Multipath information retrieval method and system based on language model and uncertainty evaluation

Abstract

The invention provides a multi-path information retrieval method and system based on a language model and uncertainty evaluation. The multi-path information retrieval method comprises the following steps: receiving a user query; performing semantic analysis on user query by using a pre-trained language model so as to output a quantitative uncertainty index in an unsupervised manner; dynamically determining a quantity ratio for controlling query generation and a balance parameter for controlling result sorting based on the index; based on the user query and the strategy parameters, generating and executing two groups of differential queries at least comprising exploratory queries and utilizable queries to obtain multiple paths of retrieval results; and combining multiple paths of results, and carrying out dynamic balance reordering between the correlation and diversity of the results according to balance parameters by adopting an ordering mechanism. According to the method, the retrieval strategy can be adaptively adjusted according to the uncertainty of query, and the coverage degree and diversity of results are remarkably improved.

Description

A Multi-Path Information Retrieval Method and System Based on Language Model and Uncertainty Assessment Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to a multi-path information retrieval method and system based on language models and uncertainty assessment. Background Technology

[0002] Information retrieval technologies, such as common search engines, are essential tools for users to obtain information. Existing information retrieval systems primarily rely on keyword matching, link analysis, and traditional machine learning models for result ranking. These methods perform well when handling user queries with clear intent and structured information. However, their limitations become increasingly apparent when faced with semantically complex, ambiguous, or multidimensional information-demanding user queries. For example, a user might enter a broad topic, hoping to understand different aspects of that topic; in such cases, keyword matching alone often fails to comprehensively cover all relevant information.

[0003] To improve search performance, the industry has proposed several optimization techniques, such as query expansion or relevance feedback. These techniques typically analyze user queries or preliminary search results to automatically add synonyms, related words, or reconstruct the query to broaden the search scope. However, these techniques largely rely on pre-built knowledge bases or large amounts of manually labeled training data, resulting in high deployment and maintenance costs, poor flexibility, and essentially remaining at a superficial level of adjusting query strings or feature weights, lacking a deep semantic understanding of the user's query intent.

[0004] With the development of large-scale language model technology, its powerful natural language understanding and generation capabilities have brought new possibilities to the field of information retrieval. Some studies have begun to explore using language models for simple query rewriting or summary generation. Furthermore, some technical solutions propose using language models to evaluate certain attributes of text, such as complexity and specialization, and selecting different levels of subsequent processing strategies based on the evaluation results. However, these existing technologies have significant shortcomings: First, their strategy adjustments are general and unidirectional; for example, they can only decide whether to perform "lightweight analysis" or "deep analysis," without revealing how to translate the evaluation results into specific, quantitative, and coordinated control over the core trade-off between "exploration" (broadening search breadth) and "utilization" (deepening search depth) in information retrieval. Second, existing technologies lack a closed-loop control framework that integrates query uncertainty assessment, multi-path query generation, and final result ranking. Specifically, existing technologies fail to provide a mechanism that can utilize a single query uncertainty metric to dynamically guide both front-end query generation strategies (e.g., the ratio of exploratory to exploitative queries) and back-end ranking and fusion strategies (e.g., balancing relevance and diversity of results), thus failing to systematically address the problems of incomplete information coverage and insufficient diversity under complex queries. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a multi-path information retrieval method and system based on language models and uncertainty assessment. This method and system are intended to solve the technical problems of incomplete information coverage and insufficient result diversity in the processing of complex and fuzzy queries, which are caused by the lack of an intelligent decision-making mechanism driven by deep semantic understanding that can dynamically and collaboratively balance the breadth and depth of the search based on the inherent uncertainty of the query.

[0006] A multi-path information retrieval method based on language models and uncertainty assessment, provided by the present invention, includes the following steps: First, receiving a user query; then, performing semantic analysis on the user query using a pre-trained language model to output a quantified uncertainty index in an unsupervised manner; next, dynamically determining at least one strategy parameter based on the quantified uncertainty index, the strategy parameter including a query generation ratio for controlling subsequent query generation and a ranking balance parameter for controlling result ranking; subsequently, generating at least two sets of differentiated queries based on the user query and the strategy parameter, including at least one set of exploratory queries aimed at broadening the search breadth and one set of utilization queries aimed at deepening the search depth, and executing the at least two sets of queries to obtain multi-path retrieval results; then, merging the multi-path retrieval results and re-ranking them using a ranking mechanism, wherein the ranking mechanism dynamically balances result relevance and diversity based on the ranking balance parameter; finally, outputting the re-ranked results.

[0007] Optionally, the pre-trained language model is a large-scale language model.

[0008] Optionally, the step of using a pre-trained language model to perform semantic analysis on the user query includes: instructing the pre-trained language model to comprehensively evaluate the semantic ambiguity, domain specialization, and breadth of information needs of the user query.

[0009] Optionally, before generating at least two sets of differentiated queries, the method further includes: performing a benchmark search using the user query to obtain a benchmark result set; and the step of generating at least two sets of differentiated queries is based on the user query and the benchmark result set. This provides richer contextual information for subsequent query generation, thereby improving the quality of the generated queries.

[0010] Optionally, the exploratory query is generated by instructing the pre-trained language model to perform semantic expansion, condition relaxation, or multi-perspective reconstruction; the exploitation query is generated by instructing the pre-trained language model to refine the query, add specific limiting conditions, or decompose the problem. By submitting differentiated instructions to the language model, the type of generated query can be effectively controlled to meet the needs of exploration and exploitation respectively.

[0011] Optionally, the ranking mechanism is the Maximum Marginal Relevance (MMR) algorithm, and the ranking balance parameter is the balance parameter λ used in the MMR algorithm to balance correlation and diversity.

[0012] Optionally, the step of dynamically determining at least one strategy parameter includes: comparing the quantified uncertainty index with at least one preset threshold to determine the query generation quantity ratio and the sorting balance parameter based on the comparison result.

[0013] Optionally, the step of determining the strategy parameters based on the comparison results includes at least one of the following: when the quantified uncertainty index is higher than a preset high uncertainty threshold, determining that the query generation ratio is more exploratory queries than exploitative queries, and determining that the ranking balance parameter is a value that favors diversity; when the quantified uncertainty index is lower than a preset low uncertainty threshold, determining that the query generation ratio is more exploitative queries than exploratory queries, and determining that the ranking balance parameter is a value that favors relevance.

[0014] A multi-path information retrieval system based on language models and uncertainty assessment, according to the present invention, comprises: a receiving unit for receiving user queries; an evaluation unit for performing semantic analysis on the user queries using a pre-trained language model to output a quantified uncertainty index in an unsupervised manner; a strategy determination unit for dynamically determining at least one strategy parameter based on the quantified uncertainty index, the strategy parameter including a query generation ratio for controlling the generation of subsequent queries and a ranking balance parameter for controlling the ranking of results; a query execution unit for generating at least two sets of differentiated queries based on the user queries and the strategy parameters, including at least one set of exploratory queries aimed at broadening the search breadth and one set of utilization queries aimed at deepening the search depth, and executing the at least two sets of queries to obtain multi-path retrieval results; a ranking unit for merging the multi-path retrieval results and re-ranking them using a ranking mechanism, wherein the ranking mechanism dynamically balances result relevance and diversity based on the ranking balance parameter; and an output unit for outputting the re-ranked results.

