A Pseudo-Document Generation Method and System Based on Multiple Models

Abstract

This disclosure provides a multi-model-based pseudo-document generation method and system, belonging to the field of data generation technology. The method includes: inputting query information into multiple language models to generate multiple pseudo-documents; decomposing each pseudo-document into multiple information bars based on its length, wherein the number of information bars is greater than or equal to the number of pseudo-documents; filtering the multiple information bars based on a relevance evaluator to obtain multiple relevant information bars; and recombining the multiple relevant information bars into a target pseudo-document. The multi-model-based pseudo-document generation method and system provided in this disclosure can improve the generation effect of pseudo-documents, thereby improving retrieval performance.

Description

Technical Field

[0001] This disclosure belongs to the field of data generation technology, and more specifically, relates to a method and system for generating pseudo-documents based on multiple models. Background Technology

[0002] In the field of information retrieval, there are generally two main approaches: sparse retrieval and dense retrieval. Sparse retrieval performs poorly in scenarios requiring strong semantic interpretation and struggles to handle word variations and synonym issues. Dense retrieval significantly improves performance for semantically complex queries, but it is still challenged by lexical inconsistencies and subtle contextual changes, leading to incomplete or inaccurate search results.

[0003] Traditional query expansion typically relies on shallow information, such as word frequency and co-occurrence relationships, limiting its ability to capture the deep semantic relationships behind the query. Furthermore, traditional methods introduce irrelevant terms, reducing retrieval accuracy. To address these issues, current research utilizes pseudo-documents generated by large language models for query expansion. However, the quality of these pseudo-documents remains inconsistent, especially when the hints are poorly designed or the model lacks domain-specific knowledge. In such cases, the generated pseudo-documents deviate from the actual semantic requirements of the query, resulting in a lack of semantic consistency between their content and the original query, and even introducing irrelevant or misleading information, leading to poor retrieval performance.

[0004] Therefore, an accurate and reliable method for generating pseudo-documents is needed to improve retrieval performance. Summary of the Invention

[0005] The purpose of this disclosure is to provide a multi-model-based pseudo-document generation method and system to improve the generation effect of pseudo-documents, thereby improving retrieval performance.

[0006] A first aspect of this disclosure provides a pseudo-document generation method based on multiple models, comprising:

[0007] The query information is input into multiple language models to generate multiple pseudo-documents;

[0008] Each pseudo-document is decomposed into multiple information bars based on its length, with the number of information bars being greater than or equal to the number of pseudo-documents.

[0009] Multiple information bars are filtered using a relevance evaluator to obtain multiple relevant information bars;

[0010] Reorganize multiple relevant information bars into a target pseudo-document.

[0011] A second aspect of this disclosure provides a multi-model-based pseudo-document generation system, comprising:

[0012] The pseudo-document generation module is used to input query information into multiple language models to generate multiple pseudo-documents.

[0013] The pseudo-document decomposition module is used to decompose each pseudo-document into multiple information bars based on the length of each pseudo-document. The number of information bars is greater than or equal to the number of pseudo-documents.

[0014] The filtering module is used to filter multiple information bars based on the relevance evaluator to obtain multiple relevant information bars;

[0015] The pseudo-document reorganization module is used to reorganize multiple related information bars into a target pseudo-document.

[0016] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the multi-model-based pseudo-document generation method described above.

[0017] A fourth aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multi-model-based pseudo-document generation method described above.

[0018] The beneficial effects of the multi-model-based pseudo-document generation method and system provided in this disclosure are as follows:

[0019] This disclosure improves the accuracy of pseudo-documents by decomposing them into multiple information bars and calculating the relevance of each information bar to the query information, thereby obtaining highly relevant information bars. This improves the accuracy of the pseudo-documents composed of these information bars and ultimately enhances the retrieval results. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a multi-model-based pseudo-document generation method provided in an embodiment of this disclosure;

[0022] Figure 2 A flowchart illustrating a second multi-model-based pseudo-document generation method provided in an embodiment of this disclosure;

[0023] Figure 3 A structural block diagram of a multi-model-based pseudo-document generation system provided in an embodiment of this disclosure;

[0024] Figure 4 This is a schematic block diagram of an electronic device provided according to an embodiment of the present disclosure. Detailed Implementation

[0025] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this disclosure with unnecessary detail.

