A general intelligent text retrieval method and system based on rule parsing

Abstract

This invention discloses a general text intelligent retrieval method and system based on rule parsing, relating to the field of data processing technology. The method includes: performing feature parsing on user-input query request data to determine data source features and query intent features; fusing historical user representation information with the data source features and query intent features using a multi-dimensional strategy to determine suitable retrieval rules and a suitability evaluation system, and parsing the suitable retrieval rules into structured retrieval conditions; after segmenting the text to be retrieved, performing preliminary information retrieval based on the structured retrieval conditions, and measuring the relevance of candidate texts according to the suitability evaluation system to output a semantic matching score set; and selecting and ranking the candidate texts according to the semantic matching score set to output a list of structured retrieval results. This invention achieves efficient and accurate semantic retrieval for multiple types of text.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a general text intelligent retrieval method and system based on rule parsing. Background Technology

[0002] With the development of natural language processing and information retrieval technologies, existing text retrieval mainly achieves basic retrieval needs through keyword matching, while large models are applied to knowledge retrieval scenarios to generate relevant content. Both types of technologies jointly support the acquisition and utilization of text information.

[0003] However, keyword matching technology is difficult to flexibly express complex search needs and has limited ability to understand context, resulting in low relevance of search results. Large models lack precise rule control and context supplementation mechanisms, which can easily lead to generalized or inaccurate search results, affecting both information acquisition efficiency and the quality of generated content. Summary of the Invention

[0004] This application provides a general text intelligent retrieval method and system based on rule parsing, aiming to solve the technical problem of insufficient relevance and accuracy of retrieval results in the prior art.

[0005] In view of the above problems, this application provides a general text intelligent retrieval method and system based on rule parsing.

[0006] Firstly, this application provides a general intelligent text retrieval method based on rule parsing, including:

[0007] Perform feature analysis on the query request data entered by the user to determine the features of the data source and the query intent.

[0008] The historical user representation information is fused with the data source features and query intent features using a multi-dimensional strategy to determine the appropriate retrieval rules and the appropriate evaluation system, and the appropriate retrieval rules are parsed into structured retrieval conditions.

[0009] After segmenting the text to be retrieved, preliminary information retrieval is performed based on the structured retrieval conditions, and the candidate texts are relevance measured according to the adaptation evaluation system to output a semantic matching score set.

[0010] The candidate texts are selected and sorted according to the semantic matching score set, and a list of structured search results is output.

[0011] Secondly, this application provides a general-purpose intelligent text retrieval system based on rule parsing, including:

[0012] The rule parsing module is used to perform feature parsing on the query request data input by the user to determine the data source features and query intent features;

[0013] The text analysis module is used to integrate historical user representation information with the data source features and query intent features using a multi-dimensional strategy to determine the appropriate retrieval rules and the appropriate evaluation system, and to parse the appropriate retrieval rules into structured retrieval conditions.

[0014] The matching and scoring module is used to perform preliminary information retrieval based on the structured retrieval conditions after segmenting the text to be retrieved, and to measure the relevance of candidate texts according to the adaptation evaluation system, and output a semantic matching score set.

[0015] The results output module is used to select and sort the candidate texts according to the semantic matching score set, and output a list of structured search results.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0017] This application provides a general text intelligent retrieval method and system based on rule parsing. By customizing retrieval rules, it effectively realizes the flexible expression of complex retrieval needs, improves the ability to understand text context, and enhances the relevance and accuracy of retrieval results. At the same time, it can be deeply integrated with large models to reduce model illusion problems, provide accurate knowledge retrieval support, optimize the text retrieval experience and generation quality, and ultimately achieve efficient and accurate semantic retrieval of multiple types of text. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a general text intelligent retrieval method based on rule parsing, provided for an embodiment of this application;

[0020] Figure 2 A schematic diagram of the structure of a general text intelligent retrieval system based on rule parsing provided in this application embodiment;

[0021] The components represented by each number in the attached diagram are explained below:

[0022] Rule parsing module 11, text analysis module 12, matching scoring module 13, and result output module 14. Detailed Implementation

[0023] This application provides a general text intelligent retrieval method and system based on rule parsing, which is used to address the technical problem of insufficient relevance and accuracy of retrieval results in the prior art.

[0024] 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 a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0026] Example 1, as Figure 1 As shown, this application provides a general text intelligent retrieval method based on rule parsing, the method comprising:

[0027] S100: Perform feature analysis on the query request data input by the user to determine the data source features and query intent features.

[0028] In this embodiment, the user-input query request data is analyzed for features to determine data source features and query intent features. A deep convolutional neural network is used to analyze the query request data. Deep convolutional neural networks possess strong local feature extraction and global feature integration capabilities, accurately separating data source features from query intent features. This provides a precise basis for subsequent adaptation of retrieval rules, addressing the problem of a disconnect between retrieval strategies and user needs from the outset.

[0029] Step S100 in the method provided in this application embodiment includes:

[0030] Using a deep convolutional neural network, feature parsing is performed on the query request data to determine data source features and query intent features. The data source features include at least text entropy, domain dictionary matching rate, structured tag coverage, average sentence length, and sentence length variance. The query intent features include at least term density, explicit occurrence rate of logical operators, query length, and structural complexity.

[0031] First, obtain the user query request data. Query request data refers to all information entered by the user to achieve the search goal, including search keywords, logical constraint symbols, domain-specific conditions, etc., and is the raw input for feature parsing. For example, obtain the complete search request entered by the user: [+Artificial Intelligence|Machine Learning], [~Deep Learning|Neural Network], [-Traditional Algorithm]. [+Artificial Intelligence|Machine Learning] indicates that it must contain artificial intelligence or machine learning, [~Deep Learning|Neural Network] indicates that it should contain deep learning or neural networks, [-Traditional Algorithm] indicates that it must not contain traditional algorithms, and | represents or. This is a precise search request belonging to the AI ​​domain.

[0032] Secondly, a Deep Convolutional Neural Network (DCNN) analytical model is constructed. DCNN is a deep learning model that excels at processing text sequences. It includes: an input layer that converts symbols such as +, -, and ~, as well as keywords, in the query request into computable vectors, e.g., + corresponds to vector [1,0,0], ~ to [0,1,0], and - to [0,0,1]; a convolutional layer that extracts combined features of logical symbols and keywords using a 2-gram sliding window, such as + for artificial intelligence and ~ for deep learning; a pooling layer that filters out key features such as the type of logical symbol and the domain of the term; and a fully connected layer that integrates features and outputs data source features and query intent features. For example, in this case, the DCNN convolutional layer focuses on capturing the association between +, ~, and - and subsequent keywords, ensuring that the correspondence between logical constraints and the search object is not lost.

