A knowledge graph-based intelligent property rights retrieval and matching system

Abstract

This invention relates to the field of intelligent retrieval and matching technology, specifically, to a knowledge graph-based intelligent property rights retrieval and matching system. It includes: a data acquisition and processing unit that collects property rights data and preprocesses it; a knowledge modeling and graph construction unit that constructs a knowledge graph from the preprocessed property rights data based on entity extraction, relation extraction, and attribute annotation; and a retrieval parsing and semantic understanding unit that receives user-inputted retrieval requests and transforms them into structured retrieval statements matching the knowledge graph based on a dynamic multi-hop reasoning intent path discovery scheme, parsing the retrieval intent through context modeling and semantic disambiguation. This invention, by introducing a dynamic multi-hop reasoning intent path discovery scheme combined with a semantic understanding mechanism that integrates context modeling and semantic disambiguation, can accurately parse complex retrieval requests input by users in natural language.

Description

Technical Field

[0001] This invention relates to the field of intelligent retrieval and matching technology, and more specifically, to a knowledge graph-based intelligent property rights retrieval and matching system. Background Technology

[0002] With the increasing complexity of socio-economic activities, property rights-related information (including property transaction records, legal clauses, asset information, etc.) is characterized by large data volume, wide sources, and heterogeneous formats. This data is typically scattered across multiple independent systems such as real estate registration, business administration, financial lending, judicial judgments, and contract management, forming serious "data silos" that hinder information exchange and correlation. Traditional property rights retrieval systems are mostly based on relational databases and keyword matching technology, supporting only simple, explicit conditional queries. They cannot effectively integrate multi-source heterogeneous data and struggle to understand complex search intents expressed in natural language, such as queries involving multi-hop relationships and deep semantics, like "property under Zhang San that is mortgaged by a bank and has not been traded in the past three years" or "equity acquired through inheritance and actually held by his son." Furthermore, property rights data commonly suffers from issues such as entity homonyms (e.g., multiple "Zhang Sans"), ambiguous referents (e.g., "he" "the property"), and diverse relationship expressions (e.g., "transfer," "sale," "ownership transfer"). Traditional methods lack contextual modeling and semantic disambiguation capabilities, easily leading to inaccurate search results or omission of key information. More importantly, many property rights statuses (such as actual control, nominee ownership, and transfer restrictions) are not directly recorded but are implicit in multi-layered legal and transactional relationships. Existing systems lack the ability to logically reason and infer paths to these implicit relationships, making it impossible to discover deep connections. Furthermore, search results often lack credibility assessments and explanations of the reasoning basis, failing to meet the interpretability and reliability requirements of high-risk scenarios such as judicial review and compliance supervision. Therefore, this paper proposes a knowledge graph-based intelligent property rights retrieval and matching system. Summary of the Invention

[0003] The purpose of this invention is to provide a knowledge graph-based intelligent property rights retrieval and matching system to solve the problems of low query accuracy and low intelligence level of traditional property rights retrieval systems mentioned in the background art, which are caused by data silos, weak semantic understanding, difficulty in mining complex relationships, insufficient reasoning ability and unreliable results.

[0004] To achieve the above objectives, the present invention aims to provide a knowledge graph-based intelligent property rights retrieval and matching system, comprising:

[0005] A data acquisition and processing unit, wherein the data acquisition and processing unit acquires property rights data and preprocesses the property rights data;

[0006] The knowledge modeling and graph construction unit constructs a knowledge graph from the preprocessed property rights data based on entity extraction, relation extraction, and attribute annotation.

[0007] The retrieval parsing and semantic understanding unit is used to receive retrieval requests input by users, and transform the retrieval requests into structured retrieval statements that match the knowledge graph based on the intent path discovery scheme of dynamic multi-hop reasoning. The retrieval intent of the retrieval request is parsed through context modeling and semantic disambiguation methods.

[0008] The knowledge graph query and reasoning unit receives structured search statements and obtains candidate results related to the search intent based on the knowledge graph.

[0009] The matching calculation unit performs semantic similarity calculation on candidate results based on cosine similarity and outputs the optimal retrieval result.

[0010] As a further improvement to this technical solution, the property rights data includes at least: property rights transaction records, legal terms, and asset information.

[0011] As a further improvement to this technical solution, the knowledge modeling and graph construction unit constructs a knowledge graph from the preprocessed property rights data, including the following steps:

[0012] S1.1 Perform text parsing on the pre-processed property transaction records, legal clauses, and asset information to identify and extract core entities related to the property rights, including at least the property owner, transaction parties, contract subjects, legal clause number, name of the property object, and asset category;

[0013] S1.2 Based on syntactic dependency analysis and deep semantic recognition methods, identify property rights relationships from the semantic context between core entities, including at least transfer relationships, mortgage relationships, constraint relationships, and inheritance relationships;

[0014] S1.3 Supplement descriptive attribute information for core entities and property rights relationships. Among them, the attributes corresponding to property rights transaction records include transaction time, transaction amount, and transaction method; the attributes corresponding to legal clauses include clause content, applicable conditions, and legal effect; and the attributes corresponding to asset information include asset location, asset valuation, and asset status.

[0015] S1.4. Take the core entities as graph nodes, property rights as graph edges, and combine attribute information to generate a weighted directed graph structure, i.e., a knowledge graph.

[0016] As a further improvement to this technical solution, the retrieval parsing and semantic understanding unit includes a retrieval parsing module and a semantic understanding module;

[0017] The retrieval parsing module performs word segmentation, part-of-speech tagging, and intent recognition on the retrieval request based on the user's input, extracts the keywords and retrieval conditions contained therein, and transforms the keywords and retrieval conditions into structured retrieval statements based on the intent path discovery scheme of dynamic multi-hop reasoning.

[0018] The semantic understanding module is based on structured search statements. It identifies the user's true intent through context modeling and semantic disambiguation methods, and maps the search request to entities, relationships and attributes in the knowledge graph.

