Training method of target state inference model and target state inference method

Abstract

This application provides a training method and a target state reasoning method for a target state reasoning model. The method includes obtaining a reasoning training set; for each temporal knowledge graph sample, processing the temporal knowledge graph sample using a structural encoder to capture the semantic dependencies between different target entities and obtain semantic relationship features corresponding to different target entities; generating a first similarity matrix and a second similarity matrix according to a preset similarity function and multiple semantic relationship features; processing the first similarity matrix and the second similarity matrix using a target-gated recurrent network to obtain the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step; and iteratively adjusting the model parameters of the initial reasoning model based on the loss result calculated according to the predicted probability to obtain a trained target state reasoning model.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a training method for a target state reasoning model and a target state reasoning method. Background Technology

[0002] Target state reasoning refers to the process of judging and optimizing the position and interrelationship of each target by comprehensively analyzing the target's movement trajectory and different target states during the target's movement.

[0003] For example, in monitoring the position of merchant ships (targets) in a fleet, it is possible to predict the probability that merchant ship A (target) in the fleet will travel to region A (state information) (entity relationship) at the next moment.

[0004] However, due to the complexity of target movement, the reasoning methods of related technologies are difficult to accurately predict the state of the target, and the predicted results differ significantly from the actual state of the target. Summary of the Invention

[0005] In view of this, this application provides a training method for a target state reasoning model, a target state reasoning method, a training device for a target state reasoning model, a target state reasoning device, an electronic device, a computer-readable storage medium, and a computer program product.

[0006] One aspect of this application provides a method for training a target state reasoning model, comprising:

[0007] Obtain a reasoning training set, wherein the reasoning training set includes multiple temporal knowledge graph samples, which are generated based on a target entity set and a set of entity relationships between different target entities in the target entity set, and the different target entities include targets and their state information;

[0008] For each of the above-mentioned temporal knowledge graph samples, a structural encoder is used to process the temporal knowledge graph samples in order to capture the semantic dependencies between different target entities and obtain the semantic relationship features corresponding to different target entities.

[0009] Based on the preset similarity function and multiple semantic relationship features mentioned above, a first similarity matrix and a second similarity matrix are generated respectively. The preset similarity function includes multiple similarity adjustment parameters. The first similarity matrix represents the potential association between two target entities that do not have a connection relationship in the temporal knowledge graph samples within the same time period. The second similarity matrix represents the potential association between different target entities within different time periods.

[0010] By using a target-gated recurrent network to process the first similarity matrix and the second similarity matrix, the predicted probability of each target in the time-series knowledge graph sample being in a different state in the next time step is obtained.

[0011] The model parameters of the initial inference model are iteratively adjusted based on the loss result calculated according to the above predicted probability to obtain a trained target state inference model, wherein the above initial inference model is generated based on the above structural encoder, the above preset similarity function and the above target gated recurrent network.

[0012] Another aspect of this application provides a target state reasoning method, comprising:

[0013] Obtain the dataset to be inferred, wherein the dataset to be inferred includes team information, trajectory information and text description information related to at least one target to be inferred;

[0014] The above dataset to be reasoned is processed by extracting entities and relations to obtain a reasoning knowledge graph;

[0015] By inputting the above reasoning knowledge graph into the target state reasoning model, the reasoning probabilities of the target to be reasoned being in different reasoning states at the next moment are obtained.

[0016] Another aspect of this application provides a training apparatus for a target state reasoning model, comprising:

[0017] The first acquisition module is used to acquire a reasoning training set, wherein the reasoning training set includes multiple temporal knowledge graph samples, which are generated based on a target entity set and a set of entity relationships between different target entities in the target entity set. The different target entities include targets and the state information of the targets.

[0018] The encoding module is used to process each of the above-mentioned temporal knowledge graph samples using a structure encoder to capture the semantic dependencies between different target entities and obtain the semantic relationship features corresponding to different target entities.

[0019] The generation module is used to generate a first similarity matrix and a second similarity matrix according to a preset similarity function and multiple semantic relationship features mentioned above. The preset similarity function includes multiple similarity adjustment parameters. The first similarity matrix represents the potential association between two target entities that do not have a connection relationship in the temporal knowledge graph samples within the same time period. The second similarity matrix represents the potential association between different target entities within different time periods.

[0020] The module is used to process the first similarity matrix and the second similarity matrix using a target-gated recurrent network to obtain the predicted probability of each target in the time-series knowledge graph sample being in a different state at the next time step.

[0021] The adjustment module is used to iteratively adjust the model parameters of the initial inference model based on the loss result calculated according to the above-mentioned predicted probability, so as to obtain the trained target state inference model, wherein the above-mentioned initial inference model is generated based on the above-mentioned structural encoder, the above-mentioned preset similarity function and the above-mentioned target gated recurrent network.

[0022] Another aspect of this application provides a target state reasoning apparatus, comprising:

[0023] The second acquisition module is used to acquire the dataset to be inferred, wherein the dataset to be inferred includes team information, trajectory information and text description information related to at least one target to be inferred;

[0024] The extraction module is used to extract entities and relations from the above dataset to be reasoned, and obtain a reasoning knowledge graph.

