Iterative concept generation information retrieval method based on large language model

Abstract

The invention discloses an iterative concept generation information retrieval method based on a large language model. The method comprises the following steps: obtaining a candidate document set; generating an internal concept set; utilizing the first instruction to obtain a first concept set; screening out a first external concept set; generating a first extended query; obtaining a second document set and a second external concept set; combining the second external concept set with the first external concept set to update the first concept complete set, generating a second extended query, and analyzing semantic consistency of the last round of extended query and the current extended query; repeating the step until the semantic consistency score is lower than a threshold value or declines, and returning to the last round of extended query as a final extended query; and obtaining a retrieval result set. According to the method, the influence caused by semantic drift and noise expansion can be effectively reduced, and the correlation, stability and interpretability of retrieval are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the technical field of combining information retrieval and large language learning models, and specifically relates to an iterative concept generation information retrieval method based on a large language model. Background Technology

[0002] In the information and digital age, the rapid development of the internet, social media, and various online platforms has led to an explosive growth in the speed and scale of data generation. Whether in daily life, scientific research, business, or entertainment, people's demand for information acquisition and processing is constantly increasing. However, with the surge in information volume, efficiently and accurately retrieving the information users need from massive amounts of data has become a significant technological challenge. Traditional information retrieval methods have limitations in terms of information accuracy, retrieval efficiency, and semantic understanding, necessitating more intelligent and adaptive solutions to meet the demands of modern information services.

[0003] Information retrieval (IR) is a key computer technology that enables the retrieval of information relevant to user queries from large amounts of unstructured or semi-structured data. This technology is widely used in search engines, question-answering systems, recommendation systems, e-commerce, knowledge management, and many other fields. A typical information retrieval system mainly includes modules such as query input, document indexing, retrieval algorithms, and matching models, among which the retrieval algorithm and matching model play a decisive role in retrieval performance. In recent years, with the development of artificial intelligence and deep learning, semantic retrieval methods based on neural networks have become a research hotspot. These methods use deep models to vectorize queries and documents, capturing the implicit semantic relationships between them, thereby overcoming the limitations of traditional keyword matching and significantly improving retrieval effectiveness and semantic understanding capabilities.

[0004] For example, Chinese invention patent CN117332782A (Subject Information Retrieval System and Method Based on Transformer) proposes to use the Query Embedding method combined with knowledge graphs for semantic expansion and relational reasoning, thereby improving the accuracy and recall of retrieval. However, such deep learning methods still have significant limitations: (1) they rely on a large amount of labeled training data, which is costly and difficult to update; (2) they have insufficient generalization performance in cross-domain or open-domain scenarios; (3) retrieval expansion strategies are mostly based on keywords or fixed templates, which makes it difficult to capture implicit concepts and semantic relationships in the query, easily leading to semantic drift and reducing retrieval accuracy and interpretability.

[0005] To address the aforementioned issues, the emergence of Large Language Models (LLMs) in recent years has provided a new direction for information retrieval. LLMs, through pre-training on massive unlabeled corpora, learn the statistical distribution and semantic patterns of human language, possessing powerful language understanding and generation capabilities. Unlike traditional deep learning models, LLMs do not require additional task-specific training; they can perform various natural language processing tasks, including information retrieval, text summarization, question answering, and logical reasoning, solely through task prompting. This flexibility significantly alleviates the dependence of traditional models on manually labeled data, providing a new technological foundation for building highly versatile information retrieval systems.

[0006] In existing research, Ouyang et al.'s "Training language models to follow instructions with human feedback" (NeurIPS 2022) introduced human feedback reinforcement learning (RLHF) technology, enabling language models to have stronger instruction comprehension and task execution capabilities. The emergence of this type of model allows large language models to perform complex reasoning and generation in open semantic spaces, providing strong technical support for information retrieval, question answering, and reasoning tasks.

[0007] However, despite the excellent performance of large language models in language understanding and generation, they still have significant shortcomings in combining conceptual-level semantic reasoning with information retrieval. First, existing LLM-based query expansion methods typically generate expanded content on a sentence or keyword basis, lacking structured modeling of the core concepts in the query. Since the generated results may not be entirely consistent with the original query semantics, this "semantic drift" can cause retrieval results to deviate from user intent, reducing the system's relevance and reliability. Second, large language models are prone to the "illusion" problem during content generation, generating inaccurate or fictitious concepts and information, thus introducing noise and weakening retrieval quality. Although some methods attempt to mitigate the model illusion problem by combining external knowledge bases for external expansion, external resources often suffer from timeliness issues, information redundancy, or low reliability, making it difficult to guarantee the accuracy and consistency of retrieval results.