[0015] Optionally, the pre-trained language model is a large-scale language model; the step of using the pre-trained language model to perform semantic analysis on the user query includes: instructing the pre-trained language model to comprehensively evaluate the semantic ambiguity, domain specialization, and breadth of information needs of the user query.

[0016] Compared with existing technologies, the present invention has the following beneficial effects: 1. The present invention constructs a closed-loop adaptive retrieval framework with query uncertainty as the unified control signal, deeply integrating the deep semantic understanding capabilities of large-scale language models with the "exploration-utilization" trade-off decision-making of information retrieval. For complex queries with fuzzy semantics, the system can automatically broaden the search scope and provide more diverse information perspectives; for specific queries with clear intent, it can automatically deepen the search depth and provide more accurate answers, improving the coverage, depth, and diversity of retrieval results, thereby improving the user experience.

[0017] 2. This invention mainly achieves complex strategy adjustments by submitting structured instructions to a pre-trained language model, replacing the complex feature engineering and model training process in traditional methods. It can be applied without complex initial configuration or training, reducing the implementation, deployment and maintenance costs of the system, and has good generalization ability and robustness for new domains and new forms of queries.

[0018] 3. This invention uses the query uncertainty assessment result as a unified control signal to collaboratively regulate the two key stages of query generation (input) and result ranking (output), achieving end-to-end strategy adaptation. This systematic integration and coupling design is a significant improvement over the unidirectional and isolated optimization strategies in existing technologies, making information retrieval systems more intelligent. Attached Figure Description

[0019] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 is a schematic diagram of the structure of an information retrieval system provided by an embodiment of the present invention; Figure 2 is a schematic flowchart of an information retrieval method provided by an embodiment of the present invention; Figure 3 is a detailed schematic flowchart of the two-stage search steps in Figure 2; Figure 4 is a timing diagram of the signaling interaction between entities in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0021] Example 1 details a multi-path information retrieval method and its basic implementation scheme based on language models and uncertainty assessment. Specifically, this scheme is applicable to scenarios where users input complex queries with high uncertainty. This example aims to fully demonstrate how this method adaptively adjusts the search strategy to emphasize "exploration" by assessing the uncertainty of the query, thereby providing users with a comprehensive and diverse set of information results.

[0022] Please refer to Figure 1, which shows a preferred structural diagram of an information retrieval system provided by an embodiment of the present invention. The system can be deployed on a cloud server, a local server cluster, or any computing device with sufficient computing power. Specifically, the system mainly includes: a query understanding and uncertainty assessment module 10, a two-stage search module 20, an adaptive strategy control module 30, and a result fusion and diversity ranking module 40. It should be noted that these modules can be implemented as software program modules, hardware circuits, or a combination thereof. The system communicates with an external pre-trained language model (e.g., a large language model with hundreds of billions or more parameters) through a large language model interface 50, and accesses the information resources to be retrieved through a data index 60. The data index 60 can be any form of structured or unstructured data collection, such as a webpage index, an internal enterprise document library, or an academic paper database.

[0023] The working process of this embodiment will be described in detail below with reference to the overall flowchart of the information retrieval method provided by the embodiment of the present invention shown in Figure 2, and the timing diagram of the signaling interaction between various entities shown in Figure 4.

[0024] The method first executes step S100, receiving a user query. In this embodiment, it is assumed that the user submits a query request to the system through a user terminal device (as shown in Figure 1), the content of which is: "Latest developments in new energy vehicle battery technology". It can be understood that this is a typical query with high uncertainty, because it involves a rapidly developing technical field, containing multiple sub-topics (such as different types of batteries, technical bottlenecks, commercial applications, etc.), and the expression "latest developments" itself is open-ended and ambiguous.

[0025] Accordingly, in step S200, the query uncertainty is evaluated. Upon receiving the query, the core task of the query understanding and uncertainty evaluation module 10 is to output a quantified uncertainty index in an unsupervised manner. In one specific implementation of the invention, this process is accomplished through interaction with a large language model (LLM). As shown in Figure 4, the system initiates an uncertainty evaluation request to the LLM. Specifically, the query understanding and uncertainty evaluation module 10 constructs a structured instruction (also called a prompt word) and sends it to the LLM through the LLM interface 50. This instruction guides the model to utilize its inherent vast world knowledge and reasoning capabilities to perform deep semantic analysis of the query. An exemplary instruction could be designed as follows: "[Role-playing and Task Instruction] You are now a top information retrieval and analysis expert. Your task is to evaluate the 'uncertainty' of the following user query. In your evaluation, please consider the following three dimensions: 1. Semantic ambiguity: Does the query contain polysemous words, ambiguous referents, or broad concepts? 2. Domain specialization: Does the query involve a domain requiring in-depth expertise to fully understand? 3. Breadth of information needs: Does the query imply that the user wants to understand multiple aspects, different perspectives, or a complete thematic picture, rather than a single fact? Based on your comprehensive judgment, output a decimal between 0.0 and 1.0 as the final quantitative uncertainty score. The higher the score, the greater the uncertainty. Note that only the score value needs to be output; no explanation or analysis process is required. [User Query to be Evaluated]: Latest Developments in New Energy Vehicle Battery Technology" Upon receiving this instruction, the large language model will analyze the query "Latest Developments in New Energy Vehicle Battery Technology". The model might recognize that "battery technology" is a broad, overarching concept, which can be further subdivided into lithium-ion batteries, solid-state batteries, sodium-ion batteries, hydrogen fuel cells, etc.; while "latest developments" implies a need for cutting-edge, dynamic information, rather than a static fact. Considering these factors, the model judges the query to have moderate semantic ambiguity, high domain specialization, and extremely broad information requirements. Therefore, the model returns a high uncertainty score through the large language model interface 50, for example, a quantified uncertainty index U(q) = 0.75. It should be noted that this process is unsupervised; it does not rely on any pre-labeled (query, uncertainty score) training sample pairs, but directly utilizes the capabilities of the pre-trained language model itself.

[0026] After determining the query uncertainty, the process proceeds to step S300 to determine the strategy parameters. The adaptive strategy control module 30 receives the uncertainty index U(q) = 0.75 from the query understanding and uncertainty assessment module 10. This module internally presets at least one, and typically two, uncertainty thresholds, such as a high uncertainty threshold θ. high(e.g., 0.7) and low uncertainty threshold θ low (e.g., 0.25). The adaptive policy adjustment module 30 compares the received U(q) with these thresholds. In this example, since 0.75 > θ high (0.7) The system determines that the current query belongs to the high uncertainty category, so the "exploration-oriented" retrieval strategy should be adopted.

