Data mining system in digital archive based on network and digital resources

Abstract

The invention relates to the field of knowledge base questions and answers, in particular to a data mining system in a digital archive based on network and digital resources, which comprises a question understanding module, an information matching module, a dynamic adjustment module and an answer generation module. By supporting a multi-hop reasoning mode, information fault and insufficient reasoning depth in complex question query are avoided, the real-time similarity between an intention recognition result and an entity is calculated, the connection number in an archive knowledge base is combined, a focus entity is selected, the pertinence is improved, and the real-time similarity between a candidate answer entity and the focus entity is obtained based on the incidence relation between the candidate answer entity and the focus entity. Factors such as policy reference frequency and business association strength are considered, a to-be-selected path is formulated, the next hop entity type is predicted through a probability model, the accuracy of path selection is ensured, the reasoning progress is monitored in real time, the reasoning strategy is dynamically adjusted through a hop strategy adjusting unit, and the adaptive capacity and reasoning efficiency of the question and answer reasoning process are improved. Effective feedback is provided for user consultation, and user experience is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of knowledge base question answering, and in particular to a data mining system in a digital archive based on a network and digital resources. BACKGROUND

[0002] With the development of computer technology and Internet technology, the retrieval of the archive database can be combined with a natural language model, for example, in the form of knowledge base question answering, to understand the user's intention and retrieve the archive database. Knowledge base question answering is widely used in search engines, voice assistants and other intelligent applications. Improving the accuracy and explainability of knowledge base question answering tasks has become a research hotspot. Knowledge base is a data structure that stores facts in the form of edges representing the relationship between entity nodes, which is represented by triples composed of head entities, tail entities and their relationships. Knowledge base question answering finds answers to questions by traversing the facts of the knowledge base, and is divided into single-hop knowledge base question answering and multi-hop knowledge base question answering. Single-hop knowledge base question answering refers to the fact that the theme entity of the question and the answer entity are one hop away, specifically, the theme entity and the answer entity are connected as head entities and tail entities in a triple. The performance of single-hop knowledge base question answering models has approached human level. Multi-hop knowledge base question answering refers to the fact that the theme entity in a complex question and the answer entity of the question are several hops away, specifically, the theme entity and the answer entity are connected in multiple triples, so complex multi-hop reasoning is needed to find the semantic relationship between the theme entity and the answer entity. At the same time, processing complex questions requires the model to have reasoning capabilities for processing multi-hop relationships, set operations and attribute comparisons. Multi-hop knowledge base questions have attracted more and more researchers due to their challenges.

[0003] Chinese patent document with publication number 202410926292.5 discloses an intelligent knowledge base question answering system and method, which includes a question receiving module, a natural language processing module, a semantic analysis module, a knowledge base management module, an inference engine module, an answer generation module, a user feedback module and a system optimization module. The question receiving module receives input questions from users and passes the questions to the natural language processing module. The natural language processing module receives the questions passed by the question receiving module, preprocesses the questions, and passes the processed data to the semantic analysis module. The semantic analysis module receives the processed data passed by the natural language processing module, analyzes the semantic relationship between the data, and passes the analyzed data to the inference engine module. However, the prior art still has the following problems: When dealing with complex questions, the question answering system can only perform single-hop reasoning and cannot effectively handle problems that require multi-step logical derivation, resulting in low feedback accuracy and affecting user experience. SUMMARY

[0004] To this end, the application provides a data mining system in a digital archive based on a network and digital resources, to overcome the problem of low efficiency of complex multi-hop reasoning in the prior art due to the lack of determining a relationship path according to a focal entity to dynamically adjust a multi-hop reasoning process.

