Archives full-text retrieval system and method based on NLP

Abstract

The present invention discloses an archive full-text retrieval system and method based on NLP, which specifically relates to the technical field of information retrieval, and comprises the following steps: analyzing the archive full-text retrieval task within the initial monitoring window, if quality risks are found, real-time monitoring of two preset retrieval modes, extracting features and conducting comprehensive analysis, calculating the comprehensive quality coefficient, judging the triggering of a switching mechanism or a restriction mechanism according to the comprehensive quality coefficient and a threshold, the switching mechanism selecting the high quality coefficient mode, the restriction mechanism sorting the retrieval request, and closing the request entrance when the comprehensive quality coefficients of the two modes are both low; the present invention can realize efficient archive full-text retrieval through real-time monitoring and comprehensive analysis, and can quickly identify and switch to a retrieval mode with better performance by performing preliminary analysis and real-time monitoring of the retrieval task, thereby ensuring the overall retrieval quality and efficiency, and comprehensively analyzing the quality of the two retrieval modes to ensure a balance between high-quality retrieval and efficient resource utilization.

Description

Technical Field

[0001] The present invention relates to the field of information retrieval technology, and more specifically, to an archive full-text retrieval system and method based on NLP. Background Art

[0002] NLP-based archive full-text retrieval is a method that uses natural language processing (NLP) technology to conduct a comprehensive search and analysis of archive documents. Among the existing archive full-text retrieval methods, the following two methods are usually used: Deep learning-based scientific and technological literature citation recommendation method (CN113239181B): This method uses deep learning technology to recommend citations to scientific and technological literature, which can capture the complex semantic relationships between documents and recommend related citations. However, this method requires high computing resources and time when processing complex semantic relationships.

[0003] Scientific research archive management method and system based on NLP (CN116909991B): This method uses natural language processing technology to manage and retrieve scientific research archives, including text parsing, information extraction, semantic understanding, etc., which can effectively improve the accuracy and efficiency of retrieval. However, there may be certain challenges when dealing with complex contexts and diverse semantics. In order to switch between two preset retrieval methods according to the usage environment and ensure the overall retrieval quality and efficiency, this paper proposes an archive full-text retrieval system and method based on NLP. Summary of the invention

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The NLP-based archive full-text retrieval method includes the following steps:

[0006] Obtain the retrieval information of all retrieval tasks in the archive full-text retrieval process under the initial monitoring window, conduct a preliminary analysis of the archive full-text retrieval process, and determine whether there is a hidden danger of insufficient retrieval quality based on the results of the preliminary analysis;

[0007] If there is a hidden danger of insufficient retrieval quality, the operation information of the two preset retrieval methods during the archive full-text retrieval process is monitored in real time, and then feature extraction and abnormal analysis are performed, and then comprehensive analysis is performed to obtain the comprehensive quality coefficients of the two preset retrieval methods;

[0008] Based on the comprehensive quality coefficients of the two preset search methods and the corresponding preset multiple levels of thresholds, it is determined whether to trigger the switching mechanism or the restriction mechanism or to trigger the switching mechanism and the restriction mechanism at the same time;

[0009] When the switching mechanism is triggered, a preset search method corresponding to a comprehensive quality coefficient with a larger value is selected for use. When the restriction mechanism is triggered, the received search requests are sorted, and the search is performed in order or the search request entry is closed;

[0010] The duration of the monitoring window is re-determined based on the result after triggering the regulation and the duration of the initial monitoring window.

[0011] The preliminary analysis of the archive full-text retrieval process refers to:

[0012] The timestamp of the retrieval request and the timestamp of the retrieval completion are received from the retrieval information of all retrieval tasks in the full-text retrieval process of the archives under the initial monitoring window, and only the retrieval tasks with the timestamp of receiving the retrieval request or the timestamp of retrieval completion are eliminated to obtain the response time data of all retained retrieval tasks. The response time data is obtained by subtracting the timestamp of receiving the retrieval request from the timestamp of retrieval completion. Different levels of response time intervals are preset. The total number of response time data of all retained retrieval tasks that fall into response time intervals of different levels is counted and marked as Dji, where j represents the level of the response time interval, and i represents the total number of data that fall into response time interval j, and the maximum and minimum values ​​of the response time data are recorded.

[0013] In a preferred embodiment, judging whether there is a hidden danger of insufficient retrieval quality based on the results of the preliminary analysis refers to:

[0014] A comparison threshold 1 and a comparison threshold 2, as well as benchmark quantity values ​​of response time intervals of different levels are preset. The maximum value and the minimum value of the response time data are compared with the comparison threshold 1 and the comparison threshold 2 respectively, and the total number Dji of response time intervals falling into different levels is compared with the benchmark quantity values ​​of response time intervals of different levels. If the maximum value of the response time data is less than or equal to the comparison threshold 1, the minimum value of the response time data is greater than or equal to the comparison threshold 2, and the total number Dji of response time intervals falling into different levels does not exceed the benchmark quantity values ​​of response time intervals of different levels, a normal signal is generated; if the maximum value of the response time data is less than or equal to the comparison threshold 1, the minimum value of the response time data is greater than or equal to the comparison threshold 2, and the total number Dji of response time intervals falling into different levels does not exceed the benchmark quantity values ​​of response time intervals of different levels, a hidden danger signal is generated. When the hidden danger signal is generated, it indicates that there is a hidden danger of insufficient retrieval quality.