[0026] The following is an explanation of the terms used in the embodiments of this disclosure:

[0027] Query expansion: This is the process of expanding and optimizing a user's original query by adding relevant words, phrases, or semantic information. The goal is to address potential ambiguity or limited expression in the initial query, enabling the retrieval system to more accurately understand the user's intent and return more comprehensive and relevant results. For example, if a user enters "artificial intelligence," the system might expand the query to include related terms like "machine learning" and "deep learning," making the search more precise.

[0028] Sparse retrieval: A method of information retrieval that represents text as sparse vectors. It treats documents and queries as sets of terms, constructing an index based on discrete information such as the frequency and position of terms within the documents. During retrieval, the relevance of the document to the query is evaluated by calculating the degree of matching and relevance weights between the query terms and the terms in the document index, thereby determining the retrieval results.

[0029] Dense retrieval: Utilizing deep learning techniques, text is transformed into a low-dimensional, dense vector representation. In vector space, semantically similar text vectors are close in distance. By calculating the similarity (such as cosine similarity) between query vectors and document vectors, the semantic relevance between texts is measured, thereby determining the retrieval results. This emphasizes a deep understanding of text semantics and the measurement of similarity in vector space.

[0030] To make the objectives, technical solutions, and advantages of this disclosure clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0031] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a multi-model-based pseudo-document generation method according to an embodiment of the present disclosure. The method includes:

[0032] S101: Input the query information into multiple language models to generate multiple pseudo-documents.

[0033] In this embodiment, the query information refers to the input information used to retrieve relevant content. The query information is preferably in text format. The language model is a deep learning algorithm model trained on a large amount of text data, capable of understanding and generating natural language, such as GPT, Qwen, Llama, GLM, or DeepSeek. The pseudo-document is not a real, original document, but rather text generated by the language model based on the query information. That is, the generated text can simulate a real document, containing content related to the query information, used to assist in information retrieval or other natural language processing tasks. The number of pseudo-documents output by a language model can be one or more; there is no explicit relationship between the number of language models and the number of pseudo-documents.

[0034] For example, if the query is "the latest breakthrough in electric vehicle battery technology," and this information is input into multiple large language models, GPT, based on its training data and algorithms, generates text containing content such as "In recent years, electric vehicle battery technology has made significant breakthroughs in energy density, and the development of solid-state batteries is accelerating, which is expected to greatly improve driving range." This text is a pseudo-document. DeepSeek generates similar pseudo-documents with content such as "In terms of battery charging speed, new fast charging technologies are constantly emerging, and some technologies can achieve charging to 80% in 15 minutes." Pseudo-documents provide more relevant text data for subsequent information retrieval and processing, helping to improve retrieval and analysis results.

[0035] S102: Based on the length of each pseudo-document, decompose each pseudo-document into multiple information bars, the number of information bars being greater than or equal to the number of pseudo-documents.

[0036] In this embodiment, the length of a pseudo-document refers to an indicator of text size, such as the number of characters, words, or sentences contained in the pseudo-document, used to determine whether and how to decompose the pseudo-document. Information bars are smaller text units split from the pseudo-document, each containing relatively independent information.

[0037] A specific decomposition method could be: based on the length of each pseudo-document, decompose each pseudo-document into multiple information bars, including:

[0038] In response to the pseudo-document's length being less than or equal to a first length threshold, the pseudo-document is treated as an information bar;

[0039] In response to a pseudo-document's length exceeding a first length threshold, the pseudo-document is decomposed into... based on heuristic rules. One message bar;

[0040] in, , The length of the pseudo-document. The first length threshold, This indicates rounding up to the nearest integer.

[0041] In this embodiment, if the length of the pseudo-document is less than or equal to the first length threshold, it is regarded as an independent information bar. When the length of the pseudo-document is greater than the first length threshold, it can be decomposed into multiple information bars by heuristic rules.

[0042] Heuristic rules are a method based on experience or specific strategies, such as splitting according to granularity, sentence boundaries, paragraph structure, topic relevance, etc.

[0043] S103: Filter multiple information bars based on the relevance evaluator to obtain multiple relevant information bars.

[0044] In this embodiment, the relevance evaluator is a tool or algorithm used to evaluate the degree of correlation between information bars and query information. Considering that the pseudo-document itself is generated data information, after splitting it, the data in the information bars may have low relevance to the initial query information. Therefore, this disclosure filters the split information bars and uses the data that meets the filtering conditions as relevant information bars.

[0045] Specifically, multiple information bars are filtered based on the relevance evaluator to obtain multiple relevant information bars, including:

[0046] If the relevance value of an information bar is greater than the first relevance value, the information bar is treated as a relevant information bar.