[0019] As a further improvement to this technical solution, the retrieval parsing module extracts the keywords and retrieval conditions contained therein, and transforms the keywords and retrieval conditions into structured retrieval statements based on a dynamic multi-hop reasoning intent path discovery scheme, including the following steps:

[0020] S2.1 Perform word segmentation on the user-input search request, dividing the search request into a sequence of words;

[0021] S2.2, Perform part-of-speech tagging on the word sequence;

[0022] S2.3 Based on word sequences and their part-of-speech tagging, an intent path discovery scheme using dynamic multi-hop reasoning is used to generate a sorted list of intent paths;

[0023] S2.4. Transform the intent path list into structured retrieval statements for the knowledge graph.

[0024] As a further improvement to this technical solution, in step S2.3, an intent path discovery scheme based on dynamic multi-hop reasoning is used to generate a sorted intent path list, including the following steps:

[0025] S2.31. Input the word sequence into the knowledge graph generated in step S1.4 through the entity linker for comparison, establish corresponding link relationships between the identified target words and the graph nodes in the knowledge graph, and obtain the anchor entity set;

[0026] S2.32. Input the verbs and relational phrases contained in the retrieval request into the relation mapping model to generate corresponding semantic embedding vectors; at the same time, encode the set of relation types defined in the knowledge graph into an embedding vector space; calculate the cosine similarity between the request relational phrase vector and the relation type vector, and filter out the candidate relations that are closest in semantics;

[0027] S2.33. Starting from the anchor entity in step S2.31, and combining the candidate relations obtained in step S2.32, perform path expansion with a finite step length in the local neighborhood of the knowledge graph to generate a candidate path set.

[0028] S2.34. Based on semantic matching degree, path confidence degree and hop count penalty factor, a ranking function is used to score and rank the candidate paths; finally, the top t paths are selected as the intent parsing results to form a ranked intent path list.

[0029] As a further improvement to this technical solution, the semantic understanding module identifies the user's true intent through context modeling and semantic disambiguation methods, and maps the retrieval request to entities, relationships, and attribute representations in a knowledge graph, including the following steps:

[0030] S3.1 By semantically encoding the retrieval request and the session history of the retrieval request, a contextual semantic vector is constructed, and the keywords, time expressions and pronouns in the retrieval request are associated with the contextual entities to generate an enhanced query representation that integrates contextual information;

[0031] S3.2 For ambiguous entity, relation, and attribute candidates in the search request, multi-factor scoring and filtering are performed by combining the context modeling results and the knowledge graph adjacency relationship;

[0032] S3.3. Fill the filtered entities, relations and attributes into predefined semantic slots according to semantic roles, and construct logical triples. Map the retrieval request to graph nodes, graph edges and attribute information in the knowledge graph to form a structured query template. Output the sorted candidate result list, and introduce evidence source credibility modeling and comparative path synthesis to optimize the sorted candidate result list.

[0033] As a further improvement to this technical solution, S3.3 introduces evidence source credibility modeling and comparative path synthesis to optimize the sorted candidate result list, including the following steps:

[0034] S3.31. For each path in the sorted candidate result list, extract its supporting evidence sources;

[0035] S3.32. Assign initial weights to different sources of supporting evidence and calculate the overall credibility score of the candidate paths. ;

[0036] S3.33, based on the overall credibility score Candidate paths that exceed the path threshold 'a' are used as anchor points, and comparison paths are generated within the local neighborhood of the knowledge graph.

[0037] S3.34. Perform a dimension-by-dimensional comparison between the anchor point path and its comparison path to evaluate the differences.

[0038] S3.35. Introduce the credibility score and the difference assessment result into the ranking function to re-score and optimize the candidate result list.

[0039] As a further improvement to this technical solution, the graph query and reasoning unit includes a graph query module and a graph reasoning module;

[0040] The graph query module is used to receive structured search statements output by the retrieval parsing and semantic understanding unit, and to perform query operations on the knowledge graph constructed in S1.4; through entity index, relation index and attribute index mechanism, it retrieves entity nodes, relation edges and their attribute information that match the search conditions from the knowledge graph, and generates a preliminary candidate result set;

[0041] The graph reasoning module, based on the preliminary candidate result set and combined with the relationship chains and reasoning rules in the knowledge graph, performs logical reasoning, path inference, and semantic expansion to generate a semantically enhanced reasoning result set.

[0042] As a further improvement to this technical solution, the matching calculation unit performs semantic similarity calculation on the candidate results based on cosine similarity and outputs the optimal retrieval result, including the following steps:

[0043] S4.1 Extract multi-dimensional feature representations of entities, relationships, and attributes from the candidate results output by the graph query and reasoning unit;

[0044] S4.2. Map the semantic representation of the user's retrieval request to the feature vectors of the candidate results into a unified representation space using the feature alignment method;

[0045] S4.3 Calculate the semantic similarity score between the user's retrieval request representation and the candidate result representation using cosine similarity;

[0046] S4.4 Sort the candidate results from high to low according to their semantic similarity scores, and select the top b results as the optimal search results.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] 1. The knowledge graph-based intelligent property rights retrieval and matching system of this invention, by introducing a dynamic multi-hop reasoning intent path discovery scheme combined with a semantic understanding mechanism that integrates context modeling and semantic disambiguation, can accurately parse complex retrieval requests input by users in natural language. In the processes of entity linking, relation mapping, and path expansion, it comprehensively considers link confidence, semantic similarity, path confidence, and contextual dependencies, and optimizes intent path generation through bundle search and ranking functions, effectively solving problems such as entity homonyms, diverse relational expressions, and ambiguous referencing in property rights data. Simultaneously, by combining conversation history for coreference resolution and context enhancement, the system can accurately identify referents such as "he" and "the property," achieving a leap from surface keyword matching to deep semantic understanding, significantly improving the ability to parse complex and implicit query intents and the relevance and accuracy of the final retrieval results.

