A keyword-based information extraction method and system

Abstract

The application relates to a keyword-based information extraction method and system, and relates to the field of information extraction, which comprises the following steps: receiving information text; extracting keywords from the information text based on a preset extraction algorithm; determining an information field to which the information text belongs based on the keywords; obtaining a context category and a context correlation of the keywords based on a preset algorithm; filling the keywords into a preset blank context module based on the context category and the context correlation to obtain a keyword context module; constructing a scene simulation based on the keyword context module to obtain key content conforming to an actual scene and outputting the key content. The application has the effect of effectively capturing deep semantic correlation by constructing a keyword context correlation network, combining semantic connection strength and statistical characteristics, and avoiding deviation caused by semantic loss in traditional weight analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information extraction, and in particular to a keyword-based information extraction method and system. BACKGROUND

[0002] With the rapid development of the Internet and digital technology, the global data volume is growing exponentially. The emergence of massive unstructured text data has led to an increasingly serious information overload problem. To meet the demand for quickly obtaining core content, keyword-based information extraction methods have emerged as the times require. The core goal is to identify keywords of interest to users from complex text through automated technology and quickly locate associated content to achieve efficient filtering and aggregation of information.

[0003] In the existing automatic information extraction process, weight analysis is performed based on the number of occurrences, position, and other attributes of keywords. The weight coefficient is used to express the closeness of the association of keywords to information content. The prior art determines the closeness of association by setting a weight coefficient, combines dynamic weight coefficients for keywords, and calculates the product of keyword weights (M = Q1 x... x Qn) within the sentence unit between punctuation marks. The top 10 sentences with the highest product are extracted as abstracts to achieve accurate removal of redundant information and improve reading efficiency. γ

[0004] In view of the above related technologies, it is difficult to accurately capture semantic associations and contexts during weight analysis of keywords, which may result in excessive reliance on statistical features of keywords such as weight products, while ignoring in-depth analysis at the semantic level, leading to deviations in the extraction of core content. SUMMARY

[0005] To solve the problem of difficulty in accurately capturing semantic associations and contexts during weight analysis of keywords, which may result in excessive reliance on statistical features of keywords such as weight products, while ignoring in-depth analysis at the semantic level, leading to deviations in the extraction of core content, the present application provides a keyword-based information extraction method and system.

[0006] In a first aspect, the present application provides a keyword-based information extraction method, which adopts the following technical solution:

[0007] A keyword-based information extraction method, comprising:

[0008] Step 1: receiving information text;

[0009] Step 2: extracting keywords from the information text based on a preset extraction algorithm;

[0010] Step 3: determining the information field to which the information text belongs based on the keywords;

[0011] ​Step 4: obtaining the context category of the keyword based on a preset correlation algorithm;

[0012] Step 5: filling the keyword into a preset blank context module based on the context category to obtain a keyword context module;

[0013] Step 6: constructing a scene simulation based on the keyword context module to obtain a key content conforming to an actual scene and outputting.

[0014] Optionally, the steps after Step 5 and before Step 6 further include:

[0015] Step 50: finding a corresponding correlation from a preset keyword database based on the keyword in any one of the keyword context modules and other keywords;

[0016] Step 51: constructing the keyword based on the correlation to form a context correlation network;

[0017] Step 52: analyzing the number of connection edges and the corresponding connection strength of the keyword in the context correlation network;

[0018] Step 53: obtaining the number of occurrences and the position parameter of the keyword in the information text based on a preset traversal algorithm;

[0019] Step 54: determining a keyword importance value of the keyword based on the number of connection edges, the connection strength, the number of occurrences and the position parameter;

[0020] Step 55: screening the keyword whose keyword importance value is lower than a preset importance threshold value, discarding it from the keyword context module to obtain a revised keyword context module, and defining the remaining keywords of the revised keyword context module as revised keywords;

[0021] Step 56: outputting the revised keyword context module as the keyword context module;

[0022] Step 57: when the number of revised keywords is less than a preset number threshold value, outputting a preset module alarm signal.

[0023] Optionally, when the number of revised keywords is less than the number threshold value, the method of outputting the module alarm signal includes:

[0024] Step 570: reducing a preset keyword judgment threshold value in the extraction algorithm to form a modified extraction algorithm, the extraction algorithm being provided with the keyword judgment threshold value;

[0025] Step 571: re-executing Steps 2 to 55 based on the modified extraction algorithm.

[0026] Step 5710: when the number of the modified keywords after the re-execution step is greater than the number threshold, execute step 6;

[0027] Step 5711: when the number of the modified keywords after the re-execution step is still less than the number threshold, output the module alarm signal.

[0028] Optionally, another method of information extraction when the number of the modified keywords after the re-execution step is still less than the number threshold, the method comprises:

[0029] Step 57110: search for a suspected information text from the Internet based on the keywords;

[0030] Step 57111: calculate a similarity value based on a preset judgment algorithm between the suspected information text and the information text;

[0031] Step 57112: when the similarity value corresponding to the suspected information text is higher than a preset similarity threshold, define the suspected information text as a similar information text;

[0032] Step 57113: re-execute steps 2 to 4 based on the similar information text to obtain a replacement keyword context module and a replacement keyword;

[0033] Step 57114: merge the replacement keyword context module and the modified keyword context module to form a hybrid keyword context module;

[0034] Step 57115: determine a hybrid number based on the replacement keyword and the modified keyword;

[0035] Step 571150: when the hybrid number is greater than the number threshold, output the hybrid keyword context module as the keyword context module;

[0036] Step 571151: when the hybrid number is less than the number threshold, discard the blank context module corresponding to the hybrid keyword context module.

[0037] Optionally, the method of constructing a scene simulation based on the keyword context module comprises:

[0038] Step 60: determine a module attribute based on the keyword context module;

[0039] Step 61: search for a corresponding scene simulation scheme, a scene simulation scheme credibility, and a scene simulation scheme credibility threshold from a preset scene simulation library based on all the module attributes and the information field;

[0040] Step 610: If the scenario simulation scheme credibility is greater than the scenario simulation scheme credibility threshold, then simulate the keyword context module corresponding to the keyword in the scenario simulation scheme;

[0041] Step 611: If the scenario simulation scheme credibility is less than the scenario simulation scheme credibility threshold, then output a preset information missing alarm signal.

[0042] Optionally, the method of simulating the keyword context module corresponding to the keyword in the scenario simulation scheme comprises:

[0043] Step 6100: Based on the scenario simulation scheme, splice the keyword context module into a complete sentence;

[0044] Step 6101: Quantitatively analyze the logic of the complete sentence to obtain a logic value of the complete sentence;

[0045] Step 6102: Based on the logic value, sort the complete sentence to obtain a credibility sorting of the complete sentence;

[0046] Step 6103: Based on the credibility sorting, output the complete sentence with the highest logic value as the key content, and select a preset number of complete sentences from the remaining complete sentences as supplementary content of the key content according to the credibility sorting.

