Method and system for cross-modal retrieval of property information based on knowledge graph

Abstract

The present application relates to the technical field of retrieval, more particularly, the present application relates to a kind of property information cross-modal retrieval method and system based on knowledge graph, comprising: the multi-modal attribute graph of static node and dynamic event node is constructed, each node in the graph is given source label according to data source, and text information and image information in the graph are processed into feature vector;The source entropy confidence of each node is calculated, and the source entropy confidence and the historical data dispersion of the corresponding source of the node in the preset time window are negatively correlated.The present application realizes the quantitative evaluation of the objectivity of different sources (such as IoT and artificial record) by constructing multi-modal attribute graph and calculating source entropy confidence, comprehensively considers historical dispersion, text and picture self-consistency and credibility factor, effectively suppresses subjective noise according to low-entropy high-power principle.

Description

A Method and System for Cross-Modal Retrieval of Property Information Based on Knowledge Graph Technical Field

[0001] This invention relates to the field of retrieval technology. More specifically, this invention relates to a method and system for cross-modal retrieval of property information based on knowledge graphs. Background Technology

[0002] With the deepening of smart community construction, the data involved in property management has experienced explosive growth, encompassing massive amounts of multimodal information such as equipment ledgers, IoT sensor logs, surveillance images, and resident repair reports. To effectively manage and utilize this fragmented and unstructured data, the industry typically employs knowledge graph technology to build a network connecting people, properties, events, and objects, and uses cross-modal retrieval technology to locate key information to assist in security early warning and fault diagnosis.

[0003] Existing technologies mainly employ deep learning models (such as BERT and ResNet) to extract and vectorize features from text and image information in the graph, mapping them to a unified feature space. Then, based on the semantic similarity between the query statement and node features, or by traversing the static topological structure of the graph, matching nodes are found as retrieval results.

[0004] However, the above methods have significant drawbacks in specific applications of property security. First, existing vectorized retrieval often treats data from different sources equally, ignoring the significant differences in objective accuracy between machine data (such as sensor records) and human data (such as resident complaints), which can easily lead to subjective noise interfering with the search results. Second, relying solely on semantic similarity or static graph traversal lacks consideration of the strict spatiotemporal logic in the physical world (such as the chronological order of causal events and the time span of event impact), which can easily lead to errors in the search results in terms of time scale and causal logic, making it difficult to meet the need for accurately locating the root cause of an incident. Summary of the Invention

[0005] This invention provides a cross-modal retrieval method and system for property information based on knowledge graphs. It aims to solve the problem that existing vectorized retrieval technologies often treat data from different sources equally, ignoring the huge difference in objective accuracy between machine data (such as sensor records) and human data (such as homeowner complaints), which easily leads to the retrieval results being interfered with by subjective noise.

[0006] In a first aspect, the present invention provides a cross-modal retrieval method for property information based on a knowledge graph, comprising: constructing a multimodal attribute graph containing static nodes and dynamic event nodes; assigning a source label to each node in the graph according to the data source; and performing feature vectorization processing on the text information and image information in the graph; calculating the source entropy confidence of each node, wherein the source entropy confidence is negatively correlated with the historical data dispersion of the corresponding source within a preset time window, and positively correlated with the cosine similarity between the text feature vector and the image feature vector within the node. The system calculates the dynamic conductivity between the current node and its neighboring nodes, which is positively correlated with the cosine similarity between the user query and the relational words between the current node and its neighboring nodes. It also calculates the final score of each node in the graph, which is positively correlated with the source entropy confidence of that node, as well as with the final score of its parent node and the dynamic conductivity between the parent node and the current node. The nodes are then sorted according to their final scores, and a preset number of nodes are output as search results. By constructing a multimodal graph and calculating the source entropy confidence of nodes, combined with historical dispersion, internal graph-text consistency, and reputation factors, the system effectively quantifies the reliability of different sources, giving higher weight to objective data (such as IoT records) and suppressing subjective noise. Simultaneously, by using dynamic conductivity to perform potential energy propagation scoring in the graph, it achieves accurate sorting based on data credibility and semantic relevance, significantly improving the accuracy of key information location in complex property scenarios.