[0047] The formula for calculating the correlation value using the correlation estimator is:

[0048] ,in, Indicates information bar Information for query The relevant values, Indicates the first relevance value. This indicates the second correlation value. The first correlation coefficient represents the weight of the first correlation value. The second correlation coefficient represents the weight of the second correlation value; Indicates the number of information bars; ,in, To retrieve information Dense embedding vectors, For information bar Dense embedding vectors. and It can be determined based on the experimental results of the experimental process or set according to experience.

[0049] In this embodiment, ,in, Indicates query information The Middle Inverse document frequency of a word, This indicates the number of terms in the query information Q. Indicates information bar Length, Indicates query information The first in One word, express In the information bar The frequency in This represents the average length of all information bars. These are the default parameters for the algorithm. , For inclusion The number of information bars Indicates the number of all information bars. This refers to inverse document frequency.

[0050] S104: Reorganize multiple related information bars into a target pseudo-document.

[0051] In this embodiment, the relevant information bars obtained in S103 can be represented as a set. ,in, This represents the first relevance value. The target pseudo-document obtained from the reorganization can be represented as... ,in, Represents a recombined set , This indicates the first message after the reorganization. This indicates the second information bar after the reorganization. Indicates the first [number] after the reorganization Each information bar, or related information bar, can be rearranged randomly or according to the logical relationships between them. For example, information bars describing events in chronological order will be arranged in that order; if the information bars elaborate on a theme from different perspectives, they can be sorted according to the theme's inherent logic, such as describing causes, current situation, and impact.

[0052] As can be seen from the above, this disclosure improves the accuracy of information bars by decomposing pseudo-documents into multiple information bars and calculating the relevance of each information bar to the query information to obtain highly relevant information bars, thereby improving the accuracy of the pseudo-documents composed of them and thus improving the retrieval effect.

[0053] Considering that different language models use different data domains and have different focuses during training—some are better at scientific and technological knowledge, while others are better at understanding humanities and social sciences—the process of generating pseudo-documents based on multilingual models should match the query information with the domains of expertise of each model. Specifically, in one embodiment of this disclosure, the query information is input into multiple language models to generate multiple pseudo-documents, including:

[0054] The selection probability of each language model is determined based on the first formula;

[0055] Based on the selection probability of each language model, the query information is input into multiple language models to generate multiple pseudo-documents;

[0056] The first formula is: ,in, To retrieve information, To retrieve information Language model in the case The probability of choosing, It involves selecting sensitivity parameters for the probabilities of each language model. For language models The corresponding data domain keywords are converted into dense vectors. For querying information and language models Data domain similarity, .

[0057] In this embodiment, the first formula can be used to calculate the given query information. In this case, select each language model probability That is, the probability of choice. In the formula... Calculate query information using vector dot product and magnitude With language models Dense vectors for keyword transformation in the data domain The similarity between them reflects the degree of semantic and domain fit. Hyperparameters Used to adjust the sensitivity of model selection, if Larger similarity values ​​mean that model selection becomes more sensitive to differences in similarity, tending to choose models with high similarity to the query information; if... The similarity is relatively small, allowing for a more lenient selection of models. The first formula calculates the probability of each model being selected by performing an exponential operation on the similarity and then normalizing it, ensuring that the sum of the probabilities is 1.

[0058] As can be seen from the above, this disclosure determines the probability of selecting each language model by calculating the similarity between the strengths of each model and the query information. This helps to ensure that the generated pseudo-documents are closer to the actual needs of the query information, improve the relevance and quality of the pseudo-documents, and thus enhance the retrieval effect.

[0059] In one embodiment of this disclosure, pseudo-documents are decomposed based on heuristic rules. The information bar includes:

[0060] Based on heuristic rules, pseudo-documents are decomposed into granular... N information bars;

[0061] ,in, Based on the granularity threshold, As a regulating factor, To query complexity metrics, , To count the number of unique words in the query information, The total number of words in the query information. The number of named entities in the query. These are the weighting coefficients.

[0062] In this embodiment, considering that if the query uses many different words (high vocabulary richness) or contains multiple entities, the granularity of the segmentation should be finer (i.e., (Make the threshold smaller). If the query is simple (such as a single keyword), the threshold can be appropriately increased (making the information bar longer). It can be determined based on prior knowledge. This indicates the number of unique words in the query information, reflecting the richness of the query vocabulary. The more unique words there are, the broader the concepts involved in the query. It is the total number of words in the query information; The number of named entities in the query represents the number of specific objects or concepts in the query. It is a weighting coefficient used to adjust the importance of the number of named entities in complexity calculation.