[0047] Optionally, the method of sorting the complete sentence based on the logic value to obtain the credibility sorting of the complete sentence comprises:

[0048] Step 61020: Based on the information field, find the corresponding logic difference threshold from a preset logic database;

[0049] Step 61021: Based on the logic difference threshold, combine the complete sentences to form a complete sentence group, and the absolute difference between the logic values of any two complete sentences in the complete sentence group is less than the logic difference threshold;

[0050] Step 61022: Based on the keyword in the complete sentence group, combine to obtain a combined keyword;

[0051] Step 61023: Based on the scenario simulation scheme, splice the combined keyword into a combined complete sentence;

[0052] Step 61024: Define the maximum logic value in the keyword context module in the complete sentence group as the maximum logic value;

[0053] Step 61025: quantitatively analyzing the logicality of the combined complete sentence to obtain a combined logical value of the combined complete sentence;

[0054] Step 610250: if the combined logical value is greater than the maximum logical value, replacing all complete sentences of the complete sentence group with the combined complete sentence and performing sorting according to step 6102;

[0055] Step 610251: if the combined logical value is less than the maximum logical value, not performing combination and directly sorting the complete sentences based on the logical value to obtain a credibility ranking of the complete sentences.

[0056] Optionally, if the combined logical value is less than the maximum logical value, a method for judging whether to not perform combination, the method comprising:

[0057] Step 6102510: analyzing the complete sentence group to obtain logical relationships and credibility values;

[0058] Step 6102511: based on the logical relationships, finding corresponding credibility value thresholds from a preset credibility value database;

[0059] Step 6102512: if the credibility value is higher than the credibility value threshold, replacing all complete sentences of the complete sentence group with the combined complete sentence;

[0060] Step 6102513: if the credibility value is lower than the credibility value threshold, not performing combination.

[0061] Optionally, based on the credibility ranking, a method for outputting the complete sentence with the highest credibility as the key content and selecting a preset number of complete sentences from the remaining complete sentences as supplementary content of the key content, the method comprising:

[0062] Step 61030: judging content relevance of the information text based on the complete sentence;

[0063] Step 61031: if the content relevance is higher than a preset reasonable relevance threshold, outputting a preset number of complete sentences with the highest credibility as the key content and selecting a preset number of complete sentences from the remaining complete sentences as supplementary content of the key content according to the credibility ranking;

[0064] Step 61032: if the content relevance is lower than a preset reasonable relevance threshold, dividing the information text based on content relevance to obtain a secondary information text;

[0065] Step 61033: classifying the complete sentence based on the secondary information text to obtain a secondary complete sentence;

[0066] Step 61034: outputting a preset number of the secondary complete sentences with the highest credibility as the respective key content, and selecting a preset number of the secondary complete sentences from the remaining secondary complete sentences as supplementary content of the key content according to the credibility.

[0067] In a second aspect, the present application provides a key information extraction system, which adopts the following technical solution:

[0068] A key information extraction system comprises:

[0069] An acquisition module is configured to acquire the information text and the suspected information text.

[0070] A memory is configured to store a program of a control method of a key information extraction method.

[0071] A processor is configured to load and execute the program in the memory to implement a key information extraction method.

[0072] In summary, the present application has at least one of the following beneficial technical effects:

[0073] By constructing a key context association network, combining semantic connection strength and statistical characteristics, deep semantic association can be effectively captured, and the deviation caused by semantic loss in traditional weight analysis can be avoided.

[0074] By presetting the framework constraint of the blank context module, the key words are forced to be embedded in a reasonable semantic position according to the context association relationship, the logical break caused by the key word stacking is avoided, and the content coherence is improved.

[0075] By dynamically combining the modular semantic units, the logical relationship of the real scene is simulated, the output content is ensured to meet the actual application scene, and the theoretical distortion is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 FIG. 1 is a flowchart of a key information extraction method and system according to an embodiment of the present application.

[0077] Figure 2 FIG. 3 is a schematic diagram of a key context module acquisition method according to an embodiment of the present application.

[0078] Figure 3 FIG. 4 is a schematic diagram of a context association network according to an embodiment of the present application.

[0079] Figure 4is a flowchart of constructing a scenario simulation according to a keyword context module in the embodiment of the present application. DETAILED DESCRIPTION

[0080] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0081] The embodiment of the present application discloses a keyword-based information extraction method. Referring to Figure 1 A keyword-based information extraction method includes:

[0082] Step 1: receiving information text.

[0083] Information text refers to text with the core purpose of transmitting information or knowledge, such as news information, industry reports, popular science articles, etc. The program can receive the input information text data through standard input, file reading or network request interface, and convert it into a string for processing.

[0084] Step 2: extracting keywords from information text based on a preset extraction algorithm.

[0085] The extraction algorithm refers to a calculation method based on statistical, semantic analysis or rule matching technology to automatically identify candidate words from unstructured text and calculate their importance, such as common term frequency statistics (TF-IDF), co-occurrence relationship analysis or deep learning models (such as TextRank).

[0086] And the keyword refers to the smallest semantic unit that can summarize the core content of the text, usually a noun or noun phrase. The extraction method of the keyword is usually to disassemble the text into a word sequence by using a word segmentation tool, then filter out the non-meaningful words using a stop word list, and finally sort the target number of words according to the word frequency weight and context correlation, and select the result.

[0087] For example, based on semantic unit, the information text is divided into title, lead, body, conclusion and other structural blocks, and a position weight gradient is established, such as title weight = 3, first paragraph = 2, body = 1, last paragraph = 1.5 (the weight parameter can be changed according to the actual situation). Secondly, the word filter strategy is adopted to remove adjectives, adverbs and other modifier words, and only nouns, verbs and professional terms are retained as the candidate set; finally, a double scoring mechanism is constructed to integrate word frequency statistics and position weighted score (such as "artificial intelligence" appears 1 time in the title, which counts 3 points, and appears 3 times in the body, which counts 3x1 = 3 points, the total score is 6 points), and the core keywords are dynamically selected according to the score threshold.

[0088] Step 3: determining the information field to which the information text belongs based on the keywords.

[0089] Information field refers to the industry, discipline or topic category involved or attributed by the text content. For example, "technology-artificial intelligence", "finance-blockchain", "medical-tumor immunity" and other sub-directions. Compare the keywords with the preset field database, extract all the information fields corresponding to the keywords, and the information field with the highest occurrence frequency is the information field of the information text. The field database is obtained by a large number of experiments by the staff. The process of the experiment is as follows: after extracting the keywords from a large number of information texts of different fields, the extracted keywords are matched with the corresponding information fields, and the matching relationship is stored in the field database.

[0090] Step 4: Obtain the context category of the keyword based on the preset association algorithm.

[0091] Context category refers to the semantic role it assumes in the text, such as subject, event, time, etc.