[0049] 2. The knowledge graph-based intelligent property rights retrieval and matching system of this invention, based on knowledge graph querying, constructs a reasoning optimization mechanism that integrates evidence source credibility modeling and comparative path synthesis. This mechanism not only retrieves matching results from the graph but also discovers implicit ownership relationships (such as inheritance chains and transfer paths) through logical reasoning. Furthermore, it calculates the comprehensive credibility of paths by weighting evidence sources (such as manual annotation and external authoritative databases), thereby improving the reliability of the results. Further, by generating semantically related but structurally different comparative paths and evaluating their differences, the system can identify potential biases or missing information in a single path, thus prioritizing results with more complete semantic coverage and more robust logic in the ranking process. This mechanism not only improves the quality and diversity of the candidate result set but also provides traceable reasoning basis and confidence scores for the final output, enhancing the system's interpretability and user trust. It is particularly suitable for high-risk decision-making scenarios such as property rights determination and legal compliance review. Attached Figure Description

[0050] Figure 1 This is an overall flowchart of the present invention;

[0051] The meanings of the labels in the diagram are as follows:

[0052] 1. Data acquisition and processing unit; 2. Knowledge modeling and graph construction unit; 3. Retrieval parsing and semantic understanding unit; 4. Graph query and reasoning unit; 5. Matching calculation unit. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example: Please refer to Figure 1 As shown, a knowledge graph-based intelligent property rights retrieval and matching system is provided, including:

[0055] Data acquisition and processing unit 1 acquires property rights data and preprocesses the property rights data;

[0056] In this embodiment, the property rights data includes at least: property rights transaction records, legal terms, and asset information.

[0057] The knowledge modeling and graph construction unit 2 constructs a knowledge graph from the preprocessed property rights data based on entity extraction, relation extraction, and attribute annotation.

[0058] In this embodiment, the knowledge modeling and graph construction unit 2 constructs a knowledge graph from the preprocessed property rights data, including the following steps:

[0059] S1.1 Perform text parsing on the pre-processed property transaction records, legal clauses, and asset information to identify and extract core entities related to the property rights, including at least the property owner, transaction parties, contract subjects, legal clause number, name of the property object, and asset category;

[0060] S1.2 Based on syntactic dependency analysis and deep semantic recognition methods, identify property rights relationships from the semantic context between core entities, including at least transfer relationships, mortgage relationships, constraint relationships, and inheritance relationships;

[0061] S1.3 Supplement descriptive attribute information for core entities and property rights relationships. Among them, the attributes corresponding to property rights transaction records include transaction time, transaction amount, and transaction method; the attributes corresponding to legal clauses include clause content, applicable conditions, and legal effect; and the attributes corresponding to asset information include asset location, asset valuation, and asset status.

[0062] S1.4. Using core entities as graph nodes and property rights relationships as graph edges, a weighted directed graph structure, i.e. a knowledge graph, is generated by combining attribute information. This enables the graph-based organization and representation of property rights data. The knowledge graph is then stored in a graph database, and an indexing mechanism is established to support subsequent retrieval, logical reasoning, and dynamic updates.

[0063] The retrieval parsing and semantic understanding unit 3 is used to receive the retrieval request input by the user and transform the retrieval request into a structured retrieval statement that matches the knowledge graph based on the dynamic multi-hop reasoning intent path discovery scheme. It parses the retrieval intent of the retrieval request through context modeling and semantic disambiguation methods.

[0064] In this embodiment, the retrieval parsing and semantic understanding unit 3 includes a retrieval parsing module and a semantic understanding module;

[0065] The retrieval parsing module performs word segmentation, part-of-speech tagging, and intent recognition on the retrieval request based on the user's input, extracts the keywords and retrieval conditions contained therein, and transforms the keywords and retrieval conditions into executable structured retrieval statements based on the intent path discovery scheme of dynamic multi-hop reasoning, thereby ensuring that subsequent queries can match the knowledge graph data format.

[0066] Furthermore, the retrieval and parsing module extracts the keywords and search conditions contained therein, and transforms the keywords and search conditions into structured search statements, including the following steps:

[0067] S2.1 Perform word segmentation on the user-input search request, dividing the search request (which is continuous natural language text) into a word sequence;

[0068] S2.2. Perform part-of-speech tagging on the word sequence to identify personal names, place names, time, numerical values, and specific words related to property rights.

[0069] S2.3 Based on word sequences and their part-of-speech tagging, an intent path discovery scheme using dynamic multi-hop reasoning is used to generate a sorted list of intent paths;

[0070] Among them, the dynamic multi-hop reasoning intent path discovery scheme addresses the problem of parsing complex intents in user natural language retrieval requests in property rights intelligent retrieval scenarios. Specifically, when a user inputs a query involving multiple entities and relationships (such as "the property that Zhang San mortgaged to Li Si last year"), the system needs to identify core entities (such as Zhang San, Li Si, and the property) and relationships (such as mortgage) from the knowledge graph, and dynamically explore potential paths connecting these entities to accurately capture the user's potentially implicit multi-hop intents (such as assets indirectly related through transfer or inheritance relationships). This solves the challenge of traditional methods struggling to handle complex semantics and indirect relationships, ensuring that retrieval requests can be transformed into executable queries that match the knowledge graph structure. Through dynamic multi-hop reasoning and path ranking mechanisms, this scheme can effectively handle complex multi-hop queries, significantly improving the accuracy and robustness of intent parsing. Existing methods are often limited to single-hop matching or simple keyword-based mapping, which cannot deeply explore the implicit relationships in knowledge graphs. This solution combines entity linking, relationship mapping, bundle search path expansion, and multi-factor ranking (such as semantic matching degree, path confidence, and hop count penalty), and introduces evidence credibility modeling and contrast path synthesis to more comprehensively capture users' deep intentions, reduce ambiguity, support intelligent processing of indirect associations and long-path reasoning, and ultimately improve the semantic understanding ability and result relevance of the retrieval system.