[0007] Furthermore, the text and image information in the graph undergoes feature vectorization processing, including: extracting text feature vectors using a pre-trained BERT model and extracting image feature vectors using a pre-trained ResNet-50 model. Using pre-trained BERT and ResNet-50 models to process the text and images respectively allows for a deeper capture of the deep semantic features of the text and the high-level visual features of the images, ensuring the richness and accuracy of the vectorized representation of multimodal data and providing a solid data foundation for subsequent high-quality cross-modal matching.

[0008] Furthermore, before calculating the cosine similarity between the text feature vector and the image feature vector within a node, a fully connected layer is used to compress the dimensions of both text and image feature vectors to unify their feature dimensions. By compressing and unifying text and image features to the same dimension through the fully connected layer, not only is the technical challenge of cross-modal feature alignment solved, but the complexity of subsequent calculations is also reduced, making it possible to calculate the consistency between text and images within a node within a unified metric space.

[0009] Furthermore, the method for calculating historical data dispersion includes: acquiring all historical records of the corresponding information source within a preset time window, extracting text and image information from them respectively, and performing vectorization processing to obtain a historical text vector set and a historical image vector set; calculating the sum of the variances of the historical text vector set and the historical image vector set to obtain the historical data dispersion. Quantifying historical data dispersion by calculating the sum of the variances of the historical text and image vector sets can effectively identify the stability of the information source, for example, distinguishing between stable camera footage and complaint records with large emotional fluctuations, thereby accurately assessing the objectivity and credibility of the information source and filtering out unstable or low-quality data interference at the source.

[0010] Furthermore, the method for calculating the historical reputation factor of a node-corresponding information source includes: obtaining all dynamic event nodes reported by the corresponding information source within a preset time window, and counting the total number of successfully verified events among all event nodes semantically related to the user's query statement, which is used as the historical reputation factor of the corresponding information source. By statistically analyzing the proportion of historically uploaded events of the information source that have been successfully verified by other modalities, such as monitoring and verification sensors, the historical reputation factor is dynamically generated, thereby establishing a trust assessment mechanism based on factual feedback, ensuring that search results point more towards data sources that have historically demonstrated honesty, accuracy, and a high verification rate.

[0011] Furthermore, the dynamic conductivity is also related to a local topological scaling factor. This local topological scaling factor is obtained by acquiring all first-order neighbor nodes of the current node and calculating the sum of the mean and standard deviation of the cosine similarity between the current node and each neighbor node's relational terms and the user's query statement. By introducing a local topological scaling factor, the threshold for energy propagation is dynamically adjusted using the mean and standard deviation of the neighbor node similarity. This allows the algorithm to lower the threshold in weakly matching regions to prevent missing weak latent clues, and raise the threshold in strongly matching regions to prevent instantaneous full energy propagation. This achieves adaptive adjustment of the local density structure of the retrieval map during the retrieval process, preserving key differentiated information.

[0012] Furthermore, calculating the dynamic conductivity between the current node and its neighboring nodes also includes: the dynamic conductivity is negatively correlated with the spatiotemporal impedance value calculated from the difference between the timestamps of the current node and its neighboring nodes and the expected time scale. Introducing a spatiotemporal impedance value related to the expected time scale suppresses unreasonable energy flow based on the deviation of the time difference between nodes. This ensures that the retrieval path strictly conforms to the chronological order of events in the physical world and a reasonable causal time span, effectively avoiding erroneous association retrieval caused by temporal logic errors.

[0013] Furthermore, the method for determining the expected time scale includes: extracting intent keywords from the user's query statement, and selecting the corresponding time length as the expected time scale based on the matching relationship between the intent keywords and preset rules. By intelligently determining the expected time scale through extracting intent keywords from the user's query statement, it is possible to automatically identify whether the user is looking for an instantaneous cause, a complete process, or a long-term trend, thereby providing a dynamic benchmark for calculating spatiotemporal impedance and enabling the retrieval strategy to flexibly adapt to different types of business query needs.

[0014] Furthermore, the preset rules include: if the intent keyword indicates instantaneous causality, the time scale is 300 seconds; if the intent keyword indicates process tracing, the time scale is 86,400 seconds; if the intent keyword indicates a long-term trend, the time scale is 30 days.

[0015] In a second aspect, a knowledge graph-based cross-modal property information retrieval system is also provided, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the knowledge graph-based cross-modal property information retrieval method described above.