[0027] Based on this judgment, the adaptive strategy control module 30 dynamically determines a set of strategy parameters. These strategy parameters include at least: 1. The query generation ratio, i.e., the number N of exploratory queries to be generated subsequently. explore And the number of utilization queries N exploit For queries with high uncertainty, the system will allocate more resources to the exploration path, for example, by setting N. explore = 3, N exploit = 1. 2. The sorting balance parameter λ, which will be used in the subsequent result fusion and sorting stages. For queries with high uncertainty, to encourage diversity of results, a smaller λ value should be set, for example, λ = 0.3. This smaller λ value means that during sorting, the system will more strongly penalize candidates similar to the selected results, thus prioritizing documents that can bring new information. These determined strategy parameters (N explore =3, N exploit =1, λ=0.3) will be distributed to subsequent related modules.

[0028] Subsequently, in step S400, a two-stage search is performed by the two-stage search module 20. The detailed process can be found in Figure 3.

[0029] The first stage is the baseline search, corresponding to step S410 in Figure 3. The two-stage search module 20 first uses the user-inputted original query "Latest Developments in New Energy Vehicle Battery Technology" to directly perform a routine keyword matching or vector retrieval in the data index 60. This aims to quickly obtain a basic set of results most relevant to the literal meaning of the query, i.e., the baseline result set. As shown in step S420 of Figure 3, the system obtains this baseline result set, which may contain, for example, 20 of the most relevant technical documents, news reports, or research papers. This baseline result set is not only part of the final result itself, but more importantly, it provides valuable contextual information for the enhanced query generation in the second stage.

[0030] The second stage is a parallel dual-path search, corresponding to the parallel generation of enhanced queries step S430 in Figure 3. The two-stage search module 20 utilizes a large language model to generate two sets of differentiated new queries in parallel, based on the original query and the baseline result set (or its summary) obtained in the first stage.

[0031] On the exploration path, according to N set by the adaptive strategy control module 30 explore = 3. The two-stage search module 20 submits an instruction specifically designed for the "exploration" task to the large language model to execute step S431. This instruction requires the model to generate three queries aimed at broadening the search boundaries based on the original query and the baseline result summary, through semantic expansion, perspective switching, or condition relaxation. For example, the instruction might be as follows: "[Exploratory Query Generation Instruction] Based on the user query 'Latest Developments in New Energy Vehicle Battery Technology' and the following preliminary search result summary, please generate three related queries with different perspectives or broader fields to help the user explore related topics. The queries should avoid being too similar to the original query. Example summary: '...Solid-state battery energy density breakthrough...thermal runaway safety research...'". The model might generate the following exploratory queries: 1. "Technological innovation of next-generation energy storage systems for electric vehicles" (expanding from "battery" to "energy storage system"); 2. "Technological breakthroughs and challenges of hydrogen fuel cells in the automotive field" (switching to alternative technology routes); 3. "Research progress of key materials in the global power battery industry chain" (expanding from the technology itself to the upstream of the industry chain). Subsequently, in step S432, the system uses these three newly generated exploratory queries to initiate searches in parallel with the data index 60, obtaining three exploratory search results.

[0032] At the same time, on the utilization path, according to the set N exploit = 1. The two-stage search module 20 submits a differentiated instruction designed for the "utilization" task to the large language model to execute step S441. This instruction requires the model to generate a query aimed at deepening the core topic by adding constraints, decomposing the problem, or refining the query. For example, the instruction might be: "[Utilization Query Generation Instruction] Based on the user query 'Latest Developments in New Energy Vehicle Battery Technology' and the preliminary result summary, please generate a more specific and in-depth query, focusing on the most core technological direction." The model might generate a utilization query like: "Analysis of the Latest Improvement Paths for Solid-State Lithium Battery Energy Density and Cycle Life." Subsequently, in step S442, the system uses this new query to initiate a search in the data index 60, obtaining a utilization search result.

[0033] At this point, step S400 is complete, and the system has obtained a total of 5 search results: 1 baseline search result, 3 exploratory search results, and 1 utilization search result.

[0034] Subsequently, in step S500, the result fusion and diversity ranking module 40 is responsible for merging the above 5 search results and performing deduplication to form a large set of candidate results.

[0035] Next, in the crucial step S600, dynamic weighted sorting is performed. The result fusion and diversity sorting module 40 uses a dynamically weighted maximum marginal relevance algorithm to reorder the candidate result set. The goal of this algorithm is to iteratively select documents d from the candidate document set D to maximize an objective function. The mathematical expression of this objective function is:

[0036] Where D is the candidate document set, S is the selected result set, and q is the original user query. Calculate the relevance score between candidate document d and the original query q, which can be obtained using the traditional BM25 algorithm or the modern vector cosine similarity. Calculate the relationship between candidate document d and a document in the selected result set S. The similarity score can also be calculated using methods such as vector similarity.

[0037] It should be noted that in the scheme of this invention, the balance parameter λ is not a fixed hyperparameter, but a value dynamically determined by the adaptive strategy control module 30 in step S300 based on the uncertainty score U(q). In this embodiment, λ is set to 0.3. This smaller λ value increases the second term in the formula, namely the diversity penalty term. The ranking process assigns weights to documents. In other words, during the ranking process, if a candidate document is very similar to the selected results, it will be heavily penalized, its final score will be lower, and it will be less likely to be selected. Conversely, documents that differ significantly from the selected results but are still somewhat relevant to the original query have a greater chance of being ranked higher. This mechanism effectively promotes diversity in the results.

[0038] Finally, in step S700, the results are output. The results fusion and diversity sorting module 40 presents the final result sequence, reordered by the dynamically weighted MMR algorithm, to the user. Due to the adoption of an "exploration-oriented" strategy and a diversity-first sorting method, the final output result list not only includes in-depth articles on core technologies such as solid-state lithium batteries (from the utilization path), but also broadly covers related or alternative technology directions such as energy storage system innovation, hydrogen fuel cells, and industrial chain materials (from the exploration path), thereby constructing a more comprehensive and multi-dimensional information landscape for users to meet their needs for exploratory understanding of a complex topic.

[0039] Example 2 This example aims to illustrate the adaptive capability of the technical solution provided by the present invention in handling low-uncertainty, factual queries. In such scenarios, the user's intent is very clear, and the system should automatically switch to a "focused utilization" strategy to provide accurate and direct answers and avoid unnecessary information diffusivity.

[0040] The system structure used in this embodiment is the same as that shown in Figure 1 of Embodiment 1, and also includes a query understanding and uncertainty assessment module 10, a two-stage search module 20, an adaptive strategy control module 30, and a result fusion and diversity ranking module 40, and interacts with a large language model interface 50 and a data index library 60. Its overall method flow also follows the steps shown in Figure 2.

[0041] In step S100, the system receives a new user query: "Who is the current CEO of Apple?" This is a typical fact-based query with extremely low uncertainty.