[0071] The intent path discovery scheme, which utilizes dynamic multi-hop reasoning, generates a sorted list of intent paths, including the following steps:

[0072] S2.31. Input the word sequence into the knowledge graph generated in step S1.4 using the entity linker for comparison. (The entity linker is used to establish a correspondence between the target words in the retrieval request and the unique entity nodes in the knowledge graph. It is implemented using publicly available entity linking tools (such as DBpedia).) The Spotlight algorithm identifies target words (words already bound to graph node IDs, such as Zhang San, Li Si, and Real Estate X) and establishes corresponding links with graph nodes (with unique identifier IDs) in the knowledge graph, resulting in a set of anchor entities. During the entity link comparison process, multiple candidate entities are generated. A matching score (confidence) is calculated between each candidate entity and the input target word. A link confidence threshold is then used for filtering. When the entity linker maps Zhang San to the knowledge graph, it may find multiple entities with the same name (e.g., Zhang San (property owner A), Zhang San (property owner B)). The entity linker calculates a confidence score for each candidate entity. If the highest score is greater than the threshold, the entity is recognized as the final identification result and a corresponding relationship is established with the node ID. If the highest score is lower than the threshold, it is considered unreliable and may enter ambiguity resolution or be discarded. In this embodiment, the link confidence threshold... A link is considered successful only if its confidence score is higher than the threshold; if multiple candidate entities exist, they are sorted from highest to lowest confidence score, and the number of candidates is retained. That is, each mentioned entity can retain a maximum of three candidate nodes as anchor entities;

[0073] S2.32. Input the verbs and relational phrases (such as "holding" and "mortgaged") contained in the retrieval request into the relation mapping model to generate corresponding semantic embedding vectors; simultaneously, encode the set of relation types defined in the knowledge graph into an embedding vector space; calculate the cosine similarity between the request relational phrase vector and the relation type vector (vectorize the request relational phrase and each relation type in the knowledge graph to obtain relational phrase vectors and relation type vectors, and calculate the similarity score between them in the semantic space using the cosine similarity formula; the score range is...). (The closer the value is to 1, the more semantically similar the relation). Candidate relations with the closest semantic similarity are selected. In this embodiment, the cosine similarity threshold is used. Filter candidate relationships (select those with similarity greater than the cosine similarity threshold) (of), retain the first one with the highest similarity. One candidate relation type;

[0074] The relation mapping model takes verbs and relational phrases in the retrieval request and their contextual entities as input. First, it models their features through a Bi-LSTM (Bilingual Long Short-Term Memory) network. (Specifically, the verbs and relational phrases in the retrieval request are segmented and vectorized into fixed-dimensional dense vector sequences. Then, they are input into a Bi-LSTM network, which models contextual dependencies through forward and backward recurrent units to obtain the semantic encoding representation of each word in the bi-directional context. Based on this, an attention mechanism is introduced to assign higher weights to the key segments in the sequence that are most relevant to the relational semantics and perform weighted aggregation to obtain a global feature vector that can highlight the core semantics and suppress redundant noise, thus realizing deep semantic feature modeling of the relational phrases in the retrieval request.) In the intermediate layer, the expression of key relational segments is enhanced using an attention mechanism or semantic alignment method, and semantic embedding vectors aligned with the relational space of the knowledge graph are generated.

[0075] S2.33. Starting from the anchor entity in step S2.31, and combining the candidate relations obtained in step S2.32, perform path expansion with a finite step length in the local neighborhood of the knowledge graph to generate a candidate path set containing multiple entities and relations.

[0076] Specifically: Based on the candidate set of anchor entities in S2.31 and the candidate set of relations in S2.32, an initial search queue is constructed; for each valid anchor entity, a first-hop candidate node pair (anchor entity ID, candidate relation type, and candidate target entity ID) is generated, and a cumulative confidence score is recorded for each candidate (initially the entity link confidence score multiplied by the relation similarity of the edge); a bundle search strategy is used to control the expansion scale, with the bundle width set to... This limits the number of candidate paths retained in each expansion layer, thereby covering high-quality paths with controllable computational cost.

[0077] Starting with the initial search queue, the path is expanded layer by layer (number of hops) until the maximum number of hops is reached. (i.e., the maximum number of layers or steps allowed to expand along relation edges in the graph starting from the initial entity) or no new paths are scalable: each expansion selects from the currently retained candidate paths. For each path, enumerate the candidate relationships provided by S2.32 at its current endpoint and retrieve the corresponding adjacent entity nodes to generate a new path;

[0078] Calculate the path confidence score for each newly generated path in real time. It is defined as the product of relation similarity and the average confidence of entity links in the path, that is:

[0079] ;

[0080] In the formula, For the first in the path The confidence level of a relation edge reflects the degree to which the relation is credible within the knowledge graph. For the first in the path The semantic relevance of each node to the input entity or context. This represents the number of relations contained in the path, i.e., the path hop count. This represents the total number of nodes involved in the path.

[0081] To support semantic matching, a semantic vector representation of each path is calculated (obtained by concatenating the nodes and relation vectors on the path in order or taking their average value), and the similarity is calculated between this representation and the semantic vector of the original query. ;

[0082] After each hop, all newly generated paths are filtered according to a temporary scoring metric, and the previous ones are retained. The number of hops is used to control the search size; the final output of this step S2.33 is a set of candidate paths generated within the hop limit (containing at least path sequences, ...). , (Number of jumps);

[0083] S2.34. Based on semantic matching degree, path confidence, and hop count penalty factor, a ranking function is used to score and rank candidate paths (according to...). (Sort all candidate paths from highest to lowest); finally select the top t paths as the intent parsing results to form a sorted intent path list;

[0084] The sorting function used for scoring is:

[0085] ;