[0015] In a preferred embodiment, feature extraction refers to: for two preset search modes, respectively extracting resource consumption features and search accuracy features;

[0016] Abnormal analysis refers to: calculating the resource consumption index through the extracted resource consumption features, and calculating the retrieval precision index through the retrieval precision features.

[0017] In a preferred embodiment, the logic for obtaining the resource consumption index is:

[0018] In a fixed time window, record the CPU usage time, memory usage, and time consumption of each retrieval of all retained retrieval tasks, and then perform standardization. Combine the standardized CPU usage time, memory usage, and time consumption to obtain m data points. The number of all retained retrieval tasks is m;

[0019] The DBSCAN algorithm is used for density clustering analysis: the maximum distance ϵ between two points that are considered neighbors and the minimum number of points minPts required to form a cluster are set. Points that do not belong to any cluster are marked as noise points and removed. Then, the average distance between points within all clusters is obtained. The resource consumption index is numerically the same as the average distance between points within all clusters.

[0020] In a preferred embodiment, the logic for obtaining the retrieval precision index is:

[0021] In a fixed time window, record the number of relevant documents TPi correctly retrieved each time, the number of irrelevant documents FPi incorrectly retrieved each time, the number of relevant documents FNi not retrieved each time, and the number of irrelevant documents TNi correctly excluded each time for all retained retrieval tasks, and then calculate the accuracy, recall and F1 score of each time respectively. The accuracy calculation formula is: ; It represents the accuracy of the retrieval task number i. The recall rate is calculated as follows: ; It represents the recall rate of the retrieval task number i. The F1 score calculation formula is: ; represents the F1 score of the retrieval task number i, and the accuracy , recall rate , F1 score The vector 1 is obtained by combining them. A standard vector 2 is preset. The Euclidean distance between the vector 1 and the standard vector 2 is calculated. Then, the average value of all the Euclidean distances is obtained. The retrieval accuracy index is numerically the same as the average value of all the Euclidean distances.

[0022] In a preferred embodiment, a comprehensive analysis is performed to obtain the comprehensive quality coefficients of the two preset search methods, which are:

[0023] The resource consumption index and retrieval precision index corresponding to the two preset retrieval methods are obtained respectively, and then substituted into the same pre-trained machine learning model to generate the comprehensive quality coefficients corresponding to the two preset retrieval methods respectively.

[0024] In a preferred embodiment, based on the comprehensive quality coefficients of the two preset search methods and the corresponding preset multiple levels of thresholds, the logic for determining whether to trigger the switching mechanism or the restriction mechanism or to trigger the switching mechanism and the restriction mechanism at the same time is:

[0025] The comprehensive quality coefficients of the two preset retrieval methods are obtained respectively, and fuzzy reasoning is applied. The comprehensive quality coefficient is defined as the input variable, and the comprehensive quality coefficient is fuzzy processed through preset multiple levels of thresholds, divided into different quality levels and converted into fuzzy sets. The output variable is the type of triggered mechanism, and the output variable is fuzzified and converted into a fuzzy set. Fuzzy rules are formulated to describe the mechanism triggering requirements under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to determine the type of triggered mechanism.

[0026] In a preferred embodiment, when the restriction mechanism is triggered, sorting the received search requests refers to:

[0027] When the restriction mechanism is triggered, the various indicator data of the impact of the received search request on the priority are obtained, and then weighted summation is performed to obtain the priority value of the search request, and the search requests are sorted according to the priority value.

[0028] In a preferred embodiment, the NLP-based archive full-text retrieval system includes a data acquisition module, a preliminary analysis module, a hidden danger judgment module, an abnormal monitoring and analysis module, a comprehensive analysis module, a decision-making module, and a control module;

[0029] The data acquisition module is responsible for collecting archival data from different data sources;

[0030] The preliminary analysis module conducts preliminary analysis on all retrieval tasks in the full-text retrieval process of archives under the initial monitoring window;

[0031] The hidden danger judgment module judges whether there are hidden dangers of insufficient retrieval quality based on the results of the preliminary analysis;

[0032] When there is a hidden danger of insufficient retrieval quality, the abnormal monitoring and analysis module monitors the operation information of the two preset retrieval methods in the full-text retrieval process of archives in real time, and then extracts features and performs abnormal analysis;

[0033] The comprehensive analysis module performs comprehensive analysis based on the results of the abnormal analysis to obtain the comprehensive quality coefficients of the two preset retrieval methods;

[0034] The decision module makes decisions on the type of mechanism triggering based on the comprehensive quality coefficients of the two preset retrieval methods and the corresponding preset multiple levels of thresholds;

[0035] The control module performs corresponding control operations according to the type of mechanism triggering and the preset control strategy.