[0005] To achieve the above object, the application provides a data mining system in a digital archive based on a network and digital resources, comprising, A question understanding module is configured to obtain demand question information input by a user end, decompose the demand question information into a plurality of entities, and perform question intention recognition on the demand question information to obtain an intention recognition result. An information matching module is connected with the question understanding module and is configured to search for matched information in an archive knowledge base based on the intention recognition result to obtain an initial relationship group, determine a reasoning matching mode based on the number of relationships in the initial relationship group, and the reasoning matching mode comprises a first reasoning mode and a second reasoning mode. A dynamic adjustment module is connected with the information matching module and is configured to, in the second reasoning mode, monitor the progress of a current reasoning process in real time, dynamically adjust a multi-hop reasoning strategy, and obtain a matching result, link the matching result with a focal entity to form an intermediate node chain. The social security field knowledge graph is updated according to a policy reference frequency and a business association strength, a candidate path is generated, the current hop answer is adjusted based on the candidate path and an association analysis result, and a matching result is obtained. An answer generation module is connected with the information matching module and is configured to, in the first reasoning mode, extract an answer based on the initial relationship group, and in the second reasoning mode, generate a plurality of answers based on the intermediate node chain. A pushing module is connected with the answer generation module and is configured to send the answers to the user end.

[0006] Further, the question understanding module comprises, The word segmentation unit is configured to decompose the demand question information into independent words or phrases to obtain a plurality of entities. The part-of-speech tagging unit is configured to determine the part of speech of each entity. The entity recognition unit is configured to identify key entities in the question. The semantic analysis unit is configured to remove redundant information and perform semantic analysis on the demand question information based on syntactic analysis and semantic role labeling. The intention recognition unit is configured to identify the intention of the user based on the semantic analysis result to obtain an intention recognition result.

[0007] Further, the information matching module comprises a matching unit and a determination unit, wherein, The matching unit is configured to perform entity matching in the archive knowledge base based on the intention recognition result to obtain an initial relation group; The determining unit is configured to acquire a real-time number of the initial relation group, and compare the real-time number with a standard number: If the real-time number is less than or equal to the standard number, the determining unit determines that the reasoning matching mode is a first reasoning mode; If the real-time number is greater than the standard number, the determining unit determines that the reasoning matching mode is a second reasoning mode; The first reasoning matching mode is single-hop reasoning, and the second reasoning matching mode is multi-hop reasoning.

[0008] Further, the dynamic adjustment module includes a monitoring unit, a jump strategy formulation unit, an association analysis unit, a verification unit, and a jump strategy adjustment unit, wherein, The monitoring unit is configured to monitor the progress of the current reasoning process and record reasoning process information; The jump strategy formulation unit is configured to determine a plurality of candidate paths based on the candidate answer entity and the focus entity in the initial relation group, and determine the path of the next-hop answer based on the entity type of the current jump and the previous jump; The association analysis unit is configured to perform association analysis on adjacent jump questions, including dynamic semantic analysis and sequential logic analysis; The verification unit is configured to monitor the multi-hop reasoning process; The jump strategy adjustment unit is configured to determine whether to adjust the current jump answer based on the association analysis result.

[0009] Further, the jump strategy formulation unit includes a focus entity selection subunit, a first path selection subunit, and a second path selection subunit, wherein, The focus entity selection subunit is configured to calculate the real-time similarity of each entity in the intention recognition result and the initial relation group, select an entity whose corresponding real-time similarity is greater than or equal to a standard similarity as a key entity, and determine the real-time link quantity of each key entity in the archive knowledge base, and select an entity whose real-time link quantity is greater than or equal to a standard link quantity as a focus entity; The first path selection subunit is configured to determine a plurality of candidate paths based on the association relationship between the candidate answer entity and the focus entity; The second path selection subunit is configured to select the path of the next-hop answer in each candidate path based on the entity type of the current jump and the previous jump.

[0010] Further, determining a plurality of candidate paths based on the association relationship between the candidate answer entity and the focus entity includes, calculating an original weight based on policy reference frequency and business association strength; The real-time relationship weight is calculated based on the original weight and a time decay factor; The initial relationship weight in the social security field knowledge graph is updated as the real-time relationship weight; For any key entity, the forward search is performed from the focus entity, the reverse search is performed from the candidate answer entity, the maximum hop limit is set, when the forward and reverse searches meet at the key entity, the to-be-selected path is generated.

[0011] Further, the path of the next-hop answer is selected from the to-be-selected paths according to the entity type of the current hop and the entity type of the last hop, comprising, The rules for selecting the path are defined, including defining the relationship between the entity types and constructing a dynamic loading mechanism; The probability model is used to predict the entity type of the next hop, and the path corresponding to the maximum probability value is selected as the path of the next-hop answer; If the selected path of the next-hop answer is not unique, the paths are prioritized according to the length, and the path with the highest priority is selected as the path of the next-hop answer.