[0092] The association algorithm is to calculate the keyword matching degree of the candidate keyword library by using the preset candidate keyword library, and to calculate the context adaptation degree of the keyword to each context type by combining the context relationship. The specific calculation method is as follows: context adaptation degree=keyword matching degree×context weight. The candidate keyword library is obtained by a large number of experiments by the staff, and the candidate keyword library includes candidate keywords, context categories and attribution probability. The process of the experiment is as follows: classify a large number of keywords of different context categories, and calculate the probability of the keyword appearing in the context, which is used as the attribution probability of the candidate keyword. Finally, the mapping relationship of candidate keywords, context categories and attribution probability is stored in the candidate keyword library. For example, the candidate keyword library can include:

[0093] Time context: day, month, future, 2025, etc. Subject context: company, team, laboratory, school, etc. Match the obtained keywords with the candidate keyword library, such as filtering the same keywords, then the keyword matching degree is the attribution probability value of the candidate keyword, and if no completely same keyword is filtered, the keyword with the most number of characters is filtered from the candidate keyword library and the attribution probability value is multiplied by a correction parameter. The correction parameter is calculated as follows: The keyword matching degree of the keyword not filtered is=correction parameter×attribution probability.

[0094] For example, the keyword "last week" exists in the candidate keyword library, and the attribution probability of "last week" to the time context is 0.9. Therefore, the module matching degree of the keyword "last week" in the information text is 0.9. However, "high-temperature superconducting energy storage device" does not exist in the candidate keyword library, but "energy storage device" exists, and the attribution probability of "energy storage device" to the object context is 0.8. The correction parameter of "energy storage device" and "high-temperature superconducting energy storage device" is 4 / 8 = 1 / 2. Therefore, the keyword library matching degree of the keyword "high-temperature superconducting energy storage device" in the information text is 0.8 x 0.5 = 0.4.

[0095] The context association relationship of the keyword can be obtained by analyzing the grammatical structure of the sentence to which the keyword belongs. For example, the keyword "school" in the information text is "school held a welcome party". Through grammatical structure analysis, it can be known that "school" is the subject in this sentence, so "school" is not suitable for "place context", and therefore the context weight of "place context" is 0. However, "school" is suitable for "subject context", so the context weight of "subject context" is 1. Similarly, for example, the keyword "school" in the information text belongs to the sentence "school held a welcome party", "held on the school playground", and "school provides party venue". In the above sentences, "school" belongs to both "subject context" and "place context". According to the frequency of occurrence in different contexts, the context weight of "school" as "subject context" is 2 / 3, and the context weight of "school" as "place context" is 1 / 3.

[0096] The context adaptation degree of the keyword to each context category is calculated, and the keyword is assigned to one or more context categories in descending order of context adaptation degree.

[0097] Step 5: Fill the keywords into the preset blank context module based on the context category to obtain the keyword context module.

[0098] The context module refers to a structured framework that can carry information according to the context. Each module is composed of keyword slots of a specific context type, which is used to systematize the keywords in the information text. Each blank context module will contain the corresponding context category, for example, "time module (when)" will contain the time context type. Save the keywords with the same context category as the context module into the blank context module to obtain the keyword context module. The specific operation process can be referred to Figure 2 .

[0099] Step 6: Construct a scene simulation for the keyword context module to obtain the key content that conforms to the actual scene and output.

[0100] Scenario simulation refers to restoring or constructing the actual application scenario described in the text based on the keyword context module, including but not limited to the structured information of Who, What, When, How, etc., combining the context association relationship and the information field, through logical deduction and element reorganization, and generating the core content expression conforming to the scenario. For example, from the information text of "enterprise releases new product", the Who (enterprise), What (product), When (time), How (technology highlight) modules are extracted and combined into a complete sentence "a certain enterprise releases a certain technology product at a certain time", which is the core content of the information text. This modular design enables scattered keywords to be dynamically combined according to logical relationships, while maintaining the semantic independence between different modules, providing a structured foundation for subsequent information reorganization and scenario application.

[0101] The step after step 5 and before step 6 further comprises:

[0102] Step 50: Find the corresponding association from the preset keyword database based on the keywords in any one keyword context module and other keywords.

[0103] Association refers to the direct or indirect contact between two keywords in terms of semantics, context or domain knowledge. The system automatically matches the association strength of the keywords through the preset keyword database.

[0104] The keyword database is set by multiple technical personnel in the field. The staff first classifies the keywords by analyzing a large number of information texts, and then determines the association between the keywords in the module and other keywords through artificial experience. If two keywords have a direct explanatory relationship in semantics, often appear together in similar texts, or belong to the same domain knowledge system, they are marked as strong association. If the association is weak or requires specific conditions to produce a connection, the two keywords are marked as weak association. At the same time, the frequency of the two keywords appearing together in information texts is saved as the association strength. After multiple reviews, the association result is recorded in the database in the format "keyword A - keyword B - association strength" for subsequent calling. When receiving a keyword in any one keyword context module and other keywords, the association between the two keywords is automatically found from the database and output.

[0105] Step 51: Construct keywords based on association to form a context association network.

[0106] Context association network refers to a structured topology system composed of keyword nodes and their association relationships. Nodes represent keywords, edges represent association, such as semantic similarity, knowledge graph path or co-occurrence frequency, and edge weights reflect association strength.

[0107] Step 52: Analyze the number of connection edges and the corresponding connection strength of the keywords in the context association network.

[0108] The number of connection edges refers to the number of direct associations of a keyword with other nodes in the context association network. The connection strength refers to the weight of each edge. Referring to Figure 3 , each keyword in the figure is a node, and the number on the line connecting each keyword is the weight of the connection strength between the two keywords, while the number of lines is the number of connection edges. For example, the connection strength between "mobile phone" and "battery" is 0.9, and the number of connection edges of "mobile phone" is 6.

[0109] Step 53: Obtain the number of occurrences and position parameters of the keywords in the information text based on the preset traversal algorithm.

[0110] The traversal algorithm refers to the process of systematically scanning each character or semantic unit in the information text, which is used to comprehensively capture the distribution characteristics of the keywords. The "number of occurrences" refers to the frequency of the cumulative occurrence of the keywords in the text, and the "position parameter" quantifies the spatial characteristics of the keywords in the text, such as whether they appear in the title, the first paragraph, the last paragraph, etc. The traversal algorithm is well known and will not be described here.

[0111] By analyzing the text structure through the traversal algorithm, the number of occurrences of the keywords is counted and their position parameters are recorded.

[0112] Step 54: Determine the keyword importance value of the keywords based on the number of connection edges, connection strength, number of occurrences, and position parameters.