[0025] The reasoning module is used to input the above reasoning knowledge graph into the target state reasoning model to obtain the reasoning probability of the target to be reasoned in different reasoning states in the next moment.

[0026] Another aspect of this application provides an electronic device comprising:

[0027] One or more processors;

[0028] Memory, used to store one or more programs.

[0029] When the one or more programs are executed by the one or more processors, the one or more processors implement the method described above.

[0030] Another aspect of this application provides a computer-readable storage medium storing computer-executable instructions that, when executed, are used to implement the method described above.

[0031] Another aspect of this application provides a computer program product comprising computer-executable instructions which, when executed, are used to implement the method described above.

[0032] According to embodiments of this application, a structural encoder is used to process temporal knowledge graph samples to capture semantic dependencies between different target entities, thereby obtaining semantic relationship features corresponding to different target entities. Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix of potential relationships within a given time period and a second similarity matrix of potential relationships across different time periods are generated. A target-gated recurrent network is used to process the first and second similarity matrices to obtain the predicted probability of the target being in different states at the next time step. The model parameters of the initial inference model are iteratively adjusted based on the loss calculated according to the predicted probabilities to obtain a trained target state inference model. Since the training phase of the model fully considers the two similarity matrices representing potential relationships between target entities within a given time period and across different time periods, and the first similarity matrix describes the potential relationships between two target entities that do not have a connection relationship in the temporal knowledge graph samples, while the second similarity matrix describes the potential relationships between different target entities across different time periods, the model can learn the complex interactive relationships between different target entities by processing these two similarity matrices, thereby improving the accuracy of the model's prediction of the target's state. Attached Figure Description

[0033] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0034] Figure 1 An exemplary system architecture for training a target state reasoning model and a target state reasoning method that can be applied according to embodiments of this application is shown;

[0035] Figure 2 A flowchart illustrating a training method for a target state reasoning model according to an embodiment of this application is shown;

[0036] Figure 3 A flowchart illustrating the generation of a potential relationship diagram according to an embodiment of this application is shown;

[0037] Figure 4 A flowchart illustrating the generation of predicted probabilities according to an embodiment of this application is shown;

[0038] Figure 5 A flowchart of a target state reasoning method according to an embodiment of this application is shown;

[0039] Figure 6 A block diagram of a training apparatus for a target state reasoning model according to an embodiment of the present application is shown;

[0040] Figure 7 A block diagram of a target state reasoning apparatus according to an embodiment of this application is shown;

[0041] Figure 8 A block diagram of an electronic device suitable for implementing the methods described above, according to an embodiment of this application, is shown. Detailed Implementation

[0042] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0043] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0044] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0045] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0046] Embodiments of this application provide a training method and a target state reasoning method for a target state reasoning model. The method includes obtaining a reasoning training set; processing the temporal knowledge graph sample using a structural encoder for each temporal knowledge graph sample to capture the semantic dependencies between different target entities and obtain semantic relationship features corresponding to different target entities; generating a first similarity matrix and a second similarity matrix according to a preset similarity function and multiple semantic relationship features; processing the first similarity matrix and the second similarity matrix using a target-gated recurrent network to obtain the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step; and iteratively adjusting the model parameters of the initial reasoning model based on the loss result calculated according to the predicted probability to obtain a trained target state reasoning model.

[0047] In the embodiments of this application, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0048] Figure 1 An exemplary system architecture for training a target state reasoning model and a target state reasoning method according to embodiments of this application is shown. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this application, in order to help those skilled in the art understand the technical content of this application, but do not mean that the embodiments of this application cannot be used in other devices, systems, environments or scenarios.

[0049] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0050] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social media platform software, etc. (for example only).

[0051] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0052] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0053] It should be noted that the training method and target state reasoning method of the target state reasoning model provided in this application embodiment can generally be executed by server 105. Correspondingly, the training device and target state reasoning device of the target state reasoning model provided in this application embodiment can generally be located in server 105. The training method and target state reasoning method of the target state reasoning model provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the training device and target state reasoning device of the target state reasoning model provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0054] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0055] Figure 2 A flowchart illustrating a training method for a target state reasoning model according to an embodiment of this application is shown.

[0056] like Figure 2 As shown, the training method for the target state reasoning model includes operations S201 to S205.

[0057] In operation S201, the inference training set is obtained. The inference training set includes multiple temporal knowledge graph samples. The temporal knowledge graph samples are generated based on the target entity set and the entity relationship set between different target entities in the target entity set. Different target entities include the target and the target's state information.

[0058] In operation S202, for each temporal knowledge graph sample, a structural encoder is used to process the temporal knowledge graph sample in order to capture the semantic dependencies between different target entities and obtain the semantic relationship features corresponding to different target entities.

[0059] In operation S203, a first similarity matrix and a second similarity matrix are generated based on a preset similarity function and multiple semantic relationship features. The preset similarity function includes multiple similarity adjustment parameters. The first similarity matrix represents the potential association between two target entities that do not have a connection relationship in the temporal knowledge graph samples within the same time period. The second similarity matrix represents the potential association between different target entities within different time periods.

[0060] In operation S204, the first similarity matrix and the second similarity matrix are processed using a target-gated recurrent network to obtain the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step.