[0012] Further, the correlation analysis unit comprises a semantic analysis subunit and a sequential logic analysis subunit, wherein, The semantic analysis subunit performs dynamic semantic analysis on the current-hop answer and the next-hop answer through context understanding; The sequential logic analysis subunit performs sequential logic analysis on the current-hop answer and the next-hop answer.

[0013] Further, the monitoring of the multi-hop reasoning process comprises, The priority of the to-be-selected path is determined according to the credibility; The initial step of the multi-hop reasoning is adjusted based on the length of the real-time reasoning process.

[0014] Further, the process of adjusting the current-hop answer is to add the last-hop information to the current-hop, so as to determine the path of the next-hop answer according to the entity type of the current-hop and the entity type of the last-hop.

[0015] Compared with the prior art, the beneficial effects of the present application are that, by constructing a plurality of intermediate nodes, i.e. constructing the intermediate nodes of each hop of the link problem, to form an intermediate node chain, finally connecting the problem and the answer, by setting a dynamic adjustment module, supporting a multi-hop reasoning mode, avoiding information gaps and insufficient reasoning depth in complex problem queries, by calculating the real-time similarity of the intention recognition result and the entity, and combining the connection quantity in the archive knowledge base, intelligently selecting the focus entity, improving the pertinence of reasoning, by considering factors such as policy reference frequency and business association strength based on the association relationship between the candidate answer entity and the focus entity, formulating a to-be-selected path, and predicting the next-hop entity type through a probability model, ensuring the accuracy of path selection, through the monitoring unit, real-time monitoring of the reasoning progress, and through the jump strategy adjustment unit, dynamically adjusting the reasoning strategy, improving the adaptability and reasoning efficiency of the question and answer reasoning process, providing effective feedback for user consultation, and improving user experience.

[0016] Further, by comprehensively considering various factors (such as policy reference frequency, business association strength, etc.) to formulate a to-be-selected path, and using a probability model to predict the next-hop entity type, the path selection and reasoning process are optimized, the intelligent level of the system is improved, and a relationship weight model with a time attenuation factor is used to automatically reduce the weight of expired policies. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 It is a structural schematic diagram of a data mining system in a digital archive based on network and digital resources according to an embodiment of the present application; Figure 2 It is a structural schematic diagram of a problem understanding module according to an embodiment of the present application; Figure 3 It is a structural schematic diagram of an information matching module according to an embodiment of the present application; Figure 4 It is a structural schematic diagram of a dynamic adjustment module according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to make the objects and advantages of the present application clearer, the present application will be further described below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the present application.

[0019] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0020] It should be noted that in the description of the present application, the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the direction or positional relationship of the terms based on the direction or positional relationship shown in the drawings, which is only for the convenience of description, and does not indicate or imply that the device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0021] In addition, it should be noted that in the description of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected, it can be mechanically connected, or it can be electrically connected, it can be directly connected, or indirectly connected through an intermediate medium, it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0022] Please refer to Figure 1 As shown in the figure, it is a structure schematic diagram of the data mining system in the digital archive based on network and digital resources according to the embodiment of the present application, the present application provides a data mining system in the digital archive based on network and digital resources, comprising, The problem understanding module is used to obtain the demand problem information input by the user end, decompose the demand problem information into several entities, and perform problem intention recognition on the demand problem information to obtain an intention recognition result; The information matching module is connected with the problem understanding module, and is used to search matching information in the archive knowledge base based on the intention recognition result to obtain an initial relationship group, determine a reasoning matching mode based on the number of relationships in the initial relationship group, including a first reasoning mode and a second reasoning mode; The dynamic adjustment module is connected with the information matching module, and is used to monitor the progress of the current reasoning process in real time under the second reasoning mode, dynamically adjust the multi-hop reasoning strategy, and obtain a matching result, link the matching result with the focus entity to form an intermediate node chain; Wherein, the social security field knowledge graph is updated according to the policy reference frequency and the business association strength, a candidate path is generated, the current hop answer is adjusted based on the candidate path and the association analysis result, and a matching result is obtained; The answer generation module is connected with the information matching module, and is used to extract an answer based on the initial relationship group under the first reasoning mode, and generate a plurality of answers based on the intermediate node chain under the second reasoning mode; The push module is connected with the answer generation module, and is used to send the answer to the user end.