[0113] The keyword importance refers to the comprehensive evaluation of the contribution of the keywords to the content theme by quantifying the connection characteristics (such as the number of connection edges and strength) of the keywords in the context association network, the frequency of occurrence in the text, and the position distribution. First, count the total number of connection edges and connection strength of the keywords in the association network, combine the number of occurrences (frequency weight) and position parameters (such as the addition of key positions such as the beginning and end of the paragraph, title, etc.) in the text, and calculate the comprehensive score through a weighted algorithm or an empirical formula.

[0114] Substitute the corresponding parameters into the empirical formula: weight x ∑ (number of edges x connection strength) + weight x ∑ (number of occurrences x position parameter) = keyword importance.

[0115] Step 55: Select the keywords whose keyword importance value is lower than the preset importance threshold, discard them from the keyword context module to obtain a revised keyword context module, and define the remaining keywords in the revised keyword context module as revised keywords.

[0116] The importance threshold refers to a numerical threshold preset according to the field requirements or task objectives, used to measure the core degree of the keyword in the context network. The keyword below the threshold is determined as a non-core keyword. The revised keyword context module refers to the structured context framework remaining after the low importance value keywords are removed, and only the keywords that significantly contribute to the content theme are retained. The revised keyword refers to the remaining keywords that significantly contribute to the content theme in the revised keyword context module after removing the non-core keywords with importance values below the preset importance threshold. The importance threshold is obtained by the staff through experiments, the keywords with weak contribution to the theme are manually screened and excluded, and finally through multiple rounds of experiments and adjustment, the relevance of the keywords and the content theme is repeatedly verified, the critical keyword importance value that can reflect the relevance is determined, and the importance threshold is set.

[0117] Step 56: output the revised keyword context module as the keyword context module.

[0118] Step 57: when the number of revised keywords is less than the preset number threshold, output a preset module alarm signal.

[0119] The module alarm signal is a signal that prompts the risk of information missing or incomplete analysis in the current context module. In information processing, if the number of revised keywords after importance screening is insufficient, it is determined that the semantic information of the information text is too sparse to support effective scene simulation or decision reasoning, and an alarm signal is output to alert data quality defects or algorithm missed detection.

[0120] In addition, when the number of revised keywords is greater than the preset number threshold, the revised keyword context module is output as the keyword context module.

[0121] When the number of revised keywords is less than the number threshold, the method for outputting the module alarm signal includes:

[0122] Step 570: reduce the preset keyword determination threshold in the extraction algorithm to form a modified extraction algorithm.

[0123] The extraction algorithm is provided with a keyword judgment threshold. The modified extraction algorithm refers to an optimized extraction algorithm that supplements more potential keywords by relaxing the screening conditions to expand the range of candidate keywords through dynamic adjustment of the judgment threshold in the original keyword extraction algorithm. When the number of revised keywords is less than the number threshold, the keyword judgment threshold reduction range is queried through the preset keyword judgment threshold table. The keyword judgment threshold table is obtained through a large number of experiments by the staff. The process of the experiment is as follows: in different information fields, the keyword judgment threshold is gradually reduced until irrelevant words appear in the extracted keywords, and then the staff records the threshold reduction range and matches it with the corresponding information field and stores it in the keyword judgment threshold table. Then, when the number of revised keywords is less than the number threshold, the system automatically reduces the judgment threshold of the extraction algorithm according to the table.

[0124] For example, when the original algorithm results in insufficient effective keywords due to a too high threshold, the keyword threshold parameter can be reduced from 0.8 to 0.6 to allow more keywords that were not originally included to enter the candidate pool and participate in subsequent steps.

[0125] Step 571: Re-execute steps 2 to 55 based on the modified extraction algorithm.

[0126] Step 5710: When the number of revised keywords after re-execution of the steps is greater than the number threshold, execute step 6.

[0127] After obtaining the revised keyword context module, it is still necessary to determine whether the keywords obtained after modifying the extraction algorithm can be applied after keyword importance value calculation. For example, when the number of revised keywords after re-execution of the steps is greater than the number threshold, it indicates that the newly extracted keywords meet the requirements, so that the corresponding revised keyword context module can be applied to the next step.

[0128] Step 5711: When the number of revised keywords after re-execution of the steps is still less than the number threshold, the output module outputs an alarm signal.

[0129] When the number of revised keywords after re-execution of the steps is still less than the number threshold, it indicates that the newly obtained keywords still cannot meet the minimum number required by the revised keyword context module, and therefore the corresponding keywords do not have sufficient relevance to the core of the information content. The information text is suspected to lack key information, so the output module outputs an alarm signal.

[0130] Another method of information extraction when the number of revised keywords after re-execution of the steps is still less than the number threshold is also included. The method comprises:

[0131] Step 57110: Search for suspected information text from the Internet based on the keywords.

[0132] Suspected information text refers to a set of candidate texts retrieved from the Internet through keyword matching, which may be related to the content of the target information text. All keywords are input into search engines such as Baidu, and a preset number of suspected information texts are downloaded and imported into the program in chronological order from recent to distant. The preset number is obtained by the staff through experiments.

[0133] Step 57111: Calculate the similarity value based on the suspected information text and the information text using a preset judgment algorithm.

[0134] The judgment algorithm refers to a mathematical model or rule used to calculate the similarity of texts, such as the cosine similarity algorithm. The cosine similarity algorithm measures the directional similarity of two vectors by calculating the cosine of the angle between them in a multi-dimensional space. The core idea is that the closer the vector direction, the more similar the content.

[0135] For example, text 1: "Machine learning is the core of artificial intelligence", text 2: "Deep learning is a branch of machine learning". Extract the vocabulary from text 1 and text 2: {"machine learning", "artificial intelligence", "core", "deep learning", "branch"}.

[0136] The vector of text 1 is {0.8, 0.4, 0.4, 0, 0}, and the vector of text 2 is {0.5, 0, 0, 0.7, 0.7}, After calculating the similarity value using the cosine similarity algorithm, compare it with the threshold value to determine whether the two texts are similar.

[0137] Similarly, the Jaccard index can also be used to determine whether two texts are similar. The above two methods are used to determine the similarity, and the specific actual means include but are not limited to the above two methods.

[0138] Step 57112: If the similarity value of the suspected information text is higher than the preset similarity threshold, define the suspected information text as similar information text.

[0139] The similarity threshold is a numerical standard preset in the information text similarity judgment, which is used to measure the similarity between two texts. The similarity threshold of different information fields can be obtained by the cosine similarity method in step 57111. When the similarity value of the suspected information text is higher than the similarity threshold, it means that the similarity between the suspected information text and the information text is high, and both of them reflect the same core information content.

[0140] This step quantifies the similarity index to objectively select texts with semantic association, avoiding the errors of subjective judgment by humans, and significantly improving the efficiency and accuracy of information deduplication or clustering.

[0141] Step 57113: Re-perform steps 2 to 4 based on the similar information text to obtain the alternative keyword context module and the alternative keyword.