[0061] In operation S205, the model parameters of the initial inference model are iteratively adjusted based on the loss result calculated according to the predicted probability to obtain the trained target state inference model. The initial inference model is generated based on the structural encoder, the preset similarity function and the target gated recurrent network.

[0062] According to embodiments of this application, the target may refer to merchant ships in a fleet and the status information between different merchant ships. The status information may refer to the area, location, or task performed by the merchant ship, such as transporting home appliances. The fleet and merchant ships can be replaced with other mobile transportation equipment, such as transport drones or trucks. The entity relationship set includes relationships such as location and execution, which characterize the relationship between the merchant ships and the status information.

[0063] According to an embodiment of this application, for each temporal knowledge graph sample in the inference training set, a structural encoder can be used to process the temporal knowledge graph sample to capture the semantic dependencies between different target entities and obtain the semantic relation features corresponding to different target entities. The structural encoder calculates the semantic relation features using formula (1):

[0064] (1)

[0065] in, The goal In temporal knowledge graph samples The set of neighbors in the middle, Indicates status information, It is the ReLU function. and It is the trainable weight parameter matrix in each layer of the structural encoder. The relationship r between different objectives is represented by the symbol r. Represents team information, initial entity embedding and Static entity embedding and Static entity embedding refers to randomly initialized low-dimensional feature vectors, with the superscript 0 indicating l=0, i.e., the initial entity embedding. After l+1 layers of convolution, in time... The obtained entity is represented as .

[0066] According to embodiments of this application, a first similarity matrix and a second similarity matrix are generated based on a preset similarity function and multiple semantic relationship features, wherein the preset similarity function can be cosine similarity as shown in formula (2):

[0067] (2)

[0068] Where T represents transpose. and These are learnable weight parameters, and x and y are input parameters, which can be semantic relationship features.

[0069] According to the embodiments of this application, by processing multiple semantic relationship features based on formula (2), a first similarity matrix representing the potential association relationship between two target entities that do not have a connection relationship in the temporal knowledge graph sample within the same time period and a second similarity matrix representing the potential association relationship between different target entities within different time periods can be obtained.

[0070] According to an embodiment of this application, the first similarity matrix and the second similarity matrix are processed by a target-gated recurrent network to obtain the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step, such as the probability that a merchant ship is in region A at the next time step or the probability that a merchant ship is performing task A.

[0071] According to embodiments of this application, the model parameters of the initial inference model are iteratively adjusted based on the loss results calculated according to the predicted probabilities to obtain a trained target state inference model. For example, this can be achieved by... And adjust the network parameters of the target gated recurrent network.

[0072] According to embodiments of this application, a structural encoder is used to process temporal knowledge graph samples to capture semantic dependencies between different target entities, thereby obtaining semantic relationship features corresponding to different target entities. Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix of potential relationships within a given time period and a second similarity matrix of potential relationships across different time periods are generated. A target-gated recurrent network is used to process the first and second similarity matrices to obtain the predicted probability of the target being in different states at the next time step. The model parameters of the initial inference model are iteratively adjusted based on the loss calculated according to the predicted probabilities to obtain a trained target state inference model. Since the training phase of the model fully considers the two similarity matrices representing potential relationships between target entities within a given time period and across different time periods, and the first similarity matrix describes the potential relationships between two target entities that do not have a connection relationship in the temporal knowledge graph samples, while the second similarity matrix describes the potential relationships between different target entities across different time periods, the model can learn the complex interactive relationships between different target entities by processing these two similarity matrices, thereby improving the accuracy of the model's prediction of the target's state.

[0073] According to embodiments of this application, obtaining a reasoning training set includes: obtaining a dataset to be processed at any acquisition time, wherein the dataset to be processed includes team information, trajectory information, and text description information related to at least one target; for any dataset to be processed, performing entity extraction processing on the dataset to be processed to obtain multiple target entity sets, wherein the target entity sets include targets and state information of different targets; performing relation extraction processing on the dataset to be processed to obtain multiple entity relation sets, wherein the entity relation sets include multiple entity relation information, and the entity relation information represents the action information of targets and state information; and generating a temporal knowledge graph sample based on the multiple target entity sets and the multiple entity relation sets.

[0074] According to embodiments of this application, fleet information can refer to grouping, for example, a fleet may have merchant ship 1, merchant ship 2 and merchant ship 3, trajectory information can refer to the movement trajectory of the merchant ships, which can be represented by a sequence of coordinate points, and text description information can be text obtained from public channels such as reports or news released by the fleet.

[0075] According to embodiments of this application, for each dataset to be processed, a large model can be used for entity extraction, or other types of named entity extraction models can be used to extract the target entity set. Similarly, an entity relation set can be extracted. For example, if the text description information S is information with N words, S={w1, w2, ..., w...} N}, w NRepresenting words or sentences, named entity recognition requires identifying all target entities in a sentence and grouping them into a target entity set E, E = {e1, e2, …, e}. k For example, target entity 1: "Merchant Ship A", "Merchant Ship B", target entity 2: locations or tasks that may be related to the target, such as "Region X", "Task 1". Simultaneously, the relationships between target entities are extracted and summarized into an entity relationship set. Among them, entity relationship information This could indicate "located in" (e.g., "Merchant Ship A is located in Region X"), entity relationship information. This could represent "execution" (e.g., "Merchant Ship B executes mission 1"), where "located" and "executed" represent action information. Through effective relation extraction, complex interaction networks between targets can be constructed. Furthermore, team information P can also be extracted from this dataset to enhance the potential associations between different targets.