[0033] Further, we analyze the data source characteristics. Data source characteristics refer to the type characteristics of the text to be retrieved indirectly reflected by the query request, including text entropy, domain dictionary matching rate, structured tag coverage, average sentence length, and sentence length variance. Text entropy is an indicator that measures the complexity and diversity of information in the text to be retrieved; the richer the vocabulary, the higher the entropy value; if the vocabulary is singular, the entropy value is lower. Domain dictionary matching rate measures the degree of matching between domain-specific terms in the query request and the preset domain dictionary. The calculation formula is: Domain dictionary matching rate = Number of matched domain terms in the query / Number of target terms in the preset domain dictionary. Structured tag coverage measures the sufficiency of the use of structured constraint tags in the query request. The calculation formula is: Structured tag coverage = Number of structured tags actually appearing in the query / Maximum number of tags that can be used in this type of retrieval scenario. Average sentence length is an indicator that measures the uniformity of sentence length in the query request. The calculation formula is: Average sentence length = Total number of characters in all sentences in the query / Number of sentences. Sentence length variance measures the degree of difference in the length of each sentence in the query request.

[0034] For example, the user-input search request is parsed as follows: Text Entropy: Contains 4 symbols (+, ~, -, |) and 5 terms (artificial intelligence, machine learning, deep learning, neural network, traditional algorithm), with diverse vocabulary types. The calculated text entropy is 2.3, which is a high entropy value, indicating that the query information is highly complex. Domain Dictionary Matching Rate: The AI ​​domain dictionary contains terms such as artificial intelligence, machine learning, deep learning, and neural network, and all queries match, resulting in a coverage rate of 100%. Structured Tag Coverage: The query uses 4 structured tags (+, ~, -, |). The maximum number of tags for this type of search request is 5, so the coverage rate is 4 / 5 = 80%, which is high, indicating that the text to be retrieved may be a highly structured technical document. Average Sentence Length: The number of characters in the 3 sentences are 8, 10, and 6 respectively, with a mean of 8; the sentence length variance is approximately 2.67, indicating that the sentence lengths are uniform. Final Output Data Source Characteristics: AI domain text tendency, high information complexity, high structured tag coverage, and uniform sentence length.

[0035] Furthermore, query intent features are analyzed. Query intent features refer to the search target tendencies conveyed by users through their query requests, including terminology density, explicit occurrence rate of logical operators, query length, and structural complexity. Terminology density refers to the proportion of technical terms in the total vocabulary of the query, reflecting whether the user is interested in content from a specific field. Explicit occurrence rate of logical operators refers to the proportion of visible logical symbols in the query, reflecting whether the user's search needs are clear. Query length refers to the total number of characters in the query. Structural complexity determines the logical hierarchy, such as whether there is nested logic, reflecting the complexity of the requirement.

[0036] For example, the user-input search request is parsed as follows: Term density: Total vocabulary: 8 words, including +, artificial intelligence, |, machine learning, ~, deep learning, -, and traditional algorithms; Number of specialized terms: 5 words, including artificial intelligence, machine learning, deep learning, and traditional algorithms; Term density = 5 / 8 = 62.5%, which is high density, indicating that the user is interested in concepts in a specific field. Explicit occurrence rate of logical operators: Total number of symbols: 4, including +, |, ~, and -; Explicit operators: 4 words, occurrence rate = 100%, high occurrence rate, indicating that the user has conveyed explicit constraints through logical operators: must contain artificial intelligence or machine learning, should contain deep learning or neural networks, must not contain traditional algorithms. Query length and structural complexity: Total number of characters: 24; Logical level: 1; Only uses OR grouping; No nesting; Combined with the high symbol occurrence rate, the user intent is determined to be: precise filtering of specific technical content in the AI ​​field. Final output query intent characteristics: Precise filtering intent, high term density, high logical operator occurrence rate, medium length and structural complexity.

[0037] Finally, the output shows the data source features and query intent features. The DCNN fully connected layer integrates these features into a structured result, explicitly annotating the impact of the logical constraints corresponding to +, -, and ~ on the intent, serving as input for subsequent multi-dimensional strategy fusion. For example, the output may include supplementary logical constraint descriptions: Query intent constraint: must include artificial intelligence or machine learning, should include deep learning or neural networks, must not include traditional algorithms.

[0038] In this embodiment, DCNN is used to accurately identify the semantics of logical constraint symbols, avoiding the constraint misjudgment caused by the prior art ignoring symbols; combined with the explicit occurrence rate of logical operators in symbol parsing, it is clear whether the user is precisely filtering or broadly acquiring, avoiding mismatch of retrieval strategies; the output features can be directly used for multi-dimensional fusion in step S200, ensuring that the subsequent adapted retrieval rules fully match the constraint requirements defined by the logical constraint symbols, thereby improving the relevance and accuracy of retrieval results from the source.

[0039] S200: The historical user representation information is fused with the data source features and query intent features using a multi-dimensional strategy to determine the appropriate retrieval rules and the appropriate evaluation system, and the appropriate retrieval rules are parsed into structured retrieval conditions.

[0040] In this embodiment, historical user representation information is fused with data source features and query intent features using a multi-dimensional strategy to determine suitable retrieval rules and a suitability evaluation system. These suitable retrieval rules are then parsed into structured retrieval conditions. By fusing historical user representation information, data source features, and query intent features, the retrieval rules are aligned with past user preferences. Furthermore, by constraining the maximum response latency, the weights of keyword retrieval and large-scale model retrieval are balanced. This dual-dimensional approach addresses the shortcomings of existing technologies, balancing user suitability with speed and accuracy, and providing strategic support for subsequent accurate retrieval.

[0041] Step S200 in the method provided in this application embodiment includes:

[0042] Based on user ID matching, historical query records are retrieved and analyzed to determine historical user representation information. The historical user representation information includes at least user value level, average term density of historical queries, average number of iterations within historical sessions, effective click-through rate of historical queries, and user identity tags.

[0043] First, historical query records are retrieved based on user ID matching to analyze and determine historical user representation information. This historical user representation information includes at least user value level, average term density in historical queries, average number of iterations within historical sessions, effective click-through rate (CTR) in historical queries, and user identity tags. User value level is a priority based on query frequency and effective feedback rate. Average term density in historical queries is the average proportion of domain-specific terms in past queries, reflecting the user's level of expertise. Average number of iterations within historical sessions is the average number of times a user modifies their query conditions in a single search; a higher number indicates more refined requirements. Effective click-through rate (CTR) is the proportion of clicks that match the user's requirements out of the total clicks, reflecting the past rule fit. User identity tags are the domain to which the user belongs, labeled based on the content of historical queries. By matching historical query behavior data in the database with the user's unique ID, historical user representation information reflecting user search preferences is extracted and analyzed from five dimensions: user priority, expertise, requirement refinement, past fit, and domain identity.

