Training methods for matching models, methods for obtaining matching scores, matching methods and devices.

Abstract

This application provides a method for training a matching model, a method for obtaining matching scores, a matching method, and an apparatus. In this application, the input to the model is event information and department information of eligible processing departments. The model is used to obtain the matching score between the event and the processing department based on the event information and the department information, and then matches the event with the appropriate processing department based on the matching score. Eligible processing departments are not restricted in the model. Subsequently, when it is necessary to update the eligible processing departments online, only the event information and the updated department information need to be input into the model, eliminating the need to retrain the model, resulting in low update cost and high update efficiency.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and apparatus for training a matching model, a method and apparatus for obtaining matching degree, and a matching method and apparatus. Background Technology

[0002] With economic development, the complexity of government departments is constantly increasing, and their management and services are continuously improving, leading to a growing workload for each department. In government settings, citizens can submit requests to various departments through multiple channels (such as hotlines or letters). Government staff receive these requests daily, organize them into events, and then manually determine which department should handle each event based on its description. The events are then manually assigned to that department for processing and feedback. In recent years, with the development of digitalization and intelligentization, there is a need to move from manual to intelligent event assignment, reducing manual operations, lowering labor costs, and indirectly improving the efficiency of social management.

[0003] For example, in one approach, each city or region has its own event classification system, a fixed number of categories, and corresponding historical data. A classification model can then be trained based on this system, the fixed number of categories, and the historical data. Alternatively, a government department can be responsible for one or more types of events. The mapping relationship between event types and government departments can be established based on the department's authority and responsibilities. Thus, when it's necessary to assign an event to a specific government department, the event type can be determined using the classification model, and then the responsible government department can be identified based on the mapping relationship between event types and government departments. The event can then be assigned to that determined department for processing.

[0004] However, statistical analysis revealed the following drawbacks to this approach: once the classification model is trained and deployed using historical data, the number of event categories is fixed, as is the mapping from event type to department. This makes it impossible to support the increasing demands for adding, deleting, and modifying event types, as well as the dynamic adjustment of departmental responsibilities, as societal management becomes more refined. It cannot support the dynamic addition, deletion, and modification of event types, quantities, and their relationships with departments. Therefore, every time the event type, quantity, and departmental relationship need to be added, deleted, or modified, the classification model must be retrained and redeployed, resulting in low efficiency and high cost. Summary of the Invention

[0005] This application illustrates a method for training a matching model, a method for obtaining matching degree, a matching method, and an apparatus.

[0006] In a first aspect, a method for training a matching model is shown, comprising: acquiring at least one first training data and an encoding network to be trained, wherein the first training data includes event information of a first sample event in a first domain; using the at least one first training data to adjust the parameters of the encoding network to be trained, thereby obtaining an encoding network in the first domain; acquiring at least one second training data and creating a matching model including the encoding network in the first domain; the second training data includes event information of a second sample event in the first domain, object information of a sample object in the first domain, and the labeled matching degree between the second sample event in the first domain and the sample object in the first domain; training the parameters in the matching model including the encoding network in the first domain using the at least one second training data until the parameters in the matching model including the encoding network in the first domain converge, thereby obtaining a first matching model; acquiring at least one third training data and creating a matching model including the first matching model; the third training data includes: event information of sample events in a target second domain in multiple second domains cascaded in the first domain, object information of sample objects in the target second domain, and the labeled matching degree between the sample events in the target second domain and the sample objects in the target second domain; using the at least one third training data to adjust the parameters of the matching model including the first matching model, thereby obtaining a second matching model.

[0007] Secondly, a method for obtaining matching degree is shown, comprising: obtaining event information of target events in a target second domain and object information of target objects in the target second domain; obtaining the matching degree between target events and target objects based on the event information of target events in the target second domain, the object information of target objects in the target second domain, and a trained second matching model; wherein, the second matching model is obtained by adjusting the parameters of a matching model including a first matching model using at least one third training data; the third training data includes: event information of sample events in the target second domain in multiple second domains cascaded in the first domain, object information of sample objects in the target second domain, and labeled matching degree between sample events and sample objects in the target second domain; the first matching model is obtained by training the parameters in a matching model including a coding network of the first domain using at least one second training data until the parameters in the matching model including the coding network of the first domain converge; the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and labeled matching degree between second sample events and sample objects in the first domain; the coding network of the first domain is obtained by adjusting the parameters of a coding network to be trained using at least one first training data, the first training data including event information of first sample events in the first domain.

[0008] Thirdly, a matching method is shown, comprising: obtaining event information of a target event, the event information including the title of the target event and the descriptive text of the target event; obtaining object information of multiple objects to be matched, the object information including the name of the object and the classification structure of the object; obtaining keywords for describing the topic of the target event based on the event information of the target event; determining a first object among the multiple objects to be matched whose object information includes the keywords; setting the sorting order of the first object to be matched before the second object among the multiple objects to be matched, excluding the first object; and selecting at least one object that matches the target event among the multiple objects to be matched according to the sorting order among the multiple objects to be matched.

[0009] Fourthly, a training apparatus for a matching model is shown, comprising: a first acquisition module for acquiring at least one first training data and an encoding network to be trained, wherein the first training data includes event information of first sample events in a first domain; a first adjustment module for adjusting the parameters of the encoding network to be trained using at least one first training data to obtain an encoding network in the first domain; a second acquisition module for acquiring at least one second training data and creating a matching model including the encoding network in the first domain; wherein the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and labeled matching degree between the second sample events in the first domain and the sample objects in the first domain; and a training module for using... At least one second training data is used to train the parameters in a matching model that includes the encoding network of the first domain until the parameters in the matching model that includes the encoding network of the first domain converge, thereby obtaining a first matching model; a third acquisition module is used to acquire at least one third training data and create a matching model that includes the first matching model; the third training data includes: event information of sample events of the target second domain in multiple second domains cascaded in the first domain, object information of sample objects of the target second domain, and labeled matching degree between sample events of the target second domain and sample objects of the target second domain; a second adjustment module is used to adjust the parameters of the matching model that includes the first matching model using at least one third training data, thereby obtaining a second matching model.

[0010] Fifthly, an apparatus for obtaining a matching degree is shown, comprising: a fourth acquisition module for acquiring event information of a target event in a target second domain and object information of a target object in the target second domain; and a fifth acquisition module for acquiring a matching degree between a target event and a target object based on the event information of the target event in the target second domain, the object information of the target object in the target second domain, and a trained second matching model; wherein the second matching model is obtained by adjusting the parameters of a matching model including a first matching model using at least one third training data; the third training data includes: event information of sample events in a target second domain in multiple second domains cascaded from a first domain, and object information of sample objects in the target second domain. The first matching model is obtained by training the parameters of a matching model including the encoding network of the first domain with at least one second training data until the parameters of the matching model including the encoding network of the first domain converge. The second training data includes: event information of the second sample event of the first domain, object information of the sample object of the first domain, and the annotation matching degree between the second sample event of the first domain and the sample object of the first domain. The encoding network of the first domain is obtained by adjusting the parameters of the encoding network to be trained with at least one first training data. The first training data includes event information of the first sample event of the first domain.