[0076] According to embodiments of this application, a temporal knowledge graph sample is constructed based on multiple sets of target entities and multiple sets of entity relationships.

[0077] According to an embodiment of this application, a temporal knowledge graph sample is generated based on multiple target entity sets and multiple entity relationship sets, including: for each target entity set, generating a target relationship tuple based on the target entity set, the entity relationship set corresponding to the target entity set, the collection time, and the grouping information of each target; and constructing a temporal knowledge graph sample between different targets based on multiple target relationship tuples.

[0078] According to an embodiment of this application, a target relationship tuple is generated based on the target entity set E, the entity relationship set R corresponding to the target entity set, the collection time t, and the grouping information P of each target. Based on multiple target relationship tuples, a temporal knowledge graph sample between different targets is constructed.

[0079] For example, a target relation tuple can be a quintuple. This quintuple represents a relationship (i.e., action information). This relationship occurs in the area belonging to The goal of grouping and status information At time t (i.e., the acquisition time), all five-tuples appearing at time t constitute a temporal knowledge graph sample. . Represents the target entity It appears at time t. (Time series knowledge graph sample) Defined as a sequence of knowledge graphs with different timestamps, i.e. .

[0080] Figure 3 A flowchart illustrating the generation of a potential relationship diagram according to an embodiment of this application is shown.

[0081] According to embodiments of this application, a first similarity matrix and a second similarity matrix are generated based on a preset similarity function and multiple semantic relationship features, including: for any semantic relationship feature, calculating a first similarity between the semantic relationship feature at the same time and the associated relationship features related to the semantic relationship feature based on the preset similarity function; generating a first similarity matrix based on multiple first similarities; calculating a second similarity between the semantic relationship feature at different times and the associated relationship features related to the semantic relationship feature based on the preset similarity function; and generating a second similarity matrix based on multiple second similarities.

[0082] According to embodiments of this application, in mining time-series knowledge graph samples... In the process of identifying potential relationships between target entities, the same target entity appearing at different points in time is considered as different target entities, for example... and Therefore, considering the number of target entities: in, Indicates in The number of target entities in the data. This represents the length of the historical sequence. Assuming no general loss, highly related target entities can be considered to exhibit similarity in the embedding space. Therefore, the similarity between entity embeddings is calculated first. There are various similarity measures to choose from. For example, it can be calculated using the cosine similarity formula in equation (2).

[0083] According to the embodiments of this application, calculations are performed at the same time. Any two target entities that appear but are not connected The first similarity between the semantic relation features can be used to obtain the first similarity matrix. First similarity The calculation is shown in formula (3):

[0084] (3)

[0085] Among them, in all In the case of and In other cases, Set the value to 0. and That is, two target entities The semantic relationship features can be calculated using formula (1).

[0086] According to the embodiments of this application, similarly, the occurrences at different times can be calculated. and Any two target entities at the location The semantic relationship features represent the second similarity between them, thus obtaining the second similarity matrix. The calculation is shown in formula (4):

[0087] (4)

[0088] in, In other cases, The value is set to 0.

[0089] According to an embodiment of this application, see Figure 3 By extracting data from the dataset containing team information, trajectory information, and text description information, a temporal knowledge graph sample is obtained. The temporal knowledge graph sample is then processed using a structural encoder. The output of the structural encoder is then aligned with potential relationships within a time period (see the calculation of the first similarity matrix) and linked with potential relationships across time periods (see the calculation of the second similarity matrix). Finally, a potential relationship graph is generated based on the first similarity matrix and the second similarity matrix.

[0090] According to an embodiment of this application, the above method further includes: performing sparse operations on the first similarity matrix and the second similarity matrix respectively to obtain a sparsed first similarity matrix and a sparsed second similarity matrix, so as to input the sparsed first similarity matrix and the sparsed second similarity matrix into the target gated recurrent network.

[0091] According to embodiments of this application, in order to preserve important potential relationships and reduce noise interference, each first similarity can be... Perform sparse operations based on k-NN (k-Nearest Neighbors), that is: for each target entity, only retain those with the previous... The potential relationship between the fractions. Thus, in time... The final sparse first similarity matrix at the point As shown in formula (5):

[0092] (5)

[0093] in, express The nth row. Each Time was recorded Important potential relationships between target entities within the same timeframe. Instantaneous potential relationships.

[0094] According to an embodiment of this application, similar to the alignment of potential relationships within time, the second similarity matrix is ​​sparsified as shown in formula (6):

[0095] (6)

[0096] Each The latent relationships between entities at different times are recorded. To learn the sparse operations for these two types of latent relationships, [the following were selected respectively]. and The value, such as , .

[0097] Figure 4 A flowchart illustrating the generation of predicted probabilities according to an embodiment of this application is shown.

[0098] According to embodiments of this application, the target gated recurrent network includes a graph attention mechanism network and an initial gated recurrent network.