[0036] Technical effects and advantages of the present invention:

[0037] The present invention provides a method for full-text archive retrieval based on NLP, which combines two preset retrieval methods and realizes efficient full-text archive retrieval through real-time monitoring and comprehensive analysis. By performing preliminary analysis and real-time monitoring on the retrieval task, it is possible to quickly identify and switch to a retrieval method with better performance, thereby improving the overall retrieval efficiency of the system.

[0038] The present invention extracts resource consumption characteristics and retrieval accuracy characteristics, uses the DBSCAN algorithm to perform density clustering analysis, ensures efficient use of system resources, and reduces resource waste. By calculating the retrieval accuracy index and resource consumption index, the quality of the two retrieval methods is comprehensively analyzed using a machine learning model to ensure that the system strikes a balance between high-quality retrieval and efficient resource utilization. According to the effect after regulation and the length of the initial monitoring window, the length of the monitoring window is dynamically adjusted to ensure the real-time and flexibility of the system and avoid the impact of frequent regulation on system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;

[0040] Figure 1 This is a schematic diagram of the NLP-based archive full-text retrieval method of the present invention.

[0041] Figure 2 This is a schematic diagram of the NLP-based archive full-text retrieval system in the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Reference Figure 1 - Figure 2 The following embodiments are obtained:

[0044] Embodiment 1: A method for full-text retrieval of archives based on NLP, comprising the following steps:

[0045] Obtain the retrieval information of all retrieval tasks in the archive full-text retrieval process under the initial monitoring window, conduct a preliminary analysis of the archive full-text retrieval process, and determine whether there is a hidden danger of insufficient retrieval quality based on the results of the preliminary analysis;

[0046] If there is a hidden danger of insufficient retrieval quality, the operation information of the two preset retrieval methods during the archive full-text retrieval process is monitored in real time, and then feature extraction and abnormal analysis are performed, and then comprehensive analysis is performed to obtain the comprehensive quality coefficients of the two preset retrieval methods;

[0047] Based on the comprehensive quality coefficients of the two preset search methods and the corresponding preset multiple levels of thresholds, it is determined whether to trigger the switching mechanism or the restriction mechanism or to trigger the switching mechanism and the restriction mechanism at the same time;

[0048] When the switching mechanism is triggered, the preset search method corresponding to the comprehensive quality coefficient with a larger value is selected for use. When the restriction mechanism is triggered, the received search requests are sorted, and the search is performed in order or the search request entry is closed. Closing the search request entry means that the comprehensive quality coefficients of the two methods are at a low level, and sorting the received search requests and searching in order cannot meet the normal operation of the search system.

[0049] It should be noted that the two preset search methods refer to:

[0050] CN113239181B: A method for recommending scientific and technological literature citations based on deep learning, Abstract: This patent introduces a method for recommending scientific and technological literature citations based on deep learning, which is mainly used for recommending and retrieving literature. By extracting information and removing noise from text data, the accuracy of recommendation and retrieval is improved. Main features: Using deep learning technology, information extraction and noise removal are performed on the content of literature. By constructing a vector representation of the literature, similarity calculation is used for recommendation and retrieval.

[0051] CN116909991B: Scientific research archive management method and system based on NLP, description: This technology is mainly used for the management and retrieval of scientific research archives, using NLP technology to process and understand document content.

[0052] CN113239181B: A deep learning-based scientific literature citation recommendation method, which mainly relies on deep learning technology to analyze the content of scientific literature and train models to recommend relevant citations. It can usually capture the complex semantic relationships and topic relevance between documents.

[0053] CN116909991B: A scientific research archive management method and system based on NLP. This method focuses on the management and retrieval of scientific research archives, and uses natural language processing (NLP) technology to process and analyze scientific research documents, including text parsing, information extraction, semantic understanding, etc., in order to more effectively manage and retrieve scientific research data.

[0054] Commonalities: Technical basis: Both methods rely on advanced artificial intelligence technologies, namely deep learning and natural language processing. Application areas: Both are suitable for processing scientific and technological literature or scientific research archives to improve the efficiency and relevance of information search.

[0055] Differences: Technical core: CN113239181B focuses on deep learning, emphasizing model training and recommendation algorithms; CN116909991B focuses on NLP, emphasizing text processing and information extraction technology.

[0056] Application focus: The former focuses more on the construction and optimization of recommendation systems, while the latter focuses more on the accuracy and efficiency of document management and full-text retrieval. The impact of different places on retrieval quality: Deep learning methods (CN113239181B) may better capture the implicit relationships and citation recommendations between documents through large-scale data and complex neural network models, but may face efficiency issues when data in specific fields is insufficient or real-time requirements are high. NLP methods (CN116909991B) can process text data more carefully, using semantic analysis and information extraction technology to improve retrieval accuracy and efficiency, but there may be certain challenges for complex contexts and semantic understanding. The choice of which method to use depends on the specific application scenario and needs.