[0086] in, To query the similarity between semantic vectors and path semantic vectors, For path confidence, Score the path. This represents the number of hops in the path. This is a weighted coefficient for similarity, with a value range of [0.4, 0.7]. Empirically, it is usually assigned the highest weight because semantic similarity to the user's query is key. The weighting coefficient for path confidence ranges from [0.2, 0.4], an empirically determined value, with weights assigned accordingly, to ensure the reliability of the path within the knowledge graph. This is a weighting factor for the number of hops on a path, ranging from -0.2 to -0.05. It's an empirically chosen value used to penalize paths with excessive hops and encourage the system to return more direct and concise answers. Sort the intent paths from highest to lowest, and output the top t intent paths as the final parsing result; each output path also includes an executable structured query template (e.g., a SPARQL template), path score, and more. The values, a list of entity node IDs involved, and their link confidence are provided for direct execution by the graph query and inference unit 4; for cases where no valid path is generated, a fallback strategy is adopted according to the following priority: reduce or Value retry, expansion Alternatively, the query may be submitted to the semantic classification module for an alternative parsing strategy;

[0087] S2.4. Transform the intent path list into an executable structured retrieval statement for the knowledge graph. The statement is constructed based on a query language (such as SPARQL) or an equivalent logical expression to drive subsequent graph queries and inferences.

[0088] The semantic understanding module is based on structured retrieval statements. It identifies the user's true intent through context modeling and semantic disambiguation methods, and maps the retrieval request to entities, relationships and attributes in the knowledge graph, thereby achieving accurate semantic understanding and intent parsing.

[0089] The semantic understanding module identifies the user's true intent through context modeling and semantic disambiguation methods, and maps the retrieval request to entities, relationships, and attributes in a knowledge graph, including the following steps:

[0090] S3.1. By semantically encoding the retrieval request and its session history, a contextual semantic vector is constructed. Keywords, time expressions, and pronouns in the retrieval request are associated with contextual entities to generate an enhanced query representation that integrates contextual information. Specifically, in constructing the contextual semantic vector, the current retrieval request and its corresponding session history text are first segmented and vectorized. Then, they are input into a semantic encoding model (such as a bidirectional Transformer) for context modeling. (The context modeling process is as follows: after segmentation and vectorization, the retrieval request and its session history are input into a semantic encoding model (such as a bidirectional Transformer). This model consists of multiple stacked encoders, each layer containing a multi-head self-attention mechanism and a feedforward neural network. First, the self-attention mechanism calculates the global dependencies between words in the sequence, then residual connections and layer normalization maintain stable information transmission, followed by nonlinear feature transformation through a feedforward network. In the data flow, the input sequence is first embedded in a vector space and then positionally encoded, passing through multiple encoders to obtain the context.) The enhanced semantic representation vector, where the output at each position not only contains the semantics of the word itself but also incorporates relevant information from the conversation history and global context, providing semantic support for subsequent entity mapping and semantic disambiguation, captures the dependencies and semantic associations between the retrieval request and the historical conversation through an attention mechanism. This generates an enhanced semantic vector that incorporates contextual information. This vector can simultaneously represent the semantic content of the current request and its contextual relationship in the conversation history, providing semantic support for subsequent entity, relation, and attribute mapping. For example, it first identifies keywords (Zhang San), time expressions (last year), and pronouns (he, the property) in the retrieval request. Then, it combines known entities in the conversation history or knowledge graph to determine the specific object or time range they refer to. For example, if the conversation history shows: "Zhang San bought property X in 2024," and the current retrieval request is "He sold the property last year," the system maps "he" to Zhang San, "the property" to property X, and "last year" to 2024 through coreference disambiguation, thereby establishing an association between natural language expressions and contextual entities and ensuring the accuracy of semantic parsing.

[0091] S3.2. For ambiguous entity, relation, and attribute candidates in the retrieval request, multi-factor scoring and screening are performed by combining context modeling results and knowledge graph adjacency relationships to ensure that the elements finally mapped to the knowledge graph are unique and accurate. Specifically: First, a semantic vector is generated for each candidate, and the semantic matching degree is calculated by combining it with the context semantic vector (each candidate entity, relation, or attribute is encoded into a fixed-dimensional vector through a pre-trained language model (such as BERT), and the retrieval request and session history are generated into a context semantic vector through the same model. Then, the matching score between the candidate vector and the context vector is calculated using a vector similarity measurement method (such as cosine similarity) to obtain the semantic matching degree). At the same time, the connection rationality between the candidate entity and related nodes is analyzed using the knowledge graph adjacency relationship (first, based on the node ID of the candidate entity in the knowledge graph, its directly adjacent relation edges and connected nodes are traversed to extract the connection rationality with other nodes in the retrieval request). The system collects adjacency information related to entities or attributes, and then evaluates whether these connections conform to semantic and logical constraints, such as whether the relationship type matches, whether the path length is reasonable, and whether there are conflicting or duplicate connections. The adjacency rationality of each candidate entity is quantified into a score (i.e., adjacency support). The semantic matching degree, adjacency support, and entity link confidence are weighted and fused to form a comprehensive score. Then, the candidates are ranked and the unique mapping result with the highest score that meets the threshold is selected. For example, if the search request is "Zhang San's invested company is located in Shanghai", and there are multiple entities with the same name "Zhang San" in the knowledge graph, then by combining semantic matching with the context and history of the session and the adjacency score of the candidate company node with "Shanghai", "Zhang San (Investor A)" can be finally determined as the unique entity mapping. At the same time, "investment" is mapped to "investment relationship", and "company" is mapped to a specific company node, thus achieving ambiguity resolution.