[0042] In step S200, the query understanding and uncertainty assessment module 10 packages this query into instructions similar to those in Example 1 and sends them to the large language model for evaluation. The large language model analyzes that the query points to a very specific, single entity attribute, is semantically clear and unambiguous, and has a very narrow information requirement. Therefore, the model returns a very low uncertainty score, for example, U(q) = 0.1.

[0043] In step S300, the adaptive strategy control module 30 receives U(q) = 0.1 and compares it with a preset low uncertainty threshold θ. low（ For example, compare it with 0.25). Since 0.1 < θ low (0.25), the system determines that the current query belongs to the low uncertainty category and should adopt the "emphasis on utilization" retrieval strategy.

[0044] Accordingly, the adaptive strategy control module 30 dynamically determines a set of strategy parameters different from those in Embodiment 1. Specifically, it sets the number of exploratory queries N. explore = 1, number of utilization queries N exploit = 3, and a ranking balancing parameter λ = 0.7 that tends to favor relevance. The logic behind these parameter settings is that for explicit queries, only minimal or even no exploration is needed, and resources should be concentrated on in-depth mining of the core topic; at the same time, the ranking should prioritize the most relevant results, hence the larger λ value, which increases the relevance factor. Weights in the MMR formula.

[0045] In step S400, the two-stage search module 20 begins to perform the search. First, in steps S410 and S420, the system performs a baseline search using "Who is the current CEO of Apple?" as the query term, and obtains a baseline result set containing information related to "Tim Cook".

[0046] Next, in the parallel dual-path search phase (step S430), the two-stage search module 20 performs the search according to the new strategy parameters (N). explore =1, Nexploit =3) Request query generation from the large language model. For the exploration path (step S431), the model is instructed to generate one lightly relevant exploratory query, such as "a list of Apple's historical CEOs". For the exploitation path (step S441), the model is instructed to generate three exploitation queries aimed at deeply mining core facts, such as: "1. From what year did Tim Cook serve?"; "2. Key achievements of Apple under Tim Cook's leadership"; "3. Tim Cook's recent public statements". Subsequently, the system executes these four new queries and the baseline query retrieval in parallel (steps S432 and S442).

[0047] In step S500, all search results are merged and deduplicated.

[0048] In step S600, the result fusion and diversity ranking module 40 applies the MMR algorithm for ranking. At this point, the balancing parameter λ is set to 0.7. This relatively large λ value makes the MMR formula more focused on maximizing the relevance score between candidate documents and the original query. The penalties for diversity are relatively small.

[0049] Finally, in step S700, the system output list will highly focus on information about Tim Cook himself and his role as CEO, with the most directly relevant documents such as his profile, tenure, and achievements placed at the top. Information about other former Apple CEOs or other aspects of the company will be relegated to a secondary position or filtered out due to its lower relevance to the original query. This result quickly and accurately meets the user's need for specific factual information, demonstrating the adaptability and efficiency of the present invention.

[0050] Example 3 illustrates a variant implementation of the technical solution of the present invention, specifically a hybrid implementation of the "parallel dual-path search" in step S400. This example aims to demonstrate that the core conceptual framework of the present invention has good openness and compatibility; its "path-based" query generation is not limited to using a large language model and can also be combined with other mature technologies in the field of information retrieval (such as pseudo-relevance feedback), thereby potentially achieving higher efficiency or effectiveness in specific scenarios.

[0051] The system structure of this embodiment is basically the same as that of embodiment 1, but in the two-stage search module 20, in addition to the ability to interact with the large language model interface 50, it also integrates the function of implementing the pseudo-correlation feedback algorithm.

[0052] As an alternative implementation, the present inventors again use the aforementioned high-uncertainty query "latest developments in new energy vehicle battery technology" as an example. The uncertainty assessment (step S200) and strategy determination (step S300) processes are exactly the same as in Example 1. That is, the system evaluates U(q) = 0.75 and determines the "exploration-oriented" strategy with parameter N. explore = 3, N exploit = 1, λ = 0.3.

[0053] The difference in this embodiment lies in the second stage of step S400, namely the parallel generation of enhanced query step S430.

[0054] The implementation of the exploration path is consistent with that of Example 1. The two-stage search module 20 still calls the large language model to generate three semantically expanded exploratory queries based on the instructions (step S431), such as "latest innovations in electric vehicle energy storage systems", and performs the retrieval (step S432).

[0055] However, this embodiment employs a different technique for exploitation paths. The two-stage search module 20 no longer invokes a large language model to generate exploitation queries; instead, it initiates a traditional pseudo-relevance feedback mechanism to execute step S441. The specific process is as follows: First, the baseline result set obtained in the first-stage baseline search (step S410) is analyzed.

[0056] Then, select the K highest-ranking documents in the result set (e.g., K=5), and assume that these K documents are highly relevant.

[0057] Next, the system performs text analysis on the full text or abstracts of these K documents, extracting frequently occurring and highly distinctive technical terms or keywords. For example, the system might extract terms such as "energy density," "thermal runaway," "silicon-carbon anode," "cycle life," and "BMS."

[0058] Finally, these extracted core terms are combined with the user's original query to automatically construct one or more exploitation queries. Based on the strategy parameter Nexploit = ​​1, the system can generate a combined query, such as: "New energy vehicle battery energy density thermal runaway safety". Alternatively, if Nexploit > 1, multiple queries can be generated, such as "New energy vehicle battery energy density improvement" and "New energy vehicle battery thermal runaway solution".

[0059] After generating the utilization query, the system performs the retrieval of the query in step S442.

[0060] The subsequent steps S500 (result fusion) and S600 (dynamic weighted sorting, λ=0.3) are exactly the same as in Example 1. The final output (step S700) also achieves a balance between exploration and utilization, providing users with comprehensive information.

[0061] This embodiment demonstrates the flexibility of the framework proposed in this invention. By replacing the generation method of exploitation queries with pseudo-relevance feedback, the cost of calling large language model APIs can be reduced. Furthermore, since the query terms generated by pseudo-relevance feedback are directly derived from high-quality documents in the data index 60, these queries may, in some cases, better match the feature distribution of the data source, thus potentially achieving comparable or better exploitation results. This indicates that the scope of protection of this invention should not be limited to using only language models to generate all enhanced queries, but rather covers scenarios where uncertainty evaluation results are used to guide the generation of any form of differentiated exploratory and exploitation queries.

[0062] Example 4: Building upon the previous examples, this example introduces an optional feedback optimization mechanism, demonstrating how the system of this invention utilizes the "reflection" and "metacognition" capabilities of a large language model for self-iteration and optimization. This mechanism aims to automatically or semi-automatically adjust the system's internal strategy parameters by analyzing historical retrieval cases, thereby continuously improving retrieval performance and reducing reliance on manual parameter tuning and long-term maintenance costs.