[0044] For example, using user ID001, the database was retrieved to show the user's historical data for the past 3 months, including a total of 65 queries. These queries primarily contained AI terminology such as neural network training and large model fine-tuning. In 12 out of 15 sessions, the user modified the constraint to include (~) type 280 results. Of these, 238 were marked as meeting the user's requirements. This user queries at least 20 times per month with a 90% effective response rate, thus the user's value level is high. Ten randomly selected historical queries were used to calculate terminology density, with an average terminology density of 60%. The user made 38 modifications in 15 sessions, resulting in an average iteration count of 38 / 15 ≈ 2.5 times per session. With 238 effective clicks out of a total of 280 clicks, the effective click-through rate for historical queries is 238 / 280 ≈ 85%. 90% of the user's historical queries revolve around AI technology, with no cross-domain content; therefore, the user's identity tag is: AI Researcher. Final output: Historical user representation information set: {User value level: high; Average term density of historical queries: 60%; Average number of iterations within historical sessions: 2.5 times; Effective click-through rate of historical queries: 85%; User identity tag: AI researcher}.

[0045] This involves fusing historical user representation information with data source features and query intent features using a multi-dimensional strategy to determine suitable retrieval rules and a suitable evaluation system, including:

[0046] Based on the historical user representation information, data source characteristics, and query intent characteristics, the urgency of the current query is assessed, and the urgency coefficient of the current query is determined.

[0047] The maximum response delay time limit is obtained by multiplying the ratio of the preset standard query urgency coefficient to the current query urgency coefficient by the preset response delay time limit;

[0048] Using the maximum response delay limit as a constraint, the appropriate retrieval rules and appropriate evaluation system are determined based on the historical user characterization information, data source characteristics, and query intent characteristics.

[0049] First, the urgency of the current query is assessed based on the historical user representation information, data source characteristics, and query intent characteristics to determine the current query urgency coefficient. The urgency coefficient is an indicator measuring the time sensitivity of the current query, ranging from 0 to 1, with higher values ​​indicating a greater need for a faster response. The urgency coefficient is calculated by weighting three types of features, with preset weights: historical user representation 40%, data source characteristics 30%, and query intent characteristics 30%. For example, historical user representation contributes a high-value level, so it's assigned a value of 0.9, and 0.9 × 40% = 0.36; data source characteristics contribute a low urgency level due to the high structure of AI-related text, so it's assigned a value of 0.6, and 0.6 × 30% = 0.18; query intent characteristics contribute a high urgency level due to precise filtering and high terminology density, so it's assigned a value of 0.8, and 0.8 × 30% = 0.24. Adding these three together, the current query urgency coefficient = 0.36 + 0.18 + 0.24 = 0.78.

[0050] Next, the ratio of the preset standard query urgency coefficient to the current query urgency coefficient is multiplied by the preset response delay limit to obtain the maximum response delay limit. The maximum response delay limit is the longest allowed retrieval response time, and the formula is: Maximum response delay limit = Preset standard query urgency coefficient / Current query urgency coefficient × Preset response delay limit. For example, if the preset standard coefficient is 0.5, the preset delay limit is 2 seconds, and the current query urgency coefficient is 0.78, the maximum response delay limit = 0.5 / 0.78 × 2 ≈ 1.28 seconds, meaning the current query must be completed within 1.28 seconds to avoid user waiting.

[0051] Furthermore, using the maximum response delay time limit as a constraint, the adaptation retrieval rules and adaptation evaluation system are determined based on the historical user characterization information, data source characteristics, and query intent characteristics.

[0052] Among them, the maximum response delay time limit is used as a constraint, and the appropriate retrieval rules and appropriate evaluation system are determined based on the historical user characterization information, data source characteristics, and query intent characteristics, including:

[0053] Using the maximum response delay time limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user characterization information, data source characteristics and query intent characteristics;

[0054] Obtain preset search rules and preset evaluation system, wherein the preset search rules include preset logical expression constraints and initial search weight ratios, the search keywords in the preset logical expression constraints include at least mandatory, should be included, and must not be included, and the preset evaluation system includes a similarity measurement algorithm and an initial evaluation weight ratio;

[0055] Based on the stated credible weight ratio, the preset retrieval rules and preset evaluation system are optimized and adjusted to determine the appropriate retrieval rules and evaluation system.

[0056] First, using the maximum response delay time limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user characterization information, data source characteristics and query intent characteristics.

[0057] Among them, the maximum response delay time limit is used as a constraint, and the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user characterization information, data source characteristics, and query intent characteristics, including:

[0058] An information query and analysis plugin is built based on deep learning, which includes a query complexity recognition branch, a query depth recognition branch, and a query breadth recognition branch.

[0059] Using the query complexity identification branch, query depth identification branch, and query breadth identification branch, the query complexity, query depth, and query breadth are identified based on the historical user representation information, data source characteristics, and query intent characteristics, and the current query complexity, query depth, and query breadth are output.

[0060] Using the maximum response delay limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth and current query breadth.

[0061] First, an information query analysis plugin is constructed based on deep learning. This plugin includes query complexity recognition branches, query depth recognition branches, and query breadth recognition branches. The plugin is a multi-branch model based on deep learning, which transforms historical user representation information, data source features, and query intent features into quantifiable query complexity, depth, and breadth metrics, providing a basis for subsequent weight setting. The plugin adopts a shared feature layer + independent branch layer structure: The shared feature layer uses a BERT pre-trained model to encode the input text features, outputting a unified feature vector; the independent branch layer contains three parallel recognition branches: query complexity recognition branch, query depth recognition branch, and query breadth recognition branch. Each branch consists of two fully connected network layers, outputting a quantified value of the corresponding metric, ranging from 0 to 1, with higher values ​​indicating stronger complexity. The query complexity recognition branch is used to determine the complexity of the logical relationships in the query request, such as whether it contains nested logic or multiple sets of constraints. Excessive complexity increases retrieval difficulty, requiring more large models to participate. The query depth identification branch is used to determine the degree of detail required by the query, such as whether specific parameters or principle descriptions are needed. Higher depth indicates that the user requires more precise and detailed matching. The query breadth identification branch is used to determine the scope of the query coverage, such as whether it spans multiple domains. Wider breadth indicates a larger range of text to be retrieved, requiring a balance between keyword coverage and the overall model's coverage capabilities.