[0142] The alternative keyword context module refers to the combination of all keywords and their corresponding context modules extracted from the similar information text. The alternative keyword refers to all keywords extracted from the similar information text. After obtaining the similar information text, steps 2 to 4 will be re-performed to extract keywords from the above text using the association algorithm, and these keywords are defined as alternative keywords.

[0143] Step 57114: Merge the alternative keyword context module and the revised keyword context module to form a hybrid keyword context module.

[0144] The hybrid keyword context module is obtained by merging the alternative keyword context module and the revised keyword context module, which has all the keywords of both.

[0145] For example, for information about solar panels, the alternative keyword context module provides the keyword "photovoltaic technology", and the revised keyword context module provides the keyword "solar panels". The "who" module in the hybrid keyword context module after merging will contain both "photovoltaic technology" and "solar panels".

[0146] Step 57115: Determine the hybrid quantity based on the alternative keyword and the revised keyword.

[0147] The hybrid quantity refers to the total number of de-duplicated words obtained by merging the alternative keyword and the revised keyword. After merging the keywords in the alternative keyword context module and the revised keyword context module, the hybrid keyword set is obtained by removing duplicates, and finally the number of words in the set is counted as the hybrid quantity.

[0148] Step 571150: When the hybrid quantity is greater than the quantity threshold, output the hybrid keyword context module as the keyword context module.

[0149] When the hybrid quantity is greater than the quantity threshold, it means that the number of alternative keywords extracted from the similar information text is sufficient to simulate according to the scenario simulation scheme, which meets the requirements of step 620, and can be output as the keyword context module.

[0150] Step 571151: When the hybrid quantity is less than the quantity threshold, discard the blank context module corresponding to the hybrid keyword context module.

[0151] When the hybrid quantity is less than the quantity threshold, it means that even if the similar information text is extracted, there is no suitable keyword for the blank context module, so the corresponding blank context module is discarded.

[0152] Referring to Figure 4 The method for constructing a keyword context module includes:

[0153] Step 60: determining a module attribute based on the keyword context module.

[0154] The module attribute refers to an attribute to which the keyword context module belongs, and is used to represent a role represented by a keyword in the keyword context module in the scenario simulation. The keyword context module is obtained by filling a keyword into a preset blank context module, and thus the keyword context module inherits a context category carried by the blank context module. For example, the keyword in the "who" module includes a module attribute of who, and the keyword in the module usually plays a role of a subject in the scenario simulation. Thus, the module attribute can be determined by querying the keyword context module.

[0155] Step 61: finding a corresponding scenario simulation scheme, a scenario simulation scheme credibility, and a scenario simulation scheme credibility threshold from a preset scenario simulation library based on all the module attributes and the information field.

[0156] The scenario simulation scheme is a standardized operation framework designed based on a scenario template, and the framework is used to combine the keywords in the keyword context module. The scenario simulation scheme credibility is a matching degree index set according to historical data or expert experience, and is used to judge a fitting degree of the module attribute and the preset scenario, so as to avoid simulation deviation caused by a low matching degree scheme. The scenario simulation library stores a mapping relationship among the module attribute, the information field, the scenario simulation scheme, the scenario simulation scheme credibility, and the scenario simulation scheme credibility threshold.

[0157] The scenario simulation library is obtained by analyzing and testing existing different types of information texts and core contents thereof by staff. In the testing process, the staff extracts core contents from a large amount of information texts, compares the contents of the information texts with contents of similar information texts in a form of expression, and classifies the information texts into different modules. Based on the correlation among the modules, a scheme is formed. Meanwhile, the staff dynamically corrects the credibility of the scheme according to a missing condition of the module. The credibility of the scheme is evaluated, and finally all the data records are stored in the scenario simulation library. When all the module attributes and the information field are input, the system finds a scenario simulation scheme formed by the input module attributes in the information field from the scenario simulation library, and finds a corresponding scenario simulation scheme credibility and a scenario simulation scheme credibility threshold in the information field.

[0158] Step 610: if the scenario simulation scheme credibility is greater than the scenario simulation scheme credibility threshold, simulating the keyword in the scenario simulation scheme corresponding to the keyword context module.

[0159] Simulation refers to substituting the qualified keyword context module into the selected scene simulation scheme to generate a structured expression or execute instructions.

[0160] When the scene simulation scheme credibility is greater than the preset scene simulation scheme credibility threshold, it means that the scheme can be implemented, and the keyword corresponding to the keyword context module is simulated in the scene simulation scheme.

[0161] Step 611: When the scene simulation scheme credibility is less than the scene simulation scheme credibility threshold, output the preset information missing alarm signal.

[0162] The information missing alarm signal refers to the alarm signal of the early warning mechanism triggered automatically by the system when the scene simulation scheme credibility is lower than the preset threshold, which is used to prompt the user that the current keyword context module has incomplete, conflicting or noisy information, and needs manual intervention or data supplement. The output method is a text output method, and the user interface outputs "information missing alarm!".

[0163] The method of simulating the keyword corresponding to the keyword context module in the scene simulation scheme includes:

[0164] Step 6100: Based on the scene simulation scheme, the keyword context module is spliced into a complete sentence.

[0165] The complete sentence refers to a semantic smooth and structurally complete expression spliced based on the logical association and grammar rules of the keyword context module (such as subject, event, time, and manner), which needs to meet the accurate transmission of core information and human language habits. When simulating the keyword corresponding to the keyword context module in the scene simulation scheme, since the fixed scene simulation scheme cannot completely and accurately meet the human language habits, a language model needs to be used for correction. The following is the use of NLP model to splice and correct the keyword context module into a complete sentence.

[0166] For example, the above-mentioned condition for simulating based on the scene simulation scheme 2, the following simulation scene is obtained:

[0167] Scene 1: "Recently, company A released car A, bringing 2w sales."

[0168] Scene 2: "This spring, company A released car A, bringing 2w sales." (Note: This is the first part of the scene simulation, and the accuracy and logic of the sentence have not been improved, such as "bringing 2w sales" in the example, which is obviously a grammatical error.)

[0169] At this time, the NLP model is used to analyze the complete sentence in scenario 1, and the NLP model is used for fine-grained segmentation: {“recently / t”, “company / n”, “A / eng”, “release / v”, “car / n”, “A / eng”, “bring / v”, “sell / v”, “2w / m”, “table / q”}. Based on the fine-grained segmentation, it is found that the key verb “bring / v” is abnormally matched with the same verb “sell / v”, and it is concluded that “bring / sell 2w tables” is a clear grammatical error. Since “sell 2w tables” is a keyword obtained from the information text, it is not modified, so “bring / v” needs to be changed. According to the Chinese grammar rule library analysis, the result noun object must be connected to the verb “bring”, and the “sell+quantity” structure should be preceded by the adverb “has / accumulated”, and the obtained sentence in the scenario is replaced with “Recently, the company A released the car A, and has sold 2w tables.”

[0170] Step 6101: Quantitative analysis of the logicality of the complete sentence to obtain the logical value of the complete sentence.

