods can only search through keyword matching or simple, fixed search statements. However, traditional information retrieval methods cannot accurately identify the user's intention, let alone perform precise retrieval based on the user's intention, resulting in low retrieval efficiency and accuracy. Summary of the invention

[0003] The embodiments of the present application provide an information retrieval method, device, electronic device, storage medium and computer program product, which can be applied at least in the field of artificial intelligence. The present application improves the retrieval accuracy by identifying the intent of prompt information, generating dynamic retrieval statements based on the identification results, and executing the dynamic retrieval statements to obtain retrieval results.

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

[0005] An embodiment of the present application provides an information retrieval method, including: obtaining a query statement for a target object and historical retrieval data of the target object; performing intent recognition on the query statement based on the historical retrieval data to obtain an intent recognition result; generating at least one retrieval condition and a retrieval result sorting sentence based on the intent recognition result; wherein each retrieval condition corresponds to a retrieval type, and the retrieval conditions under the same retrieval type include at least one retrieval dimension; the retrieval result sorting sentence is used to specify the sorting method for multiple retrieved information during information retrieval; generating a dynamic retrieval statement through the at least one retrieval condition and the retrieval result sorting sentence; performing information retrieval by executing the dynamic retrieval sentence to obtain an information retrieval result matching the query statement.

[0006] An embodiment of the present application provides an information retrieval device, including: an acquisition module, used to acquire a query statement of a target object and historical retrieval data of the target object; an intention recognition module, used to perform intention recognition on the query statement based on the historical retrieval data to obtain an intention recognition result; a first generation module, used to generate at least one retrieval condition and a retrieval result sorting sentence based on the intention recognition result; wherein each retrieval condition corresponds to a retrieval type, and the retrieval conditions under the same retrieval type include at least one retrieval dimension; the retrieval result sorting sentence is used to specify the sorting method for multiple retrieved information during information retrieval; a second generation module, used to generate a dynamic retrieval statement through the at least one retrieval condition and the retrieval result sorting sentence; a retrieval module, used to perform information retrieval by executing the dynamic retrieval sentence to obtain an information retrieval result matching the query statement.

[0007] In the above scheme, the intention recognition module is also used to: extract features from the historical retrieval data and the query statement to obtain historical retrieval data features and query statement features; splice the historical retrieval data features and the query statement features to obtain spliced ​​features; perform feature mapping on the spliced ​​features through a large language model to obtain an intention recognition probability; the intention recognition probability is used to characterize the probability that the query statement belongs to a predefined intention category; and determine the intention recognition result based on the intention recognition probability.

[0008] In the above scheme, the intention recognition module is also used to: perform a first network processing on the concatenated features through a large language model to obtain semantic recognition features of the query statement; perform a second network processing on the semantic recognition features to obtain reference resolution features used to characterize the relationship between entities and demonstrative words in the query statement; perform a third network processing on the semantic recognition features and the reference resolution features to obtain the intention recognition probability.

[0009] In the above scheme, the intention recognition result includes the intention recognition probability and metadata information; the first generation module is also used to: generate the at least one retrieval condition based on the metadata information and the intention recognition probability in the intention recognition result; the retrieval type corresponding to the retrieval condition includes at least one of the following: a condition type that should be met, a condition type that must be met, a condition type that must not be met, and a filtering condition type; based on the intention recognition probability in the intention recognition result, generate the retrieval result sorting sentence.

[0010] In the above scheme, the retrieval result sorting sentence is used to specify the sorting method including at least one of the following: relevance sorting method and timeliness sorting method; the relevance sorting method is characterized by a relevance weight, and the timeliness sorting method is characterized by a timeliness weight; the first generation module is also used to: determine the intent category corresponding to the query statement based on the intent recognition probability in the intent recognition result; regress the intent category through a regression model to obtain a relevance weight and a timeliness weight; and determine the weight description sentences corresponding to the relevance weight and the timeliness weight as the retrieval result sorting sentence.

[0011] In the above scheme, the second generation module is also used to: convert each of the search conditions into a corresponding conditional query statement to obtain at least one conditional query statement; parse the search result sorting words to obtain relevance weights and timeliness weights; construct a sorting function based on the relevance weights and the timeliness weights; and add the at least one conditional query statement and the sorting function to the dynamic search statement.

[0012] In the above scheme, the retrieval module is also used to: perform information retrieval by executing each conditional query statement in the dynamic retrieval statement, and obtain the relevance value and time sensitivity value between each candidate information in the preset information library and the dynamic retrieval statement; for each of the candidate information, call the sorting function, fuse the relevance value and the time sensitivity value, and obtain the retrieval score of the candidate information; based on the retrieval score, determine the target information that matches the query statement from the preset information library; and determine the target information as the information retrieval result.

[0013] In the above scheme, the device also includes an encapsulation module, which is used to perform data analysis on the historical retrieval data to obtain the object characteristics of the target object; encapsulate the object characteristics into a query statement; accordingly, the intention recognition module is also used to: perform intention recognition on the query statement based on the query statement to obtain the intention recognition result.

[0014] In the above scheme, the device also includes an adding module, which is used to determine the lexical interpretation of the target vocabulary in the query statement based on a pre-constructed knowledge graph; add the lexical interpretation of the target vocabulary to the query statement to obtain an interpreted query statement; accordingly, the intention recognition module is also used to: perform intention recognition on the interpreted query statement based on the historical retrieval data to obtain the intention recognition result.

[0015] An embodiment of the present application provides an electronic device, comprising: a memory for storing computer-executable instructions or computer programs; and a processor for implementing the above-mentioned information retrieval method when executing the computer-executable instructions or computer programs stored in the memory.

[0016] An embodiment of the present application provides a computer program product storing computer executable instructions or a computer program, which is used to cause a processor to execute the computer executable instructions or the computer program to implement the above-mentioned information retrieval method.

[0017] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions or a computer program, which is used to cause a processor to execute the computer-executable instructions or the computer program to implement the above-mentioned information retrieval method.

[0018] The above scheme has the following beneficial effects:

[0019] The information retrieval method in the embodiment of the present application performs intent recognition on query statements based on historical retrieval data, can understand the user's real needs, avoid shallow matching based only on keywords, and through the generated intent recognition results, can obtain retrieval conditions that are closer to user needs, thereby ensuring the accuracy of the retrieval results; moreover, the retrieval conditions can be generated according to different retrieval types and dimensions, ensuring that the system can simultaneously support comprehensive analysis of multiple dimensions; and, dynamic retrieval statements can be generated in real time according to the retrieval conditions and retrieval result sorting sentences to adapt to different retrieval scenarios, with higher flexibility; in addition, dynamic retrieval statements are generated through intent recognition results, and retrieval results are obtained by executing the dynamic retrieval statements, which can improve retrieval accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is an optional architecture diagram of the information retrieval system provided in the embodiment of the present application;

[0021] Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0022] Figure 3 This is an optional flowchart of the information retrieval method provided in the embodiment of the present application;

[0023] Figure 4 is another optional flowchart of the information retrieval method provided in the embodiment of the present application;

[0024] Figure 5 It is a flowchart of performing intent recognition on a query statement based on historical retrieval data to obtain an intent recognition result provided by an embodiment of the present application;

[0025] Figure 6It is a flowchart of generating a dynamic search statement through at least one search condition and a search result sorting clause provided by an embodiment of the present application;

[0026] Figure 7 It is a flow chart of obtaining information retrieval results matching a query statement by executing a dynamic search statement to retrieve information provided by an embodiment of the present application;

[0027] Figure 8 It is an overall architecture diagram of the information retrieval method provided in the embodiment of the present application. DETAILED DESCRIPTION

[0028] 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.

[0029] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is 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. Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meaning as those commonly understood by those skilled in the art of the technical field of the embodiments of the present application. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0030] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0031] Before describing the information retrieval method provided in the embodiment of the present application, the professional terms involved in the embodiment of the present application are first described:

[0032] 1) Query Domain Specific Language (Query DSL): QueryDSL provides a set of syntax and keywords for describing various parts of a query, such as matching conditions, filtering conditions, sorting rules, etc. The goal of Query DSL is to express query logic in a structured and readable way, making it easier to build and understand queries.

[0033] 2) Large models: Large models refer to large-scale machine learning models trained based on massive data, especially pre-trained language models in the field of natural language processing; large models usually contain hundreds of millions to hundreds of billions of parameters and can learn rich language knowledge and common sense from large amounts of text data; representative large models include the BERT series, T5, etc.; large models have powerful language understanding and generation capabilities and can be applied to various natural language processing tasks and multimodal tasks.

[0034] 3) Boolean (bool) condition: It is a commonly used query condition used to combine multiple subqueries or filter conditions; bool conditions include the following logical operations:

[0035] Must: The document must match the condition, which is equivalent to a logical AND.

[0036] Should: The document can match this condition, equivalent to a logical OR.