[0062] Secondly, utilizing the query complexity identification branch, query depth identification branch, and query breadth identification branch, the system identifies the current query complexity, current query depth, and current query breadth based on the historical user representation information, data source characteristics, and query intent characteristics. For example, inputting historical user representation information includes: user identity tag: AI domain researcher; average number of iterations within historical sessions: 2.5; average term density of historical queries: 60%; data source characteristics: domain dictionary matching rate: 100%; structured tag coverage rate: 80%; query intent characteristics: moderate structural complexity; term density: 62.5%; query content: [+Artificial Intelligence|Machine Learning, ~Deep Learning|Neural Network, -Traditional Algorithm].

[0063] For example, the identification process and output results of each branch are as follows: The query complexity identification branch model, based on the learned logical hierarchy number = 1, no nesting, and no overlapping labels, determines that the current query logic is simple, and outputs the current query complexity = 0.3, where 0.3 represents low complexity. The query depth identification branch model, combined with high terminology density and the professional user's identity, determines that the user needs matching of core concepts within the domain, rather than generalized explanations, but does not involve extreme details such as parameters and steps, and outputs the current query depth = 0.6, where 0.6 represents medium depth. The query breadth identification branch model, based on the terminology being purely AI-domain and the user having no cross-domain habits, determines that the query scope is concentrated in a single domain, and outputs the current query breadth = 0.2, where 0.2 represents narrow breadth. Through the three branch identifications of the information query analysis plugin, the final analysis results are: Current query complexity: 0.3, low complexity, simple logic; Current query depth: 0.6, medium query depth; Current query breadth: 0.2, narrow query breadth, only in the AI ​​domain.

[0064] Furthermore, with the maximum response delay time limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth and current query breadth.

[0065] Among them, the maximum response delay time limit is used as a constraint, and the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth, and current query breadth, including:

[0066] The lower limit of keyword weight ratio for keyword retrieval is determined based on the maximum response delay time limit, wherein the lower limit of keyword weight ratio is negatively correlated with the maximum response delay time limit;

[0067] Set the ratio of the preset standard query complexity to the current query complexity as the first weight compensation coefficient;

[0068] Set the ratio of the preset standard query depth to the current query depth as the second weight compensation coefficient;

[0069] Set the ratio of the preset standard query breadth to the current query breadth as the third weight compensation coefficient;

[0070] The first weight compensation coefficient, the second weight compensation coefficient, and the third weight compensation coefficient are weighted and fused to determine the comprehensive weight compensation coefficient;

[0071] The initial credibility weight of the keyword is compensated according to the comprehensive weight compensation coefficient to obtain the credibility weight of the adapted keyword, wherein the credibility weight of the adapted keyword is not greater than 0.8;

[0072] If the credibility weight of the adapted keyword is less than the lower limit of the keyword weight ratio, then it is adjusted to the lower limit of the keyword weight ratio;

[0073] The credibility weight of the adapted large model is obtained by subtracting the credibility weight of the adapted keyword from 1.

[0074] The trusted weight ratio is generated based on the trusted weight of the adapted keywords and the trusted weight of the adapted large model.

[0075] First, based on the maximum response delay time limit, a lower limit for the keyword weight ratio of keyword retrieval is determined, wherein the lower limit for the keyword weight ratio is negatively correlated with the maximum response delay time limit. The lower limit for the keyword weight ratio is the minimum weight percentage for keyword retrieval, with a value between 0 and 1. The lower limit for the keyword weight ratio is negatively correlated with the maximum response delay time limit; the shorter the delay time limit, the faster the response is required, and the more reliance is placed on fast keyword retrieval, resulting in a higher lower limit; conversely, the longer the delay time limit, the more reliance can be placed on high-precision large-model retrieval, resulting in a lower lower limit. For example, a preset maximum response delay time limit - keyword weight lower limit correspondence table is used: delay ≤ 1 second, lower limit 0.7; 1 second < delay ≤ 1.5 seconds, lower limit 0.6; 1.5 seconds < delay ≤ 2 seconds, lower limit 0.5; delay > 2 seconds, lower limit 0.4. The current maximum response delay time limit = 1.28 seconds, resulting in a matched lower limit for the keyword weight ratio of 0.6.

[0076] Secondly, the ratio of the preset standard query complexity to the current query complexity is set as the first weight compensation coefficient. This compensation coefficient is used to adjust keyword weights based on the complexity, depth, and breadth of the current query: if a dimension of the current query is lower than the standard value, it indicates simplicity and easier keyword processing, so the compensation coefficient > 1, increasing the keyword weight; if it is higher than the standard value, it indicates greater complexity and requires a larger model, so the compensation coefficient < 1. For example, if the preset standard query complexity is 1.0 and the current query complexity is 0.3, the first weight compensation coefficient = preset standard query complexity / current query complexity = 1 / 0.3 ≈ 3.33. Since the current complexity 0.3 < standard 1.0, the query logic is simple, keywords can be processed efficiently, and the weight needs to be increased.

[0077] Furthermore, the ratio of the preset standard query depth to the current query depth is set as the second weight compensation coefficient. For example, if the preset standard query depth is 1.0 and the current query depth is 0.6, the second weight compensation coefficient = preset standard query depth / current query depth = 1 / 0.6 ≈ 1.67. Since the current depth 0.6 < standard 1.0, it indicates that the requirements are not yet fully detailed, keyword matching can meet basic needs, and the weight needs to be increased.

[0078] Simultaneously, the ratio of the preset standard query breadth to the current query breadth is set as the third weight compensation coefficient. For example, if the preset standard query breadth is 1.0 and the current query breadth is 0.2, the second weight compensation coefficient = preset standard query breadth / current query breadth = 1 / 0.2 = 5.0. A current breadth of 0.2 < standard 1.0 indicates a narrow query range, making keyword retrieval easily coverable, thus requiring a significant increase in weight.

[0079] Secondly, the first, second, and third weight compensation coefficients are weighted and fused to determine the comprehensive weight compensation coefficient. The three compensation coefficients are then weighted and fused to obtain a comprehensive adjustment coefficient, balancing the impact of different dimensions on keyword weights. For example, if each weight accounts for 1 / 3, the comprehensive weight compensation coefficient = (first compensation coefficient + second compensation coefficient + third compensation coefficient) / 3 = (3.33 + 1.67 + 5.0) / 3 = 10.0 / 3 ≈ 3.33.