[0063] As an optional implementation, this embodiment adds an offline "strategy optimization module" to the structure shown in Figure 1. This module does not participate in real-time online query processing, but performs optimization tasks periodically (e.g., daily or weekly). Its working process can be described as follows: Data collection: During normal operation, the system records key information for each retrieval session, including: the user's original query, the uncertainty score U(q) evaluated by the system, and the strategy parameters (N) used at that time. explore , N exploit The strategy optimization module considers the final ranking results and user interaction behavior (such as which links were clicked, how long after which a new query was initiated). It filters out inefficient retrieval cases. Understandably, the criteria for judging inefficient cases can be varied, such as: extremely low click-through rate on the results page; users initiating new, semantically similar queries very quickly after viewing the results (indicating the initial results failed to meet their needs); or users rapidly flipping through multiple pages on the results page before making a click.

[0064] Feedback and Reflection: The strategy optimization module periodically packages a batch of collected inefficient retrieval cases. For each case, it organizes it into a structured text containing all the information collected in the first step.

[0065] Optimization suggestion generation: The strategy optimization module submits these packaged cases to the large language model through the large language model interface 50, along with a carefully designed "system optimization instruction." This instruction requires the model to act as a system diagnostic expert, analyze the possible reasons for retrieval failures, and propose specific optimization suggestions. An exemplary optimization instruction might look like this: "[System Optimization and Diagnosis Instruction] You are a senior optimization expert for an information retrieval system. Below are some recent retrieval cases with poor performance. For each case, this invention provides the [original query], [system-assessed uncertainty U(q)], [the strategy parameters used at the time (Nexplore, Nexploit, λ)], and [a description of the poor performance, such as low user click-through rate]. Your task is to: 1. Analyze the possible reasons for the poor performance. Is it due to inaccurate uncertainty assessment? Or an inappropriate query generation strategy? Or unsuitable sorting parameters? 2. Based on your analysis, please provide specific and actionable adjustment suggestions. Suggestions can address the following: a) Adjust the thresholds θhigh and θlow used to distinguish between high / low uncertainty. b) Adjust the default values ​​of the strategy parameters (Nexplore, Nexploit, λ) corresponding to different uncertainty levels (high / medium / low). Please provide your analysis and suggestions in a clear format."

[0066]

Case 1

[0067] By introducing this feedback optimization mechanism, the system of this invention possesses the ability to evolve from its own invention and continuously learn. It can learn from historical experience and continuously optimize its core decision-making logic, resulting in an increasingly higher degree of alignment between uncertainty assessment and subsequent strategy formulation. This not only gradually improves the system's retrieval performance and user satisfaction but also greatly reduces the need for tedious manual parameter tuning, thereby enhancing the system's long-term stability and maintainability.

[0068] Example 1 details a multi-path information retrieval method and its basic implementation scheme based on language models and uncertainty assessment. Specifically, this scheme is applicable to scenarios where users input complex queries with high uncertainty. This example aims to fully demonstrate how this method adaptively adjusts the search strategy to emphasize "exploration" by assessing the uncertainty of the query, thereby providing users with a comprehensive and diverse set of information results.

[0069] Please refer to Figure 1, which shows a preferred structural diagram of an information retrieval system provided by an embodiment of the present invention. The system can be deployed on a cloud server, a local server cluster, or any computing device with sufficient computing power. Specifically, the system mainly includes: a query understanding and uncertainty assessment module 10, a two-stage search module 20, an adaptive strategy control module 30, and a result fusion and diversity ranking module 40. It should be noted that these modules can be implemented as software program modules, hardware circuits, or a combination thereof. The system communicates with an external pre-trained language model (e.g., a large language model with hundreds of billions or more parameters) through a large language model interface 50, and accesses the information resources to be retrieved through a data index 60. The data index 60 can be any form of structured or unstructured data collection, such as a webpage index, an internal enterprise document library, or an academic paper database.

[0070] The working process of this embodiment will be described in detail below with reference to the overall flowchart of the information retrieval method provided by the embodiment of the present invention shown in Figure 2, and the timing diagram of the signaling interaction between various entities shown in Figure 4.

[0071] The method first executes step S100, receiving a user query. In this embodiment, it is assumed that the user submits a query request to the system through a user terminal device (as shown in Figure 1), the content of which is: "Latest developments in new energy vehicle battery technology". It can be understood that this is a typical query with high uncertainty, because it involves a rapidly developing technical field, containing multiple sub-topics (such as different types of batteries, technical bottlenecks, commercial applications, etc.), and the expression "latest developments" itself is open-ended and ambiguous.

[0072] Accordingly, in step S200, the query uncertainty is evaluated. Upon receiving the query, the core task of the query understanding and uncertainty evaluation module 10 is to output a quantified uncertainty index in an unsupervised manner. In one specific implementation of the invention, this process is accomplished through interaction with a large language model (LLM). As shown in Figure 4, the system initiates an uncertainty evaluation request to the LLM. Specifically, the query understanding and uncertainty evaluation module 10 constructs a structured instruction (also called a prompt word) and sends it to the LLM through the LLM interface 50. This instruction guides the model to utilize its inherent vast world knowledge and reasoning capabilities to perform deep semantic analysis of the query. An exemplary instruction could be designed as follows: "[Role-playing and Task Instruction] You are now a top information retrieval and analysis expert. Your task is to evaluate the 'uncertainty' of the following user query. In your evaluation, please consider the following three dimensions: 1. Semantic ambiguity: Does the query contain polysemous words, ambiguous referents, or broad concepts? 2. Domain specialization: Does the query involve a domain requiring in-depth expertise to fully understand? 3. Breadth of information needs: Does the query imply that the user wants to understand multiple aspects, different perspectives, or a complete thematic picture, rather than a single fact? Based on your comprehensive judgment, output a decimal between 0.0 and 1.0 as the final quantitative uncertainty score. The higher the score, the greater the uncertainty. Note that only the score value needs to be output; no explanation or analysis process is required. [User Query to be Evaluated]: Latest Developments in New Energy Vehicle Battery Technology" Upon receiving this instruction, the large language model will analyze the query "Latest Developments in New Energy Vehicle Battery Technology". The model might recognize that "battery technology" is a broad, overarching concept, which can be further subdivided into lithium-ion batteries, solid-state batteries, sodium-ion batteries, hydrogen fuel cells, etc.; while "latest developments" implies a need for cutting-edge, dynamic information, rather than a static fact. Considering these factors, the model judges the query to have moderate semantic ambiguity, high domain specialization, and extremely broad information requirements. Therefore, the model returns a high uncertainty score through the large language model interface 50, for example, a quantified uncertainty index U(q) = 0.75. It should be noted that this process is unsupervised; it does not rely on any pre-labeled (query, uncertainty score) training sample pairs, but directly utilizes the capabilities of the pre-trained language model itself.