[0023] In the embodiment, the intermediate nodes are constructed, that is, the intermediate nodes of each hop of the link problem are constructed to form an intermediate node chain, finally connecting the question and the answer, supporting the multi-hop reasoning mode through the setting of the dynamic adjustment module, avoiding the information fault and the insufficient reasoning depth in the complex problem query, intelligently selecting the focus entity through the calculation of the real-time similarity of the intention recognition result and the entity and the connection quantity in the archive knowledge base, improving the pertinence of reasoning, considering the policy reference frequency, the business correlation strength and other factors based on the association relationship of the candidate answer entity and the focus entity, formulating the to-be-selected path, and ensuring the accuracy of the path selection through the probability model to predict the next-hop entity type, monitoring the reasoning progress in real time through the monitoring unit, and dynamically adjusting the reasoning strategy through the jump strategy adjustment unit, improving the adaptability and the reasoning efficiency of the question and answer reasoning process.

[0024] Referring to Figure 2 As shown in the figure, it is a structural schematic diagram of the problem understanding module of the embodiment of the application; Specifically, the problem understanding module comprises a word segmentation unit, a part-of-speech tagging unit, an entity recognition unit, a semantic analysis unit and an intention recognition unit, wherein, The word segmentation unit is used to decompose the demand problem information into independent words or phrases to obtain a plurality of entities. The part-of-speech tagging unit is used to determine the part of speech of each entity. The entity recognition unit is used to identify the key entities in the question. The semantic analysis unit is used to remove redundant information and perform semantic analysis on the demand problem information based on syntactic analysis and semantic role labeling. The intention recognition unit identifies the user's intention based on the semantic analysis result to obtain an intention recognition result.

[0025] In this embodiment, parts of speech include nouns, verbs, and adjectives; key entities include names of people, places, dates, and product names; and intent recognition results include query information, comparison information, and questions. A combination of rules and machine learning is used to identify user intent. First, rules are used for initial filtering to exclude inputs that clearly do not conform to the predefined intent. Then, a machine learning model is used to perform more refined intent recognition on the remaining inputs. For example, suppose a user inputs: "What are my social security payment records from January 2023 to December 2023?" Through word segmentation, the result is: "What are my social security payment records from January 2023 to December 2023?" Part-of-speech tagging: pronouns, numerals, nouns, auxiliary words, numerals, nouns, verbs, nouns are pronouns. Entity recognition: January 2023, December 2023, payment records, semantic analysis: extracting the core concept: querying payment records from January to December 2023, intent recognition: through rule matching (containing keywords such as "payment" and "record") and / or machine learning model prediction, the system identifies that the user's intent is to query social security payment records. Semantic analysis understands that the user wants to query payment records for a specific time period, while intent recognition clarifies the user's purpose. The system can then further retrieve relevant data from the social security database based on the user's intent and return it to the user.

[0026] See Figure 3 As shown, it is a structural schematic diagram of the information matching module in an embodiment of the present invention; Specifically, the information matching module includes a matching unit and a determining unit, wherein, The matching unit is used to perform entity matching in the archive knowledge base based on the intent recognition result to obtain an initial relationship group; The determining unit is used to obtain the real-time number of initial relation groups and compare the real-time number with the standard number: If the number of real-time instances is less than or equal to the standard number of instances, the inference matching mode is determined to be the first inference mode. If the number of real-time cases is greater than the standard number of cases, the reasoning matching mode is determined to be the second reasoning mode; The first reasoning matching mode is single-hop reasoning, and the second reasoning matching mode is multi-hop reasoning.