[0037] Must not (must_not): The document must not match the condition, which is equivalent to logical NOT.

[0038] Filter condition: The document must match this condition, but it does not affect the relevance score. It is often used in filtering conditions.

[0039] By combining different bool conditions, you can build complex query logic to meet diverse retrieval needs.

[0040] 4) Sort clause: The sort clause is used to specify the sorting method of the search results; the sort clause allows the results to be sorted according to one or more fields; common sort clauses include relevance score, timeliness, etc. The sort clause also supports setting sorting priority and multi-level sorting. By setting the sort clause reasonably, the search results can be sorted according to business needs, improving the availability of search results and user satisfaction.

[0041] 5) Intent recognition: Intent recognition is a key task in natural language understanding. Intent recognition aims to identify the user's intention or purpose from the user's input text. The goal of intent recognition is to map the user's natural language expression to predefined intent categories so that the system can understand the user's needs and provide corresponding services.

[0042] 6) Codon resolution: Codon resolution is the process of identifying and resolving the entity or concept referred to by a pronoun or other pronoun in context.

[0043] 7) Query statement: Query is a query statement or keyword submitted by the user to the information retrieval system, which expresses the user's information needs; the user describes the content he wants to find through the Query, such as questions, topics, entities, etc.; the Query can be a simple keyword combination or a complex natural language question.

[0044] 8) Context: Context refers to the background information related to the current statement or interaction, including previous conversation history (search history), user's personal information, time and place, and other external factors. In natural language processing, context is crucial to understanding the true meaning of language; the same word or sentence may have completely different meanings in different contexts; considering contextual information can help the system understand the user's intentions more accurately and generate more coherent and relevant responses.

[0045] 9) Prompt: Prompt is a kind of prompt information used to guide large models to generate specific content; Prompt can contain task description, sample input and output, format description and other information to help the model understand task requirements and generation goals. By carefully designing prompts, you can control the style, theme, tone and other attributes of the content generated by the model, and achieve more flexible and controllable text generation.

[0046] 10) Term Frequency-Inverse Document Frequency (TF-IDF): TF-IDF is a statistical method used to evaluate the importance of a term to a document in a document set or a corpus. TF stands for term frequency, that is, the frequency of a term appearing in a document; IDF stands for inverse document frequency, which measures the scarcity of a term. Intuitively speaking, the higher the frequency of a term appearing in a document and the lower the frequency of its appearance in other documents in the entire corpus, the better the ability of the term to distinguish categories; TF-IDF is often used in fields such as text mining and information retrieval, such as keyword extraction and text similarity calculation.

[0047] 11) Best Match 25 (BM25): BM25 is a commonly used document scoring function. BM25 is used for relevance ranking in search engines. BM25 is based on the probabilistic retrieval model (BM, Best Match), which is an improvement on the classic TF-IDF model. BM25 introduces factors such as document length normalization and term saturation to better characterize the relevance between the query and the document. The BM25 score takes into account the frequency of occurrence of each word in the query in the document, the length of the document, the importance of the word, etc. BM25 is more complex and sophisticated than TF-IDF.

[0048] In order to better understand the information retrieval method provided in the embodiment of the present application, the information retrieval method in the related art is first described below.

[0049] In the related art, most information retrieval methods use traditional TF-IDF, BM25 or vector retrieval methods to perform information retrieval.

[0050] However, the above information retrieval methods have at least the following problems:

[0051] 1) Traditional keyword matching only considers the statistical information of terms in the document and cannot truly understand the semantics of the query statement and the document. Therefore, the relevance of the retrieval results is limited; 2) Keyword-based query conditions are usually relatively simple and it is difficult to express complex Boolean logic (such as must, should, must_not, etc.) and multi-dimensional filtering conditions (time, place, author, etc.); similarly, vector retrieval only considers the inner product of similarity and cannot handle complex bool queries; 3) The information retrieval method in the related art uses one or several sets of fixed coefficients for sorting, and cannot flexibly set weight coefficients such as relevance and timeliness, and has poor flexibility; 4) The information retrieval method in the related art cannot utilize contextual information, and cannot utilize contextual information such as the user's historical queries and browsing records, making it difficult to accurately grasp the user's query intentions.

[0052] Based on the problems existing in the related technology, the embodiment of the present application provides an information retrieval method, which performs intent recognition on query statements based on historical retrieval data, can understand the user's real needs, avoid shallow matching based only on keywords, and through the generated intent recognition results, can obtain retrieval conditions that are closer to user needs, thereby ensuring the accuracy of the retrieval results; moreover, the retrieval conditions can be generated according to different retrieval types and dimensions, ensuring that the system can simultaneously support comprehensive analysis of multiple dimensions; and, dynamic retrieval statements can be generated in real time according to the retrieval conditions and retrieval result sorting sentences to adapt to different retrieval scenarios, with higher flexibility; in addition, dynamic retrieval statements are generated through intent recognition results, and retrieval results are obtained by executing the dynamic retrieval statements, which can improve retrieval accuracy.

[0053] Specifically, an embodiment of the present application provides an information retrieval method, which first obtains a query statement of a target object and historical retrieval data of the target object; then, performs intent recognition on the query statement based on the historical retrieval data to obtain an intent recognition result; then, based on the intent recognition result, generates at least one retrieval condition and a retrieval result sorting sentence; wherein each retrieval condition corresponds to a retrieval type, and the retrieval conditions under the same retrieval type include at least one retrieval dimension; the retrieval result sorting sentence is used to specify the sorting method for multiple retrieved information during information retrieval; then, generates a dynamic retrieval statement through at least one retrieval condition and a retrieval result sorting sentence; finally, performs information retrieval by executing the dynamic retrieval statement to obtain an information retrieval result that matches the query statement.

[0054] Here, firstly, an exemplary application of the information retrieval device of the embodiment of the present application is described, and the information retrieval device is an electronic device for implementing the information retrieval method. In one implementation, the information retrieval device (i.e., electronic device) provided in the embodiment of the present application can be implemented as a terminal, or as a server. In one implementation, the electronic device provided in the embodiment of the present application can be implemented as any terminal with information retrieval function, such as a laptop computer, a tablet computer, a desktop computer, an intelligent robot, etc.; in another implementation, the information retrieval device provided in the embodiment of the present application can also be implemented as a server, wherein the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiment of the present application. Below, an exemplary application of the information retrieval device when it is implemented as a server will be described.

[0055] See also Figure 1 , Figure 1 It is an optional architecture diagram of the information retrieval system provided in the embodiment of the present application. The information retrieval system 10 in the embodiment of the present application at least includes a terminal 100, a network 200 and a server 300. An information retrieval application is deployed on the terminal 100, and the server 300 can be a background server of the information retrieval application. The server 300 can constitute the information retrieval device of the embodiment of the present application, that is, the information retrieval method of the embodiment of the present application is implemented through the server 300. The terminal 100 is connected to the server 300 through the network 200, and the network 200 can be a wide area network or a local area network, or a combination of the two.

[0056] See also Figure 1, the terminal 100 receives the information retrieval operation of the user, and the terminal 100 generates an information retrieval request in response to the information retrieval operation; then, the terminal 100 sends the information retrieval request to the server 300 through the network 200; after receiving the information retrieval request, the server 300 responds to the information retrieval request and obtains the query statement of the target object and the historical retrieval data of the target object; then, the server 300 performs intent recognition on the query statement based on the historical retrieval data to obtain the intent recognition result; then, the server 300 generates at least one retrieval condition and a retrieval result sorting sentence based on the intent recognition result; wherein each retrieval condition corresponds to a retrieval type, and the retrieval condition under the same retrieval type includes at least one retrieval dimension; the retrieval result sorting sentence is used to specify the sorting method of multiple retrieved information during information retrieval; then, the server 300 generates a dynamic retrieval sentence through at least one retrieval condition and a retrieval result sorting sentence; then, the server 300 performs information retrieval by executing the dynamic retrieval sentence to obtain the information retrieval result matching the query statement; finally, the server 300 sends the information retrieval result to the terminal 100 through the network 200, and displays the information retrieval result on the display interface of the terminal 100.

[0057] In some embodiments, the above-mentioned information retrieval method can also be executed by a terminal, that is, after receiving the user's information retrieval operation, the terminal 100 can obtain the query statement of the target object and the historical retrieval data of the target object; then, the terminal 100 performs intent recognition on the query statement based on the historical retrieval data to obtain the intent recognition result; then, the terminal 100 generates at least one retrieval condition and a retrieval result sorting sentence based on the intent recognition result; wherein each retrieval condition corresponds to a retrieval type, and the retrieval conditions under the same retrieval type include at least one retrieval dimension; the retrieval result sorting sentence is used to specify the sorting method for multiple retrieved information during information retrieval; then, the terminal 100 generates a dynamic retrieval sentence through at least one retrieval condition and a retrieval result sorting sentence; then, the terminal 100 performs information retrieval by executing the dynamic retrieval sentence to obtain information retrieval results that match the query statement; finally, the terminal 100 displays the information retrieval results on the display interface.