[0171] The logical value is a quantitative result of whether the complete sentence has logicality, which is used to measure whether the sentence conforms to the logical rigor in human cognition.

[0172] In order to obtain the logical value of the complete sentence, it is necessary to analyze the correlation strength of the complete sentence and the information text, as well as the matching degree of the keywords in each context module, the cause-effect relationship, the time consistency and other factors. Then, after weighting correction, all factors are added to obtain the logical value of the corresponding complete sentence.

[0173] For example, the logical value of the complete sentence is calculated based on the context module matching degree, the cause-effect relationship rationality and the information density. Among them, the context module matching degree can be calculated by using the cosine similarity algorithm, and the remaining factors can also be calculated by using the corresponding algorithm. Here, the related algorithms and factors are well known, and will not be described here.

[0174] Step 6102: Sort the complete sentence based on the logical value to obtain the credibility sorting of the complete sentence.

[0175] The credibility sorting refers to the order of credibility. By quantitatively evaluating the logical value of each complete sentence, the sentences are prioritized according to the logical value from high to low. After the credibility sorting, it can be directly judged whether the complete sentence is credible and which complete sentence can better reflect the core content of the information text.

[0176] Step 6103: The complete sentence with the highest logical value is output as the key content, and a preset number of complete sentences from the remaining complete sentences are selected as the supplementary content of the key content according to the credibility sorting.

[0177] The staff determines the number of complete sentences to be supplemented by analyzing a large amount of information text, combining the characteristics of the field to which the information belongs, and comprehensively considering user demand and information density. This is to enable the supplemented content to provide sufficient and reasonable information reflecting the key content of the information text.

[0178] For example, the complete sentences obtained based on the information text are "In April, enterprise A sold 2000 cars A." (credibility 0.9), "This month, enterprise A sold many cars A." (credibility 0.8), and "This month, enterprise A announced good news." (credibility 0.7). When outputting according to the rules, "In April, enterprise A sold 2000 cars A." is directly outputted as the key content of the information text, and "This month, enterprise A sold many cars A." and "This month, enterprise A announced good news." are outputted as the supplementary content of the information text below the key content and are marked as "supplementary content".

[0179] The method for sorting the complete sentences based on the logical values to obtain the credibility sorting of the complete sentences includes:

[0180] Step 61020: Based on the information field, find the corresponding logical difference threshold from the preset logical database.

[0181] The logical database is a structured storage system for saving the logical difference threshold and other logical rule parameters corresponding to different information fields. The logical difference threshold is a preset numerical parameter for measuring the maximum allowed difference between the logical values of two complete sentences. When the absolute value of the difference between the logical values of two sentences is less than the threshold, the system considers that their logical relevance is strong enough to be combined into a more complete sentence; otherwise, if the difference exceeds the threshold, it is considered that the logical relevance is weak and is not combined.

[0182] The logical database is obtained based on a large number of experiments by the staff. In the experiment, after extracting the core of the information text based on the above-mentioned keywords, it is judged whether different complete sentences reflect the core of the article. Finally, different logical difference thresholds are set for different fields. When the absolute value of the difference between the logical values of any two complete sentences is less than the threshold, it means that the two complete sentences reflect the content of the information text to some extent. When the system receives the information field, it automatically finds the logical difference threshold under the condition of the corresponding information field from the database.

[0183] Step 61021: Combine the complete sentences based on the logical difference threshold to form a complete sentence group.

[0184] The absolute difference between the logical values of any two complete sentences in the complete sentence group is less than the logical difference threshold. The complete sentence group is a set composed of a group of semantically complete and logically consistent sentences. By traversing all the sentences, all sentence groups whose logical value difference is less than the logical difference threshold are screened out from the first sentence to form a sentence group, and then the second sentence is executed to the end.

[0185] Step 61022: Combine based on keywords in the complete sentence group to obtain combined keywords.

[0186] The combined keywords are a set of keywords extracted from the complete sentence group and formed by logical association and semantic integration, aiming to integrate multiple sentence core information and cover the complete theme of the information, while avoiding redundancy. For example, the complete sentence group obtained contains "Enterprise A sells mobile phone A." and "Enterprise A sells mobile phone B.", in order to completely reflect the core content of the information text, the keywords are extracted and combined, that is, the keywords "mobile phone A" and "mobile phone B" are combined into "mobile phone A and mobile phone B".

[0187] Step 61023: Combine the combined keywords into a combined complete sentence based on the scenario simulation scheme.

[0188] The combined complete sentence is a complete sentence that is semantically complete, logically consistent and covers the core theme by integrating keywords and logical relationships in the complete sentence group. For example, in step 61022, the combined keywords are reassembled to obtain the combined complete sentence "Enterprise A sells mobile phone A and mobile phone B."

[0189] Step 61024: Define the maximum logical value as the maximum logical value in the keyword context module in the complete sentence group.

[0190] Step 61025: Quantitatively analyze the logicality of the combined complete sentence to obtain the combined logical value of the combined complete sentence.

[0191] The combined logical value is a numerical measure of the logical consistency of the combined complete sentence, which is used to quantify the overall rationality after merging multiple sentence information. The core goal is to analyze the relevance of elements such as semantics, structure, time, and place between sentences to determine whether the combination result is logical and free of contradictions.

[0192] The calculation method of the combined logical value is the same as the logical calculation step of step 6101, which is not described here.

[0193] Step 610250: If the combined logical value is greater than the maximum logical value, replace all complete sentences in the complete sentence group with the combined complete sentence and sort according to step 6102.

[0194] When the combined logic value is greater than the maximum logic value, it indicates that the combined complete sentence obtained after combination is more consistent with the information text and can better reflect the core content, and the combined complete sentence is replaced to form all complete sentences of the complete sentence group, which are sorted according to step 6102.

[0195] For example, the complete sentence "Enterprise A sells mobile phone A and mobile phone B" is obtained by combining "Enterprise A sells mobile phone A" and "Enterprise A sells mobile phone B".

[0196] Step 610251: When the combined logic value is less than the maximum logic value, do not combine, and directly sort the complete sentences based on the logic value to obtain the credibility ranking of the complete sentences.

[0197] For example, "Enterprise A sells mobile phone A" and "Enterprise B sells mobile phone B" both have high logic values and reflect the main content of the information text. However, the combined complete sentence obtained after combination is "Enterprise A and Enterprise B sell mobile phone A and mobile phone B". The combined logic value decreases (Enterprise A does not sell mobile phone B, and Enterprise B does not sell mobile phone A), so do not combine and directly sort the complete sentences based on the logic value to obtain the credibility ranking of the complete sentences.

[0198] Also included is a method for determining whether to not combine when the combined logic value is less than the maximum logic value, the method comprising:

[0199] Step 6102510: Analyze the complete sentence group to obtain the logic relationship and the credibility value.