[0057] The preliminary analysis of the archive full-text retrieval process refers to:

[0058] The timestamp of the retrieval request and the timestamp of the retrieval completion are received from the retrieval information of all retrieval tasks in the full-text retrieval process of the archives under the initial monitoring window. Only the retrieval tasks with the timestamp of the retrieval request or the timestamp of the retrieval completion are recorded and eliminated to obtain the response time data of all the retrieval tasks that are retained. The response time data is obtained by subtracting the timestamp of the retrieval request from the timestamp of the retrieval completion. Different levels of response time intervals are preset. The total number of response time data of all the retrieval tasks that fall into the response time intervals of different levels is counted and marked as Dji, where j represents the level of the response time interval, and i represents the total number of data that fall into the response time interval j, and the maximum and minimum values ​​of the response time data are recorded. Response time data is one of the key indicators for evaluating the retrieval efficiency and performance of the system. By recording and analyzing the response time of each retrieval task, it can reflect the length of time required for the system to process the request, which helps to evaluate the system's rapid response capability and efficiency. Detect potential performance problems: Preliminary analysis can identify whether there is an abnormal situation of too long response time by comparing the preset response time interval and the actual response time data. If the response time of the retrieval task exceeds the preset reasonable range, it may indicate that there is a problem with the system performance or it needs to be optimized and adjusted. Setting Service Level Agreements (SLAs): Response time data can be used as one of the bases for developing and implementing service level agreements. By monitoring and recording the response time of retrieval tasks, you can ensure that the system meets or exceeds the service level standards expected by users. Improving user experience: A fast-responding retrieval system can improve user experience and enhance user satisfaction.

[0059] Judging whether there are hidden dangers of insufficient search quality based on the results of the preliminary analysis refers to:

[0060] A comparison threshold 1 and a comparison threshold 2 are preset, as well as the benchmark quantity values ​​of response time intervals of different levels. The maximum value and the minimum value of the response time data are compared with the comparison threshold 1 and the comparison threshold 2 respectively, and the total number Dji of response time intervals falling into different levels is compared with the benchmark quantity values ​​of response time intervals of different levels. If the maximum value of the response time data is less than or equal to the comparison threshold 1, the minimum value of the response time data is greater than or equal to the comparison threshold 2, and the total number Dji of response time intervals falling into different levels does not exceed the benchmark quantity values ​​of response time intervals of different levels, a normal signal is generated; if the maximum value of the response time data is less than or equal to the comparison threshold 1, the minimum value of the response time data is greater than or equal to the comparison threshold 2, and the total number Dji of response time intervals falling into different levels does not exceed the benchmark quantity values ​​of response time intervals of different levels, a hidden danger signal is generated. When the hidden danger signal is generated, it indicates that there is a hidden danger of insufficient retrieval quality. By setting the preset comparison threshold and the benchmark quantity value of the response time interval, the system can automatically monitor and evaluate the key indicators in the retrieval process. This automated detection can send out warning signals in a timely manner when there are potential problems with retrieval quality, helping to prevent and reduce the occurrence of degraded user experience or other potential problems. By comparing the maximum and minimum values ​​of response time data and the situations falling into different levels of response time intervals, the system can quickly determine whether there are hidden dangers of insufficient retrieval quality. This early warning and problem identification can help system administrators or technical support personnel adjust system configurations or optimize algorithms in a timely manner, thereby improving retrieval efficiency and quality.

[0061] Feature extraction refers to: for two preset retrieval methods, respectively extracting resource consumption features and retrieval accuracy features; anomaly analysis refers to: calculating the resource consumption index through the extracted resource consumption features, and calculating the retrieval accuracy index through the retrieval accuracy features.

[0062] The logic of obtaining the resource consumption index is as follows: within a fixed time window, record the resource consumption characteristics of all retained retrieval tasks, such as the CPU usage time of each retrieval, the memory usage of each retrieval, and the time consumption of each retrieval. The purpose of collecting these data is to quantify the usage of system resources by each retrieval task and provide basic data for subsequent analysis. Then, standardization is performed. Standardization can eliminate the dimensional differences between different features, making the subsequent clustering analysis more reasonable and accurate. The standardized CPU usage time, memory usage, and time consumption are combined to obtain m data points. The number of all retained retrieval tasks is m, and each data point is a three-dimensional vector. The combined data points can be used for subsequent clustering analysis. Each data point represents the resource consumption characteristics of a retrieval task.