[0058] Figure 2 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, Figure 2The electronic device shown may be an information retrieval device, which includes: at least one processor 210, a memory 250, at least one network interface 220, and a user interface 230. The various components in the information retrieval device are coupled together via a bus system 240. It is understood that the bus system 240 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 240 is not described in detail. Figure 2 Various buses are labeled as bus system 240 .

[0059] The processor 210 can be an integrated circuit chip with 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., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0060] The user interface 230 includes one or more output devices 231 that enable presentation of media content, and one or more input devices 232 .

[0061] The memory 250 may be removable, non-removable or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drive, optical disk drive, etc. The memory 250 may optionally include one or more storage devices physically away from the processor 210. The memory 250 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 250 described in the embodiment of the present application is intended to include any suitable type of memory. In some embodiments, the memory 250 can store data to support various operations, and examples of these data include programs, modules, and data structures or subsets or supersets thereof, as exemplarily described below.

[0062] An operating system 251 includes 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 businesses and processing hardware-based tasks; a network communication module 252 is 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.; an input processing module 253 is used to detect one or more user inputs or interactions from one of the one or more input devices 232 and translate the detected inputs or interactions.

[0063] In some embodiments, the device provided in the embodiments of the present application can be implemented in software. Figure 2 An information retrieval device 254 stored in the memory 250 is shown. The information retrieval device 254 may be an information retrieval device in an electronic device, which may be software in the form of a program or a plug-in, and includes the following software modules: an acquisition module 2541, an intention recognition module 2542, a first generation module 2543, a second generation module 2544, and a retrieval module 2545. These modules are logical, and therefore may be arbitrarily combined or further split according to the functions implemented. The functions of each module will be described below.

[0064] In some embodiments, the device provided in the embodiments of the present application can be implemented in hardware. As an example, the device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the information retrieval method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs) or other electronic components.

[0065] The information retrieval method provided in each embodiment of the present application can be executed by an electronic device, wherein the electronic device can be a server or a terminal, that is, the information retrieval method provided in each embodiment of the present application can be executed by a server, can be executed by a terminal, or can be executed through interaction between a server and a terminal.

[0066] Figure 3This is an optional flow chart of the information retrieval method provided in the embodiment of the present application. Figure 3 The steps shown are explained as Figure 3 As shown, the information retrieval method is described by taking the execution subject of the server as an example, and the method includes the following steps S101 to S105:

[0067] Step S101, obtaining a query statement of a target object and historical search data of the target object.

[0068] Here, the target object refers to the user who performs information retrieval operations; the query statement is a request expressed in natural language form or a specific query language, which is used to retrieve the information that the target object wants to retrieve in a database or information system; wherein, the query statement can be a statement manually written by the user according to the needs, or a statement automatically constructed according to the fields or conditions selected by the user.

[0069] Historical search data refers to the search instructions that have been executed by the target object before the current moment, the search data returned according to the search instructions, and the browsing history of the target object, etc. The specific historical search data is not limited in this application. Historical search data can be obtained in the log system. Each time a search is performed, the log system will record the time, query conditions and other data of each search. By accessing the log system, historical search data can be directly obtained; historical search data can also be queried in the preset database. After each search operation is completed, the query system will store the search records and search results in the database. According to the object identifier of the target object, the corresponding historical search data can be found in the preset database.

[0070] Step S102, performing intent recognition on the query statement based on historical retrieval data to obtain an intent recognition result.

[0071] Intent recognition is a key task in natural language processing, which aims to analyze user input and infer the user's real purpose or goal (such as searching for information, purchasing goods, obtaining recommendations, etc.). Intent recognition can be implemented through traditional machine learning algorithms or deep learning algorithms. Traditional machine learning algorithms can be implemented as logistic regression algorithms, support vector machines, and decision trees, etc. Deep learning algorithms can be implemented as convolutional neural networks, large language models, etc. The specific implementation methods are not limited in this application.

[0072] The intent recognition result refers to the final output of the intent recognition operation, which is usually expressed in the form of a classification label (such as "information query" or "purchase intention") or a more detailed semantic structure (such as "query weather information: city = City A, date = tomorrow"); the intent recognition result is used to guide subsequent operation steps, such as executing retrieval, content recommendation, etc.

[0073] In some embodiments, step S102 can be implemented by the following method: first, preprocessing the historical retrieval data and the query statement, the preprocessing operation includes removing noise through a filter, removing meaningless stop words, punctuation marks or special characters through a regular expression, etc., and the specific preprocessing operation and preprocessing method are not limited in this application; then, feature extraction is performed on the historical retrieval data and the query statement through a pre-trained feature extraction model (such as BERT), and the historical retrieval data and the query statement are converted into semantic vectors; then, the semantic vector of the historical retrieval data and the semantic vector of the query statement are vector-concatenated to obtain a feature concatenation vector, and the feature concatenation vector is used as the input of the large language model; then, semantic recognition is performed on the feature concatenation vector through the semantic recognition module in the large language model to obtain semantic recognition features; next, the reference resolution feature between the demonstrative word and the entity in the query statement is determined through the reference resolution module in the large language model; then, feature mapping is performed on the semantic recognition feature and the reference resolution feature through the intention recognition module to obtain the intention recognition probability; finally, the intention recognition result is determined according to the intention recognition probability.

[0074] Here, the semantic vector refers to the numerical representation of natural language text; feature splicing refers to combining different feature vectors together to form a comprehensive feature representation. Common splicing methods include simple splicing (horizontal splicing or vertical splicing), weighted average splicing, etc. The splicing method is not limited in this application; taking horizontal splicing as an example, assuming that the first feature vector is [1, 2, 5, 4] and the second feature vector is [3, 7, 6], the feature vector obtained after splicing is [1, 2, 5, 4, 3, 7, 6].

[0075] The semantic recognition module is a submodule in the large language model, which is used to understand and extract semantic information in query statements; semantic information refers to the actual meaning or concept expressed in data, language or signals, which is used to describe the attributes, behaviors or relationships between things.

[0076] The coreference resolution module is a submodule in the large language model, which is used to determine the entity that the pronoun or pronoun in the query actually refers to. The coreference resolution feature refers to the association feature between the pronoun or pronoun and the object it refers to (such as an entity, concept or event) in the natural language processing task.

[0077] The intent recognition module is a sub-module in the large language model. The intent recognition module is a classification model used to determine the intent of the user's query statement. The output of the intent recognition module is a probability value, namely the intent recognition probability, which indicates the probability that the query statement belongs to a predefined intent category.

[0078] It should be noted that the semantic recognition module, reference resolution module and intention recognition module can all be implemented as deep learning networks, such as convolutional neural networks, BERT, long short-term memory networks, etc. The specific model is not limited in this application.

[0079] Step S103, based on the intent recognition result, generate at least one search condition and a search result sorting clause; wherein each search condition corresponds to a search type, and the search conditions under the same search type include at least one search dimension; the search result sorting clause is used to specify the sorting method for multiple retrieved information during information retrieval.

[0080] Here, the search conditions refer to the specific parameter settings used by the search system to filter relevant data, which are used to limit the scope and results of the search; the search result sorting clause is a statement that specifies the search result sorting rules, which is used to control the display order of multiple search information, for example: sorting by time, the latest published information will be displayed first.

[0081] The search type refers to the execution logic type corresponding to the search condition, including the condition type that should be met, the condition type that must be met, the condition type that must not be met, and the filter condition type; among them, the condition type that should be met refers to the condition that can be matched or not matched during the search process, that is, it is not a necessary condition; the condition type that must be met refers to the constraint condition that must be matched during the search process, and the search must be performed according to the constraint condition; the condition type that must not be met refers to the condition that must not be matched during the search process; the filter condition type means that the condition must be matched, but the condition will not affect the final search score, and is often used to filter certain conditions. For example, if the user's intention is "find data related to machine learning", the condition of the condition type that should be met can be "book name = deep learning"; the condition of the condition type that must be met can be "book name = machine learning"; the condition of the condition type that must not be met can be "book name = food"; the condition of the filter condition type can be "price range = 30-100".

[0082] Search dimensions refer to specific parameters in the search conditions, which are used to describe the detailed scope of the conditions, such as keywords, location, time, etc.

[0083] In some embodiments, step S103 can be implemented by the following method: first, perform keyword extraction on the intent recognition result to determine the keywords, qualifiers and entity information in the intent recognition result, and then generate search conditions based on the keywords; then, perform regression processing on the user intent to obtain relevance weights and timeliness weights, wherein regression processing refers to the process of predicting the weights of different sorts based on user intent; the relevance weight represents the possibility of sorting by relevance, and the timeliness weight is used to represent the possibility of sorting by time; finally, the weight description sentence corresponding to the relevance weight and the weight description sentence of the timeliness weight are determined as the search result sorting sentence; wherein the weight description sentence refers to an expression that can be executed by a search engine or database to explain the corresponding sorting method.