[0027] The number of standards in the embodiment is set to 3, based on the distribution of the number of query results under different intent and entity combinations in the statistical social security archive knowledge base, and a threshold is set. Queries exceeding the threshold are considered to require multi-hop reasoning. Data is collected from the social security system historical query log, including intent type, entity type (and corresponding query result number, data format, for example, intent type is to query personal account information, entity 1 (personal name) is Zhang San, entity 2 (ID number), entity 3 (query time period) is 2020-01-01 to 2023-12-31, and the result number is 1; The intent type is to query payment records, entity 1 (personal name) is Li Lei, entity 2 (ID number), entity 3 (query time period) is 2022-01-01 to 2022-12-31, and the result number is 12; The intent type is to query payment records, entity 1 (personal name) is Li Si, entity 2 (ID number), entity 3 (query time period) is 2019-01-01 to 2020-12-31, and the result number is 0; The intent type is to query personal account information, entity 1 (personal name) is Wang Wu, entity 2 (ID number), entity 3 (query time period) is 2020-01-01 to 2021-12-31, and the result number is 8; For different intent types and entity combinations, the distribution of the number of results is calculated. For example, the result number distribution of "query personal account information" is: mean: 1.5, standard deviation: 0.8; The threshold calculation formula is: mean ± 2*std, so for the example of "query personal account information", the threshold range is: lower limit: 1.5-2*0.8=-0.1, upper limit: 1.5+2*0.8=3.1, since the result number cannot be negative, the actual threshold is 3.1, and the integer is 3.

[0028] Referring to Figure 4 Fig. 1 is a structural schematic diagram of a dynamic adjustment module according to an embodiment of the present application; Specifically, the dynamic adjustment module comprises a monitoring unit, a jump strategy formulation unit, a correlation analysis unit, a verification unit and a jump strategy adjustment unit, wherein, The monitoring unit is used to monitor the progress of the current reasoning process and record the reasoning process information; The jump strategy formulation unit is used to determine a plurality of candidate paths based on the candidate answer entity and the focus entity in the initial relation group, and determine the path of the next jump answer according to the entity type of the current jump and the last jump; The correlation analysis unit is used to perform correlation analysis on adjacent jump problems, including dynamic semantic analysis and sequential logic analysis; The verification unit is used to monitor the multi-hop reasoning process; The jump strategy adjustment unit is configured to determine whether to adjust the current jump answer according to the association analysis result.

[0029] The inference process information in the embodiment includes information such as each step of operation, knowledge source usage, and credibility change.

[0030] By dynamically adjusting the module and the multi-jump inference strategy, the user's intention can be more accurately understood and deep information matching can be performed, thereby improving the inference accuracy and efficiency of complex problems.

[0031] Specifically, the jump strategy formulation unit includes a focus entity selection subunit, a first path selection subunit, and a second path selection subunit. The focus entity selection subunit is configured to calculate real-time similarity of each entity in the initial relationship group and the intention recognition result, select an entity with a real-time similarity greater than or equal to a standard similarity as a key entity, and determine a real-time link quantity of each key entity in the archive knowledge base, and select an entity with a real-time link quantity greater than or equal to a standard link quantity as a focus entity. The first path selection subunit is configured to determine a plurality of candidate paths based on an association relationship between the candidate answer entity and the focus entity. The second path selection subunit is configured to select a path of a next jump answer in the plurality of candidate paths according to an entity type of the current jump and a last jump.

[0032] The standard similarity in the embodiment is set to 50%-60%.

[0033] Specifically, determining a plurality of candidate paths based on an association relationship between the candidate answer entity and the focus entity includes, calculating an original weight based on a policy reference frequency and a business association strength; calculating a real-time relationship weight based on the original weight and a time decay factor; updating an initial relationship weight in the social security field knowledge graph to the real-time relationship weight; For any key entity, searching in the social security field knowledge graph, forward searching from the focus entity, and reverse searching from the candidate answer entity, setting a maximum jump number limit, when the forward and reverse searching meet at the key entity, generating a candidate path.

[0034] The original weight=(α×Wp+β×Wb)×θ; wherein, α=0.55; Wp is a policy reference weight; β=0.45; Wb is a business association weight; θ is a regional adaptation coefficient;

[0035] Rm is the number of administrative divisions matched with the query; Rt is the total number of divisions covered by the policy;

[0036] Wp is the policy reference weight; Cr is the number of times the relationship is referenced by the policy file; Cmax is the maximum reference count in the entire library (updated dynamically); Aa is the authority coefficient, ranging from 0.8 to 1.2;

[0037] i is the i-th business process; Wb is the business association weight; n is the number of business processes; Wi is the business importance coefficient; Fi is the usage frequency in the i-th business process; Fmax is the maximum usage frequency of a single business in the entire library;

[0038] e is the natural logarithm,

[0039] In this embodiment, the maximum hop limit is set to 3-5 hops, and a number of candidate paths are determined based on the association relationship between the candidate answer entity and the focus entity, for example, forward search from the focus entity and reverse search from the candidate answer entity, and the maximum hop limit is set, and when the forward and reverse searches meet at the key entity, the candidate path is generated.