[0099] According to embodiments of this application, a target-gated recurrent network is used to process a first similarity matrix and a second similarity matrix to obtain the predicted probability of each target in the temporal knowledge graph sample being in different states at the next time step. This includes: generating a latent relationship graph based on the first and second similarity matrices, wherein the latent relationship graph describes the temporal associations of different target entities; processing the latent relationship graph using a graph attention mechanism network to obtain the attention weights between each target entity and its corresponding neighboring entities; and generating a global entity vector for each target entity based on the attention weights corresponding to the target entity, the target entity vector, and the neighboring entity vectors.

[0100] By using an initial gated recurrent network to process multiple global entity vectors, the predicted probabilities of the target being in different states at the next time step are obtained.

[0101] According to embodiments of this application, a potential relationship graph P is first constructed based on the obtained first similarity matrix and second similarity matrix (or the sparsed first similarity matrix and the sparsed second similarity matrix). Specifically, when At that time, construct in the potential relationship graph P and The potential relationship within time. Similarly, when At that time, construct in the potential relationship graph P and The potential relationships across time are considered. Only potential relationships are considered in this potential relationship graph P, omitting the original relationships from the temporal knowledge graph samples. Furthermore, similar to existing relationships, both types of potential relationships can be converted into low-dimensional embedding vectors as learnable parameters to be adjusted synchronously during training.

[0102] According to embodiments of this application, after obtaining the latent relation graph P, information transfer and aggregation operations are performed on the latent relation graph P to capture the semantic dependencies between target entities under the newly learned associations. Specifically, firstly, based on the latent relation r in the latent relation graph P, a graph attention mechanism is used to calculate the attention weight between two adjacent nodes (i.e., the target entity vector and the neighbor entity vector) i and j in graph P. As shown in formula (7):

[0103] (7)

[0104] in, , The initial value is the corresponding entity embedding obtained through the structural encoder, that is, the semantic relation features output by the structural encoder. It is the latent relation embedding between node i (i.e., the target entity vector) and node j (i.e., the neighbor entity vector), which is the first similarity between the two nodes. It is the set of neighbors of node i in P. and These are the learnable weight parameters in each layer. It is the LeakyReLU activation function. This represents the join operation. Subsequently, a more comprehensive representation of each target entity is obtained by aggregating the embeddings of neighbors in the latent relation graph, i.e., the global entity vector shown in Equation (8). :

[0105] (8)

[0106] in, It is the ReLU activation function. and It is a weight parameter matrix. It is the embedding of the potential relationship between node i and node k.

[0107] According to an embodiment of this application, an initial gated recurrent network is used to process multiple global entity vectors to obtain the predicted probability of the target being in different states in the next time step.

[0108] According to an embodiment of this application, see Figure 4By extracting temporal knowledge graph samples from a dataset containing team information, trajectory information, and textual descriptions, a structural encoder is used to process these samples. The encoder's output is then aligned with latent relationships within a given time period (see the calculation of the first similarity matrix) and linked across time periods (see the calculation of the second similarity matrix) to capture dependencies between different target entities. A latent relationship graph is then generated based on the first and second similarity matrices. This graph is input into an initial gated recurrent network to output prediction results, i.e., prediction probabilities.

[0109] According to an embodiment of this application, an initial gated recurrent network is used to process multiple global entity vectors to obtain the predicted probability of the target being in different states at the next time step. This includes: for each target entity, the initial gated recurrent network is used to process the global entity vector corresponding to the target entity to obtain the target time vector; for each neighboring entity, the predicted probability is calculated based on the target time vector of the target entity, the target time vector of the neighboring entity, and the state information between the target entity and the neighboring entity.

[0110] According to an embodiment of this application, the global entity vector corresponding to the target entity is processed using an initial gated recurrent network to obtain the target time vector, and the calculation of the target time vector is shown in formula (9):

[0111] (9)

[0112] in, It is used to control each entity Two types of time information and The gate vector parameter σ(⋅) is used to constrain the value of each element to be between [0,1], and ⊙ represents element-wise multiplication. and The detailed calculation process is described below.

[0113] According to an embodiment of this application, the prediction probability is calculated based on the target time vector of the target entity, the target time vector of the neighboring entities, and the state information between the target entity and the neighboring entities.

[0114] In one specific embodiment, the target entity is first calculated at time t+1. with neighboring entities In relationship The probability of interaction with state information It can be calculated using formula (10):

[0115] (10)

[0116] in, It is the decoder function ConvTransE, the target time vector of the target entity. and the target time vector of neighboring entities It is a time representation that includes both global and local time information.

[0117] According to embodiments of this application, the potential relationship graph includes potential relationship vectors corresponding to different target entities.

[0118] According to an embodiment of this application, the process of using an initial gated recurrent network to process the global entity vector corresponding to the target entity to obtain a target time vector includes: using the initial gated recurrent network to process the global entity vector and potential relation vector corresponding to the target entity to obtain a global time vector; using the initial gated recurrent network to process the global entity vector and semantic relation features corresponding to the target entity to obtain a local time vector; and generating a target time vector based on the global time vector and the local time vector.