[0084] Step S104: Generate a dynamic search statement through at least one search condition and a search result sorting clause.

[0085] Here, the dynamic search statement refers to a search statement generated in real time based on the search conditions and search result sorting words generated by the user intent recognition results. The search conditions and search result sorting words in the dynamic search statement are not fixed, but can be flexibly adjusted according to actual conditions.

[0086] In some embodiments, step S104 can be implemented by the following method: first, each search condition is converted into a corresponding conditional query statement, where the conditional query statement refers to a query expression executable by the search engine or database after the search condition is converted; at the same time, the search result sorting words are identified, and the fields related to relevance and timeliness are determined, and the relevance weight and timeliness weight are extracted from the relevant fields, and a sorting function is constructed based on the relevance weight and timeliness weight, wherein the sorting function is a function that integrates multiple sorting weights and is used to calculate the comprehensive sorting score of the search results; finally, the conditional query statement and the sorting function are added to the field corresponding to the initial search statement to obtain a dynamic search statement, wherein the initial search statement is a search statement containing the most basic query conditions and logic.

[0087] Step S105, performing information retrieval by executing a dynamic search statement to obtain information retrieval results matching the query statement.

[0088] In some embodiments, step S105 can be implemented by the following method: first, execute each conditional query statement in the dynamic search statement, perform information search in the preset information library, and obtain at least one candidate information and the relevance value and time sensitivity value between the candidate information and the dynamic search statement; then, for each candidate information, call the sorting function to perform weighted summation on the corresponding relevance value and time sensitivity value to obtain the retrieval score of the candidate information; finally, determine the target information that matches the query statement in the preset information library according to the retrieval score, and determine the target information as the information retrieval result.

[0089] Here, the preset information base refers to a pre-set data set for storing structured and unstructured information; the relevance value represents the numerical value of the matching degree between each candidate information in the preset information base and the dynamic search statement. Specifically, the relevance value can be calculated by a preset text retrieval algorithm (such as BM25); the timeliness sensitivity value represents the release time of each candidate information in the preset information base. Specifically, the release time of each candidate information in the preset information base is obtained as the timeliness sensitivity value.

[0090] The information retrieval method of the embodiment of the present application performs intent recognition on query statements based on historical retrieval data, can understand the user's real needs, avoid shallow matching based only on keywords, and through the generated intent recognition results, can obtain retrieval conditions that are closer to user needs, thereby ensuring the accuracy of the retrieval results; moreover, the retrieval conditions can be organized according to different retrieval types and dimensions, ensuring that the system can simultaneously support comprehensive analysis of multiple dimensions; and, dynamic retrieval statements can be generated in real time according to the retrieval conditions and retrieval result sorting sentences to adapt to different retrieval scenarios, with higher flexibility; in addition, dynamic retrieval statements are generated through intent recognition results, and retrieval results are obtained by executing the dynamic retrieval statements, which can improve retrieval accuracy.

[0091] The following is an example of an application scenario of the information retrieval method provided in the embodiment of the present application. The embodiment of the present application can be applied to at least the following exemplary scenarios:

[0092] Scenario 1: In an e-commerce scenario, a user enters product information on a search page of a shopping software. The shopping software system needs to search for corresponding products based on the product information entered by the user. In order to ensure that the searched products are more in line with the needs of the user, the information retrieval method provided in the embodiment of the present application can be used. First, the user enters a query statement (product information) into the shopping software of the terminal. The terminal generates an information retrieval request based on the received query statement and the user's historical retrieval data stored in the database, and uploads the information retrieval request to the server. After obtaining the information retrieval request, the server parses the request to obtain the query statement and the historical retrieval data; then, the query statement is intentionally identified based on the historical retrieval data to obtain the intention identification result; then, based on the intention identification result, at least one search condition and a search result sorting clause are generated; wherein each search condition corresponds to a search type, and the search conditions under the same search type include at least one search dimension; the search result sorting clause is used to specify the sorting method of multiple retrieved information during information retrieval; then, a dynamic search statement is generated through at least one search condition and a search result sorting clause; finally, information retrieval is performed by executing the dynamic search statement to obtain an information retrieval result matching the query statement. Through the information retrieval results, users can prioritize products that meet their needs.

[0093] Scenario 2: In the financial field, investors need to search for investment information. In order to enable investors to retrieve financial investment information that meets their needs, the information retrieval method provided by the embodiment of the present application can be used. First, the investor inputs the query statement (financial investment information) into the financial investment software of the terminal. The terminal generates an information retrieval request based on the received query statement and the historical retrieval data of the investor stored in the database, and uploads the information retrieval request to the server. After obtaining the information retrieval request, the server parses the request to obtain the query statement and the historical retrieval data; then, the query statement is intentionally identified based on the historical retrieval data to obtain the intention identification result; then, based on the intention identification result, at least one search condition and a search result sorting sentence are generated; wherein each search condition corresponds to a search type, and the search condition under the same search type contains at least one search dimension; the search result sorting sentence is used to specify the sorting method of the retrieved multiple information during information retrieval; then, a dynamic search statement is generated through at least one search condition and a search result sorting sentence; finally, the information is retrieved by executing the dynamic search statement to obtain the information retrieval result matching the query statement. Through the information retrieval results, investors can view financial investment information that better meets their needs.

[0094] The information retrieval method of the embodiment of the present application will be described below using the above scenario 1 as an example. Figure 4is another optional flow chart of the information retrieval method provided in the embodiment of the present application, such as Figure 4 As shown, the method includes the following steps S201 to S210:

[0095] Step S201: The terminal receives an information search operation from a user.

[0096] Here, an information retrieval application may be running on the terminal, and the server constitutes a background server of the information retrieval application. The information retrieval operation may be a selection operation or an input operation inputted by a client of the information retrieval application running on the terminal, for example, the selection operation may select a query statement, or the input operation may be a user inputting a query statement on the client.

[0097] In some embodiments, the information retrieval application may provide an input interface or input box to allow the user to select or enter a query statement. The input interface may be in the form of a form, a text box, or a drop-down menu, and the specific form is not limited in this application. The user may select a query statement from a predetermined option, or manually enter a query statement.

[0098] Step S202: the terminal generates an information retrieval request in response to the information retrieval operation.

[0099] Here, after receiving the user's information retrieval operation, the terminal obtains the historical retrieval data of the target object from a preset database, and encapsulates the query statement selected or input by the user and the obtained historical retrieval data into the information retrieval request; wherein, the historical retrieval data refers to the retrieval instructions that have been executed on the target object before the current moment, the retrieval data returned according to the retrieval instructions, and the browsing history of the target object, etc.

[0100] In some embodiments, in order to ensure the security of the information retrieval request, authentication parameters are added to the information retrieval request. The authentication parameters are used to verify the legitimacy of the information retrieval request to ensure that only authorized users can access the information retrieval service and prevent access and abuse by unauthorized users. A common authentication method is to use an API key or token. An API key is a unique string of characters that identifies and verifies a user's identity. A token is a credential similar to an access token or authentication token that contains the user's identity information and permissions.

[0101] Step S203: the terminal sends an information retrieval request to the server.

[0102] In some embodiments, the terminal sends the packaged information retrieval request to the server and requests the server to perform an information retrieval operation, usually using protocols such as HTTP or Web Socket to send the information retrieval request.

[0103] Step S204: the server obtains the query statement of the target object and the historical search data of the target object in response to the information search request.

[0104] Here, after receiving the information retrieval request, the server will parse the information retrieval request. For example, for an HTTP request, the server can parse the request header and the request body. The request header is parsed to obtain relevant information of the request; the request body is parsed to obtain the main body data of the request, that is, the query statement of the target object and the historical retrieval data of the target object. The specific fields or parameters in the request body are parsed, and these fields or parameters contain the query statement of the target object and the historical retrieval data of the target object. A specific data format, such as JSON or XML, is extracted from the request body, and then the data format is parsed to obtain the query statement of the target object and the historical retrieval data of the target object.

[0105] Step S205: The server performs intent recognition on the query statement based on the historical retrieval data to obtain an intent recognition result.

[0106] A query statement refers to the natural language text that a user enters into a search engine or other query system to request specific information or perform specific operations; intent recognition refers to identifying the real purpose or need of the user entering a query statement.

[0107] In some embodiments, see Figure 5 , Figure 5 is a flow chart of performing intent recognition on a query statement based on historical search data to obtain an intent recognition result provided by an embodiment of the present application, Figure 5 It is shown that in step S205, the server performs intent recognition on the query statement based on the historical search data to obtain the intent recognition result, which can be achieved by the following steps S2051 to S2054:

[0108] Step S2051, extracting features from historical search data and query statements to obtain historical search data features and query statement features.