[0008] In information retrieval, query expansion is a crucial technique that aims to improve recall and precision by generating semantic elements (synonyms, hyponyms, or related concepts) relevant to the original query. Unlike keyword-based expansion, concept-based expansion better reflects the semantic core and contextual relationships of the user's query, thereby enhancing the system's semantic understanding. Therefore, how to leverage the generative capabilities of large language models to dynamically generate high-quality concepts while maintaining semantic consistency, and how to control the risks of illusion and semantic drift through iterative optimization mechanisms, have become pressing technical challenges in the field of information retrieval. Summary of the Invention

[0009] This invention addresses the shortcomings of existing technologies by providing an iterative concept generation information retrieval technique based on a large language model. This method combines internal concept expansion (generating concepts based on the original query) with external concept expansion (extracting concepts based on preliminary search results), and is supplemented by a chain-like reasoning mechanism. This effectively improves the semantic understanding capability and retrieval accuracy of the large language model in information retrieval tasks without requiring additional training data.

[0010] Firstly, this application provides an iterative concept generation information retrieval method based on a large language model, including: Based on the user's query, a first set of documents is obtained; a candidate set of documents is then retrieved. Utilize a large language learning model to generate an internal concept set based on query statements; We utilize a large language learning model based on a set of candidate documents to construct the first instruction for concept extraction; Input the first instruction into the large language model to obtain the first set of concepts; The concept relevance filtering prompt is constructed, and the instruction big language model filters out the most relevant first external concept set from the first concept set based on the semantics of the query and the semantic similarity of the candidate concepts. The internal concept set and the first external concept set are merged into the first concept set, and the first extended query is generated. The candidate document set is retrieved using the first extended query to obtain the second document set and the second external concept set; The second external concept set is merged with the first external concept set to update the first concept set, generating a second expanded query. The semantic consistency between the previous expanded query and the current expanded query is analyzed using a large language learning model. This step is repeated until the semantic consistency score is below the threshold or decreases, and the previous expanded query is returned as the final expanded query. The final expanded query is used to calculate document relevance scores and reorder them to obtain the set of search results.

[0011] In one possible implementation, generating an internal concept set based on a query statement using a large language learning model includes: A task template is generated based on a query statement using a large language learning model; a concept generation prompt statement is constructed by combining the query statement; the concept generation prompt statement includes a clear description of the concept generation task and the concatenation of the query statement. The concept generation prompt is input into the large language model, and concepts related to the query semantics are generated through reasoning to obtain an internal concept set.

[0012] In one possible implementation, the method for constructing the first instruction includes: firstly describing the goals and requirements of the document concept extraction task in a prompt statement, and then concatenating the candidate document set with the query statement to obtain the first instruction.

[0013] In one possible implementation, the concept relevance filtering prompt requires the large language learning model to act as an information retrieval expert, guiding the large language model to establish a mapping relationship between the query statement and candidate concepts. The construction method includes: first, describing the goals and requirements of the query statement and the concept relevance selection task in the prompt statement; then, concatenating the first concept set with the query statement to obtain the concept relevance filtering prompt.

[0014] In one possible implementation, the method for generating the first extended query includes: concatenating and combining the query statement with the first concept set in sequence to obtain the first extended query.

[0015] In one possible implementation, merging the second external concept set with the first external concept set to update the first concept set, generating a second expanded query, and using a large language learning model to analyze the semantic consistency between the previous expanded query and the previous round of expanded queries; repeating this step until the semantic consistency score is below a threshold or decreases, and returning the previous round of expanded queries as the final expanded query, includes: The second external concept set is merged with the first external concept set to update the first concept set and generate the second extended query; a query extension evaluation prompt statement is constructed, and the previous round of extended query and the current extended query are used as input by the big language learning model. Based on dimensions including semantic consistency, core concept overlap rate and information gain, the two are compared and a consistency score is output. If the current rating Lower than the previous round score or If the value falls below a preset threshold, semantic consistency is determined to have decreased, the iteration is terminated, and the previous query is returned as the final expanded query; if If the expansion is reasonable, continue to the next round of concept expansion and optimization until it stops, and then obtain the final expanded query.

[0016] Secondly, this application provides an iterative concept generation information retrieval device based on a large language model, comprising: The retrieval module is used to obtain a first set of documents based on the user's input query; and to obtain a set of candidate documents through retrieval. The first generation module is used to generate an internal concept set based on the query statement using a large language learning model; The second generation module is used to construct a first instruction for concept extraction based on a candidate document set using a large language learning model; input the first instruction into the large language model to obtain a first concept set; construct concept relevance filtering prompts; and the instruction large language model filters out the most relevant first external concept set from the first concept set based on the semantic similarity between the query semantics and the semantics of the candidate concepts. The fusion module is used to merge the internal concept set and the first external concept set into the first concept set and generate the first extended query; The extension module is used to retrieve the candidate document set using the first extended query to obtain the second document set and the second external concept set; The iterative module is used to merge the second external concept set with the first external concept set to update the first concept set, generate the second extended query, and use a large language learning model to analyze the semantic consistency between the previous extended query and the current extended query; repeat this step until the semantic consistency score is lower than the threshold or decreases, and return the previous extended query as the final extended query; The output module is used to calculate document relevance scores based on the final expanded query and reorder them to obtain a set of search results.