[0063] If a query uses many different words (high lexical richness) or contains multiple entities, it means the query complexity is high. The value is relatively large, according to The calculation formula, It will become smaller, meaning the granularity of the segmentation is finer, breaking down the pseudo-document into more and shorter information segments for more accurate matching and processing of information; if the query is simple, such as with only a single keyword, The value is small. It will become larger, with longer information bars, reducing the number of decompositions and improving processing efficiency.

[0064] As can be seen from the above, this disclosure helps to capture key information in the query by dynamically adjusting the granularity of information bar division based on the complexity of the query, reducing the omission and misunderstanding of information, thereby improving the accuracy and efficiency of information retrieval.

[0065] In one embodiment of this disclosure, the pseudo-document generation method based on multiple models further includes:

[0066] The target connection relationship is determined based on the retrieval method; the target connection relationship is the connection between the target pseudo-document and the query information.

[0067] The target pseudo-document and query information are connected based on the target connection relationship to obtain the target query information;

[0068] The search is performed based on the target query information to obtain the search results.

[0069] Determining target connections based on retrieval methods includes:

[0070] In response to the sparse retrieval method, the first connection relationship is determined as the target connection relationship;

[0071] In response to the intensive retrieval method, the second connection relationship is determined as the target connection relationship;

[0072] The number of times the queried information is repeated differs between the first and second join relationships.

[0073] In this embodiment, considering that sparse retrieval is based on the discrete representation of text, it mainly uses information such as the frequency and position of terms to build an index for retrieval. Since it relies more on precise word matching, the weight of query information is relatively important. However, in practical applications, query information is usually short while pseudo-documents are long. To ensure that query information is not overlooked during the matching process with pseudo-documents, its weight can be increased. For example, query information can be repeatedly concatenated with pseudo-documents, making the retrieval system pay more attention to the matching of query terms in pseudo-documents and improving retrieval accuracy.

[0074] Dense retrieval, on the other hand, uses deep learning to map text into a low-dimensional dense vector space representation. It measures the relevance of a document to a query by calculating the similarity between vectors. It focuses on the overall semantic relationships of the text, with less reliance on precise matching of individual words. Therefore, a simple concatenation of the original query and the pseudo-document allows the model to effectively calculate their semantic similarity in the vector space, without needing to repeatedly query information to enhance weights as in sparse retrieval.

[0075] Specifically, sparse retrieval is due to the query information Usually more than the target pseudo-document Much shorter, in order to balance the query information Between and the target pseudo-document The relative weights of the two elements are used to connect them according to the first join relationship before the query. The first join relationship can be represented as... ,in, This indicates a string concatenation operation, resulting in an expanded query. The target query information can be used as new query information for BM25 retrieval. It is a positive integer, preferably 5.

[0076] Dense retrieval: The join method is similar to sparse retrieval, a simple join between the original query q and the pseudo-document d, separated by [SEP]. Specifically, the second join relationship can be... SEP is an abbreviation for Separator, which is used to distinguish different text segments.

[0077] As can be seen from the above, for sparse retrieval, this disclosure effectively increases the weight of query information in the retrieval process by increasing the number of repetitions of query information and connecting it with the target pseudo-document. This helps ensure that query information is not ignored when matching with pseudo-documents and improves the accuracy of retrieval.

[0078] Figure 2 This is a flowchart illustrating a second multi-model-based pseudo-document generation method provided in an embodiment of this disclosure. (See reference...) Figure 2 Query information By inputting data into multiple models, multiple pseudo-documents are obtained. These pseudo-documents can then be decomposed, filtered, and recombined to obtain the target pseudo-document. Then query information and target pseudo-document By establishing a connection, you can obtain the target query information. Figure 2 The information bars obtained after filtering are the relevant information bars, which are essentially information bars. Figure 2 Choose to use information bars as the description.

[0079] Corresponding to the multi-model-based pseudo-document generation method in the above embodiments, Figure 3 This is a structural block diagram of a multi-model-based pseudo-document generation system provided in one embodiment of this disclosure. For ease of explanation, only the parts relevant to the embodiment of this disclosure are shown. References Figure 3 The multi-model-based pseudo-document generation system 20 includes: a pseudo-document generation module 21, a pseudo-document decomposition module 22, a filtering module 23, and a pseudo-document recombination module 24.