[0109] Feature extraction refers to the process of extracting information or attributes from raw data that are useful for subsequent analysis, classification, or prediction tasks. Feature extraction can be performed on data using a machine learning algorithm or a deep learning algorithm. The specific extraction method is not limited in this application.

[0110] Historical retrieval data features refer to features extracted from historical retrieval data that can reflect information such as user search habits, interest preferences, etc.; historical retrieval data features may include statistical features (such as user click-through rate, dwell time, and other statistical information), text features (such as keywords and phrases in historical query statements, etc.) and behavioral features (such as click order, repeated queries, etc.). This application does not limit the specific historical feature data.

[0111] Query statement features refer to features extracted from the query statement that can reflect information such as user intent and query subject; query statement features may include word frequency features, semantic features, etc., which are not limited in this application.

[0112] Step S2052, concatenate the historical search data features and the query statement features to obtain concatenated features.

[0113] Splicing refers to merging the features of historical retrieval data and query statement features into new features. Splicing can include horizontal splicing, vertical splicing, and weighted splicing. Taking horizontal splicing as an example, assuming that the features of historical retrieval data are [1, 2, 5, 4] and the features of query statements are [3, 7, 6], the spliced ​​features obtained after feature splicing are [1, 2, 5, 4, 3, 7, 6].

[0114] Step S2053, feature mapping is performed on the concatenated features through a large language model to obtain an intent recognition probability; the intent recognition probability is used to characterize the probability that the query statement belongs to a predefined intent category.

[0115] Feature mapping refers to the process of converting raw data from one feature space to another feature space. In the embodiment of the present application, feature mapping refers to using a large language model to convert concatenated features into a vector representation that can reflect the intent of the query statement.

[0116] In some embodiments, step S2053 can be implemented by the following method: first, the concatenated features are subjected to a first network processing through a large language model to obtain semantic recognition features of the query statement; then, the semantic recognition features are subjected to a second network processing to obtain coreference resolution features for characterizing the relationship between entities and demonstrative words in the query statement; finally, the semantic recognition features and the coreference resolution features are subjected to a third network processing to obtain an intent recognition probability.

[0117] Here, the first network processing refers to the semantic understanding of the concatenated features, and semantic understanding refers to allowing the large model to understand the specific meaning expressed by the concatenated features; specifically, the concatenated features can be feature encoded through the semantic understanding module in the large language model to extract the semantic information in the query statement and obtain the semantic recognition feature corresponding to the query statement; the semantic recognition feature is used to represent the feature vector of the semantic information in the query statement.

[0118] The second network processing refers to the operation of reference resolution on the semantic recognition features. The reference resolution operation is the key in natural language processing tasks and is used to determine the specific entities and concepts pointed to by pronouns or other demonstrative words. The reference resolution operation can be implemented by the following methods: first, according to the named entity recognition or part-of-speech tagging in the query statement and the historical retrieval data, the entities existing in the query statement and the historical retrieval data are determined; then, the position of the demonstrative words in the query statement is determined; then, by calculating the semantic similarity and context consistency between the pronouns and the entities, the specific entity pointed to by the demonstrative words is determined, and the relationship between the demonstrative words and the corresponding specific entities is determined as the reference resolution feature. For example, the query statement is "Where did Xiao Ming put the book, it looks good", then, it is determined that the entity words in the query statement are "Xiao Ming" and "book", and the demonstrative word is "it". By analyzing the context, the semantic similarity or context consistency between the demonstrative words and the entity words is determined, and it is determined that the entity that the demonstrative word "it" is more likely to refer to is "book".

[0119] The third network processing refers to performing intent recognition based on semantic recognition features and reference resolution features to obtain the intent recognition probability; it can be determined by the intent recognition module in the large language model, and the intent recognition module can be a pre-trained classification model. Through the pre-trained classification model, the probability that the user's intent belongs to a predefined intent category is output. The specific model is not limited in this application.

[0120] Step S2054, determining the intention recognition result based on the intention recognition probability.

[0121] In some embodiments, the intent recognition result refers to the final intent category selected according to the intent recognition probability, which is the model prediction result of the user's true intent. The intent with the highest probability can be directly selected as the intent recognition result according to the intent recognition probability.

[0122] In some embodiments, after acquiring the historical search data of the target object, search prompt information may be generated based on the historical search data. Specifically, first, data analysis is performed on the historical search data to obtain object features of the target object; then, the object features are encapsulated into search prompt information.

[0123] Correspondingly, after the search prompt information is generated, step S205 can also be implemented by the following method: performing intent recognition on the query statement based on the search prompt information to obtain an intent recognition result.

[0124] Here, the retrieval prompt information refers to prompt information generated based on object features and representing the user's historical query records, browsing behaviors or preference settings, which is used to optimize the intent recognition results; the retrieval prompt information is a clearer and more concise guidance information obtained by analyzing and extracting historical retrieval data, which is easier to understand and apply the large language model. The specific process of intent recognition is referred to steps S2051 to S2054, and this application will not repeat it here.

[0125] In some embodiments, after obtaining the query statement of the target object, the query statement can also be interpreted in combination with the knowledge graph. Specifically: first, based on the pre-built knowledge graph, the lexical interpretation of the target vocabulary in the query statement is determined; then, the lexical interpretation of the target vocabulary is added to the query statement to obtain an interpreted query statement.

[0126] Correspondingly, after obtaining the interpreted query statement, step S205 can also be implemented by the following method: performing intent recognition on the interpreted query statement based on historical retrieval data to obtain an intent recognition result.

[0127] Here, the knowledge graph is a structured way of storing knowledge. The knowledge graph represents entities, concepts and the relationships between them in the form of a graph. In the knowledge graph, nodes usually represent entities or concepts, while edges represent the relationships between entities or concepts.

[0128] The target vocabulary refers to the vocabulary in the query statement that needs further interpretation, usually a term, abbreviation or specific concept, etc. For example, if the query statement is "the interface of the authentication API", the "API" in the query statement can be interpreted through the pre-built knowledge graph, and the interpretation can be added to the query statement to obtain the interpreted query statement "the interface of the authentication application programming interface". After obtaining the interpreted query statement, the intent recognition can be performed based on the interpreted query statement and the historical search data. The specific process of intent recognition is shown in steps S2051 to S2054, and this application will not repeat it here.

[0129] In step S206, the server generates at least one search condition and a search result sorting clause based on the intent recognition result; wherein each search condition corresponds to a search type, and the search conditions under the same search type include at least one search dimension; the search result sorting clause is used to specify the sorting method for multiple retrieved information during information retrieval.

[0130] Here, the search conditions refer to the standards or rules used to filter information in the information retrieval system; the sorting clauses refer to the rules used to specify how to sort the multiple retrieved information after the information retrieval is completed, for example, the sorting can be based on the rules of relevance, timeliness or scoring.

[0131] In some embodiments, the intent recognition result includes intent recognition probability and metadata information; at this time, step S206 can be implemented by the following method: first, based on the metadata information and intent recognition probability in the intent recognition result, generate at least one retrieval condition; wherein the retrieval type corresponding to the retrieval condition includes at least one of the following: should meet the condition type, must meet the condition type, must not meet the condition type and filtering condition type; then, based on the intent recognition probability in the intent recognition result, generate the retrieval result sorting sentence.

[0132] Here, metadata information refers to the information contained in the intent recognition results other than the intent recognition probability, which may include time, place, person, etc. The specific information is not limited in this application.

[0133] The when-satisfied condition type refers to a condition that may or may not be matched during the search process; the must-satisfied condition type refers to a constraint condition that must be matched during the search process, and the search must be performed based on the constraint condition; the must-not-satisfied condition type refers to a condition that must not be matched during the search process; the filtering condition type means that the condition must be matched, but the condition will not affect the final search score, and is often used to filter certain conditions.

[0134] When generating retrieval conditions based on metadata information and intent recognition probability in intent recognition results, first, determine user intent in predefined intent categories based on intent recognition probability; then, determine the correlation between metadata in intent recognition results and user intent by calculating text similarity; finally, generate retrieval conditions based on the calculated correlation and a preset correlation threshold, such as: the correlation threshold is 0.8, where the correlations between location data and time data and user intent are 0.85 and 0.9 respectively, then the generated retrieval conditions that must meet the condition type include location data and time data.

[0135] In some embodiments, the retrieval result sorting sentence is used to specify a sorting method including at least one of the following: a relevance sorting method and a timeliness sorting method; wherein, the relevance sorting method is characterized by a relevance weight, and the timeliness sorting method is characterized by a timeliness weight; at this time, based on the intent recognition probability in the intent recognition result, the retrieval result sorting sentence is generated, which can be achieved by the following method: first, based on the intent recognition probability in the intent recognition result, the intent category corresponding to the query statement is determined; then, the intent category is regressed through a regression model to obtain the relevance weight and the timeliness weight; finally, the weight description sentences corresponding to the relevance weight and the timeliness weight are determined as the retrieval result sorting sentence.