[0080] Furthermore, the initial credibility weight of the keyword is compensated according to the comprehensive weight compensation coefficient to obtain the credibility weight of the adapted keyword, wherein the credibility weight of the adapted keyword is not greater than 0.8. The initial credibility weight of the keyword is adjusted using the comprehensive compensation coefficient, while ensuring that it does not exceed the upper limit of 0.8 to avoid insufficient accuracy due to over-reliance on the keyword. For example, if the initial credibility weight of the keyword is 0.5 and the comprehensive weight compensation coefficient is 3.33, the credibility weight of the adapted keyword = initial credibility weight × comprehensive weight compensation coefficient = 0.5 × 3.33 ≈ 1.66. Upper limit constraint: Since 1.66 > 0.8, it is adjusted to 0.8.

[0081] If the trusted weight of the matched keyword is less than the lower limit of the keyword weight ratio, it will be adjusted to the lower limit of the keyword weight ratio. If the trusted weight of the matched keyword is lower than the lower limit, it will be forcibly increased to the lower limit to ensure response speed. For example, if the trusted weight of the matched keyword is calculated to be 0.5 and the lower limit of the keyword weight ratio is 0.6, the trusted weight of the matched keyword will be forcibly increased to the lower limit of 0.6.

[0082] Then, the trusted weight of the adapted keyword is obtained by subtracting the trusted weight of the adapted large model from 1. The large model weight and the keyword weight are complementary and their sum is 1. For example, the trusted weight of the adapted large model = 1 - the trusted weight of the adapted keyword = 1 - 0.8 = 0.2.

[0083] Finally, a credibility weight ratio is generated based on the credibility weights of the adapted keywords and the adapted large model. For example, the credibility weight ratio is 80% for keyword retrieval and 20% for large model retrieval, meaning that keywords are used to complete 80% of the filtering during retrieval, and the large model supplements the 20% for semantic precision matching.

[0084] In addition, preset search rules and a preset evaluation system are obtained. The preset search rules include preset logical expression constraints and initial search weight ratios. The search keywords in the preset logical expression constraints include at least three categories: must-include, should-include, and must-not-include. The preset evaluation system includes a similarity measurement algorithm and an initial evaluation weight ratio. The preset search rules are a default general search framework, including preset logical expression constraints and initial search weight ratios. Preset logical expression constraints define the main logical relationships of search keywords, ensuring that basic search conditions are executable. Specifically, they include: *must*: corresponding to the + symbol, keywords that must be included in the search results, or keyword groups separated by |; if not included, they are directly excluded; *should*: corresponding to the ~ symbol, if the search results contain relevant keywords, the relevance score is improved; if not included, they are not directly excluded; *must-not*: corresponding to the - symbol, if the search results contain relevant keywords, they are directly excluded, regardless of whether other conditions are met. The initial search weight ratio refers to the weight allocation between the default keyword search and the large model search, used to balance speed and accuracy. The preset initial search weight ratio is 50% for keywords and 50% for the large model, meaning that both contribute equal weight during the search process. The default evaluation system is the standard for scoring the relevance of search results, including a similarity measurement algorithm and initial evaluation weighting. The similarity measurement algorithm calculates the degree of match between search results and query requirements. It defaults to using a cosine similarity algorithm, which measures similarity by the cosine of the angle between two vectors in vector space; a higher value indicates a higher degree of match. It also combines keyword matching and semantic matching. The initial evaluation weighting is the default allocation of weights for different logical constraints in the scoring. The default initial evaluation weighting is 50% for "must" conditions and 50% for "should" conditions, meaning that both "must" and "should" conditions contribute equally to the final score.

[0085] Finally, the preset retrieval rules and preset evaluation system are optimized and adjusted according to the trusted weight ratio to determine the appropriate retrieval rules and evaluation system. The execution entities that must be included, should be included, and must not be included in the preset rules are allocated according to the trusted weight ratio. The retrieval weights are adjusted to the trusted weight ratio, and the scoring weights of hard constraints (must / must_not) and soft constraints (should) in the evaluation system are simultaneously optimized to match the weight ratio of the execution entities.

[0086] For example, based on a weighting distribution of 80% for keyword retrieval and 20% for large-model retrieval, the retrieval rules and evaluation criteria are adjusted. Adapting retrieval rules: In retrieval rules, `must` and `must_not` are primarily handled by keyword retrieval, accounting for 80% of the weight, suitable for quick keyword filtering; `should` is handled assisted by the large-model, accounting for 20% of the weight, as the large-model is better at identifying content with similar meanings. The overall retrieval allocates the roles of the two at 80% and 20%, with keyword results serving as the basic filtering basis and large-model results used to optimize ranking. Adapting the evaluation system: In evaluation criteria, the scoring focuses on keyword matching; the fulfillment of `must` accounts for 70% of the scoring weight; failure to meet the requirement results in an immediate disqualification, while fulfillment is scored based on the completeness of the match; the matching of `should` accounts for 30% of the scoring weight, scored based on the accuracy of semantically similar matches. This adjustment aligns the retrieval strategy with the weighting distribution, balancing efficiency and accuracy.

[0087] In this embodiment, multi-dimensional feature fusion and precise strategy adaptation address the disconnect between retrieval strategies and user needs and historical habits. By combining historical user representation information with current query characteristics, retrieval rules are made to align with actual user needs, avoiding insufficient adaptability of general rules. By assessing query urgency to determine the maximum response delay, the proportion of keyword retrieval and large-model retrieval is reasonably allocated. Keyword retrieval ensures rapid response, while large-model retrieval supplements semantic understanding capabilities, avoiding the efficiency or accuracy shortcomings of a single retrieval method. By optimizing the execution entities and scoring weights of "must include," "should include," and "must not include," the retrieval logic is made more aligned with query constraint types, and the evaluation criteria are more matched to the retrieval strategy, laying the foundation for subsequent accurate filtering and ranking, and further improving the relevance and accuracy of retrieval results.

[0088] S300: After segmenting the text to be retrieved, perform preliminary information retrieval based on the structured retrieval conditions, and measure the relevance of candidate texts according to the adaptation evaluation system, and output a semantic matching score set.