[0016] Beneficial effects:

[0017] (i) By constructing a multimodal attribute map and calculating the source entropy confidence, and by comprehensively considering historical dispersion, text-image self-consistency and reputation factor, the objectivity of different sources (such as IoT and manual records) is quantitatively evaluated, and subjective noise is effectively suppressed by following the principle of low entropy and high weight.

[0018] (ii) A dynamic conductivity and spatiotemporal impedance model based on intent time-domain mapping is introduced to dynamically adjust the time scale constraint according to the user's query intent, blocking the propagation path of semantically related but temporally logically disordered; finally, a comprehensive score is generated by combining the energy propagation mechanism, realizing accurate retrieval that takes into account data credibility, semantic consistency and spatiotemporal logic, and significantly improving the accuracy and efficiency of implicit evidence mining in complex property security scenarios. Attached Figure Description

[0019] Figure 1 is a flowchart schematically illustrating a retrieval method according to an embodiment of the present invention;

[0020] Figure 2 is a schematic diagram illustrating the final filtering effect generated by the search results according to an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] As shown in Figure 1, S101: Construct a multimodal attribute map and perform feature vectorization.

[0023] In this embodiment, the first step is to address the issues of fragmented and unstructured property security data. The system collects data from various sources, including resident files and equipment ledgers in the property management system, maintenance records from the work order system, sensor logs from the Internet of Things (IoT) platform, and surveillance images captured by the security system.

[0024] Specifically, the collected spatial (e.g., building, room) and object (e.g., equipment, personnel) instances are transformed into static nodes in the graph, forming the basic framework of the graph. Next, dynamic data occurring over time is transformed into dynamic event nodes, such as maintenance, alarms, and snapshots. Based on business logic, such as a smoke alarm occurring at a specific smoke detector, the system creates connecting edges in the graph, linking dynamic event nodes to corresponding static nodes, thus forming a heterogeneous graph connecting people, rooms, events, and objects.

[0025] After the graph is constructed, key attributes are pre-embedded. The occurrence time of each dynamic event log is uniformly converted to a Unix timestamp for subsequent time difference calculations. Simultaneously, based on the objectivity of the data source, each node is tagged with a source label: IoT devices: pure machine data, highest reliability, tagged as IoT; Monitoring systems: objective records but with blind spots, tagged as Video; Property manager data entry: trained professional human data, tagged as Staff; Owner submissions: highly subjective, tagged as User.

[0026] To enable computer-computable mathematical space operations, multimodal features need to be vectorized. Specifically, a deep learning model is introduced to transform text and image information into numerical vectors, including: text vectorization: inputting the text information from nodes into a pre-trained BERT model, for example, text information describing severe water seepage in an exterior wall as a repair request. The BERT model employs a multi-layer Transformer encoder structure, capable of capturing the deep semantics of text, with an output dimension of... The text feature vector is dimensional. Image vectorization: Image attachments from the nodes are input into the pre-trained ResNet-50 model, for example, the image attachment is a snapshot taken on-site. This model extracts high-level visual features of the image through a residual network structure, and the output dimension is dimensional. Image feature vectors of dimension 1. Missing modality imputation: For nodes that are partially missing a certain modality, such as sensor logs with only text records and no images, the system automatically generates a zero-based vector of the corresponding dimension to imput the missing modality, ensuring that all nodes in the graph have a uniform dimensionality in mathematical representation.

[0027] S102: Time-domain scale mapping of user instructions.

[0028] In one embodiment, to address the problem of spatiotemporal logical inconsistencies in traditional retrieval, the system needs to parse the implicit time scale in the user's query command. Through dual semantic parsing, the starting point and radius of the retrieval are determined.

[0029] The first layer of analysis involves explicit time-based localization. The system identifies explicit time terms in the command, such as "yesterday" or "October 1, 1949," and combines these with entity nouns in the command to locate the specific source entity node in the graph. For example, for the command "What was the cause of the fire in Building 3 yesterday?", the system locates the fire event node with a timestamp of yesterday.

[0030] The second layer of parsing involves temporal mapping of intent. The system extracts implicit intent words from the command, matches them with preset mapping rules, and defines the expected time scale. The specific rules are as follows: Short time sequence: Keywords include "cause," "trigger," "instantaneous," etc., reflecting causal relationships that occur immediately in the physical world, and are set... Seconds. Mid-time sequence: Keywords include "process," "investigation," and "entire process," needing to cover the complete activity cycle before and after the incident. Seconds, or 24 hours. Long-term time series: Keywords include "historical," "trend," "long-term," and "aging," reflecting the aging process of equipment or the accumulation of potential problems. Seconds, or 30 days. If no keywords are matched, the system's default experience value will be used. Second.