[0011] A sixth aspect is shown: a matching apparatus comprising: a sixth acquisition module for acquiring event information of a target event, the event information including a title of the target event and a descriptive text of the target event; a seventh acquisition module for acquiring object information of a plurality of objects to be matched, the object information including: the name of the object and the classification structure of the object; an eighth acquisition module for acquiring keywords describing the subject of the target event based on the event information of the target event; a determination module for determining a first object whose object information includes the keywords among the plurality of objects to be matched; a setting module for setting the sorting order of the first object to be placed before a second object other than the first object among the plurality of objects to be matched; and a selection module for selecting at least one object that matches the target event among the plurality of objects to be matched according to the sorting order among the plurality of objects to be matched.

[0012] In a seventh aspect, this application discloses an electronic device comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to perform the methods shown in any of the foregoing aspects.

[0013] Eighthly, this application discloses a non-transitory computer-readable storage medium that, when the instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to perform the methods shown in any of the foregoing aspects.

[0014] Ninthly, this application discloses a computer program product in which, when the instructions in the computer program product are executed by a processor of an electronic device, the electronic device is enabled to perform the methods shown in any of the foregoing aspects.

[0015] Compared with the prior art, this application has the following advantages:

[0016] On the one hand, the traditional matching method involves inputting event information into the model so that the model can output processing departments that match the event. The processing departments that the model can output with matching qualifications are fixed, and it is not possible to update the processing departments with matching qualifications online in the model later. If it is necessary to update the processing departments with matching qualifications, the model needs to be retrained, which is costly and inefficient.

[0017] In this application, the input to the model is event information and object information of eligible objects. The model is used to obtain the matching degree between the event and the object based on the event information and the object information, and then matches the event with the object based on the matching degree. Eligible objects are not restricted in the model. Subsequently, when it is necessary to update eligible objects online, only the event information and the updated object information need to be input into the model, eliminating the need to retrain the model, resulting in low update cost and high update efficiency.

[0018] On the other hand, this application employs a step-by-step learning approach when training the model. For example, it acquires at least one first training data and an encoding network to be trained, wherein the first training data includes event information of first sample events in a first domain; it uses at least one first training data to adjust the parameters of the encoding network to be trained, thereby obtaining an encoding network for the first domain; it acquires at least one second training data and creates a matching model that includes the encoding network for the first domain; the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the label matching degree between the second sample events in the first domain and the sample objects in the first domain; it uses at least one second training data to train the parameters in the matching model that includes the encoding network for the first domain until the parameters in the matching model that includes the encoding network for the first domain converge, thereby obtaining a first matching model; it acquires at least one third training data and creates a matching model that includes the first matching model; the third training data includes: event information of sample events in a target second domain in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the label matching degree between the sample events in the target second domain and the sample objects in the target second domain; it uses at least one third training data to adjust the parameters of the matching model that includes the first matching model, thereby obtaining a second matching model. The first matching model has the ability to match events and objects within the first domain. Furthermore, the second training data used to train the first matching model is readily available; for example, labeled data in the second training data is readily available. This makes training the first matching model easier and increases its matching accuracy. If a matching model is subsequently needed that can match events in a second domain cascaded from the first domain with objects in that second domain cascaded from the first domain, a matching model with higher matching accuracy can be trained using a small amount of second domain training data, reducing the difficulty and time required to collect second domain training data.

[0019] On the other hand, when it is necessary to match a target event with a target object among multiple objects to be matched, the objects can be divided into two categories based on whether the object information contains the keyword of the target event. The matching degree between the target event and the object information containing the keyword of the target event is higher than that between the target event and the object information not containing the keyword of the target event. Thus, the sorting order of the objects containing the keyword of the target event can be set to be before the sorting order of the objects not containing the keyword of the target event, thereby improving the accuracy of sorting and thus improving the matching accuracy between the event and the object. Attached Figure Description

[0020] Figure 1This is a schematic diagram of a scenario illustrated in an exemplary embodiment of this application.

[0021] Figure 2 This is a schematic flowchart illustrating a training method for a matching model according to an exemplary embodiment of this application.

[0022] Figure 3 This is a structural block diagram of a matching model illustrated in an exemplary embodiment of this application.

[0023] Figure 4 This is a structural block diagram of a matching model illustrated in an exemplary embodiment of this application.

[0024] Figure 5 This is a structural block diagram of a matching model illustrated in an exemplary embodiment of this application.

[0025] Figure 6 This is a structural block diagram of a matching model illustrated in an exemplary embodiment of this application.

[0026] Figure 7 This is a structural block diagram of a matching model illustrated in an exemplary embodiment of this application.

[0027] Figure 8 This is a structural block diagram of a matching model illustrated in an exemplary embodiment of this application.

[0028] Figure 9 This is a flowchart illustrating a method for obtaining matching degree according to an exemplary embodiment of this application.

[0029] Figure 10 This is a flowchart illustrating a matching method according to an exemplary embodiment of this application.

[0030] Figure 11 This is a structural block diagram of a matching model training device illustrated in an exemplary embodiment of this application.

[0031] Figure 12 This is a structural block diagram of an apparatus for obtaining matching degree, as illustrated in an exemplary embodiment of this application.

[0032] Figure 13 This is a structural block diagram of a matching device illustrated in an exemplary embodiment of this application.

[0033] Figure 14 This is a structural block diagram of an apparatus illustrated in an exemplary embodiment of this application. Detailed Implementation

[0034] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] Before describing this application, some terms used in this application will be explained.

[0036] Event: This article refers to the demands or suggestions made by citizens to relevant institutions and units through various channels in social management involving government affairs. These are generally recorded in written form or converted into written form for record-keeping.

[0037] Event Classification: These requests and suggestions are classified according to certain rules (so that the events can be transferred to the appropriate government departments for handling based on their type).

[0038] Incident assignment: These requests and suggestions are forwarded to the corresponding government departments for handling based on the incident description and classification.

[0039] Step-by-step learning: In education, step-by-step learning generally refers to the method by which students develop good learning habits step by step. Here, it is used to refer to the step-by-step learning of algorithm models. Specifically, this article refers to the learning process of first obtaining a large pre-trained model based on unsupervised pre-training on general data, then further training a domain-specific pre-trained model based on historical data in a large domain (data constructed by humans or machines, which may contain a lot of erroneous data), and finally fine-tuning a high-quality model based on accurate data labeled by humans in a subdivided domain.

[0040] Departmental authority and responsibility: A department refers to a relevant department, institution or individual established in social management. Departmental authority and responsibility correspond to the scope of jurisdiction (time, space, etc.) and the types of events that these departments, institutions or individuals are responsible for.

[0041] BERT: A language representation model representing bidirectional encoder representations from Transformers. Transformers are deep learning components that, unlike recurrent neural networks (RNNs) which depend on sequence, can process entire sequences in parallel. This allows for the analysis of larger datasets and accelerates model training. Transformers use attention mechanisms to gather contextual information about words and encode it as rich vectors representing that context, thus processing (rather than processing individually) words related to all other words in a sentence simultaneously. Transformer models learn how to derive the meaning of a given word from every other word in a sentence. BERT is essentially a bidirectional feature encoder for Transformers, representing the current input as feature representations (or encoding, or representation), i.e., word vector representations.

[0042] To illustrate this application with an example scenario, see [link to example]. Figure 1 This scenario includes citizens, electronic devices, and processing departments, among others. The electronic devices may include servers. A citizen has a target event and needs to request a processing department to handle it. The citizen can upload the event information to the electronic device. Based on this information, the electronic device can determine the matching processing department from among multiple departments and assign the target event to that department for processing.

[0043] See Figure 2 The diagram illustrates a flowchart of a training method for a matching model according to this application. The method includes:

[0044] In step S101, at least one first training data and an encoding network to be trained are obtained. The first training data includes event information of first sample events in the first domain.