[0092] S3.3. Fill the predefined semantic slots with the filtered entities, relations, and attributes according to their semantic roles (first, determine the type of each slot based on the semantic structure of the retrieval request (e.g., subject slot corresponds to entity A, predicate slot corresponds to relation type, object slot corresponds to entity B or attribute constraint), then match the candidate entities, relations, and attributes filtered through ambiguity resolution and scoring with the corresponding slots one by one according to their semantic roles), and construct logical triples or plurals (e.g., <entity A, relation, entity B> or <entity, attribute, value>), mapping the retrieval request to graph nodes (e.g., Zhang San, real estate X) and graph edges (here, graph edges correspond to the relationships between entities in the knowledge graph). The relationship category (such as holding, mortgaging, i.e., the unique identifier of the edge type) and attribute information (the attribute values ​​or restrictions of the corresponding entity or relationship, such as transaction amount, transaction time, asset status, etc., which can be used to accurately match the conditions in the retrieval request) form a structured query template. A confidence score and source basis are attached to each semantic mapping result, and a sorted candidate result list is output. The template with high confidence is selected first for query execution, while the low confidence results are labeled to support subsequent human-computer clarification interaction. Evidence source credibility modeling and comparative path synthesis are introduced to optimize the sorted candidate result list to improve the reliability and semantic accuracy of the candidate path results.

[0093] Furthermore, this method introduces evidence source credibility modeling and comparative path synthesis, primarily addressing the core issues of uneven candidate path quality and ambiguous credibility in property rights retrieval scenarios. In serious fields like property rights, which involve legal validity and asset transactions, automatically generated candidate paths may originate from data of varying credibility (such as high-quality data annotated manually or noisy data extracted automatically), and a single path may not fully reflect complex property rights relationships (such as assets with competing ownership). This method directly solves the key problem of how to identify the most reliable and comprehensive property rights relationship chain among multiple paths by quantitatively evaluating the credibility of evidence sources and generating comparative paths, effectively preventing retrieval bias caused by relying on low-quality or one-sided evidence. Existing technologies typically rely solely on semantic similarity or a single confidence level for ranking. This scheme, through the synergistic mechanism of evidence credibility quantification and comparative path synthesis, achieves a qualitative leap from finding relevant paths to finding reliable and comprehensive paths. Existing technologies often treat triples from different sources (such as manual verification and automatic extraction) in the knowledge graph equally and are prone to getting stuck in local optima. In contrast, this method prioritizes the adoption of highly authoritative data by assigning weights to evidence sources and calculating comprehensive credibility. At the same time, it actively generates differential comparison paths, which can expose potential contradictions and discover complementary information, thereby systematically improving the reliability, robustness and decision support capabilities of retrieval results, especially in property disputes or complex ownership reasoning.

[0094] The ranking of candidate results is optimized by introducing evidence source credibility modeling and comparative path synthesis, including the following steps:

[0095] S3.31. For each path in the sorted candidate result list, extract its supporting evidence sources, including at least the triple relationships within the knowledge graph, external knowledge base references, and historical interaction log records.

[0096] S3.32. Assign initial weights to different sources of supporting evidence (e.g., high weights for manually labeled knowledge bases, and low weights for automatically extracted or weakly supervised knowledge bases), and calculate the overall credibility score of candidate paths. ( In the formula, The amount of evidence upon which the path depends For the first The weight of each piece of evidence (high credibility sources have higher weight, such as manually labeled knowledge bases). For the first The credibility of the evidence itself (between [0,1]);

[0097] S3.33, based on the overall credibility score Candidate paths exceeding the path threshold 'a' are used as anchor points. Semantically related but differentiated comparative paths are generated within the local neighborhood of the knowledge graph to ensure that different paths complement or contrast each other in terms of entity coverage, relational semantics, and hop length. Specifically, this involves first using a comprehensive credibility score... Starting with key entities and relationships in anchor paths exceeding a path threshold 'a', the adjacency relationships and sets of adjacent entities of each entity are extracted. Candidate relationships with similar but different relationship types to the original path are selected and combined with adjacent entities to generate new path extensions. Then, hop count limits and semantic vector calculations are applied to the new paths to ensure that the paths are semantically related to the original anchor paths while differing in entity coverage or relationship combinations. Next, a difference score (e.g., entity overlap rate) is calculated for each newly generated path, and paths with excessive semantic deviation or logical conflicts are eliminated. Finally, the set of comparison paths that meet the criteria is output for subsequent comparison and optimization ranking with anchor paths.

[0098] S3.34. Perform a dimension-by-dimensional comparison between the anchor path and its comparison path to conduct a difference assessment, which measures the degree of difference between the two in terms of semantic relevance, consistency of evidence sources, and logical reasoning. First, align the entity and relation vectors at corresponding positions in the two paths in order. Then, calculate the difference index in each dimension, including whether the entity IDs match, the similarity of relation types, the difference in attribute constraints, and the difference in hop count, etc. Next, perform weighted summation or normalization on the difference results of each dimension to obtain the overall difference score for each pair of paths. This score reflects the degree of difference between the two paths in terms of entity coverage, relation combination, and structural logic. The larger the value, the greater the difference. It can be directly used in the ranking function to optimize the candidate result list.

[0099] S3.35. Introduce the credibility score and the difference assessment result into the ranking function to re-score and optimize the candidate result list;

[0100] The ranking function, which incorporates the credibility score and the difference assessment results, is as follows:

[0101]

[0102] In the formula, The overall credibility score of the candidate paths. The result of the difference assessment between the anchor path and the comparison path, i.e., the overall difference score, This is a weighting coefficient for the overall confidence score, empirically chosen with a value ranging from [0.1, 0.3]. As a supplement and correction to the basic ranking function, the weight should not be too high to avoid completely overturning the dominant roles of semantics and confidence. The weighting coefficient for the difference assessment results is empirically selected, with a value range of [0.05, 0.15]. Diversity is encouraged, and the weight is usually the lowest, aiming to fine-tune the results without significantly affecting the main ranking.

[0103] In this embodiment, steps S3.3 to S3.34 mainly improve the quality of the initial candidate path list. Through evidence source credibility modeling and comparative path generation, the system re-evaluates the reliability and semantic coverage of each candidate path, thereby adjusting the priority of candidate results or eliminating low-quality paths to ensure that the candidate set itself is more credible and complete. After the candidate set is optimized, it is still necessary to map the paths to the final executable query ranking. Here, the credibility score and the difference evaluation result are combined with the original semantic matching degree, path confidence, and hop count index to calculate the final path score. This is used for final path sorting and optimization.