[0089] In this embodiment, after segmenting the text to be retrieved, a preliminary information retrieval is performed based on the structured retrieval conditions, and the candidate texts are relevance-measured according to the adaptation evaluation system, outputting a semantic matching score set. The text to be retrieved often has complex structures and uneven information density. If the retrieval rules are directly matched as a whole, key information may be missed due to contextual fragmentation, and it is difficult to accurately correspond to the structured retrieval conditions. At the same time, the preliminary retrieval lacks a relevance measurement logic linked to the adaptation evaluation system, which may lead to a disconnect between the retrieval results and the scoring criteria, resulting in insufficient accuracy in candidate text selection. Therefore, step S300 first decomposes the text structure through segmentation, then performs targeted retrieval based on the structured retrieval conditions, and finally quantifies the relevance using the adaptation evaluation system. This solves the problem of matching confusion caused by complex text structures and ensures that the scores of candidate texts are consistent with the retrieval strategy, laying the foundation for subsequent accurate selection.

[0090] First, the text to be retrieved is segmented. Segmentation involves breaking down the complete text into independent semantic units, each segment serving as a separate matching object. A contextual sliding window is also supported to supplement the semantic meaning of each segment and avoid biases caused by isolated matching. For example, the technical document to be retrieved in the file is segmented into four independent paragraphs: Paragraph 1: "Artificial intelligence technology is developing rapidly." Paragraph 2: "Machine learning is an important branch of artificial intelligence." Paragraph 3: "Deep learning uses neural networks." Paragraph 4: "Traditional algorithms such as decision trees." A contextual sliding window is also enabled, automatically associating each segment with the preceding and following paragraphs during matching, improving the completeness of the match.

[0091] Secondly, preliminary information retrieval is performed based on the structured retrieval conditions. Preliminary information retrieval refers to the process of screening the segmented text according to the structured retrieval conditions determined in S200. Paragraphs that clearly do not meet the requirements are excluded, and candidate texts are retained, based on the rules of "must / must_not" handled by keyword retrieval and "should" assisted by large-scale model retrieval. For example, the keyword retrieval process for the "must" condition: segment-by-segment matching [must contain artificial intelligence or machine learning]. Paragraph 1 contains [artificial intelligence], and paragraph 2 contains [machine learning], both satisfying the "must" condition; paragraphs 3 and 4 do not contain these two keywords and do not satisfy the "must" condition. The keyword retrieval process for the "mustnot" condition: segment-by-segment matching [must not contain traditional algorithms]. Paragraph 4 contains [traditional algorithms], violating the "mustnot" condition and is directly excluded; paragraphs 1, 2, and 3 do not contain this keyword and meet the requirements. The large-scale model retrieval process for the "should" condition: for paragraphs 1 and 2 that satisfy the "must" condition and do not violate the "must_not" condition, semantic matching is performed [should contain deep learning or neural networks]. The results show that neither paragraph contains relevant semantics. Preliminary search results: Paragraphs 1 and 2 were retained as candidate texts; paragraph 3 was excluded because it did not meet the must condition; and paragraph 4 was excluded because it violated must_not.

[0092] Furthermore, the candidate texts are relevance measured according to the aforementioned adaptation evaluation system. The relevance measurement is based on the scoring rules of the adaptation evaluation system, and the formula is: Final Score = Base Score 0.5 + Number of Should Matching Groups / Total Number of Should Groups. Here, the base score of 0.5 is only obtained when the must condition is met; otherwise, 0 points are awarded. The number of should matching groups refers to the number of groups that satisfy the should constraint, and the total number of should groups refers to the total number of groups for the user-defined should condition. In this case, [~Deep Learning|Neural Network] is 1 group, so the total number of should groups = 1. For example, the relevance measurement for paragraph 1: Must condition: Contains [Artificial Intelligence], satisfying the must condition, obtaining a base score of 0.5; Should condition: Must contain [Deep Learning|Neural Network]. The first paragraph does not contain [Deep Learning] or [Neural Network], so the number of should matching groups = 0; Final Score = 0.5 + 0 / 1 = 0.5. Paragraph 2 relevance measurement: Must condition: Contains "[Machine Learning]", satisfying must, receiving a base score of 0.5; Should condition: Number of matching groups should = 0; Final score = 0.5 + 0 / 1 = 0.5. If there exists a paragraph "Machine Learning combined with Deep Learning improves efficiency.", satisfying must, the number of matching groups should = 1, then the final score = 0.5 + 1 / 1 = 1.0.

[0093] Finally, the semantic matching score set is output. For example, the output score set is as follows: [{paragraph identifier: 1; semantic matching score: 0.5}, {paragraph identifier: 2; semantic matching score: 0.5}].

[0094] In this embodiment, text semantic units are segmented and supplemented with semantic information using a contextual sliding window, avoiding the omission of key information due to overall matching and improving search targeting. Candidate texts are selectively filtered based on structured search conditions and quantitatively scored according to an adaptation evaluation system, ensuring that the scoring logic is consistent with the execution rules and weighting of must / should criteria, thus improving the accuracy of candidate text selection. The output semantic matching score set clearly defines the degree of fit of candidate texts and can be directly used for subsequent sorting and filtering of low-relevance results, further improving the relevance and accuracy of search results.

[0095] S400: Select and sort the candidate texts according to the semantic matching score set, and output a list of structured search results.

[0096] In this embodiment, the candidate texts are selected and sorted according to the semantic matching score set, and a structured search result list is output. Step S300 outputs the semantic matching score set of candidate texts, but the results are not filtered or sorted, which presents two problems: first, it may include low-relevance candidate texts, interfering with the acquisition of effective information; second, the candidate texts lack clear priorities, requiring users to manually filter highly relevant content, which is inefficient. Furthermore, the original score set lacks a structured presentation, with key information scattered and difficult to quickly access. Therefore, step S400 needs to eliminate low-value results through selective filtering, clarify priorities by score sorting, and then organize the results into a structured list to improve the practicality and ease of use of the search results.

[0097] First, set a selection threshold. The selection threshold is the lowest semantic matching score required to retain candidate texts. It ranges from 0 to 1 and is used to filter out invalid results with extremely low relevance. For example, based on general search requirements, a preset selection threshold of 0.3 is used, meaning only candidate texts with a semantic matching score ≥ 0.3 are retained, and results below this value are directly discarded.

[0098] Secondly, a selective screening process is performed on the candidate texts. This selective screening involves removing candidate texts that do not meet the semantic matching score set based on a set screening threshold, retaining only valid results that meet the relevance requirements. For example, the candidate text scores are checked one by one: if the candidate text scores are 0.5, 0.5, and 1.0 respectively, all are greater than the screening threshold of 0.3; the screening result is: all candidate texts meet the requirements, no texts are removed, and 3 valid candidate texts are retained. If there is a candidate text with a score of 0.2, "There are many artificial intelligence-related technologies," it will be directly removed because it is below the threshold of 0.3.