[0017] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the iterative concept generation information retrieval method based on a large language model as described in the first aspect.

[0018] Fourthly, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the iterative concept generation information retrieval method based on a large language model as described in the first aspect.

[0019] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the iterative concept generation information retrieval method based on a large language model as described in the first aspect.

[0020] The beneficial effects of this application are as follows: The iterative concept generation information retrieval method based on a large language model provided in this application achieves deep modeling and adaptive expansion of query semantics by automatically extracting and optimizing semantic concepts during the query process, and by using a two-layer generation mechanism of internal and external semantic concepts and an iterative optimization strategy driven by a large language model. This significantly improves the semantic relevance and accuracy of information retrieval results.

[0021] The method provided in this application not only optimizes the retrieval performance of general search engines but also provides effective support for vertical domain knowledge retrieval. It has significant research value and broad application prospects, and can be widely applied to high-precision retrieval in scenarios such as enterprise knowledge bases, academic literature, laws and regulations, and medical information. By introducing the method provided in this application into recommendation systems and professional databases, it is possible to understand the user's deep intent and expand the association of concepts, thereby realizing cross-domain semantic retrieval and personalized information recommendation. Attached Figure Description

[0022] Figure 1 A flowchart of an iterative concept generation information retrieval scheme based on a large language model provided in this application embodiment; Figure 2 A flowchart illustrating the framework of an iterative concept generation information retrieval scheme based on a large language model, provided in this application embodiment. Figure 3 This is a schematic diagram of the structure of the iterative concept generation information retrieval device based on a large language model provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0025] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0026] Existing retrieval methods based on large language learning models, while offering technical solutions to mitigate the model illusion problem by combining external knowledge bases for external expansion, still suffer from insufficient timeliness, information redundancy, or low reliability, making it difficult to guarantee the accuracy and consistency of retrieval results.

[0027] Based on this, this application provides an iterative concept generation and information retrieval method for large language models.

[0028] This application provides an iterative concept generation information retrieval method based on a large language model. This method achieves a semantic closed-loop process from user query input to retrieval result optimization through a three-stage collaborative process of "internal concept generation—external concept extraction—iterative query optimization." Specifically, firstly, internal concepts are generated from the user-input query, extracting its core semantic units and potential semantic associations; then, based on the candidate documents obtained from the initial retrieval, external concepts are extracted to supplement semantically relevant concepts that are not explicitly present in the query; finally, through an iterative optimization mechanism driven by a large language model, internal and external concepts are fused and filtered, gradually correcting and strengthening the query expression until the retrieval performance converges or reaches a set threshold. Through the above multi-stage concept generation and dynamic optimization process, this method can significantly improve the semantic matching accuracy and result relevance of information retrieval tasks.

[0029] See Figure 1 , Figure 2 The iterative concept generation information retrieval method for a large language model provided in this application includes the following steps: S101. Based on the query statement input by the user, obtain the first document set; and obtain the candidate document set through retrieval.

[0030] In one possible implementation, S101 includes: S101a, Query statement based on user input Obtain the first document collection ; S101b, Use a basic retrieval model (such as BM25 or a vector retrieval model) to search the document collection. A preliminary search was conducted to obtain the previous results. A collection of candidate documents This collection of documents serves as an external knowledge source, providing a basis for subsequent concept extraction and semantic expansion.

[0031] Furthermore, the query statement based on user input Obtain the first document collection The methods include: (1) Identify the source of at least one document resource for obtaining candidate documents; (2) Based on query statement q A preliminary search is performed on the document resources to obtain a first set of documents related to the query semantics. D This collection of documents serves as an external knowledge source, providing a basis for subsequent extraction of external concepts.

[0032] Specifically, the document resources include: structured or unstructured documents in an enterprise's internal knowledge base, database, file system, or document management platform; public network resources, such as web pages, academic literature databases, open datasets, or online encyclopedias; and other accessible collections of structured, semi-structured, or unstructured documents. In this embodiment, the document resources are the TREC-COVID dataset and the NFCorpus dataset.

[0033] For example, the original query q The question is "Can drinking green tea lower blood pressure?". The retrieval model uses the BM25 algorithm to perform a preliminary search on the medical and health corpus D (which includes nutritional research, clinical reviews, health Q&A, etc.), selecting the top k=5 documents as the candidate document set. These documents cover topics such as "tea polyphenols and blood pressure," "green tea intake research," and "risk factors for hypertension," constituting external knowledge sources and providing a basis for subsequent concept extraction and semantic expansion.

[0034] S102. Utilize a large language learning model to generate an internal concept set based on the query statement.