[0200] The logic relationship refers to the semantic association pattern between the sentences in the complete sentence group, which is used to describe the connection mode between the information. Common types include: cause-effect relationship, parallel relationship, etc. The credibility value is a quantitative evaluation of the probability of the logic relationship, which is used to measure the rationality and reliability of the combined sentence.

[0201] For example, the probability of analyzing the cause-effect relationship between the complete sentences can be calculated by weighted combination from three aspects of semantic similarity, event time sequence, and verb cause-effect implication.

[0202] ① Semantic similarity (S): Calculate the cosine similarity of two complete sentences. If the cosine similarity is higher, it tends to be a parallel relationship, and if it is lower, it may be a cause-effect relationship. The specific cosine similarity calculation method is mentioned in step 57111 and will not be repeated here.

[0203] ② Event time sequence (T): Extract the keywords of "time (when)" in the two complete sentences, and judge the time clues of the complete sentences based on the implicit clues of the time keywords. If event A occurs before event B, it may be a cause-effect relationship. The specific parameters of event time sequence can be determined according to the field of the information text.

[0204] ③ Verb causality (C): Find the words containing causality probability such as "because", "so", "result in" in the information text, calculate the cosine similarity of the two complete sentences of the judgment relationship respectively before and after the causality probability, and the higher the cosine similarity of the sentence after "because" in the information text, the more the complete sentence tends to be the "cause" sentence. Similarly, the higher the cosine similarity of the sentence after "so" in the information text, the more the complete sentence tends to be the "result" sentence. The larger value of the cosine similarity obtained by each complete sentence is retained as a specific parameter representing verb causality.

[0205] The causality probability is calculated by combining the above features with weights:

[0206] Where α, β, γ are preset weight parameters obtained by the staff after multiple tests, S is the semantic similarity, T is the event time sequence, and C is the verb causality.

[0207] Step 6102511: Find the corresponding confidence value threshold from the preset confidence value database based on the logical relationship.

[0208] The confidence value threshold refers to a critical value for determining whether two or more complete sentences satisfy a specific logical relationship based on the preset logical rules and the confidence value database. The confidence value database is a structured knowledge base that stores the confidence value distribution of different logical relationships (such as causality, parallelism, etc.) in a large number of experimental scenarios. Its core is to quantify the establishment probability of different logical relationships in actual applications through historical experimental data. The experimenter extracts and verifies the logical relationships in a large number of real information texts through manual annotation, statistical analysis or machine learning methods, and compares the obtained probability with the actual causality relationship to obtain the confidence value after calculating according to the calculation method in step 6102510.

[0209] When two or more complete sentences are analyzed to obtain a suspected logical relationship, the corresponding confidence value threshold of the logical relationship is automatically found from the database.

[0210] Step 6102512: When the confidence value is higher than the confidence value threshold, replace all complete sentences in the complete sentence group with the combined complete sentence.

[0211] The combined complete sentence refers to merging two or more independent complete sentences into a logically coherent and semantically complete statement based on the logical relationship through conjunction, semantic integration or information supplementation.

[0212] Step 6102513: When the confidence value is lower than the confidence value threshold, do not combine.

[0213] When the trust value is lower than the trust value threshold, it means that there is no logical relationship between the two or more independent complete sentences, and therefore the two or more independent complete sentences are not combined to ensure the accuracy of the core content.

[0214] Based on the trust ranking, the complete sentence with the highest trust value is output as the key content, and the method for selecting a preset number of complete sentences from the remaining complete sentences as supplementary content of the key content according to the trust ranking includes:

[0215] Step 61030: Determine the content relevance of the information text based on the complete sentence.

[0216] Content relevance refers to whether multiple complete sentences belong to the same information text in terms of semantics, theme, entity, or event logic, which is used as a judgment standard. The content relevance can be obtained by methods including but not limited to calculating the text similarity or cosine similarity of two complete sentences.

[0217] Step 61031: When the content relevance is higher than the preset reasonable association threshold, the preset number of complete sentences with the highest trust value are output as the key content, and a preset number of complete sentences are selected from the remaining complete sentences as supplementary content of the key content according to the trust ranking.

[0218] The reasonable association threshold is an index for determining whether the information content is highly relevant. The reasonable association threshold can be obtained by searching a preset reasonable association threshold table. The reasonable association threshold table is obtained by workers through multiple tests. The workers analyze a large number of information texts in different fields, calculate the similarity between information texts reflecting the same content, obtain the minimum relevance value representing the similarity between the two, and set it as the reasonable association threshold. When the content relevance is higher than the preset reasonable association threshold, it means that the content in the information text is highly relevant, and there is no problem of mistakenly inputting other information texts together, so the content output can be normally performed.

[0219] Step 61032: When the content relevance is lower than the preset reasonable association threshold, the information text is segmented based on the content relevance to obtain a secondary information text.

[0220] Based on content relevance segmentation refers to segmenting and splitting the information text into multiple independent information texts by analyzing the semantic similarity, entity relevance, event logic chain, and other indicators between sentences, to ensure that each independent information text only reflects a single core content.

[0221] The secondary information text is an independent unit obtained by content correlation segmentation of the original information text, and the complete sentences in each unit have high correlation, that is, around the same theme, entity or event, but have low correlation with other secondary information texts. The appearance of the secondary information text represents that there are actually multiple single complete information in the input text, which may be the result of the user mistakenly mixing multiple information.

[0222] Step 61033: classifying the complete sentences based on the secondary information text to obtain secondary complete sentences.

[0223] The secondary complete sentence refers to the complete sentence reflecting the core content of the secondary information text after content correlation segmentation. The classification based on the secondary information text refers to the process of further classifying the complete sentences in the segmented secondary information text according to semantics, functions or information types. Because the user mistakenly mixed multiple information, different complete sentences may correspond to different secondary information texts after segmentation, so it is necessary to classify based on the core content reflected by the complete sentences to ensure the rationality of the information.

[0224] Step 61034: outputting the respective preset number of secondary complete sentences with the highest credibility as the respective key content, and selecting a preset number of complete sentences from the secondary remaining complete sentences according to the credibility to output as the supplementary content of the key content.

[0225] From the classification results of each secondary information text, the most representative sentences are extracted according to the preset number (such as 2-5 sentences per category) as the key content output of the category. At the same time, the complete sentences with slightly lower logical values are also output as the key content of the information text after classification.

[0226] Based on the same inventive concept, the embodiment of the present application provides a kind of information extraction system based on keyword.

[0227] A kind of information extraction system based on keyword, comprising:

[0228] The acquisition module is used to acquire information text and suspected information text. The memory is used to store the program of the control method of the information extraction method of a kind of keyword. The processor loads and executes the program in the memory.