[0045] In one embodiment, the domain can be divided by location region, such as urban neighborhoods and rural domains. The urban domain includes the domains where events can occur within a city. Events occurring in a city can include: events that violate morality, laws, or public order and good morals, or events that have been complained about by citizens, such as wage arrears, illegal construction, and noise pollution. In another embodiment, the domain can also be distinguished by industry, such as education, gaming, sports, scientific research, and tourism. Event information for the first sample event in the first domain can be automatically collected from the internet. Event information includes the event name and a textual description of the event. The encoding network to be trained can include a vectorized encoder. The vectorized encoder can include LSTM (Long Short-Term Memory), GRU (Gate Recurrent Unit), RNN (Recurrent Neural Network), and BERT-based encoders.

[0046] In step S102, the parameters of the encoding network to be trained are adjusted using at least one first training data to obtain the encoding network of the first domain.

[0047] In this application, when adjusting the parameters of the encoding network to be trained using at least one first training data, an unsupervised adjustment method can be used. For specific adjustment methods, please refer to the existing unsupervised adjustment methods, which will not be detailed here.

[0048] In step S103, at least one second training data is acquired and a matching model including a coding network of the first domain is created; the second training data includes: event information of the second sample event of the first domain, object information of the sample object of the first domain, and the annotation matching degree between the second sample event of the first domain and the sample object of the first domain.

[0049] The first sample event in the first domain and the second sample event in the first domain can be the same or different. The event information of the second sample event in the first domain can be automatically collected from the network, etc. The second sample event in the first domain can include historical events that have occurred beforehand and have actually matched objects, etc.

[0050] In one embodiment, the sample object may include a sample processing department, and the object information may include department information of the sample processing department, etc.

[0051] The system can automatically retrieve event information and department information of matched processing departments from historical events. These historical events are used as second sample events in the first domain, their event information as event information of the second sample events in the first domain, and their matched processing departments as sample processing departments in the first domain. In practice, the matching relationship between historical events and their matched processing departments is often done manually. Therefore, the matching degree (e.g., accuracy) between historical events and their matched processing departments is often very high. Thus, a high matching degree can be set as the labeled matching degree between the second sample events and their matched processing departments in the first domain. Then, the event information of the second sample events in the first domain, the department information of the sample processing departments in the first domain, and the labeled matching degree between the second sample events and their matched processing departments in the first domain can be used to form the second training data.

[0052] Additionally, event information and department information of matched processing departments for historical events can be automatically retrieved from the matched processing departments of historical events. Historical events can be used as second sample events in the first domain, their event information as event information of the second sample events in the first domain, and unmatched processing departments of historical events can be used as sample processing departments in the first domain. The department information of unmatched processing departments of historical events can also be used as department information of sample processing departments in the first domain. In practice, the matching relationship between historical events and their matched processing departments is often done manually. Therefore, the matching degree between historical events and their matched processing departments is often very high, while the matching degree between historical events and their unmatched processing departments is often very low. Thus, a very low matching degree can be set and used as the labeled matching degree between historical events and unmatched sample processing departments in the first domain. Then, the event information of the second sample event in the first domain, the department information of the sample processing department in the first domain, and the label matching degree between the second sample event in the first domain and the sample processing department in the first domain can be used to form the second training data.

[0053] In one embodiment of this application, the sample object includes a sample event type, and the object information includes type information of the sample event type, etc.

[0054] This involves automatically obtaining event information and matched event types of historical events from the matching department that matches historical events. The historical events can be used as the second sample events in the first domain, their event information as the event information of the second sample events in the first domain, their matched event types as the sample event types in the first domain (which can be understood as the event types involved in the first domain), and their type information as the type information of the matched event types in the first domain. In practice, the matching relationship between historical events and their matched event types is often done manually. Therefore, the matching degree (e.g., matching accuracy) between historical events and their matched event types is often very high. Thus, a high matching degree can be set and used as the labeled matching degree between the second sample events and the sample event types in the first domain. Then, the event information of the second sample events in the first domain, the type information of the sample event types of the second sample events in the first domain, and the labeled matching degree between the second sample events and the sample event types in the first domain can be used to form the second training data.

[0055] Specifically, the system can automatically obtain event information and matched event types of historical events from the matching department that matches historical events. Historical events can be used as second sample events in the first domain, their event information as event information for the second sample events in the first domain, and unmatched event types from historical events as sample event types for the second sample events in the first domain. In practice, the matching relationship between historical events and their matched event types is often done manually. Therefore, the matching degree between historical events and their matched event types is often very high, while the matching degree between historical events and their unmatched processing departments is often very low. Thus, a very low matching degree can be set and used as the labeled matching degree between historical events and unmatched event types in the first domain. Then, the event information of the second sample events in the first domain, the sample event types of the second sample events in the first domain, and the labeled matching degree between the second sample events and their sample event types in the first domain can be used to form the second training data.

[0056] In this application, the matching model including the encoding network of the first domain may include: an encoding network of the first domain and a matching network. The encoding network of the first domain is used to encode the event information of the second sample event in the first domain to obtain the event features of the second sample event in the first domain, and to encode the object information of the sample object in the first domain to obtain the object features of the sample object in the first domain. The matching network is used to obtain the matching degree between the second sample event in the first domain and the sample object in the first domain based on the event features of the second sample event in the first domain and the object features of the sample object in the first domain.

[0057] In one embodiment, the sample object may include a sample processing department. A higher degree of matching between sample events in the first domain and the sample processing department in the first domain indicates that, based on the authority and responsibility attributes of the sample processing department in the first domain, it is more suitable to assign the sample events in the first domain to the sample processing department for processing. Conversely, a lower degree of matching between sample events in the first domain and the sample processing department in the first domain indicates that, based on the authority and responsibility attributes of the sample processing department in the first domain, it is less suitable to assign the sample events in the first domain to the sample processing department for processing.

[0058] In another embodiment, the sample object may include a sample event type. A higher degree of matching between the sample event in the first domain and the sample event type in the first domain indicates that the event type of the sample event in the first domain matches the sample event type in the first domain better. Conversely, a lower degree of matching between the sample event in the first domain and the sample event type in the first domain indicates that the event type of the sample event in the first domain does not match the sample event type in the first domain better.

[0059] Matching networks can include dual-tower matching networks, etc.

[0060] In one embodiment, see Figure 3 The matching model, which includes the encoding network of the first domain, has its input terminal connected to the encoding network of the first domain. The output terminal of the matching model, which includes the encoding network of the first domain, is connected to the input terminal of the matching network.

[0061] In one embodiment, see Figure 4The matching network can include a feature computation layer, a feature concatenation layer, and a prediction layer. The inputs of the matching network include the inputs of the feature computation layer and the feature concatenation layer. The output of the feature computation layer is also connected to the input of the feature concatenation layer. The output of the feature concatenation layer is connected to the input of the prediction layer. The output of the matching network includes the output of the prediction layer. The feature computation layer performs cross-operations on the event features of the second sample event in the first domain and the object features of the sample objects in the first domain to obtain computational features. The feature concatenation layer concatenates the computational features, the event features of the second sample event in the first domain, and the object features of the sample objects in the first domain to obtain concatenated features. The prediction layer predicts the matching degree between the second sample event in the first domain and the sample objects in the first domain based on the concatenated features.