[0073] After determining the query uncertainty, the process proceeds to step S300 to determine the strategy parameters. The adaptive strategy control module 30 receives the uncertainty index U(q) = 0.75 from the query understanding and uncertainty assessment module 10. This module internally presets at least one, and typically two, uncertainty thresholds, such as a high uncertainty threshold θ. high(e.g., 0.7) and low uncertainty threshold θ low (e.g., 0.25). The adaptive policy adjustment module 30 compares the received U(q) with these thresholds. In this example, since 0.75 > θ high (0.7) The system determines that the current query belongs to the high uncertainty category, so the "exploration-oriented" retrieval strategy should be adopted.

[0074] Based on this judgment, the adaptive strategy control module 30 dynamically determines a set of strategy parameters. These strategy parameters include at least: 1. The query generation ratio, i.e., the number N of exploratory queries to be generated subsequently. explore And the number of utilization queries N exploit For queries with high uncertainty, the system will allocate more resources to the exploration path, for example, by setting N. explore = 3, N exploit = 1. 2. The sorting balance parameter λ, which will be used in the subsequent result fusion and sorting stages. For queries with high uncertainty, to encourage diversity of results, a smaller λ value should be set, for example, λ = 0.3. This smaller λ value means that during sorting, the system will more strongly penalize candidates similar to the selected results, thus prioritizing documents that can bring new information. These determined strategy parameters (N explore =3, N exploit =1, λ=0.3) will be distributed to subsequent related modules.

[0075] Subsequently, in step S400, a two-stage search is performed by the two-stage search module 20. The detailed process can be found in Figure 3.

[0076] The first stage is the baseline search, corresponding to step S410 in Figure 3. The two-stage search module 20 first uses the user-inputted original query "Latest Developments in New Energy Vehicle Battery Technology" to directly perform a routine keyword matching or vector retrieval in the data index 60. This aims to quickly obtain a basic set of results most relevant to the literal meaning of the query, i.e., the baseline result set. As shown in step S420 of Figure 3, the system obtains this baseline result set, which may contain, for example, 20 of the most relevant technical documents, news reports, or research papers. This baseline result set is not only part of the final result itself, but more importantly, it provides valuable contextual information for the enhanced query generation in the second stage.

[0077] The second stage is a parallel dual-path search, corresponding to the parallel generation of enhanced queries step S430 in Figure 3. The two-stage search module 20 utilizes a large language model to generate two sets of differentiated new queries in parallel, based on the original query and the baseline result set (or its summary) obtained in the first stage.

[0078] On the exploration path, according to N set by the adaptive strategy control module 30 explore = 3. The two-stage search module 20 submits an instruction specifically designed for the "exploration" task to the large language model to execute step S431. This instruction requires the model to generate three queries aimed at broadening the search boundaries based on the original query and the baseline result summary, through semantic expansion, perspective switching, or condition relaxation. For example, the instruction might be as follows: "[Exploratory Query Generation Instruction] Based on the user query 'Latest Developments in New Energy Vehicle Battery Technology' and the following preliminary search result summary, please generate three related queries with different perspectives or broader fields to help the user explore related topics. The queries should avoid being too similar to the original query. Example summary: '...Solid-state battery energy density breakthrough...thermal runaway safety research...'". The model might generate the following exploratory queries: 1. "Technological innovation of next-generation energy storage systems for electric vehicles" (expanding from "battery" to "energy storage system"); 2. "Technological breakthroughs and challenges of hydrogen fuel cells in the automotive field" (switching to alternative technology routes); 3. "Research progress of key materials in the global power battery industry chain" (expanding from the technology itself to the upstream of the industry chain). Subsequently, in step S432, the system uses these three newly generated exploratory queries to initiate searches in parallel with the data index 60, obtaining three exploratory search results.

[0079] At the same time, on the utilization path, according to the set N exploit = 1. The two-stage search module 20 submits a differentiated instruction designed for the "utilization" task to the large language model to execute step S441. This instruction requires the model to generate a query aimed at deepening the core topic by adding constraints, decomposing the problem, or refining the query. For example, the instruction might be: "[Utilization Query Generation Instruction] Based on the user query 'Latest Developments in New Energy Vehicle Battery Technology' and the preliminary result summary, please generate a more specific and in-depth query, focusing on the most core technological direction." The model might generate a utilization query like: "Analysis of the Latest Improvement Paths for Solid-State Lithium Battery Energy Density and Cycle Life." Subsequently, in step S442, the system uses this new query to initiate a search in the data index 60, obtaining a utilization search result.

[0080] At this point, step S400 is complete, and the system has obtained a total of 5 search results: 1 baseline search result, 3 exploratory search results, and 1 utilization search result.

[0081] Subsequently, in step S500, the result fusion and diversity ranking module 40 is responsible for merging the above 5 search results and performing deduplication to form a large set of candidate results.

[0082] Next, in the crucial step S600, dynamic weighted sorting is performed. The result fusion and diversity sorting module 40 uses a dynamically weighted maximum marginal relevance algorithm to reorder the candidate result set. The goal of this algorithm is to iteratively select documents d from the candidate document set D to maximize an objective function. The mathematical expression of this objective function is:

[0083] Where D is the candidate document set, S is the selected result set, and q is the original user query. Calculate the relevance score between candidate document d and the original query q, which can be obtained using the traditional BM25 algorithm or the modern vector cosine similarity. Calculate the relationship between candidate document d and a document in the selected result set S. The similarity score can also be calculated using methods such as vector similarity.

[0084] It should be noted that in the scheme of this invention, the balance parameter λ is not a fixed hyperparameter, but a value dynamically determined by the adaptive strategy control module 30 in step S300 based on the uncertainty score U(q). In this embodiment, λ is set to 0.3. This smaller λ value increases the second term in the formula, namely the diversity penalty term. The ranking process assigns weights to documents. In other words, during the ranking process, if a candidate document is very similar to the selected results, it will be heavily penalized, its final score will be lower, and it will be less likely to be selected. Conversely, documents that differ significantly from the selected results but are still somewhat relevant to the original query have a greater chance of being ranked higher. This mechanism effectively promotes diversity in the results.

[0085] Finally, in step S700, the results are output. The results fusion and diversity sorting module 40 presents the final result sequence, reordered by the dynamically weighted MMR algorithm, to the user. Due to the adoption of an "exploration-oriented" strategy and a diversity-first sorting method, the final output result list not only includes in-depth articles on core technologies such as solid-state lithium batteries (from the utilization path), but also broadly covers related or alternative technology directions such as energy storage system innovation, hydrogen fuel cells, and industrial chain materials (from the exploration path), thereby constructing a more comprehensive and multi-dimensional information landscape for users to meet their needs for exploratory understanding of a complex topic.

[0086] Example 2 This example aims to illustrate the adaptive capability of the technical solution provided by the present invention in handling low-uncertainty, factual queries. In such scenarios, the user's intent is very clear, and the system should automatically switch to a "focused utilization" strategy to provide accurate and direct answers and avoid unnecessary information diffusivity.