[0099] Further, the filtered candidate texts are sorted. The sorting is done in descending order of semantic matching score, with higher scores appearing first. This prioritizes results, allowing users to view highly relevant content first. If scores are the same, they are arranged according to the order of appearance in the original document being retrieved. For example, the scores of the extracted candidate texts are: 1.0, 0.5, 0.5; sorted in descending order: the text with a score of 1.0 is ranked first, and the two texts with scores of 0.5 are ranked second and third according to their original order of appearance; the sorted results are: "Machine learning combined with deep learning improves efficiency" (score 1.0); "Artificial intelligence technology is developing rapidly" (score 0.5); "Machine learning is an important branch of artificial intelligence" (score 0.5).

[0100] Finally, a structured search results list is output. The structured search results list is a data set that organizes the sorted valid candidate texts into a structured presentation format based on key information. It includes paragraph identifiers, matched must-keywords, semantic matching scores, and text content, facilitating quick browsing and location for users. For example, the output structured list is as follows: [{Paragraph identifier: 3; Matching must-keyword: Machine Learning; Semantic matching score: 1.0; Text content: Machine Learning combined with deep learning improves efficiency}, {Paragraph identifier: 1; Matching must-keyword: Artificial Intelligence; Semantic matching score: 0.5; Text content: Artificial intelligence technology is developing rapidly}, {Paragraph identifier: 2; Matching must-keyword: Machine Learning; Semantic matching score: 0.5; Text content: Machine Learning is an important branch of Artificial Intelligence}].

[0101] In this embodiment, texts with substandard scores are filtered out using a threshold to ensure that all output results meet basic relevance requirements and reduce interference from invalid information. Sorting by score in descending order prioritizes highly relevant content, eliminating the need for manual filtering and improving information retrieval efficiency. The structured list integrates key information, presenting it clearly and intuitively, facilitating quick access to crucial content and providing a well-organized data foundation for subsequent secondary searches or in-depth analysis, further enhancing the user experience.

[0102] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects:

[0103] This application provides a general text intelligent retrieval method and system based on rule parsing. By relying on feature analysis and historical user information fusion before retrieval, the retrieval rules and evaluation system are precisely adapted to user needs, text characteristics, and historical preferences, avoiding the limitations of general rules. Through the complementary advantages of keyword retrieval and large-model retrieval, combined with text segmentation and contextual supplementation, the retrieval response speed and result accuracy are balanced, compensating for the shortcomings of single retrieval methods. Quantitative scoring rules, low-relevance result filtering, priority ranking, and structured presentation enhance the relevance and interpretability of retrieval results, reducing the cost of information acquisition for users. Simultaneously, it supports flexible logical constraint expression and multi-type text processing, can be embedded in large-model retrieval processes, adapts to complex retrieval needs and multi-scenario applications, and caters to the retrieval demands of different users, achieving a generalized and intelligent upgrade of text retrieval.

[0104] Example 2, as Figure 2 As shown, this application provides a general-purpose intelligent text retrieval system based on rule parsing, the system comprising:

[0105] Rule parsing module 11 is used to perform feature parsing on the query request data input by the user to determine the data source features and query intent features;

[0106] Text analysis module 12 is used to integrate historical user representation information with the data source features and query intent features using multi-dimensional strategies to determine the appropriate retrieval rules and the appropriate evaluation system, and to parse the appropriate retrieval rules into structured retrieval conditions.

[0107] The matching scoring module 13 is used to perform preliminary information retrieval based on the structured retrieval conditions after segmenting the text to be retrieved, and to measure the relevance of the candidate texts according to the adaptation evaluation system, and output a semantic matching score set.

[0108] The result output module 14 is used to select and sort the candidate texts according to the semantic matching score set, and output a list of structured search results.

[0109] In one embodiment, the rule parsing module 11 is further configured to:

[0110] Using a deep convolutional neural network, feature parsing is performed on the query request data to determine data source features and query intent features. The data source features include at least text entropy, domain dictionary matching rate, structured tag coverage, average sentence length, and sentence length variance. The query intent features include at least term density, explicit occurrence rate of logical operators, query length, and structural complexity.

[0111] In one embodiment, the text analysis module 12 is further configured to:

[0112] Based on user ID matching, historical query records are retrieved and analyzed to determine historical user representation information. The historical user representation information includes at least user value level, average term density of historical queries, average number of iterations within historical sessions, effective click-through rate of historical queries, and user identity tags.

[0113] This involves fusing historical user representation information with data source features and query intent features using a multi-dimensional strategy to determine suitable retrieval rules and a suitable evaluation system, including:

[0114] Based on the historical user representation information, data source characteristics, and query intent characteristics, the urgency of the current query is assessed, and the urgency coefficient of the current query is determined.

[0115] The maximum response delay time limit is obtained by multiplying the ratio of the preset standard query urgency coefficient to the current query urgency coefficient by the preset response delay time limit;

[0116] Using the maximum response delay limit as a constraint, the appropriate retrieval rules and appropriate evaluation system are determined based on the historical user characterization information, data source characteristics, and query intent characteristics.

[0117] Among them, the maximum response delay time limit is used as a constraint, and the appropriate retrieval rules and appropriate evaluation system are determined based on the historical user characterization information, data source characteristics, and query intent characteristics, including:

[0118] Using the maximum response delay time limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user characterization information, data source characteristics and query intent characteristics;

[0119] Obtain preset search rules and preset evaluation system, wherein the preset search rules include preset logical expression constraints and initial search weight ratios, the search keywords in the preset logical expression constraints include at least mandatory, should be included, and must not be included, and the preset evaluation system includes a similarity measurement algorithm and an initial evaluation weight ratio;

[0120] Based on the stated credible weight ratio, the preset retrieval rules and preset evaluation system are optimized and adjusted to determine the appropriate retrieval rules and evaluation system.

[0121] Among them, the maximum response delay time limit is used as a constraint, and the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user characterization information, data source characteristics, and query intent characteristics, including:

[0122] An information query and analysis plugin is built based on deep learning, which includes a query complexity recognition branch, a query depth recognition branch, and a query breadth recognition branch.

[0123] Using the query complexity identification branch, query depth identification branch, and query breadth identification branch, the query complexity, query depth, and query breadth are identified based on the historical user representation information, data source characteristics, and query intent characteristics, and the current query complexity, query depth, and query breadth are output.

[0124] Using the maximum response delay limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth and current query breadth.