[0040] By considering various factors (such as policy reference frequency, business association strength, etc.) to develop candidate paths, and using a probability model to predict the next hop entity type, the path selection and reasoning process is optimized, and the intelligence level of the system is improved, and a relationship weight model with a time decay factor is used to automatically reduce the weight of expired policies.

[0041] Specifically, the path of the next hop answer is selected from each candidate path according to the entity type of the current hop and the last hop, which includes, defining the rules for selecting the path, including defining the relationship between entity types and constructing a dynamic loading mechanism; using a probability model to predict the entity type of the next hop, and selecting the path corresponding to the maximum probability value as the path of the next hop answer; If the selected path of the next hop answer is not unique, the paths are prioritized according to their lengths, and the path with the highest priority is selected as the path of the next hop answer.

[0042] In the present embodiment, a special entity type system for the social security field is defined, including: Amount: ["premium amount", "pension amount", "reimbursement amount"] Time: ["premium time", "policy effective date", "treatment calculation base year"] Policy: ["regulatory provisions", "local implementation rules"] Business entity: ["insured person", "employer", "social security agency"] Record type: ["premium record", "medical record", "work injury identification paper"] A dynamic loading mechanism is constructed to determine path selection requests according to whether the business is cross-provincial. If it is a cross-provincial business, load the cross-provincial transfer rule set, and if it is not a cross-provincial business, load the regular rule set. The probability model can be a Bayesian network or a Markov chain.

[0043] Specifically, the correlation analysis unit includes a semantic analysis subunit and a sequential logic analysis subunit, wherein, The semantic analysis subunit performs dynamic semantic analysis on the current jump answer and the next jump answer through context understanding; The sequential logic analysis subunit performs sequential logic analysis on the current jump answer and the next jump answer.

[0044] In this embodiment, pre-trained BERT, RoBERTa and other models are used, and combined with the knowledge graph (KG) in the social security field for fine-tuning, for example, to identify "insured person's name", "payment date", "payment amount", "medical insurance type" and other entities, and label their types, accurately identify and distinguish different types of entities, such as "2023-01-15" and "2023-01-15", which should be identified as "date", and through semantic relationship extraction: build a KG in the social security field, including entities and their relationships. For example, there is an "association" relationship between "insured person" and "medical insurance type"; there is a "correspondence" relationship between "payment date" and "payment amount", and GNN (Graph Neural Network) is used to learn the relationship between entities and consider the weight and direction of the relationship. In the field of social security, the order relationship such as "insured person"→"payment record"→"medical insurance card number" is very important, and by inputting the context information of the current jump and the last jump, including the user's input sentence and historical interaction records, into the model. For example, if the last jump is "query insured person information" and the current jump is "get payment record", the context needs to be considered, such as "query 2023 medical insurance payment record"; through dynamic semantic analysis to evaluate the semantic similarity between the candidate answer and the user question, the cosine similarity can be used to calculate the semantic similarity, for example, user question: "How to apply for unemployment insurance?" Candidate answer 1: "You can apply for unemployment insurance online." Candidate answer 2: "The application process for unemployment insurance includes filling out a form." Use word embedding (such as Word2Vec or BERT) to convert these sentences into vectors: User question vector: [0.1, 0.2, 0.3]; Candidate answer 1 vector: [0.1, 0.2, 0.4]; Candidate answer 2 vector: [0.2, 0.1, 0.3]; The cosine similarity of candidate answer 1 is 0.98, and the cosine similarity of candidate answer 2 is 0.87, and the threshold is set to 0.90, then candidate answer 1 meets the user's intention; Sequential logic analysis checks the logical relationship between each jump in multi-hop reasoning, which can use simple rules to determine the entity type relationship between the current jump and the last jump, for example, the current jump: "apply for unemployment insurance online." (entity type: application method) last jump: "unemployment insurance application process." (entity type: application process), set the rule as if the entity type of the current jump is "application method" and the entity type of the last jump is "application process", then the logical relationship is established.