[0062] In one example, the encoding network for the first domain is BERT, and the prediction layer is a sigmoid function, etc. The feature concatenation layer can include fully connected layers, etc. See also Figure 5The encoding network of the first domain encodes the event information of the second sample event in the first domain, resulting in an event feature vector Vc for the second sample event in the first domain. It also encodes the object information of the sample object in the first domain, resulting in an object feature vector Vt for the sample object in the first domain. Vectors Vc and Vt are input to the feature processing layer and the fully connected layer, respectively. The feature processing layer subtracts corresponding elements from vectors Vc and Vt to obtain |Vc-Vt|, where the dimensions of |Vc-Vt|, Vc, and Vt are the same. The feature processing layer then multiplies corresponding elements from vectors Vc and Vt to obtain |Vc·Vt|, where the dimensions of |Vc·Vt|, Vc, and Vt are the same. Finally, the feature processing layer concatenates |Vc·Vt| with |Vc-Vt| to obtain the vector (|Vc·Vt|, |Vc-Vt|). The dimension of the vector (|Vc·Vt|, |Vc-Vt|) is greater than the dimension of both |Vc·Vt| and |Vc-Vt|. The feature processing layer then inputs the vector (|Vc·Vt|, |Vc-Vt|) into the fully connected layer. The fully connected layer concatenates the vectors Vc, Vt, and (|Vc·Vt|, |Vc-Vt|) to obtain the vector [Vc,Vt,(|Vc·Vt|, |Vc-Vt|)]. The dimension of the vector [Vc,Vt,(|Vc·Vt|, |Vc-Vt|)] is greater than the dimension of Vc, greater than the dimension of Vt, and greater than the dimension of (|Vc·Vt|, |Vc-Vt|). The fully connected layer then inputs the vector [Vc,Vt,(|Vc·Vt|, |Vc-Vt|)] into the prediction layer. The prediction layer obtains the matching degree between the second sample event in the first domain and the sample object in the first domain based on [Vc,Vt,(|Vc·Vt|,|Vc-Vt|)], and outputs the matching degree between the second sample event in the first domain and the sample object in the first domain.

[0063] In step S104, the parameters in the matching model including the encoding network of the first domain are trained using at least one second training data until the parameters in the matching model including the encoding network of the first domain converge, thereby obtaining the first matching model.

[0064] The second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the labeled matching degree between the second sample events and sample objects in the first domain. In this application, after obtaining at least one second training data and a matching model including the encoding network of the first domain, the network parameters in the matching model including the encoding network of the first domain can be trained based on at least one second training data. During training, for any second training data, the event information of the second sample events in the first domain and the object information of sample objects in the first domain can be input into the matching model including the encoding network of the first domain, so that the matching model including the encoding network of the first domain processes the event information of the second sample events in the first domain and the object information of sample objects in the first domain to obtain the matching degree between the second sample events in the first domain and sample objects in the first domain. Then, the network parameters in the matching model including the encoding network of the first domain can be adjusted using a loss function and based on the output matching degree and the labeled matching degree between the second sample events in the first domain and sample objects in the first domain. For each other second training data, the above operation is performed in the same way until the network parameters in the matching model including the encoding network of the first domain converge, thereby completing the training and obtaining the first matching model.

[0065] In step S105, at least one third training data is acquired and a matching model including the first matching model is created; the third training data includes: event information of sample events in multiple second domains cascaded in the first domain, object information of sample objects in the target second domain, and the labeled matching degree between sample events in the target second domain and sample objects in the target second domain.

[0066] The first matching model can be applied to matching objects and events in the first domain. The first domain is cascaded with multiple second domains. The matching ability of objects and events in each second domain of the first matching model is relatively balanced. In order to improve the matching accuracy of specific events in the first cascaded second domain and specific objects in the first cascaded second domain, steps S105 and S106 can also be executed.

[0067] In this application, a first domain is cascaded with multiple second domains, and the target second domain is a subset of the multiple second domains cascaded from the first domain. The target second domain can be one or more second domains. The event information of the sample events in the target second domain can be automatically collected from the network, etc. The sample events in the target second domain can include historical events that have occurred beforehand and have actually matched the processing departments of the target second domain.

[0068] In one embodiment, the sample object may include a sample processing department, and the object information includes the department information of the sample processing department.

[0069] The system can automatically retrieve event information and department information of matched processing departments for historical events in the target second domain. Historical events in the target second domain can be used as second sample events, their event information as event information for second sample events, their matched processing departments as sample processing departments, and their department information as department information for sample processing departments. In practice, the matching relationship between historical events and their matched processing departments in the target second domain is often done manually. Therefore, the matching degree (e.g., accuracy) between historical events and their matched processing departments in the target second domain is often very high. Thus, a very high matching degree can be set and used as the labeled matching degree between the second sample events and their matched processing departments in the target second domain. Then, the event information of the second sample event in the second target domain, the department information of the sample processing department in the second target domain, and the label matching degree between the second sample event in the second target domain and the sample processing department in the second target domain can be used to form the second training data.

[0070] The system can automatically obtain event information of historical events that have occurred in the target second domain and department information of the matching processing departments of historical events that have occurred in the target second domain. It can use historical events that have occurred in the target second domain as second sample events in the target second domain, use the event information of historical events that have occurred in the target second domain as event information of second sample events in the target second domain, use a processing department that has not been matched with a historical event that has occurred in the target second domain as a sample processing department in the target second domain, and use the department information of a processing department that has not been matched with a historical event that has occurred in the target second domain as the department information of the sample processing department in the target second domain. In practice, the matching relationship between historical events in the target second domain and the processing departments that have matched those historical events is often done manually. This results in a high degree of matching between historical events and their matched processing departments, and conversely, a low degree of matching between historical events and their unmatched processing departments. Therefore, a very low matching degree can be set and used as the labeled matching degree between historical events and their unmatched sample processing departments in the target second domain. Then, the event information of the second sample event in the target second domain, the department information of the sample processing department in the target second domain, and the labeled matching degree between the second sample event and its sample processing department in the target second domain can be used to form the second training data.

[0071] In one embodiment, the sample object may include a sample event type, and the object information includes type information of the sample event type.

[0072] The system can automatically obtain event information of historical events that have occurred in the target second domain, as well as the type information of the event types that have been matched with those historical events, from the matching department that matches historical events. The historical events that have occurred in the target second domain can be used as the second sample events in the target second domain, the event information of those historical events can be used as the event information of those second sample events, the event types that have been matched with those historical events can be used as the sample event types in the target second domain (which can be understood as the event types involved in the target second domain), and the type information of those matching with those historical events can be used as the type information of those sample event types. In practice, the matching relationship between historical events and their matched event types in the target second domain is often done manually. Therefore, the matching degree (e.g., matching accuracy) between these historical events and their matched event types is often very high. Thus, a very high matching degree can be set and used as the labeled matching degree between the second sample events and their sample event types in the target second domain. Then, the event information of the second sample event in the second target domain, the type information of the sample event type of the second sample event in the second target domain, and the label matching degree between the second sample event and the sample event type in the second target domain can be used to form the second training data.

[0073] The system can automatically obtain event information of historical events that have occurred in the target second domain, as well as the event types that have been matched with those historical events, from the matching department that matches historical events. Historical events that have occurred in the target second domain can be used as second sample events in the target second domain, and their event information can be used as the event information of the second sample events in the target second domain. An unmatched event type from historical events in the target second domain can be used as a sample event type for the second sample events in the target second domain. In practice, the matching relationship between historical events and their matched event types in the target second domain is often done manually. Therefore, the matching degree between historical events and their matched event types in the target second domain is often very high, while the matching degree between historical events and unmatched event types in the target second domain is often very low. Thus, a very low matching degree can be set and used as the labeled matching degree between historical events and unmatched event types in the target second domain. Then, the event information of the second sample event in the second target domain, the sample event type of the second sample event in the second target domain, and the labeled matching degree between the second sample event and the sample event type in the second target domain can be used to form the second training data. In one embodiment, the matching model including the first matching model can be the first matching model, and the model structure can be found in [reference needed]. Figure 3-5 As shown, details will not be elaborated here.