[0125] Among them, the maximum response delay time limit is used as a constraint, and the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth, and current query breadth, including:

[0126] The lower limit of keyword weight ratio for keyword retrieval is determined based on the maximum response delay time limit, wherein the lower limit of keyword weight ratio is negatively correlated with the maximum response delay time limit;

[0127] Set the ratio of the preset standard query complexity to the current query complexity as the first weight compensation coefficient;

[0128] Set the ratio of the preset standard query depth to the current query depth as the second weight compensation coefficient;

[0129] Set the ratio of the preset standard query breadth to the current query breadth as the third weight compensation coefficient;

[0130] The first weight compensation coefficient, the second weight compensation coefficient, and the third weight compensation coefficient are weighted and fused to determine the comprehensive weight compensation coefficient;

[0131] The initial credibility weight of the keyword is compensated according to the comprehensive weight compensation coefficient to obtain the credibility weight of the adapted keyword, wherein the credibility weight of the adapted keyword is not greater than 0.8;

[0132] If the credibility weight of the adapted keyword is less than the lower limit of the keyword weight ratio, then it is adjusted to the lower limit of the keyword weight ratio;

[0133] The credibility weight of the adapted large model is obtained by subtracting the credibility weight of the adapted keyword from 1.

[0134] The trusted weight ratio is generated based on the trusted weight of the adapted keywords and the trusted weight of the adapted large model.

[0135] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0136] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0137] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A general intelligent text retrieval method based on rule parsing, characterized in that, The methods include: Perform feature analysis on the query request data entered by the user to determine the features of the data source and the query intent. The historical user representation information is fused with the data source features and query intent features using a multi-dimensional strategy to determine the appropriate retrieval rules and the appropriate evaluation system, and the appropriate retrieval rules are parsed into structured retrieval conditions. After segmenting the text to be retrieved, preliminary information retrieval is performed based on the structured retrieval conditions, and the candidate texts are relevance measured according to the adaptation evaluation system to output a semantic matching score set. The candidate texts are selected and sorted according to the semantic matching score set, and a list of structured search results is output. This involves fusing historical user representation information with data source features and query intent features using a multi-dimensional strategy to determine suitable retrieval rules and a suitable evaluation system, including: Based on the historical user representation information, data source characteristics, and query intent characteristics, the urgency of the current query is assessed, and the urgency coefficient of the current query is determined. The maximum response delay time limit is obtained by multiplying the ratio of the preset standard query urgency coefficient to the current query urgency coefficient by the preset response delay time limit; Using the maximum response delay limit as a constraint, and based on the historical user representation information, data source characteristics, and query intent characteristics, the appropriate retrieval rules and appropriate evaluation system are determined, including: Using the maximum response delay time limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user characterization information, data source characteristics and query intent characteristics; Obtain preset search rules and preset evaluation system, wherein the preset search rules include preset logical expression constraints and initial search weight ratios, the search keywords in the preset logical expression constraints include at least mandatory, should be included, and must not be included, and the preset evaluation system includes a similarity measurement algorithm and an initial evaluation weight ratio; Based on the stated credible weight ratio, the preset retrieval rules and preset evaluation system are optimized and adjusted to determine the appropriate retrieval rules and evaluation system.

2. The general text intelligent retrieval method based on rule parsing according to claim 1, characterized in that, Using a deep convolutional neural network, feature parsing is performed on the query request data to determine data source features and query intent features. The data source features include at least text entropy, domain dictionary matching rate, structured tag coverage, average sentence length, and sentence length variance. The query intent features include at least term density, explicit occurrence rate of logical operators, query length, and structural complexity.

3. The general intelligent text retrieval method based on rule parsing according to claim 1, characterized in that, Based on user ID matching, historical query records are retrieved and analyzed to determine historical user representation information. The historical user representation information includes at least user value level, average term density of historical queries, average number of iterations within historical sessions, effective click-through rate of historical queries, and user identity tags.

4. The general intelligent text retrieval method based on rule parsing according to claim 1, characterized in that, Using the maximum response delay limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set according to the historical user representation information, data source characteristics, and query intent characteristics, including: An information query and analysis plugin is built based on deep learning, which includes a query complexity recognition branch, a query depth recognition branch, and a query breadth recognition branch. Using the query complexity identification branch, query depth identification branch, and query breadth identification branch, the query complexity, query depth, and query breadth are identified based on the historical user representation information, data source characteristics, and query intent characteristics, and the current query complexity, query depth, and query breadth are output. Using the maximum response delay limit as a constraint, the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth and current query breadth.

5. The general text intelligent retrieval method based on rule parsing according to claim 4, characterized in that, Constrained by the maximum response latency limit, the credibility weight ratio of keyword retrieval and large model retrieval is set based on the current query complexity, current query depth, and current query breadth, including: The lower limit of keyword weight ratio for keyword retrieval is determined based on the maximum response delay time limit, wherein the lower limit of keyword weight ratio is negatively correlated with the maximum response delay time limit; Set the ratio of the preset standard query complexity to the current query complexity as the first weight compensation coefficient; Set the ratio of the preset standard query depth to the current query depth as the second weight compensation coefficient; Set the ratio of the preset standard query breadth to the current query breadth as the third weight compensation coefficient; The first weight compensation coefficient, the second weight compensation coefficient, and the third weight compensation coefficient are weighted and fused to determine the comprehensive weight compensation coefficient; The initial credibility weight of the keyword is compensated according to the comprehensive weight compensation coefficient to obtain the credibility weight of the adapted keyword, wherein the credibility weight of the adapted keyword is not greater than 0.8; If the credibility weight of the adapted keyword is less than the lower limit of the keyword weight ratio, then it is adjusted to the lower limit of the keyword weight ratio; The credibility weight of the adapted large model is obtained by subtracting the credibility weight of the adapted keyword from 1. The trusted weight ratio is generated based on the trusted weight of the adapted keywords and the trusted weight of the adapted large model.

6. A general-purpose intelligent text retrieval system based on rule parsing, characterized in that, The system is used to implement the general text intelligent retrieval method based on rule parsing as described in any one of claims 1-5, the system comprising: The rule parsing module is used to perform feature parsing on the query request data input by the user to determine the data source features and query intent features; The text analysis module is used to integrate historical user representation information with the data source features and query intent features using a multi-dimensional strategy to determine the appropriate retrieval rules and the appropriate evaluation system, and to parse the appropriate retrieval rules into structured retrieval conditions. The matching and scoring module is used to perform preliminary information retrieval based on the structured retrieval conditions after segmenting the text to be retrieved, and to measure the relevance of candidate texts according to the adaptation evaluation system, and output a semantic matching score set. The results output module is used to select and sort the candidate texts according to the semantic matching score set, and output a list of structured search results.

Images