[0104] Knowledge graph query and reasoning unit 4 receives structured search statements and obtains candidate results related to the search intent based on the knowledge graph;

[0105] In this embodiment, the graph query and reasoning unit 4 includes a graph query module and a graph reasoning module;

[0106] The graph query module receives structured search statements output by the retrieval parsing and semantic understanding unit 3 and performs query operations on the knowledge graph constructed in S1.4. Through entity indexing, relation indexing, and attribute indexing mechanisms, it retrieves entity nodes, relation edges, and their attribute information that match the search conditions from the knowledge graph, generating a preliminary candidate result set to provide basic data support for subsequent reasoning. Specifically, it first parses the entity, relation, and attribute conditions from the structured search statement, and then uses the entity index, relation index, and attribute index of the knowledge graph to quickly locate graph nodes and edges that meet the conditions: the entity index is used to find graph nodes that match the search keywords or entity IDs; the relation index is used to match the relation type or relation ID specified in the request; and the attribute index is used to filter nodes or edges that satisfy attribute constraints (such as transaction amount or asset status). During the retrieval process, Boolean matching, range query, or fuzzy matching strategies can be combined to generate candidate triples or multi-tuples for entity-relationship-attribute combinations that meet the conditions, and these candidate results are summarized to form a preliminary candidate result set.

[0107] The graph reasoning module, based on a preliminary candidate result set and combined with the relationship chains and reasoning rules in the knowledge graph, performs logical reasoning, path inference, and semantic expansion to generate a semantically enhanced reasoning result set. Specifically, starting with the preliminary candidate result set, it combines the relationship chains and predefined reasoning rules in the knowledge graph (the predefined reasoning rules are used to characterize typical logical relationships in the property rights field: if an asset is registered under A's name, and A and B have an inheritance relationship, then it can be deduced that the actual control of the asset may belong to B; if A mortgages a property to B, and B transfers the property to C, then it can be deduced that there is an indirect transfer relationship between A and C; if the asset has been frozen, then its transfer relationship should be...). (Marked as invalid) for inference. In the inference process, firstly, the candidate entity nodes and their adjacency relationships are expanded and combined to identify implicit or indirect relationships (such as the final ownership in a transfer chain), while correcting low-confidence matching results; then, based on semantic embedding vectors, semantic matching and relevance evaluation are performed on the expanded paths to supplement missing attributes or related entities, thereby expanding semantic information; finally, the logical reasoning, path inference, and semantically enhanced entities, relationships, and attributes are combined and organized into a reasoning result set, with each result accompanied by a confidence score and reasoning basis, providing high-quality input to the matching calculation unit to improve the accuracy and intelligence level of the retrieval results.

[0108] The matching calculation unit 5 performs semantic similarity calculation on the candidate results based on cosine similarity and outputs the optimal retrieval result;

[0109] In this embodiment, the matching calculation unit 5 performs semantic similarity calculation on the candidate results based on cosine similarity and outputs the optimal retrieval result, including the following steps:

[0110] S4.1 For the candidate results output by the graph query and reasoning unit 4, extract multi-dimensional feature representations of entities, relations and attributes respectively, including text description features, semantic embedding features, numerical attribute features and context structure features;

[0111] S4.2. By using the feature alignment method (linear transformation), the semantic representation of the user's retrieval request and the feature vector of the candidate results are mapped to a unified representation space, laying the foundation for similarity calculation and ranking.

[0112] S4.3 Calculate the semantic similarity score between the user's search request representation and the candidate result representation using cosine similarity (by converting the semantic vector of the user's search request...). and the feature vectors of candidate results Input their respective linear mapping matrices and The mapped vector is obtained through matrix multiplication. , Then, the mapping vectors are normalized to ensure they reside in the same vector space and have the same scale.

[0113] S4.4 Sort the candidate results from high to low according to the semantic similarity score, and select the top b results as the optimal search result output. The optimal search result output includes entity, relation, attribute information and corresponding comprehensive score, so that users can obtain accurate matching results or enter the subsequent interactive confirmation process.

[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A knowledge graph-based intelligent property rights retrieval and matching system, characterized in that, include: Data acquisition and processing unit (1), wherein the data acquisition and processing unit (1) acquires property rights data and preprocesses the property rights data; The knowledge modeling and graph construction unit (2) constructs a knowledge graph from the preprocessed property data based on entity extraction, relation extraction and attribute annotation. The retrieval parsing and semantic understanding unit (3) is used to receive the retrieval request input by the user and transform the retrieval request into a structured retrieval statement that matches the knowledge graph based on the dynamic multi-hop reasoning intent path discovery scheme. The retrieval intent of the retrieval request is parsed through context modeling and semantic disambiguation methods. The graph query and reasoning unit (4) receives structured search statements and obtains candidate results related to the search intent based on the knowledge graph. The matching calculation unit (5) performs semantic similarity calculation on the candidate results based on cosine similarity and outputs the optimal retrieval result; The retrieval parsing and semantic understanding unit (3) includes a retrieval parsing module and a semantic understanding module; The retrieval parsing module performs word segmentation, part-of-speech tagging, and intent recognition on the retrieval request based on the user's input, extracts the keywords and retrieval conditions contained therein, and transforms the keywords and retrieval conditions into structured retrieval statements based on the intent path discovery scheme of dynamic multi-hop reasoning. The semantic understanding module is based on structured retrieval statements. It identifies the user's true intent through context modeling and semantic disambiguation methods, and maps the retrieval request to entities, relationships and attributes in the knowledge graph. The retrieval parsing module extracts the keywords and search conditions contained therein, and transforms the keywords and search conditions into structured search statements based on a dynamic multi-hop reasoning intent path discovery scheme, including the following steps: S2.1 Perform word segmentation on the user-input search request, dividing the search request into a sequence of words; S2.2, Perform part-of-speech tagging on the word sequence; S2.3 Based on word sequences and their part-of-speech tagging, an intent path discovery scheme using dynamic multi-hop reasoning is used to generate a sorted list of intent paths; S2.