[0074] In another embodiment, the matching model including the first matching model may also include an auxiliary feature extraction network in addition to the first matching model. The auxiliary feature extraction network is used to extract auxiliary information associated with "event information of sample events in the target second domain and object information of sample objects in the target second domain" and encode the auxiliary information to obtain auxiliary features. The auxiliary information may include keywords in the event information of sample events in the target second domain, the character overlap rate between the event information of sample events in the target second domain and the object information of sample objects in the target second domain, the vocabulary overlap rate between the event information of sample events in the target second domain and the object information of sample objects in the target second domain, and the hypernyms and hyponyms of the event information of sample events in the target second domain, etc.

[0075] In this application, the matching model, including the first matching model, may include: an auxiliary feature extraction network, a first-domain encoding network, and a matching network. The first-domain encoding network is used to encode event information of sample events in the target second domain to obtain event features of the sample events in the target second domain, and to encode object information of sample objects in the target second domain to obtain object features of the sample objects in the target second domain. The matching network is used to obtain the matching degree between sample events in the target second domain and sample objects in the target second domain based on the auxiliary features, the event features of sample events in the target second domain, and the object features of sample objects in the target second domain.

[0076] In one embodiment, the sample object may include a sample processing department. Therefore, a higher degree of matching between the sample event in the target second domain and the sample object in the target second domain indicates that, based on the authority and responsibility attributes of the sample object in the target second domain, it is more suitable to assign the sample event in the target second domain to the sample processing department for processing. Conversely, a lower degree of matching between the sample event in the target second domain and the sample object in the target second domain indicates that, based on the authority and responsibility attributes of the sample object in the target second domain, it is less suitable to assign the sample event in the target second domain to the sample processing department for processing.

[0077] In another embodiment, the sample object may include a sample event type. A higher degree of matching between a sample event in the target second domain and a sample object in the target second domain indicates a better match between the event type of the sample event in the target second domain and the sample event type in the target second domain. Conversely, a lower degree of matching between a sample event in the target second domain and a sample object in the target second domain indicates a less good match between the event type of the sample event in the target second domain and the sample event type in the target second domain.

[0078] Matching networks can include dual-tower matching networks, etc.

[0079] In one embodiment, see Figure 6 The matching model, including the first matching model, has inputs to a first-domain encoding network and an auxiliary encoding network. The output of the first-domain encoding network is connected to the input of the matching network. The output of the auxiliary encoding network is also connected to the input of the matching network. The output of the matching model, including the first matching model, includes the output of the matching network.

[0080] In one embodiment, see Figure 7The matching network can include a feature computation layer, a feature concatenation layer, and a prediction layer. The inputs of the matching network include the inputs of the feature computation layer and the feature concatenation layer. The output of the feature computation layer is also connected to the input of the feature concatenation layer. The output of the feature concatenation layer is connected to the input of the prediction layer. The output of the matching network includes the output of the prediction layer. The feature computation layer performs cross-operations on the event features of sample events in the target second domain and the object features of sample objects in the target second domain to obtain computational features. The feature concatenation layer concatenates the computational features, the event features of sample events in the target second domain, the object features of sample objects in the target second domain, and auxiliary features to obtain concatenated features. The prediction layer predicts the matching degree between sample events and sample objects in the target second domain based on the concatenated features.

[0081] In one example, the encoding network for the first domain is BERT, and the prediction layer is a sigmoid function, etc. The feature concatenation layer can include fully connected layers, etc. See also Figure 8The encoding network of the first domain encodes the event information of sample events in the target second domain, resulting in an event feature vector Vc for the sample events in the target second domain. It also encodes the object information of sample objects in the target second domain, resulting in an object feature vector Vt for the sample objects in the target second domain. The auxiliary feature extraction network extracts auxiliary information associated with the event information of sample events in the target second domain and the object information of sample objects in the target second domain, and encodes this auxiliary information, resulting in an auxiliary feature vector Vwf. Vectors Vc and Vt are input to the feature processing layer and the fully connected layer, respectively. Vector Vwf is input to the fully connected layer. The feature processing layer subtracts corresponding elements from vectors Vc and Vt to obtain |Vc-Vt|, where the dimensions of |Vc-Vt|, Vc, and Vt are the same. The feature processing layer also multiplies corresponding elements from vectors Vc and Vt to obtain |Vc·Vt|, where the dimensions of |Vc·Vt|, Vc, and Vt are the same. The feature processing layer then concatenates vectors |Vc·Vt| and |Vc-Vt| to obtain vectors (|Vc·Vt|, |Vc-Vt|). The dimensions of vectors (|Vc·Vt|, |Vc-Vt|) are greater than the dimensions of both vectors |Vc·Vt| and |Vc-Vt|. The feature processing layer then inputs vectors (|Vc·Vt|, |Vc-Vt|) into a fully connected layer. The fully connected layer concatenates vectors Vc, Vt, (|Vc·Vt|, |Vc-Vt|), and Vwf to obtain vectors [Vc, Vt, (|Vc·Vt|, |Vc-Vt|), Vwf]. The dimension of the vector [Vc, Vt, (|Vc·Vt|,|Vc-Vt|)Vwf] is greater than the dimension of vector Vc, greater than the dimension of vector Vt, greater than the dimension of vector Vwf, and greater than the dimension of vector (|Vc·Vt|,|Vc-Vt|). The fully connected layer then inputs the vector [Vc, Vt, (|Vc·Vt|,|Vc-Vt|), Vwf] into the prediction layer. The prediction layer obtains the matching degree between sample events and sample objects in the target second domain based on [Vc, Vt, (|Vc·Vt|,|Vc-Vt|), Vwf], and outputs the matching degree between sample events and sample objects in the target second domain.

[0082] In step S106, the parameters of the matching model including the first matching model are adjusted using at least one third training data to obtain the second matching model.

[0083] In one embodiment, the second matching model can be an end-to-end model, such as a one-step method that goes directly from event description to assignment to the corresponding department, without the need for a two-step method of "first classifying by fixed event types and then assigning to departments according to mapping relationships".

[0084] The third training data includes: event information of sample events in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the annotation matching degree between sample events and sample objects in the target second domain.

[0085] In this application, after obtaining at least one third training data and a matching model including the first matching model, the network parameters in the matching model including the first matching model can be adjusted based on the at least one third training data. During parameter adjustment, for any third training data, event information of sample events in the target second domain and object information of sample objects in the target second domain can be input into the matching model including the first matching model. This allows the matching model to process the event information of sample events in the target second domain and the object information of sample objects in the target second domain, obtaining the matching degree between the sample events and sample objects in the target second domain. Then, using a loss function, the network parameters in the matching model including the first matching model can be adjusted based on the output matching degree and the labeled matching degree between the sample events and sample objects in the target second domain. The same operation is performed for each other third training data until the network parameters in the matching model including the first matching model converge, thus completing the parameter adjustment and obtaining the second matching model.