[0080] Among them, the pseudo-document generation module 21 is used to input query information into multiple language models to generate multiple pseudo-documents;

[0081] The pseudo-document decomposition module 22 is used to decompose each pseudo-document into multiple information bars based on the length of each pseudo-document, wherein the number of information bars is greater than or equal to the number of pseudo-documents.

[0082] Filtering module 23 is used to filter multiple information bars based on the relevance evaluator to obtain multiple relevant information bars;

[0083] The pseudo-document reorganization module 24 is used to reorganize multiple related information bars into a target pseudo-document.

[0084] In one embodiment of this disclosure, the pseudo-document generation module 21 is specifically used to determine the selection probability of each language model based on a first formula;

[0085] Based on the selection probability of each language model, the query information is input into multiple language models to generate multiple pseudo-documents;

[0086] The first formula is: ,in, To retrieve information, Language model in the case of querying information Q The probability of choosing, It involves selecting sensitivity parameters for the probabilities of each language model. For language models The corresponding data domain keywords are converted into dense vectors. For querying information and language models Data domain similarity, .

[0087] In one embodiment of this disclosure, the pseudo-document decomposition module 22 is specifically used to treat the pseudo-document as an information bar in response to the length of the pseudo-document being less than or equal to a first length threshold.

[0088] In response to a pseudo-document's length exceeding a first length threshold, the pseudo-document is decomposed into... based on heuristic rules. One message bar;

[0089] in, , The length of the pseudo-document. The first length threshold, This indicates rounding up to the nearest integer.

[0090] In one embodiment of this disclosure, the pseudo-document decomposition module 22 is further configured to decompose the pseudo-document into granularities based on heuristic rules. N information bars;

[0091] in, ,in, Based on the granularity threshold, As a regulating factor, To query complexity metrics, , To count the number of unique words in the query information, The total number of words in the query information. The number of named entities in the query. These are the weighting coefficients.

[0092] In one embodiment of this disclosure, the filtering module 23 is specifically used to select an information bar as a relevant information bar in response to the information bar's correlation value being greater than a first correlation value.

[0093] The formula for calculating the correlation value using the correlation estimator is:

[0094] ,in, Indicates information bar Information for query The correlation, Indicates the first relevance value. This indicates the second correlation value. The first correlation coefficient represents the weight of the first correlation value. The second correlation coefficient represents the weight of the second correlation value; Indicates the number of information bars; ,in, To retrieve information Dense embedding vectors, For information bar Dense embedding vectors.

[0095] In one embodiment of this disclosure, the multi-model-based pseudo-document generation system 20 further includes: an information connection module, used to determine the target connection relationship based on the retrieval method; the target connection relationship is the connection relationship between the target pseudo-document and the query information;

[0096] The target pseudo-document and query information are connected based on the target connection relationship to obtain the target query information;

[0097] The search is performed based on the target query information to obtain the search results.

[0098] In one embodiment of this disclosure, the information connection module is specifically used to determine the first connection relationship as the target connection relationship in response to the retrieval method being sparse retrieval;

[0099] In response to the intensive retrieval method, the second connection relationship is determined as the target connection relationship;

[0100] The number of times the queried information is repeated differs between the first and second join relationships.

[0101] See Figure 4 , Figure 4 This is a schematic block diagram of an electronic device provided according to an embodiment of the present disclosure. Figure 4 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of each module / unit in the above system embodiments, for example... Figure 3 The functions of modules 21 to 24 are shown.

[0102] It should be understood that, in the embodiments of this disclosure, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0103] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0104] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0105] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this disclosure can execute the implementation methods described in the first and second embodiments of the pseudo-document generation method based on multiple models provided in the embodiments of this disclosure, or they can execute the implementation methods of the electronic devices described in the embodiments of this disclosure, which will not be repeated here.

[0106] In another embodiment of this disclosure, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to implement these processes. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0107] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0108] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0109] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0111] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this disclosure, depending on actual needs.