[0063] Use the DBSCAN algorithm to perform density clustering analysis: set the maximum distance ϵ between two points that are considered neighbors and the minimum number of points minPts required to form a cluster. Points that do not belong to any cluster are marked as noise points and removed. Then, the average distance between points in all clusters is obtained. The resource consumption index is numerically the same as the average distance between points in all clusters. Specifically, set the maximum distance ϵ between two points that are considered neighbors and the minimum number of points minPts required to form a cluster. Use the DBSCAN algorithm to perform density clustering analysis on data points: for each data point p, find the points in its ϵ neighborhood. If the number of points in the neighborhood is greater than or equal to minPts, mark p as a core point and start a new cluster. Recursively add all points in the neighborhood to the cluster. Mark points that do not belong to any cluster as noise points and remove them. The DBSCAN algorithm can identify task clusters with similar resource consumption characteristics and filter out noise points. These clusters represent different patterns of system resource consumption. For each formed cluster, the average distance between the points in the cluster is calculated. The Euclidean distance can be used to calculate the distance between any two points in the cluster. The average distance between the points in the cluster reflects the similarity of the resource consumption characteristics in the cluster. The smaller the distance, the more consistent the resource consumption. The resource consumption index is defined as the average of the average distances between the points in all clusters. Assuming that k clusters are formed, it is the average of the average distances between the points in these k clusters. The resource consumption index is numerically the same as the average distance between the points in all clusters, reflecting the resource utilization efficiency of the system in the retrieval task. A lower resource consumption index indicates that the system uses resources more concentratedly and efficiently when processing retrieval tasks. A low resource consumption index indicates that the resource consumption characteristics of the system are relatively consistent when processing tasks, the system operates efficiently, and the resource utilization rate is good. A high resource consumption index indicates that the resource consumption characteristics of the system are relatively scattered when processing tasks, and there may be problems of resource waste or unbalanced use, which requires further optimization and adjustment.

[0064] The logic for obtaining the retrieval precision index is:

[0065] In a fixed time window, record the retrieval accuracy characteristics of all retained retrieval tasks: the number of relevant documents TPi correctly retrieved each time, the number of irrelevant documents FPi incorrectly retrieved each time, the number of relevant documents not retrieved each time FNi, and the number of irrelevant documents correctly excluded each time TNi, and then calculate the accuracy, recall and F1 score of each time respectively. The accuracy calculation formula is: ; It represents the accuracy of the retrieval task number i. The recall rate is calculated as follows: ; It represents the recall rate of the retrieval task number i. The F1 score calculation formula is: ; represents the F1 score of the retrieval task number i, and the accuracy , recall rate , F1 score Combine them to get vector 1, and preset standard vector 2. Define a preset standard vector 2 as the ideal retrieval performance indicator. Standard vector 2 represents the accuracy, recall rate and F1 score in the ideal situation, that is, all retrieval results are completely correct and comprehensive. Calculate the Euclidean distance between vector 1 and standard vector 2. The Euclidean distance measures the difference between the actual retrieval performance and the ideal retrieval performance. The smaller the distance, the closer the retrieval performance is to the ideal value. Then calculate the average value of all Euclidean distances. The retrieval precision index is numerically the same as the average value of all Euclidean distances. The retrieval precision index reflects the overall retrieval performance of the system in a fixed time window. The smaller the value, the better the retrieval performance of the system. The retrieval precision index (SPI) is the average difference between the actual retrieval performance and the ideal retrieval performance. The smaller the value, the closer the accuracy, recall rate and F1 score of the system in the retrieval task are to the ideal value, that is, the higher the retrieval quality; the larger the value, the greater the gap between the system's performance on these indicators and the ideal value, and the lower the retrieval quality.

[0066] After comprehensive analysis, the comprehensive quality coefficients of the two preset search methods are:

[0067] Respectively obtain the resource consumption index and retrieval precision index corresponding to the two preset retrieval methods, and then substitute them into the same pre-trained machine learning model to generate the comprehensive quality coefficients corresponding to the two preset retrieval methods. The machine learning model is not specifically limited here. Any model that can perform a comprehensive analysis of the resource consumption index and the retrieval precision index to generate a comprehensive quality coefficient is acceptable. In order to implement the technical solution of the present invention, the present invention provides a specific implementation method: for each preset retrieval method, collect and calculate its resource consumption index and retrieval precision index within a fixed time window. The above-mentioned indicator data are standardized for analysis by the machine learning model. Data collection and preprocessing are the basis of model analysis, ensuring the standardization and consistency of input data, improving the prediction accuracy of the model, and selecting a suitable machine learning model, such as a linear regression model. ; Indicates the comprehensive quality coefficient of the preset search method 1, ; Indicates the comprehensive quality coefficient of the preset search method 2. is a constant term, which means that when all independent variables (here, resource consumption index and retrieval precision index) are zero, the value of the predicted comprehensive quality coefficient is the baseline value of the model. is the coefficient of the resource consumption index (RCI), which indicates the influence of the resource consumption index on the comprehensive quality coefficient. A negative value means that an increase in the resource consumption index will lead to a decrease in the comprehensive quality coefficient. is the coefficient of the retrieval precision index (SPI), which indicates the influence of the retrieval precision index on the comprehensive quality coefficient. is a positive value, which means that an increase in the retrieval precision index will lead to an increase in the comprehensive quality coefficient. Indicates the resource consumption index after normalization of preset search method 1. Represents the retrieval precision index after standardized processing of preset retrieval method 1, Indicates the resource consumption index after normalization of preset search method 2. Represents the normalized retrieval precision index of the preset retrieval method 2. The comprehensive quality coefficient is a comprehensive evaluation indicator. It quantifies the overall quality of the retrieval method by considering the resource consumption and retrieval precision of the system within a fixed time window. It reflects the comprehensive performance of the system in two aspects and can be used to compare and optimize different retrieval methods. Overall performance evaluation: The comprehensive quality coefficient integrates the resource consumption index and the retrieval precision index, providing a unified indicator to measure the overall performance of the system, which helps to understand the performance of the system in balancing resource utilization and retrieval accuracy. Optimization and improvement direction: By analyzing the comprehensive quality coefficient, it can be determined which retrieval method has better overall performance, thereby guiding system optimization and improvement work. A higher comprehensive quality coefficient indicates that the retrieval method has achieved a better balance between resource consumption and retrieval precision. Decision support: In practical applications, the comprehensive quality coefficient can be used to automate decision-making processes, such as selecting between different retrieval methods and triggering switching mechanisms to ensure the best retrieval quality and efficiency.