[0086] Once the second matching model is obtained, it can be deployed online. For example, when an event is received and a department needs to be matched to handle it, the second matching model can be used to obtain the matching degree between the event and the department for any given department. The same process applies to all other departments. Then, the department is determined based on the matching degree between the event and each department, and the event can be assigned to that department. Similarly, when an event is received and its event type needs to be matched, the second matching model can be used to obtain the matching degree between the event and the event type for any given event type. The same process applies to all other event types. Then, the event type is determined based on the matching degree between the event and each event type, and the event can be processed according to its type. For example, a suitable processing method can be found based on the event type, and then the event can be processed according to that method.

[0087] In this application, the input data of the second matching model may include event information of the event and object information of the object. The second matching model can calculate the matching degree between the event and the object based on the event information and object information input into the second matching model. If it is necessary to update the object that needs to be matched for the event, the object information of the updated object can be directly input into the second matching model without retraining the second matching model. The update cost is low and the update efficiency is high.

[0088] For example, refer to Figure 9 This application illustrates a method for obtaining a matching degree, the method comprising:

[0089] In step S201, the event information of the target event in the second domain of the target and the object information of the target object in the second domain of the target are obtained.

[0090] In one example, the object information of the target object in the second domain can be stored locally beforehand, allowing direct retrieval of the object information from the local storage. Event information includes the event name and descriptive text. In one embodiment, the object includes a processing department, and the object information includes the department's information. The department's information includes the object's name, responsibilities, and classification structure. In another embodiment, the object includes an event type, and the object information includes the event type's type information. The event type's type information includes the event type's related name and type structure.

[0091] The event information of the target event in the second target domain can be submitted by a large number of users (such as the general public).

[0092] In step S202, the matching degree between the target event and the target object is obtained based on the event information of the target event in the second target domain, the object information of the target object in the second target domain, and the trained second matching model.

[0093] Specifically, the event information of the target event in the second domain and the object information of the target object in the second domain can be input into the trained second matching model so that the trained second matching model processes the event information of the target event in the second domain and the object information of the target object in the second domain to obtain the matching degree between the target event and the target object.

[0094] The second matching model is obtained by adjusting the parameters of a matching model including the first matching model using at least one third training data. The third training data includes: event information of sample events in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the labeled matching degree between sample events and sample objects in the target second domain. The first matching model is obtained by training the parameters of a matching model including the encoding network of the first domain using at least one second training data until the parameters of the matching model including the encoding network of the first domain converge. The second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the labeled matching degree between second sample events and sample objects in the first domain. The encoding network of the first domain is obtained by adjusting the parameters of the encoding network to be trained using at least one first training data, and the first training data includes event information of first sample events in the first domain.

[0095] On the one hand, the traditional matching method involves inputting event information into the model so that the model can output processing departments that match the event. The processing departments that the model can output with matching qualifications are fixed, and it is not possible to update the processing departments with matching qualifications online in the model later. If it is necessary to update the processing departments with matching qualifications, the model needs to be retrained, which is costly and inefficient.

[0096] In this application, the input to the model is event information and object information of eligible objects. The model is used to obtain the matching degree between the event and the object based on the event information and the object information, and then matches the event with the object based on the matching degree. Eligible objects are not restricted in the model. Subsequently, when it is necessary to update eligible objects online, only the event information and the updated object information need to be input into the model, eliminating the need to retrain the model, resulting in low update cost and high update efficiency.

[0097] On the other hand, this application employs a step-by-step learning approach when training the model. For example, it acquires at least one first training data and an encoding network to be trained, wherein the first training data includes event information of first sample events in a first domain; it uses at least one first training data to adjust the parameters of the encoding network to be trained, thereby obtaining an encoding network for the first domain; it acquires at least one second training data and creates a matching model that includes the encoding network for the first domain; the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the label matching degree between the second sample events in the first domain and the sample objects in the first domain; it uses at least one second training data to train the parameters in the matching model that includes the encoding network for the first domain until the parameters in the matching model that includes the encoding network for the first domain converge, thereby obtaining a first matching model; it acquires at least one third training data and creates a matching model that includes the first matching model; the third training data includes: event information of sample events in a target second domain in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the label matching degree between the sample events in the target second domain and the sample objects in the target second domain; it uses at least one third training data to adjust the parameters of the matching model that includes the first matching model, thereby obtaining a second matching model. The first matching model has the ability to match events and objects within the first domain. Furthermore, the second training data used to train the first matching model is readily available; for example, labeled data in the second training data is readily available. This makes training the first matching model easier and increases its matching accuracy. If a matching model is subsequently needed that can match events in a second domain cascaded from the first domain with objects in that second domain cascaded from the first domain, a matching model with higher matching accuracy can be trained using a small amount of second domain training data, reducing the difficulty and time required to collect second domain training data.

[0098] When an event is received and it is necessary to match a processing department to handle the event, at least one processing department can be matched from multiple processing departments to be matched in the following way, and then the event can be assigned to the matched processing department so that the matched processing department can handle the event.

[0099] For example, refer to Figure 10 This application illustrates a matching method, which includes:

[0100] In step S301, the event information of the target event is obtained.

[0101] Event information includes the title and description of the target event. Event information for the target event in the second target domain can be submitted by a broad range of users (e.g., the general public).

[0102] In step S302, department information of multiple processing departments to be matched is obtained.

[0103] The department information includes: the department name and the department classification structure; in one example, the department information of each processing department to be matched can be stored locally in advance, so that the department information of multiple processing departments to be matched can be obtained directly from the local machine.

[0104] In step S303, keywords used to describe the subject of the target event are obtained based on the event information of the target event.

[0105] In this application, existing keyword extraction methods can be used to extract keywords describing the theme of the target event from the event information of the target event. This application does not limit the keyword extraction method.

[0106] In step S304, among the multiple processing departments to be matched, the first processing department whose department information includes the keyword is determined.

[0107] In step S305, the sorting order of the first processing department is set to precede that of the second processing department (excluding the first processing department) among the multiple processing departments to be matched.

[0108] If a processing department's department information includes the keyword, it often indicates a high degree of match between that department and the target event. This match is generally higher than that between other processing departments whose department information does not include the keyword. Furthermore, based on the department's responsibilities and authority, it is more suitable to assign the target event to that department. The second processing department is the first processing department whose department information does not include the keyword.

[0109] In step S306, at least one processing department that matches the target event is selected from among the multiple processing departments to be matched, according to the sorting order among the multiple processing departments to be matched.

[0110] In this embodiment, multiple processing departments to be matched are sorted according to the department information including the keyword, and the order of the first processing department is before the order of the second processing department.

[0111] In one embodiment, when there are two or more first processing departments, for any one first processing department, based on the event information of the target event and the department information of the first processing department, the text matching degree between the target event and the first processing department is obtained, the scene matching degree between the target event and the first processing department is obtained, and the keyword matching degree between the target event and the first processing department is obtained; a matching score between the target event and the first processing department is obtained based on the text matching degree, scene matching degree, and keyword matching degree, and the same process is repeated for each other first processing department. The two or more first processing departments are sorted in descending order of their matching scores with the target event. And / or, in another embodiment, when there are two or more second processing departments, for any one second processing department, based on the event information of the target event and the department information of the second processing department, the text matching degree between the target event and the second processing department is obtained, the scene matching degree between the target event and the second processing department is obtained, and the keyword matching degree between the target event and the second processing department is obtained; a matching score between the target event and the second processing department is obtained based on the text matching degree, scene matching degree, and keyword matching degree, and the same process is repeated for each other second processing department. The second processing departments are sorted in descending order of their matching scores with the target event.