[0035] In one possible implementation, the large language learning model includes: LLaMA series models, Qwen series models, GPT series models, Claude series models, and other pre-trained or fine-tunable large-scale language models. The large language model is used to understand natural language queries and generate corresponding internal concept sets.

[0036] In one possible implementation, S102 includes: S102a. Utilize a large language learning model to generate task templates based on query statement q; combine query statement q to construct concepts and generate prompt statements. The concept generates prompt statements. This includes a clear description of the concept generation task and the concatenation of query statements; S102b, Generate prompt statements for concepts Inputting a large language learning model, it generates concepts semantically related to the query through reasoning, resulting in an internal concept set. m q This represents the number of internal concepts generated by the large language model.

[0037] Furthermore, the method for constructing the concept generation prompt statement includes: firstly, clearly defining the requirements and description of the concept generation task in the prompt statement, and then concatenating the query statement into the prompt statement to obtain the concept generation prompt statement.

[0038] Specifically, the requirements and descriptions of the concept generation task include: (1) Task instruction: Instruct the large language model to identify the core concepts of a given query as an expert in a professional field (such as biomedicine) and explain the reasoning process of concept generation; (2) Guided concept extraction: First, extract core concepts directly from the query statement, such as diseases, intervention measures, biomarkers, etc.; then expand the core concepts, including synonyms, near-synonyms or semantically related variants, to enrich the concept set; (3) Prompt language specifications and guidelines: Analyze the main focus and context of the query statement; avoid generating redundant or overly broad concepts; control the number of concepts and prioritize the most important and context-related concepts; provide step-by-step explanations for each step of reasoning (Chain-of-Thought, CoT) to clarify the basis for concept generation; (4) Output format requirements: The generated internal concept set should be a unique and relevant list of key concepts, accompanied by corresponding step-by-step reasoning explanations; (5) Example description: Selectively query the example and generate the corresponding concepts and reasoning steps to clarify the actual application of the prompt.

[0039] Understandably, through the above construction method, concept generation prompts are generated. It can guide a large language model to generate a set of internal concepts that are highly related to the semantics of the input query. .

[0040] For example, concept generation prompt statements as follows: "As a biomedical concept analysis expert, please analyze the core semantics of the following query and generate key concepts or keywords that reflect the meaning of the query."

[0041] Query Can drinking green tea lower blood pressure? Requirements: (1) Identify the core concepts in the query (such as disease, intervention, physiological process, etc.); (2) Generate several synonyms or semantically related terms for each core concept (such as synonyms, common expressions, abbreviations, etc.); (3) Retain the most representative and semantically related concepts, not exceeding 8.

[0042] The large language model analyzes the semantic structure of the query based on the prompts, and generates an internal concept set through contextual understanding and semantic expansion: ={Green tea, tea polyphenols, antioxidant effects, blood pressure regulation, cardiovascular health, caffeine effects, epidemiological studies}.

[0043] While concepts such as "antioxidant effect" and "cardiovascular health" did not appear directly in the original query, the large language model, based on semantic relevance and biomedical knowledge, reflected the implicit semantics and reasoning logic behind the query. This process yielded interpretable internal semantic representations, providing a semantic foundation for subsequent external concept extraction and query optimization.

[0044] S103. Utilize a large language learning model based on a set of candidate documents to construct the first instruction for concept extraction.

[0045] In one possible implementation, the first instruction The construction method includes: first, describing the goals and requirements of the document concept extraction task in the prompt statement; then concatenating the candidate document set with the query statement q to obtain the first instruction. .

[0046] Furthermore, the objectives and requirements of the document concept extraction task include: (1) Task instruction: Instruct the large language model to identify the most relevant concepts from the candidate document set as a biomedical expert; (2) Guided concept extraction: For each document in the candidate document set, extract key concepts related to the document topic and query semantics; concepts include, but are not limited to, core terms such as disease, intervention measures, physiological processes, and biomarkers; (3) Concept quantity control: Limit the number of concepts generated per document (e.g., top-k) to ensure the conciseness and representativeness of the concept set; (4) Repeated checks and accuracy: If a concept cannot be clearly extracted, the model should re-analyze the document content to ensure the completeness and accuracy of concept extraction; (5) Output format requirements: The generated concept set should be a unique and interpretable list format to facilitate subsequent comparison with internal concept sets. Integration or iterative optimization.

[0047] Understandably, based on the above requirements and statements, the first instruction... It can guide large language models to generate a set of concepts related to query semantics from a set of candidate documents, providing external knowledge support for subsequent iterative query optimization.

[0048] For example, the query q is: "Can drinking green tea lower blood pressure?", and a set of candidate documents is obtained using the basic retrieval model. And input several representative documents into the large language model, and execute the first instruction ( ).