[0068] Based on the comprehensive quality coefficients of the two preset search methods and the corresponding preset multiple levels of thresholds, the logic for determining whether to trigger the switching mechanism or the restriction mechanism or to trigger the switching mechanism and the restriction mechanism at the same time is:

[0069] Obtain the comprehensive quality coefficients of the two preset retrieval methods respectively, apply fuzzy reasoning, define the comprehensive quality coefficient as the input variable, fuzzify the comprehensive quality coefficient through the preset multiple levels of thresholds, divide it into different quality levels and convert it into a fuzzy set, the output variable is the type of triggered mechanism, fuzzify the output variable, convert the output variable into a fuzzy set, formulate fuzzy rules, describe the mechanism triggering requirements under different data type combinations, and reason the fuzzified input variables through fuzzy rules to determine the type of triggered mechanism. From the results of the previous calculation, obtain the comprehensive quality coefficients of the two preset retrieval methods respectively. For example, assume that the quality coefficients are Q1 and Q2. Define input variables and output variables: Input variables: comprehensive quality coefficients Q1 and Q2. Output variables: type of triggered mechanism (switching mechanism, restriction mechanism, simultaneous triggering of switching and restriction mechanism).

[0070] Fuzzy processing: preset multiple levels of thresholds, fuzzify the comprehensive quality coefficient, divide it into different quality levels and convert it into fuzzy sets. Assume that the comprehensive quality coefficient is divided into three levels: low, medium and high, and the corresponding division standards are: low: [0,0.4], medium: [0.3,0.7], high: [0.6,1]. Fuzzy rules are formulated based on actual needs and experience to describe the mechanism triggering requirements under different data type combinations. Example rule: If Q1 is high and Q2 is low and the search method corresponding to Q2 is in use at this time, the switching mechanism is triggered and the search method corresponding to Q1 is used. If Q1 is medium and Q2 is low and the search method corresponding to Q2 is in use at this time, the restriction mechanism and switching mechanism are triggered, and the search method corresponding to Q1 is used, and the received search requests are sorted and searched. If Q1 is low and Q2 is low, no matter which search method is in use at this time, the restriction mechanism is triggered, and the search request entrance is closed at this time. Low: The comprehensive quality coefficient is in the range of [0,0.4], indicating that the search quality is low and may need to be optimized or restricted. Medium: The comprehensive quality coefficient is in the range of [0.3, 0.7], indicating that the retrieval quality is medium and acceptable within a certain range, but potential performance issues need to be noted. High: The comprehensive quality coefficient is in the range of [0.6, 1], indicating that the retrieval quality is high and the system runs well; ultimately, the mechanism trigger logic based on the comprehensive quality coefficient and fuzzy reasoning can be realized, which can flexibly process complex system performance data, provide accurate mechanism trigger decisions, and improve the system's operating efficiency and stability.

[0071] When the restriction mechanism is triggered, sorting the received search requests means: when the restriction mechanism is triggered, obtaining the various indicator data of the received search requests that affect the priority, and then performing weighted summation to obtain the priority value of the search request, and sorting the search requests according to the priority value; when the restriction mechanism is triggered, it is necessary to calculate the priority value of each search request based on multiple indicators. These indicators may include but are not limited to the following items: Urgency: Indicates the urgency of the search request. For example, requests from VIP users or tasks that must be completed within a specific time can be given a higher degree of urgency. Importance of the request: Indicates the importance of the request to the business. For example, requests involving critical business operations can be given a higher importance. Identity of the requester: The role or identity of the requester. For example, requests from senior management can be given a higher priority, and the formula is illustrated as follows: ; YXi is the priority value of the retrieval request, N is the total number of indicators, Yi is the data corresponding to each indicator, and wi is the preset proportional coefficient corresponding to each indicator, ensuring that the system gives priority to high-priority retrieval requests, thereby improving the efficiency and responsiveness of the system.