[0112] The method for obtaining the text matching degree between the event and the handling department can be found in [link to relevant documentation]. Figure 9 The illustrated embodiments will not be described in detail here. When obtaining the keyword matching degree between an event and a processing department, one can refer to the number of keywords related to the event included in the department information of the processing department. The more keywords there are, the higher the keyword matching degree between the event and the processing department; the fewer keywords there are, the lower the keyword matching degree between the event and the processing department.

[0113] When it is necessary to match a target event with a target object among multiple objects to be matched, the objects can be divided into two categories based on whether the object information contains the keyword of the target event. The matching degree between the target event and the object information containing the keyword of the target event is higher than that between the target event and the object information containing the keyword of the target event. Thus, the sorting order of the object information containing the keyword of the target event can be set to be ahead of the sorting order of the object information not containing the keyword of the target event, thereby improving the sorting accuracy and thus improving the matching accuracy between the event and the object.

[0114] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions involved are not necessarily required by this application.

[0115] Reference Figure 11 The diagram illustrates a structural block diagram of a training apparatus for a matching model according to this application. The apparatus includes: a first acquisition module 11, configured to acquire at least one first training data and an encoding network to be trained, wherein the first training data includes event information of first sample events in a first domain; a first adjustment module 12, configured to adjust the parameters of the encoding network to be trained using at least one first training data to obtain an encoding network in the first domain; a second acquisition module 13, configured to acquire at least one second training data and create a matching model including the encoding network in the first domain; wherein the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the labeled matching degree between the second sample events in the first domain and the sample objects in the first domain; and a training module. 14, used to train the parameters in a matching model including a first-domain encoding network using at least one second training data until the parameters in the matching model including the first-domain encoding network converge, thereby obtaining a first matching model; a third acquisition module 15, used to acquire at least one third training data and create a matching model including the first matching model; the third training data includes: event information of sample events in multiple second domains cascaded in the first domain, object information of sample objects in the target second domain, and labeled matching degree between sample events in the target second domain and sample objects in the target second domain; a second adjustment module 16, used to adjust the parameters of the matching model including the first matching model using at least one third training data, thereby obtaining a second matching model.

[0116] In one optional implementation, the matching model including the encoding network of the first domain includes: an encoding network of the first domain and a matching network; the encoding network of the first domain is used to encode the event information of the second sample event of the first domain to obtain the event features of the second sample event of the first domain, and to encode the object information of the sample object of the first domain to obtain the object features of the sample object of the first domain; the matching network is used to obtain the matching degree between the second sample event of the first domain and the sample object of the first domain based on the event features of the second sample event of the first domain and the object features of the sample object of the first domain.

[0117] In one optional implementation, the matching network includes a feature computation layer, a feature concatenation layer, and a prediction layer. The feature computation layer is used to perform cross-operations on the event features of the second sample event in the first domain and the object features of the sample object in the first domain to obtain computational features. The feature concatenation layer is used to concatenate the computational features, the event features of the second sample event in the first domain, and the object features of the sample object in the first domain to obtain concatenated features. The prediction layer is used to predict the matching degree between the second sample event in the first domain and the sample object in the first domain based on the concatenated features.

[0118] In one optional implementation, the matching model, including the first matching model, comprises: an auxiliary feature extraction network, a first domain encoding network, and a matching network. The auxiliary feature extraction network is used to extract auxiliary information associated with the event information of sample events in the target second domain and the object information of sample objects in the target second domain, and to encode the auxiliary information to obtain auxiliary features. The auxiliary information includes: keywords in the event information of sample events in the target second domain, the character overlap rate between the event information of sample events in the target second domain and the object information of sample objects in the target second domain, the vocabulary overlap rate between the event information of sample events in the target second domain and the object information of sample objects in the target second domain, and the hypernyms and hyponyms of the event information of sample events in the target second domain. The first domain encoding network is used to encode the event information of sample events in the target second domain to obtain event features of sample events in the target second domain, and to encode the object information of sample objects in the target second domain to obtain object features of sample objects in the first domain. The matching network is used to obtain the matching degree between sample events in the target second domain and sample objects in the target second domain based on the auxiliary features, the event features of sample events in the target second domain, and the object features of sample objects in the target second domain.

[0119] The matching network includes a feature computation layer, a feature concatenation layer, and a prediction layer. The feature computation layer performs cross-operations on the event features of sample events in the second target domain and the object features of sample objects in the second target domain to obtain computational features. The feature concatenation layer concatenates the computational features, the event features of sample events in the second target domain, the object features of sample objects in the second target domain, and auxiliary features to obtain concatenated features. The prediction layer predicts the matching degree between sample events in the second target domain and sample objects in the second target domain based on the concatenated features.

[0120] On the one hand, traditional matching methods input event information into the model so that the model can output processing departments that match the event. The processing departments that the model can output with matching qualifications are fixed, and it is impossible to update the matching qualifications online within the model later. If an update is needed, the model needs to be retrained, resulting in high update costs and low efficiency. In this application, however, the input to the model is event information and object information of objects with matching qualifications. The model is used to obtain the matching degree between the event and the object based on the event information and the object information, and then matches the event with the object based on the matching degree. The objects with matching qualifications are not restricted in the model. If an update of matching qualifications is needed later, only the event information and the object information of the updated object need to be input into the model, eliminating the need for model retraining, resulting in low update costs and high update efficiency.

[0121] On the other hand, this application employs a step-by-step learning approach when training the model. For example, it acquires at least one first training data and an encoding network to be trained, wherein the first training data includes event information of first sample events in a first domain; it uses at least one first training data to adjust the parameters of the encoding network to be trained, thereby obtaining an encoding network for the first domain; it acquires at least one second training data and creates a matching model that includes the encoding network for the first domain; the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the label matching degree between the second sample events in the first domain and the sample objects in the first domain; it uses at least one second training data to train the parameters in the matching model that includes the encoding network for the first domain until the parameters in the matching model that includes the encoding network for the first domain converge, thereby obtaining a first matching model; it acquires at least one third training data and creates a matching model that includes the first matching model; the third training data includes: event information of sample events in a target second domain in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the label matching degree between the sample events in the target second domain and the sample objects in the target second domain; it uses at least one third training data to adjust the parameters of the matching model that includes the first matching model, thereby obtaining a second matching model. The first matching model has the ability to match events and objects within the first domain. Furthermore, the second training data used to train the first matching model is readily available; for example, labeled data in the second training data is readily available. This makes training the first matching model easier and increases its matching accuracy. If a matching model is subsequently needed that can match events in a second domain cascaded from the first domain with objects in that second domain cascaded from the first domain, a matching model with higher matching accuracy can be trained using a small amount of second domain training data, reducing the difficulty and time required to collect second domain training data.

[0122] Reference Figure 12This application discloses an apparatus for obtaining a matching degree, the apparatus comprising: a fourth acquisition module 21, configured to acquire event information of target events in a target second domain and object information of target objects in the target second domain; and a fifth acquisition module 22, configured to acquire a matching degree between a target event and a target object based on the event information of the target events in the target second domain, the object information of the target objects in the target second domain, and a trained second matching model; wherein the second matching model is obtained by adjusting the parameters of a matching model including a first matching model using at least one third training data; the third training data includes: event information of sample events in the target second domain in multiple second domains cascaded from the first domain, and sample objects in the target second domain. The first matching model is obtained by training the parameters of the matching model including the encoding network of the first domain using at least one second training data until the parameters of the matching model including the encoding network of the first domain converge. The second training data includes: event information of the second sample event of the first domain, object information of the sample object of the first domain, and the annotation matching degree between the second sample event of the first domain and the sample object of the first domain. The encoding network of the first domain is obtained by adjusting the parameters of the encoding network to be trained using at least one first training data. The first training data includes event information of the first sample event of the first domain.