[0045] The jump strategy adjustment unit determines whether to adjust the current jump answer according to the correlation analysis result: The results of dynamic semantic analysis and sequential logic analysis are integrated to form a comprehensive evaluation index; The comprehensive evaluation index is set as: If the similarity is greater than 0.90 and the logical relationship is established, the current jump answer is valid; Otherwise, adjust the current jump answer.

[0046] Specifically, monitoring the multi-hop reasoning process includes, According to the credibility, the priority of the candidate path is determined; Based on the real-time reasoning process time length, the step length of multi-hop reasoning is adjusted.

[0047] In the embodiment, the real-time reasoning process time length is the cumulative time length of the reasoning process. If the real-time reasoning process time length is greater than or equal to the standard reasoning process time length, the initial step length of multi-hop reasoning is adjusted to reduce the amount of calculation, otherwise it is not adjusted. The standard reasoning process time length is set to 5 seconds, and the real-time reasoning process time length refers to the cumulative time length from the start of reasoning to the current time. The credibility represents the reliability of the candidate answer or path, and the credibility is calculated by weighted average, credibility = semantic similarity × A1 + historical accuracy × A2 + context consistency × A3; A1 is the weight coefficient of semantic similarity; A2 is the weight coefficient of historical accuracy; A3 is the weight coefficient of context consistency; A1+A2+A3=1; A1=0.5, A2=0.3, A3=0.2; Semantic similarity, historical accuracy and context consistency are all normalized to (0, 1); For example, when the semantic similarity is 0.85, the historical accuracy is 0.90, and the context consistency is 0.80, the credibility is calculated as 0.855, and the candidate path is sorted according to the calculated credibility; The initial step length is 3, which means that each time the reasoning jumps 3 entities. When the real-time reasoning process time length is greater than or equal to the standard reasoning process time length, the initial step length of multi-hop reasoning is adjusted to: new step length = current step length - 1, but the step length cannot be less than 1. For example, if the initial step length is 3 and the real-time reasoning process time length is 6 seconds (greater than 5 seconds), the adjustment process is as follows: New step length calculation: new step length = 3-1 = 2 If the real-time reasoning process time length is monitored again as 4 seconds (less than 5 seconds), the step length remains unchanged, still 2.

[0048] Specifically, the process of adjusting the current jump answer is to add the last jump information to the current jump to determine the path of the next jump answer according to the entity types of the current jump and the last jump.

[0049] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0050] The above only describes the preferred embodiments of the present application and is not intended to limit the present application; the present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A data mining system in a network and digitized resource based digital archive, characterized by, Comprising, a question understanding module configured to obtain demand question information input by a user, decompose the demand question information into a plurality of entities, and perform question intent recognition on the demand question information to obtain an intent recognition result; an information matching module connected to the question understanding module and configured to search for matching information in an archive knowledge base based on the intent recognition result to obtain an initial relationship group, determine a reasoning matching mode based on a number of relationships in the initial relationship group, the reasoning matching mode including a first reasoning mode and a second reasoning mode; a dynamic adjustment module connected to the information matching module and configured to, in the second reasoning mode, monitor progress of a current reasoning process in real time, dynamically adjust a multi-hop reasoning strategy, and obtain a matching result, link the matching result with a focus entity to form an intermediate node chain; wherein the social security field knowledge graph is updated according to a policy reference frequency and a business association strength, a to-be-selected path is generated, a current hop answer is adjusted based on the to-be-selected path and an association analysis result, and a matching result is obtained; an answer generation module connected to the information matching module and configured to, in the first reasoning mode, extract an answer based on the initial relationship group, and in the second reasoning mode, generate an answer based on the constructed intermediate node chain.

2. The data mining system in a web and digitized resource based digital archive of claim 1, wherein, The question understanding module includes a word segmentation unit, a part-of-speech tagging unit, an entity recognition unit, a semantic analysis unit, and an intent recognition unit, wherein the word segmentation unit is configured to decompose the demand question information into independent words or phrases to obtain a plurality of entities; the part-of-speech tagging unit is configured to determine the part of speech of each entity; the entity recognition unit is configured to identify key entities in the question; the semantic analysis unit is configured to remove redundant information and perform semantic analysis on the demand question information based on syntactic analysis and semantic role labeling; the intent recognition unit is configured to identify the user's intent based on the semantic analysis result to obtain an intent recognition result.