[0136] Here, the output of the regression model is a floating point number between 0 and 1. The intent category is input into the regression model for regression processing to obtain a set of floating point numbers [w1, w2], where w1 represents the relevance weight and w2 represents the timeliness weight. The relevance weight represents the relevance sorting method, and the timeliness weight represents the timeliness sorting method. Finally, the relevance sorting method represented by the relevance weight and the timeliness sorting method represented by the timeliness weight are converted into expressions that can be executed by the retrieval engine or database as retrieval sorting terms.

[0137] Step S207: the server generates a dynamic search statement through at least one search condition and a search result sorting clause.

[0138] Here, the server generates a dynamic search statement by combining the search dimensions specified by the set search conditions and the sorting method determined by the search result sorting words.

[0139] In some embodiments, see Figure 6 , Figure 6 It is a flowchart of generating a dynamic search statement through at least one search condition and a search result sorting clause provided by an embodiment of the present application; Figure 6 It is shown that in step S207, the server generates a dynamic search statement through at least one search condition and a search result sorting clause, which can be implemented through steps S2071 to S2074:

[0140] Step S2071, convert each search condition into a corresponding conditional query statement to obtain at least one conditional query statement.

[0141] Here, the conditional query statement refers to a query expression executable by a search engine or database obtained after the search conditions are converted; wherein the expression of the conditional query statement is shown in Formula 1:

[0142] R={d|d∈D,d satisfies q} Formula 1

[0143] Wherein, R represents the set of search results that meet the search condition q; d represents the data in the database D.

[0144] Step S2072, parse the search result ranking words to obtain the relevance weight and timeliness weight.

[0145] Here, a preset dynamic search statement generation module is used to identify fields related to relevance and timeliness in the search result sorting sentences, and then the relevance weight and timeliness weight are extracted from the relevant fields; wherein, the preset dynamic search statement generation module is used to generate dynamic search statements, and the preset dynamic search statement generation module can be implemented as a deep learning model, such as a convolutional neural network, a recurrent neural network, etc., which is not limited in this application.

[0146] Step S2073, constructing a ranking function based on the relevance weight and the timeliness weight.

[0147] Here, the ranking function is a function that integrates multiple ranking weights and is used to calculate the comprehensive ranking score of the search results; the calculation formula of the ranking function is shown in Formula 2:

[0148] score(r) = f_s(r) Formula 2

[0149] Wherein, score(r) represents the sorting result obtained by sorting the data in the search result set R that meets the search condition q; f_s() represents the sorting function.

[0150] Step S2074: adding at least one conditional query statement and a sorting function to the dynamic search statement.

[0151] Here, the basic statement structure of the initial search statement is obtained, the fields corresponding to the conditional query statement and sorting function in the basic statement structure are determined, and at least one conditional query statement and sorting function are added to the corresponding fields to obtain a dynamic search statement; wherein the initial search statement is a search statement containing the most basic query conditions and logic.

[0152] Step S208: The server performs information retrieval by executing the dynamic retrieval statement to obtain information retrieval results that match the query statement.

[0153] Here, the server executes dynamic search statements to filter out information retrieval results that match the query statements from the database or information system.

[0154] In some embodiments, see Figure 7 , Figure 7 It is a flow chart of obtaining information retrieval results matching a query statement by executing a dynamic search statement to retrieve information provided by an embodiment of the present application; Figure 7 It is shown that in step S208, the server performs information retrieval by executing a dynamic search statement to obtain information retrieval results matching the query statement, which can be achieved through steps S2081 to S2084:

[0155] Step S2081, performing information retrieval by executing each conditional query statement in the dynamic search statement, and obtaining the relevance value and time sensitivity value between each candidate information in the preset information library and the dynamic search statement.

[0156] Here, the relevance value represents the numerical value of the matching degree between each candidate information in the preset information base and the dynamic search statement; the time sensitivity value represents the release time of each candidate information in the preset information base.

[0157] In some embodiments, step S2081 can be implemented by the following method: first, execute the dynamic search statement through the preset execution retrieval module; then, calculate the relevance value between each candidate information in the preset information library and the dynamic search statement according to the preset text retrieval algorithm (such as BM25, vector space model), etc.; at the same time, obtain the release time of each candidate information in the preset information library as the timeliness sensitivity value.

[0158] Step S2082: for each candidate information, call the ranking function, perform a fusion calculation on the relevance value and the timeliness sensitivity value, and obtain the retrieval score of the candidate information.

[0159] Here, fusion calculation refers to weighted summation of relevance value and time sensitivity value. The calculation formula of fusion calculation is shown in Formula 3:

[0160] S=w1*relevant_score+w2*time_sensitivity Formula 3

[0161] Among them, S represents the retrieval score of candidate information; w1 represents the relevance weight; w2 represents the timeliness weight; relevant_score represents the relevance value, and time_sensitivity represents the timeliness sensitivity value.

[0162] Step S2083, based on the search score, determine the target information matching the query statement from the preset information library.

[0163] Here, based on the retrieval score, the candidate information with the highest retrieval score can be selected from the preset information library as the target information matching the query statement; or, based on the set retrieval score threshold, at least one candidate information with a retrieval score greater than the retrieval score threshold is screened out, and the information is sorted from high to low according to the retrieval score, and the sorting result is determined as the target information matching the query statement.

[0164] Step S2084, determining the target information as the information retrieval result.

[0165] Step S209: the server sends the information retrieval result to the terminal.

[0166] Step S210: the terminal displays the information retrieval result on the current interface.

[0167] After receiving the information retrieval results, users can perform further processing based on the information retrieval results, such as browsing and filtering, purchasing, saving, etc.

[0168] The information retrieval method of the embodiment of the present application performs intent recognition on query statements based on historical retrieval data, can understand the user's real needs, avoid shallow matching based only on keywords, and through the generated intent recognition results, can obtain retrieval conditions that are closer to user needs, thereby ensuring the accuracy of the retrieval results; and the retrieval conditions can be organized according to different retrieval types and dimensions, ensuring that the system can simultaneously support comprehensive analysis of multiple dimensions; and, dynamic retrieval statements can be generated in real time according to the retrieval conditions and retrieval result sorting sentences to adapt to different retrieval scenarios, with higher flexibility; in addition, dynamic retrieval statements are generated through intent recognition results, and retrieval results are obtained by executing the dynamic retrieval statements, which can improve retrieval accuracy.

[0169] The following is an explanation of an exemplary application of the embodiments of the present application in a practical application scenario.

[0170] The information retrieval method provided in the embodiment of the present application, in addition to the query statement, also needs to combine the user's contextual information (such as historical queries, browsing records, etc.) and the content of multiple rounds of retrieval (the above-mentioned historical retrieval data) to more comprehensively understand the user's retrieval needs and intentions, and identify the query keywords, entities, intentions and other information. Moreover, the information retrieval method provided in the embodiment of the present application dynamically constructs Boolean query conditions (the above-mentioned retrieval conditions) based on the semantic understanding of the query statement, including must (the above-mentioned must meet the condition type), should (the above-mentioned should meet the condition type), must_not (the above-mentioned must not meet the condition type) and filter (the above-mentioned filter condition type). Boolean query conditions can cover multiple dimensions, such as keywords, locations, times, author levels, etc.; at the same time, according to the query intent and context, the weight coefficients of relevance and timeliness in the sorting words sort (the above-mentioned search result sorting words) are dynamically adjusted.

[0171] The embodiment of the present application provides a personalized and flexible information retrieval method; fully utilize the natural language understanding ability of the large model, combine context information and multiple rounds of retrieval content (the above historical retrieval data), accurately identify user intent, and generate customized query DSL (the above dynamic retrieval statement) according to different query statements to improve the relevance and accuracy of the retrieval results. For example, "Today's News", the "Today" in it means that the user requires high timeliness, so time filtering should be performed during the retrieval; in addition, for news and information retrieval, the timeliness coefficient generally needs to be increased.

[0172] See also Figure 8 , Figure 8 is the overall architecture diagram of the information retrieval method provided in the embodiment of the present application; Figure 8 The information retrieval method provided in the embodiment of the present application is introduced.

[0173] First, the user inputs a query statement 801 through the query interface. After that, the query statement is input into the large model analysis module 802. The core of the large model analysis module 802 is a pre-trained large language model. The main task of the large model analysis module 802 is to perform semantic understanding, reference resolution and intent recognition on the query statement, obtain the intent recognition result, rewrite the query statement, generate metadata such as time, place, and person, and provide a basis for the subsequent generation of dynamic search statements.