[0087] The system structure used in this embodiment is the same as that shown in Figure 1 of Embodiment 1, and also includes a query understanding and uncertainty assessment module 10, a two-stage search module 20, an adaptive strategy control module 30, and a result fusion and diversity ranking module 40, and interacts with a large language model interface 50 and a data index library 60. Its overall method flow also follows the steps shown in Figure 2.

[0088] In step S100, the system receives a new user query: "Who is the current CEO of Apple?" This is a typical fact-based query with extremely low uncertainty.

[0089] In step S200, the query understanding and uncertainty assessment module 10 packages this query into instructions similar to those in Example 1 and sends them to the large language model for evaluation. The large language model analyzes that the query points to a very specific, single entity attribute, is semantically clear and unambiguous, and has a very narrow information requirement. Therefore, the model returns a very low uncertainty score, for example, U(q) = 0.1.

[0090] In step S300, the adaptive strategy control module 30 receives U(q) = 0.1 and compares it with a preset low uncertainty threshold θ. low（ For example, compare it with 0.25). Since 0.1 < θ low (0.25), the system determines that the current query belongs to the low uncertainty category and should adopt the "emphasis on utilization" retrieval strategy.

[0091] Accordingly, the adaptive strategy control module 30 dynamically determines a set of strategy parameters different from those in Embodiment 1. Specifically, it sets the number of exploratory queries N. explore = 1, number of utilization queries N exploit = 3, and a ranking balancing parameter λ = 0.7 that tends to favor relevance. The logic behind these parameter settings is that for explicit queries, only minimal or even no exploration is needed, and resources should be concentrated on in-depth mining of the core topic; at the same time, the ranking should prioritize the most relevant results, hence the larger λ value, which increases the relevance factor. Weights in the MMR formula.

[0092] In step S400, the two-stage search module 20 begins to perform the search. First, in steps S410 and S420, the system performs a baseline search using "Who is the current CEO of Apple?" as the query term, and obtains a baseline result set containing information related to "Tim Cook".

[0093] Next, in the parallel dual-path search phase (step S430), the two-stage search module 20 performs the search according to the new strategy parameters (N). explore =1, Nexploit =3) Request query generation from the large language model. For the exploration path (step S431), the model is instructed to generate one lightly relevant exploratory query, such as "a list of Apple's historical CEOs". For the exploitation path (step S441), the model is instructed to generate three exploitation queries aimed at deeply mining core facts, such as: "1. From what year did Tim Cook serve?"; "2. Key achievements of Apple under Tim Cook's leadership"; "3. Tim Cook's recent public statements". Subsequently, the system executes these four new queries and the baseline query retrieval in parallel (steps S432 and S442).

[0094] In step S500, all search results are merged and deduplicated.

[0095] In step S600, the result fusion and diversity ranking module 40 applies the MMR algorithm for ranking. At this point, the balancing parameter λ is set to 0.7. This relatively large λ value makes the MMR formula more focused on maximizing the relevance score between candidate documents and the original query. The penalties for diversity are relatively small.

[0096] Finally, in step S700, the system output list will highly focus on information about Tim Cook himself and his role as CEO, with the most directly relevant documents such as his profile, tenure, and achievements placed at the top. Information about other former Apple CEOs or other aspects of the company will be relegated to a secondary position or filtered out due to its lower relevance to the original query. This result quickly and accurately meets the user's need for specific factual information, demonstrating the adaptability and efficiency of the present invention.

[0097] Example 3 illustrates a variant implementation of the technical solution of the present invention, specifically a hybrid implementation of the "parallel dual-path search" in step S400. This example aims to demonstrate that the core conceptual framework of the present invention has good openness and compatibility; its "path-based" query generation is not limited to using a large language model and can also be combined with other mature technologies in the field of information retrieval (such as pseudo-relevance feedback), thereby potentially achieving higher efficiency or effectiveness in specific scenarios.

[0098] The system structure of this embodiment is basically the same as that of embodiment 1, but in the two-stage search module 20, in addition to the ability to interact with the large language model interface 50, it also integrates the function of implementing the pseudo-correlation feedback algorithm.

[0099] As an alternative implementation, the present inventors again use the aforementioned high-uncertainty query "latest developments in new energy vehicle battery technology" as an example. The uncertainty assessment (step S200) and strategy determination (step S300) processes are exactly the same as in Example 1. That is, the system evaluates U(q) = 0.75 and determines the "exploration-oriented" strategy with parameter N. explore = 3, N exploit = 1, λ = 0.3.

[0100] The difference in this embodiment lies in the second stage of step S400, namely the parallel generation of enhanced query step S430.

[0101] The implementation of the exploration path is consistent with that of Example 1. The two-stage search module 20 still calls the large language model to generate three semantically expanded exploratory queries based on the instructions (step S431), such as "latest innovations in electric vehicle energy storage systems", and performs the retrieval (step S432).

[0102] However, this embodiment employs a different technique for exploitation paths. The two-stage search module 20 no longer invokes a large language model to generate exploitation queries; instead, it initiates a traditional pseudo-relevance feedback mechanism to execute step S441. The specific process is as follows: First, the baseline result set obtained in the first-stage baseline search (step S410) is analyzed.

[0103] Then, select the K highest-ranking documents in the result set (e.g., K=5), and assume that these K documents are highly relevant.

[0104] Next, the system performs text analysis on the full text or abstracts of these K documents, extracting frequently occurring and highly distinctive technical terms or keywords. For example, the system might extract terms such as "energy density," "thermal runaway," "silicon-carbon anode," "cycle life," and "BMS."

[0105] Finally, these extracted core terms are combined with the user's original query to automatically construct one or more exploitation queries. Based on the strategy parameter Nexploit = ​​1, the system can generate a combined query, such as: "New energy vehicle battery energy density thermal runaway safety". Alternatively, if Nexploit > 1, multiple queries can be generated, such as "New energy vehicle battery energy density improvement" and "New energy vehicle battery thermal runaway solution".

[0106] After generating the utilization query, the system performs the retrieval of the query in step S442.

[0107] The subsequent steps S500 (result fusion) and S600 (dynamic weighted sorting, λ=0.3) are exactly the same as in Example 1. The final output (step S700) also achieves a balance between exploration and utilization, providing users with comprehensive information.

[0108] This embodiment demonstrates the flexibility of the framework proposed in this invention. By replacing the generation method of exploitation queries with pseudo-relevance feedback, the cost of calling large language model APIs can be reduced. Furthermore, since the query terms generated by pseudo-relevance feedback are directly derived from high-quality documents in the data index 60, these queries may, in some cases, better match the feature distribution of the data source, thus potentially achieving comparable or better exploitation results. This indicates that the scope of protection of this invention should not be limited to using only language models to generate all enhanced queries, but rather covers scenarios where uncertainty evaluation results are used to guide the generation of any form of differentiated exploratory and exploitation queries.