[0049] The instructions require the large language model to act as an expert in the biomedical field, identifying the most important thematic concepts from a given text, with a focus on core terms such as disease, physiological processes, and intervention methods.

[0050] For example, the first instruction As shown in Table 1: Table 1 First Instruction

[0051]

[0052] S104. Input the first instruction into the large language model to obtain the first concept set.

[0053] In one possible implementation, the first instruction After inputting into the large language learning model, from each document Extracting key concept sets from documents And merge all document concept sets to form the first concept set. .

[0054] For example, large language learning models are based on instructions From each document Extracting concept sets After merging, we obtain the first set of concepts: ={Catechins, hypertension, diastolic blood pressure, systolic blood pressure, human trials, metabolic syndrome, dietary intervention}.

[0055] S105. Construct concept relevance filtering suggestions. The instruction large language model filters out the most relevant first external concept set from the first concept set based on the query semantics and the semantic similarity of candidate concepts. .

[0056] In one possible implementation, the concept relevance filtering prompt The requirement is for the large language learning model to act as an information retrieval expert, guiding the large language model to establish a mapping relationship between the query statement q and candidate concepts.

[0057] Furthermore, the method for constructing the concept relevance filtering prompt includes: firstly, describing the goals and requirements of the query statement and the concept relevance selection task in the prompt statement, and then concatenating the first concept set with the query statement to obtain the concept relevance filtering prompt.

[0058] Specifically, the goals and requirements of the query statement and the concept relevance selection task are as follows: (1) Task objective: Based on the query statement q input by the user, select the concept most relevant to the query semantics from the candidate concept set; (2) Relevance assessment requirements: The large language model should assess the semantic similarity or relevance between each candidate concept and the query statement, and select the concept that best reflects the query intent; (3) Concept quantity control: Limit the number of output concepts to a preset value. To ensure that the selected set of concepts is concise and highly representative; (4) Output format: The generated set of related concepts should be in a unique list format to facilitate subsequent fusion or iterative optimization; (5) Accuracy requirements: During the screening process, concepts that are irrelevant to the query semantics or redundant should be avoided, and the concept set should be highly focused on the core intent of the query.

[0059] Understandably, through the above task objectives and requirements, this step can guide the large language model to generate the most relevant set of concepts from the candidate concept set based on query semantics, providing a semantic foundation for subsequent concept fusion and iterative query optimization.

[0060] It should be noted that the candidate concepts refer to the set of concepts obtained from the aforementioned steps that are related to the query semantics, including but not limited to: Internal concept set : A set of concepts generated by a large language model based on the query statement q, including the core concepts that appear directly in the query and related concepts generated by the model based on semantic expansion; Document concept set : A set of concepts extracted by a large language model from a set of candidate documents, including explicit or implicit key terms in the documents that are related to the query semantics; Combination of candidate concept sets: This can be the union of the internal concept set and the document concept set, or other forms of combination, used for the subsequent task of selecting the most relevant concept.

[0061] Based on the above definition, candidate concepts not only cover the core concepts directly involved in user queries, but also include related concepts generated by the model based on semantic reasoning and external document knowledge, providing a semantic basis for query relevance selection and iterative optimization.

[0062] For example, constructing concept relevance filtering suggestions This guides the large language model to match the query statement q with the first concept set. Establish semantic mapping relationships between them. The requirement is for the large language model to act as an information retrieval expert, based on the semantic similarity between the query semantics q and candidate concepts, from a set... The most relevant first set of external concepts was selected from the data. .

[0063] Specifically, concept relevance filtering prompts As shown in Table 2: Table 2 Concept Relevance Screening Tips

[0064]

[0065] After filtering, the first subset of external concepts was obtained: ={Catechins, Hypertension, Dietary Intervention}. While "catechins" and "dietary intervention" did not appear directly in the original query, they were identified as knowledge concepts closely related to the topic of "green tea lowering blood pressure" through literature content and semantic association. These external concepts effectively supplemented the knowledge background not covered in the internal concept generation stage, making the extended query more complete and interpretable in the semantic space.

[0066] S106. Merge the internal concept set and the first external concept set into the first concept set and generate the first extended query.

[0067] In one possible implementation, the method for generating the first extended query includes: combining the query statement with the first concept set according to a set concatenation rule to obtain the first extended query.

[0068] In one possible implementation, in S106, the internal concept set is... With the first external concept set By merging them, we obtain the complete set of first concepts. And generate extended queries based on the concatenation rules. .

[0069] Furthermore, the concatenation rules refer to the original query statement. With the complete set of first concepts Connect sequentially without adding extra sorting or labels, .in, This indicates that the query statement and the set of concepts will be combined in a preset order to form a single extended query text.

[0070] The first extended query generated using the above direct concatenation rules It also includes the semantic information of the original query, as well as the semantic information supplemented by the internal concepts and the first external concepts, thus providing semantically complete input for subsequent document retrieval and iterative concept generation based on extended queries.