[0072] Embodiment 2: An archive full-text retrieval system based on NLP, comprising a data collection module, a preliminary analysis module, a hidden danger judgment module, an abnormality monitoring and analysis module, a comprehensive analysis module, a decision-making module, and a control module;

[0073] The data acquisition module is responsible for collecting archival data from different data sources;

[0074] The preliminary analysis module conducts preliminary analysis on all retrieval tasks in the full-text retrieval process of archives under the initial monitoring window;

[0075] The hidden danger judgment module judges whether there are hidden dangers of insufficient retrieval quality based on the results of the preliminary analysis;

[0076] When there is a hidden danger of insufficient retrieval quality, the abnormal monitoring and analysis module monitors the operation information of the two preset retrieval methods in the full-text retrieval process of archives in real time, and then extracts features and performs abnormal analysis;

[0077] The comprehensive analysis module performs comprehensive analysis based on the results of the abnormal analysis to obtain the comprehensive quality coefficients of the two preset retrieval methods;

[0078] The decision module makes decisions on the type of mechanism triggering based on the comprehensive quality coefficients of the two preset retrieval methods and the corresponding preset multiple levels of thresholds;

[0079] The control module performs corresponding control operations according to the type of mechanism triggering and the preset control strategy. The control operation also includes re-determining the duration of the monitoring window according to the result after the triggering control and the duration of the initial monitoring window. Specifically, according to the type of the triggered mechanism (switching mechanism, restriction mechanism or both), select the corresponding control strategy. Switching mechanism triggering: Strategy: Select a retrieval method with a higher comprehensive quality coefficient and switch to this retrieval method for processing. Restriction mechanism triggering: Strategy: Sort according to the priority value of the retrieval request and process it in order from high to low according to the priority value. It may be necessary to restrict requests with low priority values ​​or close the retrieval request entrance. Simultaneously trigger the switching and restriction mechanisms: Strategy: While switching to the retrieval method with a higher comprehensive quality coefficient, prioritize the received retrieval requests and restrict requests with low priority values ​​or close the retrieval request entrance. Re-determine the duration of the monitoring window according to the result after the triggering control and the duration of the initial monitoring window. The specific method for re-determining the duration of the monitoring window is as follows: Evaluate the control effect: Monitor the performance indicators of the system after the control operation, such as response time, compare the system performance before and after the control, and evaluate the control effect. Adjust the monitoring window duration: If the control effect is significant and the system performance is significantly improved, the monitoring window duration can be appropriately extended to reduce frequent control operations. If the control effect is not obvious and the system performance improvement is limited, the monitoring window duration can be shortened to allow for more frequent monitoring and control. Formula for adjusting the monitoring window: ; TN represents the new monitoring window duration, TC represents the initial monitoring window duration, α is the adjustment coefficient, which is set according to the actual situation, ΔP is the change in system performance before and after regulation, and PC represents the value corresponding to the initial system performance indicator.

[0080] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0081] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0082] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0083] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0084] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. The NLP-based archive full-text retrieval method is characterized by: The following steps are involved: Obtain the retrieval information of all retrieval tasks in the archive full-text retrieval process under the initial monitoring window, conduct a preliminary analysis of the archive full-text retrieval process, and determine whether there is a hidden danger of insufficient retrieval quality based on the results of the preliminary analysis; If there is a hidden danger of insufficient retrieval quality, the operation information of the two preset retrieval methods during the archive full-text retrieval process is monitored in real time, and then feature extraction and abnormal analysis are performed, and then comprehensive analysis is performed to obtain the comprehensive quality coefficients of the two preset retrieval methods; Based on the comprehensive quality coefficients of the two preset search methods and the corresponding preset multiple levels of thresholds, it is determined whether to trigger the switching mechanism or the restriction mechanism or to trigger the switching mechanism and the restriction mechanism at the same time; When the switching mechanism is triggered, a preset search method corresponding to a comprehensive quality coefficient with a larger value is selected for use. When the restriction mechanism is triggered, the received search requests are sorted, and the search is performed in order or the search request entry is closed; The preliminary analysis of the archive full-text retrieval process refers to: The timestamp of the retrieval request and the timestamp of the retrieval completion are received from the retrieval information of all retrieval tasks in the full-text retrieval process of the archives under the initial monitoring window, and only the retrieval tasks with the timestamp of the retrieval request received or the timestamp of the retrieval completion are eliminated to obtain the response time data of all the retained retrieval tasks. The response time data is obtained by subtracting the timestamp of the retrieval request received from the timestamp of the retrieval completion. Different levels of response time intervals are preset. The total number of response time data of all the retained retrieval tasks that fall into response time intervals of different levels is counted and marked as Dji, where j represents the level of the response time interval, and i represents the total number of data that fall into the response time interval j, and the maximum and minimum values ​​of the response time data are recorded; Judging whether there are hidden dangers of insufficient search quality based on the results of the preliminary analysis refers to: A comparison threshold 1 and a comparison threshold 2, as well as benchmark quantity values ​​of response time intervals of different levels are preset. The maximum value and the minimum value of the response time data are compared with the comparison threshold 1 and the comparison threshold 2 respectively, and the total number Dji of response time intervals falling into different levels is compared with the benchmark quantity values ​​of response time intervals of different levels. If the maximum value of the response time data is less than or equal to the comparison threshold 1, the minimum value of the response time data is greater than or equal to the comparison threshold 2, and the total number Dji of response time intervals falling into different levels does not exceed the benchmark quantity values ​​of response time intervals of different levels, a normal signal is generated; if the maximum value of the response time data is less than or equal to the comparison threshold 1, the minimum value of the response time data is greater than or equal to the comparison threshold 2, and the total number Dji of response time intervals falling into different levels does not exceed the benchmark quantity values ​​of response time intervals of different levels, a hidden danger signal is generated. When the hidden danger signal is generated, it indicates that there is a hidden danger of insufficient retrieval quality.