[0119] According to embodiments of this application, an initial gated recurrent network is used to process the global entity vector and potential relation vector corresponding to the target entity to obtain a global time vector. As shown in formula (11):

[0120] (11)

[0121] in, It is the latent relation vector of the target entity s at time t. GRU is the initial gated recurrent network. It is the global entity vector at time t, and its initial value is the static embedding of entity s.

[0122] According to an embodiment of this application, an initial gated recurrent network is used to process the global entity vector and semantic relationship features corresponding to the target entity to obtain a local time vector, as shown in formula (12):

[0123] (12)

[0124] in, These are the semantic relationship features of the corresponding target entities obtained through a structural encoder. It is a local time vector at time t, and its initial value is the static embedding of entity s.

[0125] According to an embodiment of this application, based on formula (9), a learnable gating function is used to adaptively integrate the global time vector and the local time vector into a unified time representation, thereby obtaining the target time vector.

[0126] According to embodiments of this application, by integrating global and local time vectors, the target entity can be focused on in the time dimension, enabling the model to more flexibly reflect the state of the target entity at different points in time and adapt to the dynamic changes in trajectory.

[0127] According to an embodiment of this application, the model parameters of the initial inference model are iteratively adjusted based on the loss result calculated according to the predicted probability to obtain a trained target state inference model, including: processing the predicted probability based on minimizing the negative log-likelihood function to obtain the loss result; and iteratively adjusting the model parameters of the initial inference model according to the loss result to obtain the target state inference model.

[0128] According to embodiments of this application, the predicted probability is processed based on minimizing the negative log-likelihood function to obtain the loss result. As shown in formula (13):

[0129] (13)

[0130] in, This refers to the duration of the time.

[0131] According to embodiments of this application, the target includes at least one of any type of ship, vehicle, and drone, the state information is used to describe the target's environment or purpose, the entity relationship information in the entity relationship set includes at least one of execution, location, being, going to, arriving, etc., and the temporal knowledge graph sample is used to describe the correlation between the target, the state information, and the entity relationship information.

[0132] Figure 5 A flowchart of a target state reasoning method according to an embodiment of this application is shown.

[0133] like Figure 5 As shown, the target state reasoning method includes operations S501 to S503.

[0134] In operation S501, a dataset to be inferred is obtained, wherein the dataset to be inferred includes team information, trajectory information and text description information related to at least one target to be inferred.

[0135] In operation S502, entity and relation extraction processing is performed on the dataset to be reasoned to obtain a reasoning knowledge graph.

[0136] In operation S503, the reasoning knowledge graph is input into the target state reasoning model to obtain the reasoning probability of the target being in different reasoning states in the next moment.

[0137] According to the embodiments of this application, the inference dataset can be information collected at the current moment. The method of extracting entities and relations from it to generate an inference knowledge graph is the same as the method of generating temporal knowledge graph samples during the training phase, and will not be described in detail here.

[0138] According to an embodiment of this application, the reasoning knowledge graph is input into a trained target state reasoning model to obtain the reasoning probability of the target to be reasoned about being in different reasoning states in the next moment.

[0139] According to embodiments of this application, a structural encoder is used to process temporal knowledge graph samples to capture semantic dependencies between different target entities, thereby obtaining semantic relationship features corresponding to different target entities. Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix of potential associations within a given time period and a second similarity matrix of potential associations across time periods are generated. A target-gated recurrent network is used to process the first and second similarity matrices to obtain the predicted probabilities of the target being in different states at the next time step. Based on the loss calculated according to the predicted probabilities, the model parameters of the initial inference model are iteratively adjusted to obtain a trained target state inference model. Since the two similarity matrices, representing potential associations between target entities within a given time period and across time periods, are fully considered during the model training phase, the model can learn complex interactions between different target entities, thereby improving the accuracy of the model's prediction of the target's state.

[0140] Figure 6 A block diagram of a training apparatus for a target state reasoning model according to an embodiment of the present application is shown.

[0141] like Figure 6 As shown, the training device 600 for the target state reasoning model includes a first acquisition module 610, an encoding module 620, a generation module 630, an acquisition module 640, and an adjustment module 650.

[0142] The first acquisition module 610 is used to acquire the inference training set, wherein the inference training set includes multiple temporal knowledge graph samples. The temporal knowledge graph samples are generated based on the target entity set and the entity relationship set between different target entities in the target entity set. Different target entities include targets and target state information.

[0143] The encoding module 620 is used to process the temporal knowledge graph sample with a structure encoder for each temporal knowledge graph sample in order to capture the semantic dependencies between different target entities and obtain the semantic relationship features corresponding to different target entities.

[0144] The generation module 630 is used to generate a first similarity matrix and a second similarity matrix according to a preset similarity function and multiple semantic relationship features. The preset similarity function includes multiple similarity adjustment parameters. The first similarity matrix represents the potential association between two target entities that do not have a connection relationship in the temporal knowledge graph samples within the same time period. The second similarity matrix represents the potential association between different target entities within different time periods.

[0145] The module 640 is used to process the first similarity matrix and the second similarity matrix using the target-gated recurrent network to obtain the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step.

[0146] The adjustment module 650 is used to iteratively adjust the model parameters of the initial inference model based on the loss result calculated according to the predicted probability, so as to obtain the trained target state inference model, wherein the initial inference model is generated based on the structural encoder, the preset similarity function and the target gated recurrent network.