[0071] S107. The candidate document set is retrieved using the first extended query to obtain the second document set and the second external concept set.

[0072] In one possible implementation, in S107, the first extended query... For document collections A search was performed to obtain the second document set. and the corresponding second external concept set .

[0073] S108. Merge the second external concept set with the first external concept set to update the first concept set, generate the second extended query, and use the big language learning model to analyze the semantic consistency between the current extended query and the previous round of extended queries; repeat this step until the semantic consistency score is lower than the threshold or decreases, and return the previous round of extended queries as the final extended query.

[0074] In one possible implementation, S108 includes: S108a, Set the second external concept With the first external concept set Merge to update the complete set of first concepts Generate a second extended query; construct query extension evaluation hint statements. Utilizing a large language learning model to expand queries in the previous round Compared with the current extended query As input, the two are compared based on dimensions including semantic consistency, core concept overlap rate, and information gain, and a consistency score is output. S108b, If the current rating Lower than the previous round score or If the value falls below a preset threshold, semantic consistency is determined to have decreased, the iteration is terminated, and the previous round of expanded query is returned as the final expanded query; if If the expansion is reasonable, continue to the next round of concept expansion and optimization until it stops, and then obtain the final expanded query.

[0075] Furthermore, in S108a, the method for generating the extended query is the same as that in S106.

[0076] Furthermore, in S108a, the query expands the evaluation prompt statement. This includes instructing the large language model to evaluate the previous round of expanded queries in the role of an information retrieval expert. Compared with the current extended query The semantic relevance between them; specify the scoring criteria or dimensions, including coverage of core concepts, semantic coherence, and information value; specify that the scoring output is an integer, which can be set from 0 to 5. This is used to quantify the semantic consistency of extended queries; it requires the model to output only the scoring results without providing additional explanations or text, so that it can be used for iterative judgment.

[0077] For example, As shown in Table 3: Table 3 Query Extended Evaluation Prompt Statements

[0078]

[0079] Furthermore, in S108a, the consistency score is determined by a large language model using preset scoring prompts. The process is executed under constraints, wherein the scoring prompts are used to limit the scoring scale, scoring dimensions and output format; the large language model includes, but is not limited to, LLaMA series models, Qwen series models, GPT series models, Claude series models and other pre-trained or fine-tunable large-scale language models.

[0080] (1) Core concept coverage: Calculate the degree of overlap between the expanded query and the original query in terms of core concepts, reflecting the semantic preservation. (2) Semantic coherence: Evaluate whether the extended query naturally extends the original query intent and maintains semantic consistency; (3) Information value: measure whether the extended query provides useful supplementary information and avoids redundant or irrelevant content; (4) Scoring judgment and iteration control: integrate the various scoring dimensions into a consistent score. ;like < or If the value is below a preset threshold, the iteration stops and the previous round of expanded query is returned as the final expanded query; if... ≥ Then, the next round of concept expansion iterations will continue.

[0081] For example, the internal concept set With external concept set By merging, we obtain the complete set of concepts: ={Green tea, tea polyphenols, antioxidant effects, blood pressure regulation, cardiovascular health, caffeine effects, epidemiological studies, catechins, hypertension, dietary intervention}.

[0082] Generate extended queries based on the concatenation rules: = "Can drinking green tea lower blood pressure? Antioxidant effects of green tea polyphenols and catechins, and dietary intervention for hypertension."

[0083] The system expands queries For document collections A second round of retrieval was performed to obtain a new set of documents. .

[0084] For example, to prevent query semantic drift, this application embodiment introduces a semantic consistency evaluation mechanism based on a large language model. In the third iteration, the query is expanded: =“Green tea catechins dietary intervention for blood pressure regulation and health”, compared to the previous query: =The consistency score for "Green tea catechins for blood pressure health" is When the new concept of "caffeine effect" is further introduced: =“Green tea catechin dietary intervention for caffeine-induced blood pressure health”, the consistency score decreased to .because The system determines that semantic consistency has decreased, terminates the iteration, and returns to the third round of queries. As the final optimized query.

[0085] S109. Calculate the document relevance score based on the final expanded query and re-rank them to obtain the retrieval result set.

[0086] In one possible implementation, in S109, the final expanded query... The data is input into the information retrieval system, the document relevance scores are recalculated, and the results are reordered to obtain the final optimized set of retrieval results. .

[0087] It should be noted that the information retrieval system referred to here refers to traditional information retrieval systems based on inverted indexes and probabilistic models, such as retrieval systems based on BM25; semantic retrieval systems based on vector representations, such as similarity retrieval systems based on word vectors, sentence vectors, or document vectors; hybrid information retrieval systems that combine keyword retrieval and semantic retrieval; and other information retrieval systems that can calculate document relevance scores based on query statements and output ranking results.