[0031] S103: Calculate the source entropy confidence level.

[0032] In one embodiment, to reflect the value judgment that objective data is superior to subjective description, this invention constructs a modal capacity calculation model based on source entropy. This model is used to evaluate the storage capacity of nodes, i.e., the benchmark of weights, and follows the principle of low entropy and high weight, that is, amplifying the capacity of data with high certainty (such as IoT) and suppressing the capacity of data with low certainty (such as subjective user complaints).

[0033] First, due to the output of BERT dimensional vectors and the output of ResNet-50 Since the dimensions of text vectors differ, similarity cannot be directly calculated. Therefore, two independent fully connected layers are used as transformers to compress the text vectors to [a specific size / dimensionality]. The image vector is also compressed to dimensionality. dimension.

[0034] Calculate the source entropy confidence level based on the compressed vector. The calculation formula is: In the formula, Represents a node The source entropy confidence level; It is a node Cosine similarity between internal text feature vectors and image feature vectors; This is a quality sensitivity factor used to control the system's sensitivity to image-text matching; in this embodiment, it is set to a value of [value missing]. ; It is a historical dispersion index used to measure the dispersion of nodes. Corresponding source The confidence level within a past time window, for example, a time window of 30 days; It is the global average variance, used for normalization; It is the historical credibility factor, representing the source of information. The probability that historical data is corroborated by other modalities (such as surveillance); It is a very small positive number, such as To prevent the denominator from being 0; It is a natural constant.

[0035] It should be noted that for a given type of information source, the four types of information sources previously identified are labeled IoT, Video, Staff, and User. In the past time window, these sources will generate not only text records but also image records. When calculating their variance, they need to be calculated separately. In this example, the time window is the past 30 days.

[0036] For the information source corresponding to node i We retrieve all records within a past time window, extract text and image sets separately, and feed them into pre-trained BERT and ResNet-50 models to obtain text vector sets and image vector sets. We then calculate the variance of each set to obtain the text semantic variance. and image visual variance Among them, for the information source corresponding to node i Its historical dispersion index The calculation is as follows: In the formula, This represents the text semantic variance of the information source within a past time window. Let V be the image visual variance. As the formula shows, if the information sources are highly consistent, such as the image from a fixed camera with minimal variation, ... Approaching 0 makes the denominator smaller and the logarithmic term larger, ultimately... A larger variance indicates higher confidence; conversely, if the information source is disorganized, such as emotional complaints, the variance will be large, leading to... It becomes smaller. Meanwhile, if the image-text matching degree... The higher, The larger the value, the more consistent the content of that node is.

[0037] Calculate the global average variance Similarly, the historical dispersion index is calculated for each of the four labels, and the average value is taken to obtain the global average variance. .

[0038] For historical reputation factors : Get the source corresponding to node i Within the past time window, all dynamic event nodes reported are filtered to identify all events related to the keywords in the user's query, and the total number of successfully verified events is counted.

[0039] S104: Calculate dynamic conductivity.

[0040] In one embodiment, to introduce spatiotemporal logic into the graph search, this embodiment abandons static graph traversal and constructs a dynamic energy propagation model. The core lies in calculating the dynamic conductivity between nodes, which controls the flow of energy along logically rigorous and spatiotemporally consistent paths.

[0041] Let the user query be For nodes Its neighboring nodes Relationship words Calculate dynamic conductivity The calculation formula is: In the formula, The user's query statement is The dynamic conductivity between node i and its neighbor node j; It is a query statement With relational words Cosine similarity; It is a local topology signal scaling factor; The user's query statement is The spatiotemporal impedance value is calculated based on the degree of deviation of the time difference between node i and its neighbor node j relative to the expected time scale. This is the sigmoid function.

[0042] For local topology signal scaling factor When the algorithm reaches node i, the system obtains all its first-order neighbor nodes, extracts all relation words between node i and these neighbor nodes, and calculates them respectively. Finally, the sum of the mean and standard deviation of all calculated results (using the subtraction truncation threshold principle in binarization) is used as the local topological signal scaling factor. This method achieves the following: in weakly matching regions, the mean is low... Smaller A larger mean value improves the algorithm's sensitivity to weak latent clues and prevents missed detections; in strong matching regions, the mean value is high. Larger Smaller size prevents energy from flowing through an entire area instantly, forcing the system to retain minute differences and selectively propagate energy.