[0109] Example 4: Building upon the previous examples, this example introduces an optional feedback optimization mechanism, demonstrating how the system of this invention utilizes the "reflection" and "metacognition" capabilities of a large language model for self-iteration and optimization. This mechanism aims to automatically or semi-automatically adjust the system's internal strategy parameters by analyzing historical retrieval cases, thereby continuously improving retrieval performance and reducing reliance on manual parameter tuning and long-term maintenance costs.

[0110] As an optional implementation, this embodiment adds an offline "strategy optimization module" to the structure shown in Figure 1. This module does not participate in real-time online query processing, but performs optimization tasks periodically (e.g., daily or weekly). Its working process can be described as follows: Data collection: During normal operation, the system records key information for each retrieval session, including: the user's original query, the uncertainty score U(q) evaluated by the system, and the strategy parameters (N) used at that time. explore , N exploit The strategy optimization module considers the final ranking results and user interaction behavior (such as which links were clicked, how long after which a new query was initiated). It filters out inefficient retrieval cases. Understandably, the criteria for judging inefficient cases can be varied, such as: extremely low click-through rate on the results page; users initiating new, semantically similar queries very quickly after viewing the results (indicating the initial results failed to meet their needs); or users rapidly flipping through multiple pages on the results page before making a click.

[0111] Feedback and Reflection: The strategy optimization module periodically packages a batch of collected inefficient retrieval cases. For each case, it organizes it into a structured text containing all the information collected in the first step.

[0112] Optimization suggestion generation: The strategy optimization module submits these packaged cases to the large language model through the large language model interface 50, along with a carefully designed "system optimization instruction." This instruction requires the model to act as a system diagnostic expert, analyze the possible reasons for retrieval failures, and propose specific optimization suggestions. An exemplary optimization instruction might look like this: "[System Optimization and Diagnosis Instruction] You are a senior optimization expert for an information retrieval system. Below are some recent retrieval cases with poor performance. For each case, this invention provides the [original query], [system-assessed uncertainty U(q)], [the strategy parameters used at the time (Nexplore, Nexploit, λ)], and [a description of the poor performance, such as low user click-through rate]. Your task is to: 1. Analyze the possible reasons for the poor performance. Is it due to inaccurate uncertainty assessment? Or an inappropriate query generation strategy? Or unsuitable sorting parameters? 2. Based on your analysis, please provide specific and actionable adjustment suggestions. Suggestions can address the following: a) Adjust the thresholds θhigh and θlow used to distinguish between high / low uncertainty. b) Adjust the default values ​​of the strategy parameters (Nexplore, Nexploit, λ) corresponding to different uncertainty levels (high / medium / low). Please provide your analysis and suggestions in a clear format."

[0113]

Case 1

[0114] By introducing this feedback optimization mechanism, the system of this invention possesses the ability to evolve from its own invention and continuously learn. It can learn from historical experience and continuously optimize its core decision-making logic, resulting in an increasingly higher degree of alignment between uncertainty assessment and subsequent strategy formulation. This not only gradually improves the system's retrieval performance and user satisfaction but also greatly reduces the need for tedious manual parameter tuning, thereby enhancing the system's long-term stability and maintainability.

[0115] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0116] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A multi-path information retrieval method based on language models and uncertainty assessment, characterized in that, Includes the following steps: The system receives user queries; performs semantic analysis on the user queries using a pre-trained language model to output a quantified uncertainty index in an unsupervised manner; dynamically determines at least one strategy parameter based on the quantified uncertainty index, the strategy parameter including a query generation ratio for controlling the generation of subsequent queries and a ranking balance parameter for controlling the ranking of results; generates at least two sets of differentiated queries based on the user queries and the strategy parameters, including at least one set of exploratory queries aimed at broadening the search breadth and one set of utilization queries aimed at deepening the search depth, and executes the at least two sets of queries to obtain multi-path retrieval results; merges the multi-path retrieval results and re-ranks them using a ranking mechanism, wherein the ranking mechanism dynamically balances result relevance and diversity based on the ranking balance parameter; and outputs the re-ranked results.

2. The method according to claim 1, characterized in that, The pre-trained language model is a large-scale language model.

3. The method according to claim 1 or 2, characterized in that, The step of using a pre-trained language model to perform semantic analysis on the user query includes instructing the pre-trained language model to comprehensively evaluate the semantic ambiguity, domain specialization, and breadth of information needs of the user query.

4. The method according to claim 1, characterized in that, Before the step of generating at least two sets of differentiated queries, the method further includes: performing a benchmark search using the user query to obtain a benchmark result set; and the step of generating at least two sets of differentiated queries is based on the user query and the benchmark result set.

5. The method according to claim 4, characterized in that, The exploratory query is generated by instructing the pre-trained language model to perform semantic expansion, condition relaxation, or multi-perspective reconstruction; the exploitation query is generated by instructing the pre-trained language model to perform precision, add specific limiting conditions, or decompose the problem.

6. The method according to claim 1, characterized in that, The ranking mechanism is the Maximum Marginal Relevance (MMR) algorithm, and the ranking balance parameter is the balance parameter λ used in the MMR algorithm to balance relevance and diversity.

7. The method according to claim 1, characterized in that, The step of dynamically determining at least one strategy parameter includes: comparing the quantified uncertainty index with at least one preset threshold to determine the query generation quantity ratio and the sorting balance parameter based on the comparison result.

8. The method according to claim 7, characterized in that, Includes at least one of the following: when the quantitative uncertainty index is higher than a preset high uncertainty threshold, determine that the query generation ratio is more exploratory queries than exploitative queries, and determine that the ranking balance parameter is a value that favors diversity; when the quantitative uncertainty index is lower than a preset low uncertainty threshold, determine that the query generation ratio is more exploitative queries than exploratory queries, and determine that the ranking balance parameter is a value that favors relevance.

9. A multi-path information retrieval system based on language models and uncertainty assessment, characterized in that, include: The receiving unit is used to receive user queries; An evaluation unit is used to perform semantic analysis on the user query using a pre-trained language model to output a quantitative uncertainty index in an unsupervised manner. The strategy determination unit is used to dynamically determine at least one strategy parameter based on the quantitative uncertainty index. The strategy parameter includes a query generation ratio for controlling the generation of subsequent queries and a sorting balance parameter for controlling the sorting of results. The query execution unit is used to generate at least two sets of differentiated queries based on the user query and the strategy parameters, including at least one set of exploratory queries aimed at broadening the search scope and one set of utilization queries aimed at deepening the search depth, and execute the at least two sets of queries to obtain multi-path retrieval results; the sorting unit is used to merge the multi-path retrieval results and re-sort them using a sorting mechanism, wherein the sorting mechanism dynamically balances result relevance and diversity based on the sorting balance parameter; the output unit is used to output the re-sorted results.

10. The system according to claim 9, characterized in that, The pre-trained language model is a large-scale language model; the semantic analysis of the user query using the pre-trained language model includes: instructing the pre-trained language model to comprehensively evaluate the semantic ambiguity, domain specialization, and breadth of information needs of the user query.

Images