[0088] This application's embodiments, without requiring additional model training, utilize a large language model to sequentially complete steps such as query concept generation, document concept extraction, iterative query optimization, and expansion evaluation, achieving adaptive optimization of query intent and high-precision retrieval result output. By introducing a multi-stage semantic expansion mechanism combining internal concept generation and external concept extraction, and a dynamic iterative control strategy based on semantic consistency scoring, this invention can effectively reduce interference caused by semantic drift and invalid expansion while maintaining the stability of the original query semantics, significantly improving the relevance, stability, and interpretability of the retrieval process.

[0089] The above process can be applied to semantic retrieval scenarios such as enterprise knowledge base retrieval, academic literature retrieval, question-answering systems, and recommendation systems. Through automated concept generation and iterative optimization mechanisms, it achieves accurate modeling and semantic expansion of user query intent. This solution can effectively improve the semantic matching degree between queries and documents in complex, multi-topic retrieval tasks, reduce manual intervention, and enhance the system's intelligence and interpretability.

[0090] The above process can also be applied to scenarios such as government information services, medical literature analysis, patent search, legal text retrieval, and intelligent decision support. By introducing an adaptive optimization and semantic consistency evaluation mechanism of a large language model, it ensures that the extended query continuously improves the relevance and accuracy of the search while maintaining semantic stability, thereby achieving highly reliable and controllable semantic information retrieval optimization.

[0091] To verify the effectiveness of the method provided in the embodiments of this application, comparative experiments were conducted on the public datasets TREC-COVID and NFCorpus. The experiments were divided into two groups: one group used the sparse retrieval methods BM25 and BM25 MF and the dense retrieval method Contriever (without concept expansion and iterative optimization), and the other group used the iterative concept generation scheme proposed in this invention, using the large language model Llama3-8B-Instruct.

[0092] The test results are presented below.

[0093] To verify the effectiveness of the method of this invention, comparative experiments were conducted on the publicly available datasets TREC-COVID and NFCorpus. Two schemes were set up for the experiment: one scheme used the traditional sparse retrieval methods BM25, BM25 MF, and the dense retrieval method Contriever, without concept expansion and iterative optimization; the other scheme used the iterative concept generation method based on a large language model proposed in the embodiments of this application, using the Llama3-8B-Instruct large language model. The evaluation metrics included mean accuracy (MAP), top 1k recall (R@1k), top 10 precision (P@10), and top 10 normalized depreciation cumulative gain (nDCG@10). The experimental results are shown in Table 1: Table 1. Performance comparison of iterative concept generation methods based on large language models on information retrieval tasks.

[0094] As shown in Table 1, the method provided in this application significantly outperforms traditional sparse and dense retrieval models on both publicly available datasets, especially in terms of MAP and nDCG@10 metrics. The results demonstrate that the method provided in this application, by introducing iterative concept generation and semantic consistency control mechanisms, can significantly improve the performance of large language models in information retrieval scenarios.

[0095] The iterative concept generation information retrieval device based on a large language model provided in this application is described below. The iterative concept generation information retrieval device based on a large language model described below and the iterative concept generation information retrieval method based on a large language model described above can be referred to in correspondence.

[0096] Figure 3 This is a schematic diagram of the structure of the iterative concept generation information retrieval device based on a large language model provided in the embodiments of this application, such as... Figure 3 As shown, it includes: a retrieval module 31, a first generation module 32, a second generation module 33, a fusion module 34, an extension module 35, an iteration module 36, and an output module 37, wherein: The retrieval module 31 is used to obtain a first document set based on the query statement input by the user; and to obtain a candidate document set through retrieval. The first generation module 32 is used to generate an internal concept set based on a query statement using a large language learning model; The second generation module 33 is used to construct a first instruction for concept extraction based on a candidate document set using a large language learning model; input the first instruction into the large language model to obtain a first concept set; construct concept relevance filtering prompts; and the instruction large language model filters out the most relevant first external concept set from the first concept set based on the semantic similarity between the query semantics and the semantics of the candidate concepts. The fusion module 34 is used to merge the internal concept set and the first external concept set into a first concept set and generate a first extended query; Extension module 35 is used to retrieve the candidate document set using the first extended query to obtain the second document set and the second external concept set; Iteration module 36 is used to merge the second external concept set with the first external concept set to update the first concept set, generate the second extended query, and use the big language learning model to analyze the semantic consistency between the previous extended query and the current extended query; repeat this step until the semantic consistency score is lower than the threshold or decreases, and return the previous extended query as the final extended query; Output module 37 is used to calculate document relevance scores based on the final expanded query and reorder them to obtain a set of search results.

[0097] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communications bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communications bus 440. The processor 410 can call logical instructions from the memory 430 to execute an iterative concept generation information retrieval method based on a large language model.

[0098] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the iterative concept generation information retrieval method based on a large language model provided by the above methods.