[0147] According to embodiments of this application, a structural encoder is used to process temporal knowledge graph samples to capture semantic dependencies between different target entities, thereby obtaining semantic relationship features corresponding to different target entities. Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix of potential associations within a given time period and a second similarity matrix of potential associations across time periods are generated. A target-gated recurrent network is used to process the first and second similarity matrices to obtain the predicted probabilities of the target being in different states at the next time step. Based on the loss calculated according to the predicted probabilities, the model parameters of the initial inference model are iteratively adjusted to obtain a trained target state inference model. Since the two similarity matrices, representing potential associations between target entities within a given time period and across time periods, are fully considered during the model training phase, the model can learn complex interactions between different target entities, thereby improving the accuracy of the model's prediction of the target's state.

[0148] Figure 7 A block diagram of a target state reasoning apparatus according to an embodiment of this application is shown.

[0149] like Figure 7 As shown, the target state reasoning device 700 includes a second acquisition module 710, an extraction module 720, and a reasoning module 730.

[0150] The second acquisition module 710 is used to acquire the dataset to be inferred, wherein the dataset to be inferred includes team information, trajectory information and text description information related to at least one target to be inferred.

[0151] The extraction module 720 is used to extract entities and relations from the dataset to be reasoned, thereby obtaining a reasoning knowledge graph.

[0152] The reasoning module 730 is used to input the reasoning knowledge graph into the target state reasoning model to obtain the reasoning probability of the target being in different reasoning states in the next time step.

[0153] According to embodiments of this application, a structural encoder is used to process temporal knowledge graph samples to capture semantic dependencies between different target entities, thereby obtaining semantic relationship features corresponding to different target entities. Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix of potential associations within a given time period and a second similarity matrix of potential associations across time periods are generated. A target-gated recurrent network is used to process the first and second similarity matrices to obtain the predicted probabilities of the target being in different states at the next time step. Based on the loss calculated according to the predicted probabilities, the model parameters of the initial inference model are iteratively adjusted to obtain a trained target state inference model. Since the two similarity matrices, representing potential associations between target entities within a given time period and across time periods, are fully considered during the model training phase, the model can learn complex interactions between different target entities, thereby improving the accuracy of the model's prediction of the target's state.

[0154] Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or implemented by hardware or firmware in any other reasonable manner by integrating or packaging circuits, or implemented in any one of software, hardware, and firmware, or in a suitable combination of any of these. Alternatively, one or more of the modules, submodules, units, and subunits according to the embodiments of this application can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0155] For example, any and multiple of the first acquisition module 610, encoding module 620, generation module 630, obtaining module 640, adjustment module 650, or the second acquisition module 710, extraction module 720, and reasoning module 730 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least some of the functions of one or more of these modules / units / subunits can be combined with at least some of the functions of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this application, at least one of the first acquisition module 610, encoding module 620, generation module 630, obtaining module 640, adjustment module 650, or the second acquisition module 710, extraction module 720, and inference module 730 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the first acquisition module 610, encoding module 620, generation module 630, obtaining module 640, adjustment module 650, or the second acquisition module 710, extraction module 720, and inference module 730 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0156] It should be noted that the training device and target state reasoning device of the target state reasoning model in the embodiments of this application correspond to the training method and target state reasoning method of the target state reasoning model in the embodiments of this application. For a detailed description of the training device and target state reasoning device of the target state reasoning model, please refer to the training method and target state reasoning method of the target state reasoning model, which will not be repeated here.

[0157] Figure 8 A block diagram of an electronic device suitable for implementing the methods described above, according to an embodiment of this application, is shown. Figure 8 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0158] like Figure 8As shown, an electronic device 800 according to an embodiment of this application includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage portion 808 into a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0159] RAM 803 stores various programs and data required for the operation of electronic device 800. Processor 801, ROM 802, and RAM 803 are interconnected via bus 804. Processor 801 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 802 and / or RAM 803. It should be noted that the programs may also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0160] According to embodiments of this application, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output (I / O) interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0161] According to embodiments of this application, the method flow according to embodiments of this application can be implemented as a computer software program. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by processor 801, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0162] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0163] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0164] For example, according to embodiments of this application, a computer-readable storage medium may include the ROM 802 and / or RAM 803 described above and / or one or more memories other than ROM 802 and RAM 803.

[0165] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this application. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the methods provided in the embodiments of this application.

[0166] When the computer program is executed by the processor 801, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0167] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 809, and / or installed from a removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0168] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0169] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. This application does not depart from its scope, and those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A training method for a target state reasoning model, characterized in that, include: Obtain a reasoning training set, wherein the reasoning training set includes multiple temporal knowledge graph samples, the temporal knowledge graph samples are generated based on a target entity set and a set of entity relationships between different target entities in the target entity set, and the different target entities include targets and the state information of the targets; For each of the time-series knowledge graph samples, a structural encoder is used to process the time-series knowledge graph samples to capture the semantic dependencies between different target entities and obtain the semantic relationship features corresponding to different target entities. Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix and a second similarity matrix are generated respectively. The preset similarity function includes multiple similarity adjustment parameters. The first similarity matrix represents the potential association between two target entities that do not have a connection relationship in the temporal knowledge graph sample within the same time period. The second similarity matrix represents the potential association between different target entities within different time periods. By using a target-gated recurrent network to process the first similarity matrix and the second similarity matrix, the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step is obtained; The model parameters of the initial inference model are iteratively adjusted based on the loss result calculated according to the predicted probability to obtain a trained target state inference model, wherein the initial inference model is generated based on the structural encoder, the preset similarity function and the target gated recurrent network.