[0174] Then, the analysis result of the large model analysis module 802 is input into the dynamic search statement generation module 803 to generate a dynamic search statement. The dynamic search statement usually includes two main parts: a Boolean condition (at least one of the above search conditions) and a sorting clause (the above search result sorting clause). The generation of the Boolean condition and the sorting clause is introduced below:

[0175] (1) Generate Boolean conditions: Boolean conditions can be generated based on the analysis results of large language models. The logical operations of Boolean conditions include should, must, must_not and filter, among which: should: means "should be satisfied", which is used to specify the keywords or conditions that should be included in the information retrieval results. must: means "must be satisfied", which is used to specify the keywords or conditions that must be included in the information retrieval results. must_not: means "must not be satisfied", which is used to specify the keywords or conditions that must not be included in the information retrieval results. filter: represents the filtering condition, which is used to filter the information retrieval results according to dimensions such as location, time, and author level.

[0176] (2) Generate ranking terms: Ranking terms can be generated based on user intent (the above-mentioned intent recognition results). Ranking terms are used to specify the ranking method of search results, including relevance and timeliness. Relevance refers to the degree of relevance between the search results and the query statement, which is usually calculated using algorithms such as TF-IDF and BM25. Timeliness refers to the temporal freshness of the search results. For queries with high timeliness requirements such as news and hot topics, a higher weight needs to be given.

[0177] In some embodiments, the generation of ranking words can be modeled as a regression problem (the output of the regression model is a floating point number between 0 and 1). Assuming that the semantic information obtained by the large model analysis module is x, the regression model outputs a set of floating point numbers [w1, w2], which respectively represent the coefficients of each factor of the ranking module; the calculation formula of the final retrieval score S is shown in the above formula 3. The large model needs to comprehensively consider factors such as the user's search intent and query semantics to determine which ranking strategy should be used; for example, if the user query contains words such as "latest" and "real-time", then the weight of timeliness ranking should be increased.

[0178] In addition, the large model can also dynamically adjust the sorting strategy based on the query context and historical search records (the above historical search data). For example, if the user's previous query prefers timeliness sorting, then the weight of timeliness sorting can be appropriately increased in this query.

[0179] Finally, the generated dynamic search statement is input into the search execution module 804, and the Boolean condition and the sorting clause are executed by the search execution module 804 to generate the final information search result 805. The task of the search execution module 804 is to search the database according to the generated dynamic search statement and return the result that meets the condition; the condition execution process is introduced below:

[0180] (1) Executing Boolean conditions: For the generated Boolean conditions, the Boolean conditions need to be converted into specific query statements and executed in the database. Different query conditions correspond to different query statements, such as term query, range query, fuzzy query, etc. In some embodiments, assuming that the generated Boolean condition is q and the document set in the database is D, the execution of the Boolean condition can be formalized as shown in the above formula 1.

[0181] (2) Execute the sorting clause: For the generated sorting clause, it is necessary to convert the sorting clause into a specific sorting function and sort the search results.

[0182] In some embodiments, assuming that the generated sorting clause is s and the search result set is R, a sorting function f_s can be defined and applied to each document d in the search result set R. The formula of the sorting function is shown in Formula 2 above to obtain the final sorting result.

[0183] It should be noted that the specific form of the sorting function f_s depends on the content of the sorting clause s; for example, if the sorting clause s specifies relevance sorting, then the sorting function f_s can be a relevance score calculated based on BM25 or a vector space model; if the sorting clause s specifies timeliness sorting, then the sorting function f_s can be a freshness score calculated based on the document release time.

[0184] In practical applications, a ranking clause usually contains multiple ranking conditions, each condition corresponds to a ranking function, and the final ranking score is a weighted combination of these functions. The weighted calculation formula is shown in Formula 4:

[0185] score(d) = Σ w_i * f_i(d) Formula 4

[0186] Among them, w_i represents the weight coefficient of each ranking function f_i, and w_i can be dynamically determined by the large model when generating ranking sentences.

[0187] The information retrieval method provided in the embodiment of the present application utilizes the powerful natural language understanding ability of the large model, deeply analyzes the query statement, combines the context information and multiple rounds of search content, accurately identifies the user's true intention, and clarifies the user's true intention through reference resolution and query statement rewriting. In addition, the information retrieval method provided in the embodiment of the present application can dynamically generate customized dynamic query statements for different query statements, including Boolean conditions and sorting clauses; wherein, the Boolean conditions can flexibly set filtering conditions of various dimensions, such as keywords, locations, time, author levels, etc.; sorting clauses can adjust the weight coefficients of relevance and timeliness according to needs. In this way, it can meet the diverse search needs of users and provide search results that are more in line with user intentions.

[0188] In addition, the information retrieval method provided in the embodiment of the present application can also construct the user's object characteristics by analyzing the user's historical query records, browsing behavior, preference settings and other data. Combined with the object characteristics, the large model can better understand the user's search intention and generate more personalized dynamic search statements.

[0189] The information retrieval method provided in the embodiment of the present application can also combine the big model with the knowledge graph. The knowledge graph contains rich structured knowledge, which can assist the big model to better understand the query semantics. For example, for the query sentence "XX2", the knowledge graph can tell the big model that this is the second season of a TV series. By combining the knowledge graph, the big model's ability to understand proper nouns, domain terms, etc. can be improved, thereby generating more accurate dynamic search sentences.

[0190] In addition, it should be noted that since the reasoning of large models is usually time-consuming and may affect the real-time response capability of the retrieval system, it is necessary to optimize the reasoning efficiency of large models. Specific optimization methods may include model compression, asynchronous reasoning, GPU acceleration and other technologies to ensure that the real-time generation of dynamic retrieval statements does not become a system bottleneck.

[0191] It is understandable that in the embodiments of the present application, if data related to user information or corporate information is involved, when the embodiments of the present application are applied to specific products or technologies, it is necessary to obtain user permission or consent, or to blur this information to eliminate the correspondence between this information and the user; and the relevant data collection and processing should be strictly in accordance with the requirements of relevant laws and regulations when applied in examples, and the informed consent or separate consent of the subject of personal information should be obtained, and subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the subject of personal information.

[0192] The following is a description of an exemplary structure of the information retrieval device 254 provided in the present application as a software module. In some embodiments, Figure 2As shown, the information retrieval device 254 includes: an acquisition module 2541, which is used to acquire the query statement of the target object and the historical retrieval data of the target object; an intention recognition module 2542, which is used to perform intention recognition on the query statement based on the historical retrieval data to obtain an intention recognition result; a first generation module 2543, which is used to generate at least one retrieval condition and a retrieval result sorting sentence based on the intention recognition result; wherein each retrieval condition corresponds to a retrieval type, and the retrieval conditions under the same retrieval type include at least one retrieval dimension; the retrieval result sorting sentence is used to specify the sorting method for multiple retrieved information during information retrieval; a second generation module 2544, which is used to generate a dynamic retrieval statement through the at least one retrieval condition and the retrieval result sorting sentence; a retrieval module 2545, which is used to perform information retrieval by executing the dynamic retrieval sentence to obtain an information retrieval result matching the query statement.

[0193] In some embodiments, the intention recognition module is also used to: extract features from the historical retrieval data and the query statement to obtain historical retrieval data features and query statement features; concatenate the historical retrieval data features and the query statement features to obtain concatenated features; perform feature mapping on the concatenated features through a large language model to obtain an intention recognition probability; the intention recognition probability is used to characterize the probability that the query statement belongs to a predefined intention category; and determine the intention recognition result based on the intention recognition probability.

[0194] In some embodiments, the intention recognition module is also used to: perform a first network processing on the concatenated features through a large language model to obtain semantic recognition features of the query statement; perform a second network processing on the semantic recognition features to obtain reference resolution features used to characterize the relationship between entities and demonstrative words in the query statement; perform a third network processing on the semantic recognition features and the reference resolution features to obtain the intention recognition probability.

[0195] In some embodiments, the intent recognition result includes the intent recognition probability and metadata information; the first generation module is also used to: generate the at least one retrieval condition based on the metadata information and the intent recognition probability in the intent recognition result; the retrieval type corresponding to the retrieval condition includes at least one of the following: a condition type that should be met, a condition type that must be met, a condition type that must not be met, and a filtering condition type; based on the intent recognition probability in the intent recognition result, generate the retrieval result sorting sentence.

[0196] In some embodiments, the retrieval result sorting sentence is used to specify a sorting method including at least one of the following: a relevance sorting method and a timeliness sorting method; the relevance sorting method is characterized by a relevance weight, and the timeliness sorting method is characterized by a timeliness weight; the first generation module is also used to: determine the intent category corresponding to the query statement based on the intent recognition probability in the intent recognition result; regress the intent category through a regression model to obtain a relevance weight and a timeliness weight; and determine the weight description sentences corresponding to the relevance weight and the timeliness weight as the retrieval result sorting sentence.