[0043] The formula for calculating the spatiotemporal impedance value is as follows: In the formula, The user's query statement is At that time, the spatiotemporal impedance value is calculated based on the degree of deviation of the time difference between node i and its neighbor node j relative to the expected time scale. It is a user query statement With relational words Cosine similarity; It is a user query statement The expected timescale is obtained by step S102; It is the difference between the timestamps of node i and its neighbor node j.

[0044] Analysis of the formula shows that: The item represents the user's query statement Q and related terms. The degree of similarity between the nodes indicates that the higher the degree of similarity, the more reliable the path to that neighboring node is. The term is determined by calculating the difference between the timestamps of node i and its neighbor node j to determine whether it falls within a reasonable expected time scale. Within a certain range, the greater the deviation, the lower the reliability; but if The tolerance level is relatively high, and time deviations can be tolerated to a certain extent.

[0045] It should be noted that nodes i and j may be dynamic event nodes, such as fire, alarm, or repair requests, which have specific timestamps; or they may be static entity nodes, such as Building 3, Room 101, or smoke detector 01, which are spatial or physical objects without specific timestamps. If either node i or node j is a static entity node, the spatiotemporal impedance value... The conductivity is 0, and it depends only on semantic similarity; if both nodes are dynamic event nodes, the calculation proceeds normally. That's all;

[0046] S105: Energy propagation and retrieval result generation.

[0047] For node i, it has several first-order neighbor nodes. For the j-th neighbor node, the energy it can receive from its parent node i, combined with its own energy storage capacity, yields a final score. The specific calculation formula is as follows: In the formula, It is the source entropy confidence of node j; It is the final score of the previous node i, and the initial node score is preset to 1; The user's query statement is The dynamic conductivity between node i and its neighbor node j.

[0048] It should be noted that if the user's query Q involves nodes i and k, and the j-th neighbor of node i is also a neighbor of node k, then both nodes i and k need to be considered to obtain the final score for node j. The specific calculation formula is as follows: ; It is a node The final score; The user's query statement is The dynamic conductivity between node x and its neighbor node j; This indicates that if there are multiple predecessor nodes (e.g.) and If a node is connected to node j, then the energy transmitted between them is summed up, meaning the propagation energy needs to be accumulated. It can be seen that through this layer-by-layer scoring framework, nodes highly relevant to the user's query Q can be clearly selected.

[0049] Although the final score of each node has been calculated, simply selecting the path of the node with the highest final score as the final query result would likely overlook the real core evidence. Therefore, by combining the calculation logic of potential energy conduction field in physics, the initial query node is regarded as a high-potential energy source. The energy will be conducted layer by layer along the graph network to obtain the final score of each path node. The results are sorted in descending order of score, and at most 20 paths are selected as the final search results.

[0050] As shown in Figure 2, this figure visually illustrates the final filtering effect of energy propagation and retrieval results generated in step S105 of the scheme: the system regards the user query Q as a high-potential energy source, and calculates the energy conduction efficiency based on the source confidence in step S103 and the dynamic conductivity in step S104. In the figure, the red high-scoring path (such as node A1 with a flow score of 0.95) represents smooth energy flow, filtering out the core evidence chain that is logically rigorous, spatiotemporally consistent, and has a reliable source; while the gray low-scoring path (such as node C1 with a flow score of 0.31) indicates energy attenuation, reflecting the system's ability to automatically suppress subjective noise and spatiotemporally disordered information, and accurately lock in the optimal retrieval result.

[0051] This invention also provides a knowledge graph-based cross-modal property information retrieval system. The system includes a processor and a memory, the memory storing computer program instructions. When the processor executes the computer program instructions, it implements the knowledge graph-based cross-modal property information retrieval method according to the first aspect of this invention.