2. The method of claim 1, wherein, Obtain the inference training set, including: Obtain the dataset to be processed at any acquisition time, wherein the dataset to be processed includes team information, trajectory information and text description information related to at least one target; For any of the datasets to be processed, entity extraction processing is performed on the datasets to be processed to obtain multiple sets of target entities, wherein the sets of target entities include the targets and the state information of different targets; The dataset to be processed is subjected to relation extraction processing to obtain multiple entity relation sets, wherein the entity relation sets include multiple entity relation information, and the entity relation information represents the action information of the target and the state information; The temporal knowledge graph sample is generated based on the multiple sets of target entities and the multiple sets of entity relationships.

3. The method of claim 2, wherein, Based on multiple sets of target entities and multiple sets of entity relationships, the temporal knowledge graph sample is generated, including: For each target entity set, a target relationship tuple is generated based on the target entity set, the entity relationship set corresponding to the target entity set, the collection time, and the grouping information of each target. Based on multiple target relationship tuples, construct temporal knowledge graph samples between different targets.

4. The method of claim 1, wherein, Based on a preset similarity function and multiple semantic relationship features, a first similarity matrix and a second similarity matrix are generated, including: For any of the semantic relationship features, a first similarity is calculated between the semantic relationship feature and the associated relationship feature related to the semantic relationship feature at the same time based on the preset similarity function; The first similarity matrix is ​​generated based on multiple first similarities; The second similarity between the semantic relationship features and the association relationship features related to the semantic relationship features at different times is calculated based on the preset similarity function. The second similarity matrix is ​​generated based on multiple second similarities; The target includes at least one of any type of ship, vehicle, and drone; the state information is used to describe the environmental or purpose information of the target; any entity relationship information in the entity relationship set includes at least one of executing, located, in, heading to, and arriving; and the temporal knowledge graph sample is used to describe the correlation between the target, the state information, and the entity relationship information.

5. The method of claim 4, wherein, Also includes: The first similarity matrix and the second similarity matrix are subjected to sparse operations to obtain a sparse first similarity matrix and a sparse second similarity matrix, which are then input into the target gated recurrent network.

6. The method of claim 1, wherein, The target gated recurrent network includes a graph attention mechanism network and an initial gated recurrent network; Specifically, the first similarity matrix and the second similarity matrix are processed using a target-gated recurrent network to obtain the predicted probability of each target in the temporal knowledge graph sample being in a different state at the next time step, including: Based on the first similarity matrix and the second similarity matrix, a potential relationship graph is generated, wherein the potential relationship graph is used to describe the temporal association relationship between different target entities; The latent relationship graph is processed using the graph attention mechanism network to obtain the attention weights between each target entity and its corresponding neighbor entities. For each target entity, a global entity vector of the target entity is generated based on the attention weight corresponding to the target entity, the target entity vector, and the neighbor entity vectors. The initial gated recurrent network is used to process multiple global entity vectors to obtain the predicted probability of the target being in different states in the next time step.

7. The method of claim 6, wherein, The initial gated recurrent network is used to process multiple global entity vectors to obtain the predicted probabilities of the target being in different states at the next time step, including: For each target entity, the initial gated recurrent network is used to process the global entity vector corresponding to the target entity to obtain the target time vector; For each neighboring entity, the predicted probability is calculated based on the target time vector of the target entity, the target time vector of the neighboring entity, and the state information between the target entity and the neighboring entity.

8. The method of claim 7, wherein, The potential relationship graph includes potential relationship vectors corresponding to different target entities; Specifically, the initial gated recurrent network is used to process the global entity vector corresponding to the target entity to obtain the target time vector, including: The global time vector is obtained by processing the global entity vector and the potential relation vector corresponding to the target entity using the initial gated recurrent network. The initial gated recurrent network is used to process the global entity vector corresponding to the target entity and the semantic relationship features to obtain the local time vector; The target time vector is generated based on the global time vector and the local time vector.

9. The method according to claim 1, characterized in that, The model parameters of the initial inference model are iteratively adjusted based on the loss result calculated according to the predicted probability to obtain a trained target state inference model, including: The predicted probability is processed by minimizing the negative log-likelihood function to obtain the loss result; The model parameters of the initial inference model are iteratively adjusted based on the loss results to obtain the target state inference model.

10. A target state reasoning method, characterized in that, include: Obtain the dataset to be reasoned, wherein the dataset to be reasoned includes team information, trajectory information and text description information related to at least one target to be reasoned; The dataset to be reasoned is processed by extracting entities and relations to obtain a reasoning knowledge graph; The reasoning knowledge graph is input into the target state reasoning model to obtain the reasoning probability of the target to be reasoned being in different reasoning states in the next moment. The target state reasoning model is trained using the method described in any one of claims 1 to 9.

Images