[0100] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the iterative concept generation information retrieval method based on a large language model provided by the above methods.

[0101] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0102] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An iterative concept generation information retrieval method based on a large language model, characterized in that, include: Based on the user's query, obtain the first set of documents; A set of candidate documents was obtained through retrieval; Utilize a large language learning model to generate an internal concept set based on query statements; We utilize a large language learning model based on a set of candidate documents to construct the first instruction for concept extraction; Input the first instruction into the large language model to obtain the first set of concepts; The concept relevance filtering prompt is constructed, and the instruction big language model filters out the most relevant first external concept set from the first concept set based on the semantics of the query and the semantic similarity of the candidate concepts. The internal concept set and the first external concept set are merged into the first concept set, and the first extended query is generated. The candidate document set is retrieved using the first extended query to obtain the second document set and the second external concept set; The second external concept set is merged with the first external concept set to update the first concept set, generating the second expanded query, and the semantic consistency between the previous expanded query and the current expanded query is analyzed using a large language learning model. Repeat this step until the semantic consistency score is below the threshold or decreases, and return the previous round of expanded queries as the final expanded query; The final expanded query is used to calculate document relevance scores and reorder them to obtain the set of search results.

2. The iterative concept generation information retrieval method according to claim 1, characterized in that, The method of generating an internal concept set based on a query statement using a large language learning model includes: A task template is generated based on a query statement using a large language learning model; a concept generation prompt statement is constructed by combining the query statement; the concept generation prompt statement includes a clear description of the concept generation task and the concatenation of the query statement. The concept generation prompt is input into the large language model, and concepts related to the query semantics are generated through reasoning to obtain an internal concept set.

3. The iterative concept generation information retrieval method according to claim 1, characterized in that, The method for constructing the first instruction includes: firstly describing the goals and requirements of the document concept extraction task in the prompt statement, and then concatenating the candidate document set with the query statement to obtain the first instruction.

4. The iterative concept generation information retrieval method according to claim 1, characterized in that, The concept relevance filtering prompt requires the large language learning model to act as an information retrieval expert, guiding the large language model to establish a mapping relationship between the query statement and candidate concepts. The construction method includes: first, describing the goals and requirements of the query statement and the concept relevance selection task in the prompt statement; then, concatenating the first concept set with the query statement to obtain the concept relevance filtering prompt.

5. The iterative concept generation information retrieval method according to claim 1, characterized in that, The method for generating the first extended query includes: concatenating and combining the query statement with the complete set of first concepts in sequence to obtain the first extended query.

6. The iterative concept generation information retrieval method according to claim 1, characterized in that, The process involves merging the second external concept set with the first external concept set to update the first concept set, generating a second expanded query, and using a large language learning model to analyze the semantic consistency between the previous expanded query and the previous round of expanded queries. Repeat this step until the semantic consistency score falls below the threshold or decreases, then return the previous round's expanded query as the final expanded query, including: Merge the second external concept set with the first external concept set to update the complete first concept set, generating the second extended query; Construct query expansion evaluation prompts, use a large language learning model to take the previous round of expanded queries and the current expanded queries as inputs, and compare the two based on dimensions including semantic consistency, core concept overlap rate and information gain, and output a consistency score; If the current rating Lower than the previous round score or If the value falls below a preset threshold, semantic consistency is determined to have decreased, the iteration is terminated, and the previous query is returned as the final expanded query; if If the expansion is reasonable, continue to the next round of concept expansion and optimization until it stops, and then obtain the final expanded query.

7. An information retrieval device based on iterative concept generation using a large language model, characterized in that, include: The retrieval module is used to obtain the first set of documents based on the query statement entered by the user; A set of candidate documents was obtained through retrieval; The first generation module is used to generate an internal concept set based on the query statement using a large language learning model; The second generation module is used to construct a first instruction for concept extraction based on a candidate document set using a large language learning model; input the first instruction into the large language model to obtain a first concept set; construct concept relevance filtering prompts; and the instruction large language model filters out the most relevant first external concept set from the first concept set based on the semantic similarity between the query semantics and the semantics of the candidate concepts. The fusion module is used to merge the internal concept set and the first external concept set into a first concept set and generate a first extended query; The extension module is used to retrieve the candidate document set using the first extended query to obtain the second document set and the second external concept set; The iterative module is used to merge the second external concept set with the first external concept set to update the first concept set, generate the second extended query, and use a large language learning model to analyze the semantic consistency between the previous extended query and the current extended query. Repeat this step until the semantic consistency score is below the threshold or decreases, then return the previous round of expanded query as the final expanded query; The output module is used to calculate document relevance scores based on the final expanded query and reorder them to obtain a set of search results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the iterative concept generation information retrieval method based on a large language model as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the iterative concept generation information retrieval method based on a large language model as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the iterative concept generation information retrieval method based on a large language model as described in any one of claims 1 to 6.

Images