[0112] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0113] The above are merely specific embodiments of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this disclosure, and these modifications or substitutions should all be covered within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A pseudo-document generation method based on multiple models, characterized in that, include: The query information is input into multiple language models to generate multiple pseudo-documents; Each pseudo-document is decomposed into multiple information bars based on its length, and the number of information bars is greater than or equal to the number of pseudo-documents. Multiple information bars are obtained by filtering the multiple information bars based on the relevance evaluator; Reorganize the multiple relevant information bars into a target pseudo-document; The step of inputting query information into multiple language models to generate multiple pseudo-documents includes: The selection probability of each language model is determined based on the first formula; Based on the selection probability of each language model, the query information is input into multiple language models to generate multiple pseudo-documents; The first formula is: ,in, To retrieve information, Language model in the case of querying information Q The probability of choosing, It involves selecting sensitivity parameters for the probabilities of each language model. For language models The corresponding data domain keywords are converted into dense vectors. For querying information and language models Data domain similarity, ; The process of decomposing each pseudo-document into multiple information bars based on its length includes: In response to the pseudo-document's length being less than or equal to a first length threshold, the pseudo-document is treated as an information bar; In response to a pseudo-document's length exceeding a first length threshold, the pseudo-document is decomposed into... based on heuristic rules. One message bar; in, , The length of the pseudo-document. The first length threshold, Indicates rounding up; The heuristic rules-based decomposition of the pseudo-document into... The information bar includes: Based on heuristic rules, the pseudo-document is decomposed into granularity. N information bars; in, ,in, Based on the granularity threshold, As a regulating factor, To query complexity metrics, , To count the number of unique words in the query information, The total number of words in the query information. The number of named entities in the query. These are the weighting coefficients; The process of filtering the multiple information bars based on the relevance evaluator yields multiple relevant information bars, including: In response to the information bar having a correlation value greater than a first correlation value, the information bar is designated as a relevant information bar; The formula for calculating the correlation value of the correlation estimator is as follows: ,in, Indicates information bar With query information The correlation, Indicates the first relevance value. This indicates the second correlation value. The first correlation coefficient represents the weight of the first correlation value. The second correlation coefficient represents the weight of the second correlation value; Indicates the number of information bars; ,in, To retrieve information Dense embedding vectors, For information bar Dense embedding vectors.

2. The pseudo-document generation method based on multiple models as described in claim 1, characterized in that, Also includes: The target connection relationship is determined based on the retrieval method; the target connection relationship is the connection relationship between the target pseudo-document and the query information. Based on the target connection relationship, the target pseudo-document and the query information are connected to obtain the target query information; The search is performed based on the target query information to obtain the search results.

3. The pseudo-document generation method based on multiple models as described in claim 2, characterized in that, The determination of target connection relationships based on retrieval methods includes: In response to the sparse retrieval method, the first connection relationship is determined as the target connection relationship; In response to the retrieval method being dense retrieval, the second connection relationship is determined as the target connection relationship; The number of times the queried information is repeated differs between the first connection relationship and the second connection relationship.

4. A pseudo-document generation system based on multiple models, characterized in that, include: The pseudo-document generation module is used to input query information into multiple language models to generate multiple pseudo-documents. The pseudo-document decomposition module is used to decompose each pseudo-document into multiple information bars based on the length of each pseudo-document, wherein the number of information bars is greater than or equal to the number of pseudo-documents. The filtering module is used to filter the multiple information bars based on the relevance evaluator to obtain multiple relevant information bars; The pseudo-document reorganization module is used to reorganize the multiple related information bars into a target pseudo-document; The pseudo-document generation module is specifically used to determine the selection probability of each language model based on the first formula. Based on the selection probability of each language model, the query information is input into multiple language models to generate multiple pseudo-documents; The first formula is: ,in, To retrieve information, Language model in the case of querying information Q The probability of choosing, It involves selecting sensitivity parameters for the probabilities of each language model. For language models The corresponding data domain keywords are converted into dense vectors. For querying information and language models Data domain similarity, ; The pseudo-document decomposition module is specifically used to treat the pseudo-document as an information bar in response to the length of the pseudo-document being less than or equal to a first length threshold. In response to a pseudo-document's length exceeding a first length threshold, the pseudo-document is decomposed into... based on heuristic rules. One message bar; in, , The length of the pseudo-document. The first length threshold, Indicates rounding up; The pseudo-document decomposition module is further used to decompose the pseudo-document into granularities based on heuristic rules. N information bars; in, ,in, Based on the granularity threshold, As a regulating factor, To query complexity metrics, , To count the number of unique words in the query information, The total number of words in the query information. The number of named entities in the query. These are the weighting coefficients; The filtering module is specifically used to select an information bar as a relevant information bar in response to the information bar having a relevance value greater than a first relevance value. The formula for calculating the correlation value of the correlation estimator is as follows: ,in, Indicates information bar Information for query The correlation, Indicates the first relevance value. This indicates the second correlation value. The first correlation coefficient represents the weight of the first correlation value. The second correlation coefficient represents the weight of the second correlation value; Indicates the number of information bars; ,in, To retrieve information Dense embedding vectors, For information bar Dense embedding vectors.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 3.

Images