[0197] In some embodiments, the second generation module is also used to: convert each of the search conditions into a corresponding conditional query statement to obtain at least one conditional query statement; parse the search result sorting terms to obtain relevance weights and timeliness weights; construct a sorting function based on the relevance weights and the timeliness weights; and add the at least one conditional query statement and the sorting function to the dynamic search statement.

[0198] In some embodiments, the retrieval module is also used to: perform information retrieval by executing each conditional query statement in the dynamic retrieval statement to obtain the relevance value and time sensitivity value between each candidate information in the preset information library and the dynamic retrieval statement; for each of the candidate information, call the ranking function to fuse the relevance value and the time sensitivity value to obtain the retrieval score of the candidate information; based on the retrieval score, determine the target information that matches the query statement from the preset information library; and determine the target information as the information retrieval result.

[0199] In some embodiments, the device also includes an encapsulation module, which is used to perform data analysis on the historical retrieval data to obtain object features of the target object; and encapsulate the object features into a query statement; accordingly, the intention recognition module is also used to: perform intention recognition on the query statement based on the query statement to obtain the intention recognition result.

[0200] In some embodiments, the device also includes an adding module, which is used to determine the lexical interpretation of the target vocabulary in the query statement based on a pre-constructed knowledge graph; add the lexical interpretation of the target vocabulary to the query statement to obtain an interpreted query statement; accordingly, the intention recognition module is also used to: perform intention recognition on the interpreted query statement based on the historical retrieval data to obtain the intention recognition result.

[0201] It should be noted that the description of the device of the embodiment of the present application is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment, so it is not repeated. For technical details not disclosed in the embodiment of the device, please refer to the description of the method embodiment of the present application for understanding.

[0202] An embodiment of the present application provides an electronic device, comprising: a memory for storing computer-executable instructions or computer programs; and a processor for implementing the above-mentioned information retrieval method when executing the computer-executable instructions or computer programs stored in the memory.

[0203] An embodiment of the present application provides a computer program product, which includes computer-executable instructions or a computer program, and the computer-executable instructions or the computer program are stored in a computer-readable storage medium; wherein, when a processor of an electronic device reads the computer-executable instructions or the computer program from the computer-readable storage medium and executes the computer-executable instructions or the computer program, the above-mentioned information retrieval method is implemented.

[0204] The present application embodiment provides a computer-readable storage medium storing computer-executable instructions or computer programs. When the computer-executable instructions or computer programs are executed by a processor, the processor will be caused to execute the above-mentioned information retrieval method, for example, Figure 3 The method shown.

[0205] In some embodiments, the storage medium can be a computer-readable storage medium, for example, a ferroelectric random access memory (FRAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPR OM), a flash memory, a magnetic surface memory, an optical disk, or a compact disk read-only memory (CD-ROM), etc.; it can also be various devices including one or any combination of the above memories.

[0206] 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.

[0207] As an example, executable instructions may, but need not correspond to, a file in a file system, may be stored as part of a file storing other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions). As an example, executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed at multiple locations and interconnected by a communication network.

[0208] 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. An information retrieval method, characterized in that: The method comprises: Obtaining a query statement of a target object and historical search data of the target object; Performing intent recognition on the query statement based on the historical search data to obtain an intent recognition result; Based on the intention recognition result, at least one search condition and a search result sorting clause are generated; wherein each search condition corresponds to a search type, and the search conditions under the same search type include at least one search dimension; the search result sorting clause is used to specify the sorting method of the multiple retrieved information during information retrieval; Generate a dynamic search statement through the at least one search condition and the search result sorting clause; By executing the dynamic search statement to perform information search, an information search result matching the query statement is obtained.

2. The method according to claim 1, characterized in that The performing intent recognition on the query statement based on the historical search data to obtain an intent recognition result includes: Extracting features from the historical search data and the query statement to obtain historical search data features and query statement features; Splicing the historical search data features and the query statement features to obtain splicing features; Performing feature mapping on the concatenated features through a large language model to obtain an intent recognition probability; the intent recognition probability is used to characterize the probability that the query statement belongs to a predefined intent category; Based on the intention recognition probability, the intention recognition result is determined.

3. The method according to claim 2, characterized in that The step of performing feature mapping on the concatenated features through a large language model to obtain an intent recognition probability includes: Performing a first network processing on the concatenated features through a large language model to obtain semantic recognition features of the query statement; Performing a second network processing on the semantic recognition feature to obtain a coreference resolution feature for characterizing the relationship between the entity and the demonstrative word in the query sentence; The semantic recognition feature and the coreference resolution feature are processed by a third network to obtain the intention recognition probability.

4. The method according to claim 2, characterized in that: The intention recognition result includes the intention recognition probability and metadata information; The generating at least one search condition and a search result sorting clause based on the intention recognition result includes: generating the at least one search condition based on the metadata information and the intent recognition probability in the intent recognition result; The search type corresponding to the search condition includes at least one of the following: a condition type that should be satisfied, a condition type that must be satisfied, a condition type that must not be satisfied, and a filtering condition type; The search result ranking phrase is generated based on the intention recognition probability in the intention recognition result.

5. The method according to claim 4, characterized in that The sorting method specified by the search result sorting phrase includes at least one of the following: a relevance sorting method and a timeliness sorting method; the relevance sorting method is characterized by a relevance weight, and the timeliness sorting method is characterized by a timeliness weight; The step of generating the search result ranking phrase based on the intention recognition probability in the intention recognition result includes: Determining the intent category corresponding to the query statement based on the intent recognition probability in the intent recognition result; Regressing the intent categories through a regression model to obtain a relevance weight and a timeliness weight; The weight description sentences corresponding to the relevance weight and the timeliness weight are determined as the search result ranking sentences.

6. The method according to claim 1, characterized in that The step of generating a dynamic search statement by using the at least one search condition and the search result sorting clause includes: Convert each of the search conditions into a corresponding conditional query statement to obtain at least one conditional query statement; Parsing the search result ranking words to obtain relevance weights and timeliness weights; Constructing a ranking function based on the relevance weight and the timeliness weight; The at least one conditional query statement and the ranking function are added to the dynamic search statement.

7. The method according to claim 6, characterized in that The information retrieval is performed by executing the dynamic search statement to obtain information retrieval results matching the query statement, including: By executing each conditional query statement in the dynamic search statement to perform information retrieval, a relevance value and a time sensitivity value between each candidate information in the preset information database and the dynamic search statement are obtained; For each of the candidate information, calling the ranking function, performing a fusion calculation on the relevance value and the timeliness sensitivity value, and obtaining a retrieval score of the candidate information; Based on the search score, determining target information matching the query statement from the preset information library; The target information is determined as the information retrieval result.

8. The method according to any one of claims 1 to 7, characterized in that: After acquiring the historical search data of the target object, the method further includes: Performing data analysis on the historical search data to obtain object features of the target object; Encapsulating the object features into retrieval prompt information; Accordingly, performing intent recognition on the query statement based on the historical search data to obtain an intent recognition result includes: The query statement is subjected to intent recognition based on the search prompt information to obtain the intent recognition result.

9. The method according to any one of claims 1 to 7, characterized in that: After obtaining the query statement of the target object, the method further includes: Determine the vocabulary interpretation of the target vocabulary in the query sentence based on the pre-built knowledge graph; Adding the vocabulary interpretation of the target vocabulary to the query sentence to obtain an interpreted query sentence; Accordingly, performing intent recognition on the query statement based on the historical search data to obtain an intent recognition result includes: The intention recognition is performed on the interpreted query statement based on the historical retrieval data to obtain the intention recognition result.

10. An information retrieval device, characterized in that: The device comprises: An acquisition module, used to acquire a query statement of a target object and historical search data of the target object; An intention recognition module, used to perform intention recognition on the query statement based on the historical search data to obtain an intention recognition result; A first generation module is used to generate at least one search condition and a search result sorting sentence based on the intention recognition result; wherein each search condition corresponds to a search type, and the search conditions under the same search type include at least one search dimension; the search result sorting sentence is used to specify the sorting method of multiple retrieved information during information retrieval; A second generating module, used to generate a dynamic search statement through the at least one search condition and the search result sorting phrase; The retrieval module is used to perform information retrieval by executing the dynamic retrieval statement to obtain information retrieval results matching the query statement.

11. An electronic device, characterized in that: include: Memory for storing computer executable instructions or computer programs; A processor, configured to implement the information retrieval method according to any one of claims 1 to 9 when executing the computer executable instructions or computer programs stored in the memory.

12. A computer-readable storage medium, characterized in that: Computer executable instructions or computer programs are stored, which are used to cause a processor to execute the computer executable instructions or the computer program to implement the information retrieval method described in any one of claims 1 to 9.

13. A computer program product comprising computer executable instructions or a computer program, characterized in that: When the computer executable instructions or computer program are executed by a processor, the information retrieval method according to any one of claims 1 to 9 is implemented.

Images