3. The data mining system in a web and digitized resource based digital archive of claim 2, wherein, The information matching module includes a matching unit and a determination unit; the matching unit is configured to perform entity matching in the archive knowledge base based on the intent recognition result to obtain an initial relationship group; the determination unit is configured to obtain a real-time number of the initial relationship group, compare the real-time number with a standard number, and: if the real-time number is less than or equal to the standard number, the determination unit determines that the reasoning matching mode is the first reasoning mode; if the real-time number is greater than the standard number, the determination unit determines that the reasoning matching mode is the second reasoning mode; wherein the first reasoning matching mode is single-hop reasoning, and the second reasoning matching mode is multi-hop reasoning.

4. The data mining system in a web and digitized resources based digital archive of claim 1, wherein, The dynamic adjustment module includes a monitoring unit, a jump strategy formulation unit, an association analysis unit, a verification unit, and a jump strategy adjustment unit, wherein the monitoring unit is configured to monitor progress of a current reasoning process and record reasoning process information; the jump strategy formulation unit is configured to determine a plurality of to-be-selected paths based on candidate answer entities and a focus entity in the initial relationship group, and determine a path of a next hop answer according to an entity type of a current hop and a previous hop; the association analysis unit is configured to perform association analysis on adjacent hop questions, including dynamic semantic analysis and sequential logic analysis; The inspection unit is configured to monitor the multi-hop reasoning process. The jump strategy adjustment unit is configured to determine whether to adjust the current jump answer according to the association analysis result.

5. The data mining system in a web and digitized resource based digital archive of claim 4, wherein, The jump strategy formulation unit comprises a focus entity selection subunit, a first path selection subunit and a second path selection subunit, wherein, The focus entity selection subunit is configured to calculate the real-time similarity of each entity in the initial relationship group and the intention recognition result, select an entity with a real-time similarity greater than or equal to a standard similarity as a key entity, and determine the real-time link quantity of each key entity in the archive knowledge base, and select an entity with a real-time link quantity greater than or equal to a standard link quantity as a focus entity. The first path selection subunit is configured to determine a plurality of candidate paths based on the association relationship between the candidate answer entity and the focus entity. The second path selection subunit is configured to select a path of the next jump answer in each candidate path according to the entity type of the current jump and the last jump.

6. The data mining system in a web and digitized resource based digital archive of claim 5, wherein, Determining a plurality of candidate paths based on the association relationship between the candidate answer entity and the focus entity comprises, calculating an original weight based on the policy reference frequency and the business association strength; calculating a real-time relationship weight based on the original weight and a time decay factor; updating the initial relationship weight in the social security field knowledge graph to the real-time relationship weight; For any key entity, forward search from the focus entity and backward search from the candidate answer entity, set a maximum jump number limit, and when the forward and backward searches meet at the key entity, generate a candidate path.

7. The data mining system in a web and digitized resources based digital archive of claim 5, wherein, Selecting a path of the next jump answer in each candidate path according to the entity type of the current jump and the last jump comprises, defining rules for selecting paths, including defining relationships between entity types and constructing a dynamic loading mechanism; using a probability model to predict the entity type of the next jump, and selecting a path corresponding to the maximum probability value as the path of the next jump answer; wherein, if the selected path of the next jump answer is not unique, the paths are prioritized according to the length, and the path with the highest priority is selected as the path of the next jump answer.

8. The data mining system in a web and digitized resources based digital archive of claim 4, wherein, The association analysis unit comprises a semantic analysis subunit and a sequential logic analysis subunit, wherein, The semantic analysis subunit performs dynamic semantic analysis on the current jump answer and the next jump answer through context understanding; The sequential logic analysis subunit performs sequential logic analysis on the current jump answer and the next jump answer.

9. The data mining system in a web and digitized resources based digital archive of claim 4, wherein, Monitoring the multi-hop reasoning process comprises, determining the priority of the candidate path according to the credibility; adjusting the initial step length of the multi-hop reasoning based on the real-time reasoning process time length.

10. The data mining system in a web and digitized resources based digital archive of claim 4, wherein, The process of adjusting the current jump answer is to add the last jump information to the current jump to determine the path of the next jump answer according to the entity type of the current jump and the last jump.

Images