4. Transform the intent path list into a structured retrieval statement for the knowledge graph; In step S2.3, the intent path discovery scheme using dynamic multi-hop reasoning generates a sorted list of intent paths, including the following steps: S2.

31. Input the word sequence into the knowledge graph generated in step S1.4 through the entity linker for comparison, establish corresponding link relationships between the identified target words and the graph nodes in the knowledge graph, and obtain the anchor entity set; S2.

32. Input the verbs and relational phrases contained in the retrieval request into the relation mapping model to generate corresponding semantic embedding vectors; at the same time, encode the set of relation types defined in the knowledge graph into an embedding vector space; calculate the cosine similarity between the request relational phrase vector and the relation type vector, and filter out the candidate relations that are closest in semantics; S2.

33. Starting from the anchor entity in step S2.31, and combining the candidate relations obtained in step S2.32, perform path expansion with a finite step length in the local neighborhood of the knowledge graph to generate a candidate path set. S2.

34. Based on semantic matching degree, path confidence degree and hop count penalty factor, a ranking function is used to score and rank the candidate paths; finally, the top t paths are selected as the intent parsing results to form a ranked intent path list.

2. The knowledge graph-based intelligent property rights retrieval and matching system according to claim 1, characterized in that: The property rights data includes at least: property rights transaction records, legal terms, and asset information.

3. The knowledge graph-based intelligent property rights retrieval and matching system according to claim 2, characterized in that: The knowledge modeling and graph construction unit (2) constructs a knowledge graph from the preprocessed property rights data, including the following steps: S1.1 Perform text parsing on the pre-processed property transaction records, legal clauses, and asset information to identify and extract core entities related to the property rights, including at least the property owner, transaction parties, contract subjects, legal clause number, name of the property object, and asset category; S1.2 Based on syntactic dependency analysis and deep semantic recognition methods, identify property rights relationships from the semantic context between core entities, including at least transfer relationships, mortgage relationships, constraint relationships, and inheritance relationships; S1.3 Supplement descriptive attribute information for core entities and property rights relationships. Among them, the attributes corresponding to property rights transaction records include transaction time, transaction amount, and transaction method; the attributes corresponding to legal clauses include clause content, applicable conditions, and legal effect; and the attributes corresponding to asset information include asset location, asset valuation, and asset status. S1.

4. Take the core entities as graph nodes, property rights as graph edges, and combine attribute information to generate a weighted directed graph structure, i.e., a knowledge graph.

4. The knowledge graph-based intelligent property rights retrieval and matching system according to claim 1, characterized in that: The semantic understanding module identifies the user's true intent through context modeling and semantic disambiguation methods, and maps the retrieval request to entities, relationships, and attribute representations in a knowledge graph, including the following steps: S3.1 By semantically encoding the retrieval request and the session history of the retrieval request, a contextual semantic vector is constructed, and the keywords, time expressions and pronouns in the retrieval request are associated with the contextual entities to generate an enhanced query representation that integrates contextual information; S3.2 For ambiguous entity, relation, and attribute candidates in the search request, multi-factor scoring and filtering are performed by combining the context modeling results and the knowledge graph adjacency relationship; S3.

3. Fill the filtered entities, relations and attributes into predefined semantic slots according to semantic roles, and construct logical triples. Map the retrieval request to graph nodes, graph edges and attribute information in the knowledge graph to form a structured query template. Output the sorted candidate result list, and introduce evidence source credibility modeling and comparative path synthesis to optimize the sorted candidate result list.

5. The knowledge graph-based intelligent property rights retrieval and matching system according to claim 4, characterized in that: In step S3.3, evidence source credibility modeling and comparative path synthesis are introduced to optimize the ranked candidate result list, including the following steps: S3.

31. For each path in the sorted candidate result list, extract its supporting evidence sources; S3.

32. Assign initial weights to different sources of supporting evidence and calculate the overall credibility score of the candidate paths. ; S3.33, based on the overall credibility score Candidate paths that exceed the path threshold 'a' are used as anchor points, and comparison paths are generated within the local neighborhood of the knowledge graph. S3.

34. Perform a dimension-by-dimensional comparison between the anchor point path and its comparison path to evaluate the differences. S3.

35. Introduce the credibility score and the difference assessment result into the ranking function to re-score and optimize the candidate result list.

6. The knowledge graph-based intelligent property rights retrieval and matching system according to claim 3, characterized in that: The graph query and reasoning unit (4) includes a graph query module and a graph reasoning module; The graph query module is used to receive the structured search statement output by the retrieval parsing and semantic understanding unit (3), and perform query operations on the knowledge graph constructed in S1.4; through the entity index, relation index and attribute index mechanism, it retrieves the entity nodes, relation edges and their attribute information that match the search conditions from the knowledge graph and generates a preliminary candidate result set; The graph reasoning module, based on the preliminary candidate result set and combined with the relationship chains and reasoning rules in the knowledge graph, performs logical reasoning, path inference, and semantic expansion to generate a semantically enhanced reasoning result set.

7. The knowledge graph-based intelligent property rights retrieval and matching system according to claim 1, characterized in that: The matching calculation unit (5) performs semantic similarity calculation on the candidate results based on cosine similarity and outputs the optimal retrieval result, including the following steps: S4.1 Extract multidimensional feature representations of entities, relationships and attributes from the candidate results output by the graph query and reasoning unit (4); S4.

2. Map the semantic representation of the user's retrieval request to the feature vectors of the candidate results into a unified representation space using the feature alignment method; S4.3 Calculate the semantic similarity score between the user's retrieval request representation and the candidate result representation using cosine similarity; S4.4 Sort the candidate results from high to low according to their semantic similarity scores, and select the top b results as the optimal search results.

Images