Claims

1. A keyword-based information extraction method, characterized by, The method comprises the following steps: Step 1: receiving an information text; Step 2: extracting keywords from the information text based on a preset extraction algorithm; Step 3: determining an information field to which the information text belongs based on the keywords; Step 4: obtaining a context category of the keywords based on a preset association algorithm; Step 5: filling the keywords into a preset blank context module based on the context category to obtain a keyword context module; Step 6: constructing a scene simulation based on the keyword context module to obtain key content conforming to an actual scene and outputting the key content; The steps between step 5 and step 6 further comprise the following steps: Step 50: finding corresponding relevance from a preset keyword database based on the keywords in any one of the keyword context modules and other keywords; Step 51: constructing the keywords based on the relevance to form a context association network; Step 52: analyzing the number of connection edges and corresponding connection strength of the keywords in the context association network; Step 53: obtaining the number of occurrences and position parameters of the keywords in the information text based on a preset traversal algorithm; Step 54: determining a keyword importance value of the keywords based on the number of connection edges, the connection strength, the number of occurrences and the position parameters; Step 55: screening the keywords with a keyword importance value lower than a preset importance threshold value, discarding the keywords from the keyword context module to obtain a modified keyword context module, and defining the remaining keywords in the modified keyword context module as modified keywords; Step 56: outputting the modified keyword context module as the keyword context module; Step 57: outputting a preset module alarm signal when the number of the modified keywords is less than a preset number threshold value; The method of constructing a scene simulation based on the keyword context module comprises the following steps: Step 60: determining module attributes based on the keyword context module; Step 61: finding corresponding scene simulation schemes, scene simulation scheme reliabilities and scene simulation scheme reliability thresholds from a preset scene simulation library based on all the module attributes and the information field; Step 610: simulating the keywords corresponding to the keyword context module in the scene simulation scheme when the scene simulation scheme reliability is greater than the scene simulation scheme reliability threshold value; Step 611: outputting a preset information missing alarm signal when the scene simulation scheme reliability is less than the scene simulation scheme reliability threshold value; The method of simulating the keywords corresponding to the keyword context module in the scene simulation scheme comprises the following steps: Step 6100: splicing the keyword context module into a complete sentence based on the scene simulation scheme; Step 6101: quantitatively analyzing the logic of the complete sentence to obtain a logic value of the complete sentence; Step 6102: sorting the complete sentence based on the logic value to obtain a reliability sorting of the complete sentence; Step 6103: based on the credibility ranking, output the complete sentence with the highest logical value as the key content, and select a preset number of complete sentences from the remaining complete sentences according to the credibility ranking as supplementary content of the key content; wherein, based on the credibility ranking, the complete sentence with the highest credibility is output as the key content, and a preset number of complete sentences from the remaining complete sentences are selected according to the credibility ranking as supplementary content of the key content, the method comprises: Step 61030: judging the content relevance of the information text based on the complete sentence; Step 61031: when the content relevance is higher than a preset reasonable relevance threshold, a preset number of complete sentences with the highest credibility are output as the key content, and a preset number of complete sentences from the remaining complete sentences are selected according to the credibility ranking as supplementary content of the key content; Step 61032: when the content relevance is lower than a preset reasonable relevance threshold, the information text is segmented based on content relevance to obtain a secondary information text; Step 61033: classifying the complete sentences based on the secondary information text to obtain secondary complete sentences; Step 61034: outputting a respective preset number of the secondary complete sentences with the highest credibility as the respective key content, and selecting a preset number of the secondary complete sentences from the remaining secondary complete sentences according to the credibility ranking as supplementary content of the key content.

2. The keyword-based information extraction method according to claim 1, wherein, When the number of modified keywords is less than the number threshold, the method for outputting the module alarm signal comprises: Step 570: reducing a preset keyword judgment threshold in the extraction algorithm to form a modified extraction algorithm, wherein the keyword judgment threshold is set in the extraction algorithm; Step 571: re-executing steps 2 to 55 based on the modified extraction algorithm; Step 5710: when the number of modified keywords after re-execution is greater than the number threshold, step 6 is executed; Step 5711: when the number of modified keywords after re-execution is still less than the number threshold, the module alarm signal is output.

3. The keyword-based information extraction method according to claim 2, wherein, Another method for information extraction when the number of modified keywords after re-execution is still less than the number threshold is also included, which comprises: Step 57110: searching for a suspected information text from the Internet based on the keyword; Step 57111: calculating a similarity value based on the suspected information text and the information text according to a preset judgment algorithm; Step 57112: when the similarity value corresponding to the suspected information text is higher than a preset similarity threshold, the suspected information text is defined as a similar information text; Step 57113: re-executing steps 2 to 4 based on the similar information text to obtain an alternative keyword context module and an alternative keyword; Step 57114: merging the alternative keyword context module and the modified keyword context module to form a hybrid keyword context module; Step 57115: determining a mixed number based on the alternative keyword and the corrected keyword; Step 571150: when the mixed number is greater than the number threshold, outputting the mixed keyword context module as the keyword context module; Step 571151: when the mixed number is less than the number threshold, discarding the blank context module corresponding to the mixed keyword context module.

4. The keyword-based information extraction method of claim 1, wherein, The method for sorting the complete sentences based on the logical values to obtain the credibility sorting of the complete sentences comprises: Step 61020: finding a corresponding logical difference threshold from a preset logical database based on the information field; Step 61021: combining the complete sentences based on the logical difference threshold to form a complete sentence group, wherein the absolute difference between the logical values of any two complete sentences in the complete sentence group is less than the logical difference threshold; Step 61022: combining the keywords in the complete sentence group to obtain a combined keyword; Step 61023: splicing the combined keyword into a combined complete sentence based on the scenario simulation scheme; Step 61024: defining the maximum logical value in the keyword context module in the complete sentence group as a maximum logical value; Step 61025: quantitatively analyzing the logicality of the combined complete sentence to obtain a combined logical value of the combined complete sentence; Step 610250: when the combined logical value is greater than the maximum logical value, replacing all complete sentences in the complete sentence group with the combined complete sentence and then sorting the complete sentences according to step 6102; Step 610251: when the combined logical value is less than the maximum logical value, not combining and directly sorting the complete sentences based on the logical values to obtain the credibility sorting of the complete sentences.

5. The keyword-based information extraction method according to claim 4, wherein, Further comprising a judgment method for whether to not combine when the combined logical value is less than the maximum logical value, the method comprising: Step 6102510: analyzing the complete sentence group to obtain a logical relationship and a credibility value; Step 6102511: finding a corresponding credibility value threshold from a preset credibility value database based on the logical relationship; Step 6102512: when the credibility value is higher than the credibility value threshold, replacing all complete sentences in the complete sentence group with the combined complete sentence; Step 6102513: when the credibility value is lower than the credibility value threshold, not combining.

6. A keyword information extraction system characterized by comprising: Comprise: The acquisition module is used for acquiring the information text and the suspected information text; The memory is used for storing the program of the keyword information extraction method according to any one of claims 1 to 5; The processor loads and executes the program in the memory.

Images