2. The NLP-based archive full-text retrieval method according to claim 1, characterized in that: Feature extraction refers to: for two preset retrieval methods, respectively extracting resource consumption features and retrieval accuracy features; Abnormal analysis refers to: calculating the resource consumption index through the extracted resource consumption features, and calculating the retrieval precision index through the retrieval precision features.

3. The NLP-based archive full-text retrieval method according to claim 2, characterized in that: The logic for obtaining the resource consumption index is: In a fixed time window, record the CPU usage time, memory usage, and time consumption of each retrieval of all retained retrieval tasks, and then perform standardization. Combine the standardized CPU usage time, memory usage, and time consumption to obtain m data points. The number of all retained retrieval tasks is m; The DBSCAN algorithm is used for density clustering analysis: the maximum distance ϵ between two points that are considered neighbors and the minimum number of points minPts required to form a cluster are set. Points that do not belong to any cluster are marked as noise points and removed. Then, the average distance between points within all clusters is obtained. The resource consumption index is numerically the same as the average distance between points within all clusters.

4. The NLP-based archive full-text retrieval method according to claim 3, characterized in that: The logic for obtaining the retrieval precision index is: In a fixed time window, record the number of relevant documents TPi correctly retrieved each time, the number of irrelevant documents FPi incorrectly retrieved each time, the number of relevant documents FNi not retrieved each time, and the number of irrelevant documents TNi correctly excluded each time for all retained retrieval tasks, and then calculate the accuracy, recall and F1 score of each time respectively. The accuracy calculation formula is: ; It represents the accuracy of the retrieval task number i. The recall rate is calculated as follows: ; It represents the recall rate of the retrieval task number i. The F1 score calculation formula is: ; represents the F1 score of the retrieval task number i, and the accuracy , recall rate , F1 score The vector 1 is obtained by combining them. A standard vector 2 is preset. The Euclidean distance between the vector 1 and the standard vector 2 is calculated. Then, the average value of all the Euclidean distances is obtained. The retrieval accuracy index is numerically the same as the average value of all the Euclidean distances.

5. The NLP-based archive full-text retrieval method according to claim 4, characterized in that: After comprehensive analysis, the comprehensive quality coefficients of the two preset search methods are: The resource consumption index and retrieval precision index corresponding to the two preset retrieval methods are obtained respectively, and then substituted into the same pre-trained machine learning model to generate the comprehensive quality coefficients corresponding to the two preset retrieval methods respectively.

6. The NLP-based archive full-text retrieval method according to claim 5, characterized in that: Based on the comprehensive quality coefficients of the two preset search methods and the corresponding preset multiple levels of thresholds, the logic for determining whether to trigger the switching mechanism or the restriction mechanism or to trigger the switching mechanism and the restriction mechanism at the same time is: The comprehensive quality coefficients of the two preset retrieval methods are obtained respectively, and fuzzy reasoning is applied. The comprehensive quality coefficient is defined as the input variable, and the comprehensive quality coefficient is fuzzy processed through preset multiple levels of thresholds, divided into different quality levels and converted into fuzzy sets. The output variable is the type of triggered mechanism, and the output variable is fuzzified and converted into a fuzzy set. Fuzzy rules are formulated to describe the mechanism triggering requirements under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to determine the type of triggered mechanism.

7. The NLP-based archive full-text retrieval method according to claim 6, characterized in that: When the restriction mechanism is triggered, the order of the received retrieval requests refers to: When the restriction mechanism is triggered, the various indicator data of the impact of the received search request on the priority are obtained, and then weighted summation is performed to obtain the priority value of the search request, and the search requests are sorted according to the priority value.

8. An NLP-based archive full-text retrieval system, based on the NLP-based archive full-text retrieval method according to any one of claims 1 to 7, characterized in that: It includes data collection module, preliminary analysis module, hidden danger judgment module, abnormal monitoring and analysis module, comprehensive analysis module, decision-making module and control module; The data acquisition module is responsible for collecting archival data from different data sources; The preliminary analysis module performs preliminary analysis on all retrieval tasks in the full-text retrieval process of archives under the initial monitoring window; The hidden danger judgment module judges whether there are hidden dangers of insufficient retrieval quality based on the results of the preliminary analysis; When there is a hidden danger of insufficient retrieval quality, the abnormal monitoring and analysis module monitors the operation information of the two preset retrieval methods in the full-text retrieval process of archives in real time, and then extracts features and performs abnormal analysis; The comprehensive analysis module performs comprehensive analysis based on the results of the abnormal analysis to obtain the comprehensive quality coefficients of the two preset retrieval methods; The decision module makes decisions on the type of mechanism triggering based on the comprehensive quality coefficients of the two preset retrieval methods and the corresponding preset multiple levels of thresholds; The control module performs corresponding control operations according to the type of mechanism triggering and the preset control strategy.

Images