[0052] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0053] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0054] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A cross-modal retrieval method for property information based on knowledge graphs, characterized in that, include: A multimodal attribute graph containing static nodes and dynamic event nodes is constructed. Each node in the graph is assigned a source label according to the data source, and the text and image information in the graph are processed by feature vectorization. Calculate the source entropy confidence of each node. The source entropy confidence is negatively correlated with the historical data dispersion of the corresponding source within a preset time window, positively correlated with the cosine similarity between the text feature vector and the image feature vector within the node, and positively correlated with the historical reputation factor of the corresponding source. Calculate the dynamic conductivity between the current node and its neighboring nodes. The dynamic conductivity is positively correlated with the cosine similarity between the user query and the relational words between the current node and its neighboring nodes. The final score of each node in the graph is calculated. The final score of each node is positively correlated with the source entropy confidence of the node, and also positively correlated with the final score of its parent node and the dynamic conductivity between the parent node and the node. The nodes are sorted according to the final scores, and a preset number of nodes are output as search results. The formula for calculating source entropy confidence is: ； Represents a node The source entropy confidence level; It is a node Cosine similarity between internal text feature vectors and image feature vectors; Quality-sensitive factor; It is a historical dispersion index used to measure the dispersion of nodes. Corresponding source Confidence level over the past time window; It is the global average variance, used for normalization; It is the historical credibility factor, representing the source of information. The probability that historical data is successfully corroborated by other modalities; It is a very small positive number; It is a natural constant; The method for calculating the historical reputation factor of a node's corresponding information source includes: obtaining all dynamic event nodes reported by the node's corresponding information source within a preset time window; counting the total number of successfully corroborated events among all event nodes semantically related to the user's query statement, and using this count as the historical reputation factor of the corresponding information source; the formula for calculating dynamic conductivity is as follows: ； The user's query statement is The dynamic conductivity between node i and its neighbor node j; It is a query statement With relational words Cosine similarity; It is a local topology signal scaling factor; The user's query statement is The spatiotemporal impedance value is calculated based on the degree of deviation of the time difference between node i and its neighbor node j relative to the expected time scale. The sigmoid function is used; the final score is calculated using the following formula: ； It is the source entropy confidence of node j; It is the final score of the previous node i, and the initial node score is preset to 1.

2. The cross-modal retrieval method for property information based on knowledge graphs according to claim 1, characterized in that, The text and image information in the graph are processed into feature vectors, including: extracting text feature vectors using a pre-trained BERT model and extracting image feature vectors using a pre-trained ResNet-50 model.

3. The cross-modal retrieval method for property information based on knowledge graphs according to claim 1, characterized in that, Before calculating the cosine similarity between the text feature vector and the image feature vector within the node, the algorithm further includes: using fully connected layers to compress the dimensions of the text feature vector and the image feature vector respectively, in order to unify their feature dimensions.

4. The cross-modal retrieval method for property information based on knowledge graphs according to claim 1, characterized in that, The method for calculating the dispersion of historical data includes: acquiring all historical records of the corresponding information source within a preset time window, extracting the text information and image information from them respectively and performing vectorization processing to obtain a set of historical text vectors and a set of historical image vectors; calculating the sum of the variance of the set of historical text vectors and the variance of the set of historical image vectors to obtain the dispersion of the historical data.

5. The cross-modal retrieval method for property information based on knowledge graphs according to claim 1, characterized in that, The dynamic conductivity is also related to the local topology scaling factor, which is obtained by: obtaining all first-order neighbor nodes of the current node, and calculating the sum of the mean and standard deviation of the cosine similarity between the relation words and user query statements between the current node and each neighbor node.

6. The cross-modal retrieval method for property information based on knowledge graphs according to claim 1, characterized in that, Calculating the dynamic conductivity between the current node and its neighboring nodes further includes: the dynamic conductivity is also negatively correlated with the spatiotemporal impedance value calculated from the difference between the timestamps of the current node and its neighboring nodes and the expected time scale.

7. The cross-modal retrieval method for property information based on knowledge graphs according to claim 6, characterized in that, The method for determining the expected time scale includes: extracting intent keywords from the user's query statement, and selecting the corresponding time length as the expected time scale based on the matching relationship between the intent keywords and preset rules.

8. The cross-modal retrieval method for property information based on knowledge graphs according to claim 7, characterized in that, The preset rules include: if the intent keyword indicates instantaneous causality, the time scale is 300 seconds; if the intent keyword indicates process tracing, the time scale is 86,400 seconds; if the intent keyword indicates a long-term trend, the time scale is 30 days.

9. A cross-modal property information retrieval system based on knowledge graphs, comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to implement the knowledge graph-based cross-modal property information retrieval method as described in any one of claims 1-8.

Images