[0123] On the one hand, traditional matching methods input event information into the model so that the model can output processing departments that match the event. The processing departments that the model can output with matching qualifications are fixed, and it is impossible to update the matching qualifications online within the model later. If an update is needed, the model needs to be retrained, resulting in high update costs and low efficiency. In this application, however, the input to the model is event information and object information of objects with matching qualifications. The model is used to obtain the matching degree between the event and the object based on the event information and the object information, and then matches the event with the object based on the matching degree. The objects with matching qualifications are not restricted in the model. If an update of matching qualifications is needed later, only the event information and the object information of the updated object need to be input into the model, eliminating the need for model retraining, resulting in low update costs and high update efficiency. On the other hand, this application employs a step-by-step learning approach when training the model. For example, it acquires at least one first training data and an encoding network to be trained, wherein the first training data includes event information of first sample events in a first domain; it uses at least one first training data to adjust the parameters of the encoding network to be trained, thereby obtaining an encoding network for the first domain; it acquires at least one second training data and creates a matching model that includes the encoding network for the first domain; the second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the label matching degree between the second sample events in the first domain and the sample objects in the first domain; it uses at least one second training data to train the parameters in the matching model that includes the encoding network for the first domain until the parameters in the matching model that includes the encoding network for the first domain converge, thereby obtaining a first matching model; it acquires at least one third training data and creates a matching model that includes the first matching model; the third training data includes: event information of sample events in a target second domain in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the label matching degree between the sample events in the target second domain and the sample objects in the target second domain; it uses at least one third training data to adjust the parameters of the matching model that includes the first matching model, thereby obtaining a second matching model. The first matching model has the ability to match events and objects in the first domain, and the second training data used to train the first matching model is relatively easy to obtain, such as the labeled data in the second training data. Thus, the training of the first matching model is easier and the matching accuracy of the trained first matching model is higher.If a matching model is needed later that can match events in a second domain that are cascaded with the first domain with objects in the same second domain that are cascaded with the first domain, a matching model with higher matching accuracy can be trained using a small amount of training data from the second domain, reducing the difficulty and time spent collecting training data from the second domain.

[0124] Reference Figure 13 This application illustrates a matching device, comprising: a sixth acquisition module 31, configured to acquire event information of a target event, the event information including the title of the target event and a description of the target event; a seventh acquisition module 32, configured to acquire object information of multiple objects to be matched, the object information including the object's name and object classification structure; an eighth acquisition module 33, configured to acquire keywords describing the theme of the target event based on the event information of the target event; a determination module 34, configured to determine a first object whose object information includes the keywords among the multiple objects to be matched; a setting module 35, configured to set the sorting order of the first object to be placed before a second object (excluding the first object) among the multiple objects to be matched; and a selection module 36, configured to select at least one object matching the target event from among the multiple objects to be matched according to the sorting order among the multiple objects to be matched.

[0125] In an optional implementation, the apparatus further includes: a first sorting module, configured to, when there are two or more first objects, for any one first object, obtain a text matching degree between the target event and the first object, obtain a scene matching degree between the target event and the first object, and obtain a keyword matching degree between the target event and the first object based on the event information of the target event and the object information of the first object; obtain a matching score between the target event and the first object based on the text matching degree, scene matching degree, and keyword matching degree, and sort the two or more first objects in descending order of their respective matching scores with the target event; and / or, a second sorting module, configured to, when there are two or more second objects, for any one second object, obtain a text matching degree between the target event and the second object, obtain a scene matching degree between the target event and the second object, and obtain a keyword matching degree between the target event and the second object based on the event information of the target event and the object information of the second object; obtain a matching score between the target event and the second object based on the text matching degree, scene matching degree, and keyword matching degree, and sort the two or more second objects in descending order of their respective matching scores with the target event.

[0126] When it is necessary to match a target event with a target object among multiple objects to be matched, the objects can be divided into two categories based on whether the object information contains the keyword of the target event. The matching degree between the target event and the object information containing the keyword of the target event is higher than that between the target event and the object information containing the keyword of the target event. Thus, the sorting order of the object information containing the keyword of the target event can be set to be ahead of the sorting order of the object information not containing the keyword of the target event, thereby improving the sorting accuracy and thus improving the matching accuracy between the event and the object.

[0127] This application also provides a non-volatile readable storage medium storing one or more modules (programs). When these modules are applied to a device, they enable the device to execute the instructions for the method steps in this application.

[0128] This application provides one or more machine-readable media storing instructions that, when executed by one or more processors, cause an electronic device to perform one or more methods as described in the above embodiments. In this application, the electronic device includes a server, a gateway, sub-devices, etc., and the sub-devices are devices such as Internet of Things (IoT) devices.

[0129] Embodiments of this disclosure can be implemented as an apparatus with any suitable hardware, firmware, software, or any combination thereof, configured as desired. This apparatus may include electronic devices such as servers (clusters) and terminal devices such as IoT devices.

[0130] Figure 14 An exemplary apparatus 1300 that can be used to implement the various embodiments of this application is schematically illustrated. For one embodiment, Figure 14 An exemplary device 1300 is illustrated, comprising one or more processors 1302, a control module (chipset) 1304 coupled to at least one of the processors 1302, a memory 1306 coupled to the control module 1304, a non-volatile memory (NVM) / storage device 1308 coupled to the control module 1304, one or more input / output devices 1310 coupled to the control module 1304, and a network interface 1312 coupled to the control module 1304. The processors 1302 may include one or more single-core or multi-core processors, and may include any combination of general-purpose processors or special-purpose processors (e.g., graphics processors, application processors, baseband processors, etc.). In some embodiments, the device 1300 can function as a server device such as a gateway in the embodiments of this application.

[0131] In some embodiments, device 1300 may include one or more computer-readable media (e.g., memory 1306 or NVM / storage device 1308) having instructions 1314 and one or more processors 1302 configured in conjunction with the one or more computer-readable media to execute instructions 1314 to implement a module thereby performing the actions of this disclosure. In one embodiment, control module 1304 may include any suitable interface controller to provide any suitable interface to at least one of the processors 1302 and / or any suitable device or component communicating with control module 1304. Control module 1304 may include a memory controller module to provide an interface to memory 1306. The memory controller module may be a hardware module, a software module, and / or a firmware module. Memory 1306 may be used, for example, to load and store data and / or instructions 1314 for device 1300. In one embodiment, memory 1306 may include any suitable volatile memory, such as suitable DRAM. In some embodiments, memory 1306 may include double data rate quad synchronous dynamic random access memory (DDR4 SDRAM).

[0132] In one embodiment, control module 1304 may include one or more input / output controllers to provide an interface to NVM / storage device 1308 and (one or more) input / output devices 1310. For example, NVM / storage device 1308 may be used to store data and / or instructions 1314. NVM / storage device 1308 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable (one or more) non-volatile storage devices (e.g., one or more hard disk drive (HDD), one or more optical disc (CD) drives, and / or one or more digital universal optical disc (DVD) drives). NVM / storage device 1308 may include storage resources that are physically part of a device on which device 1300 is mounted, or that are accessible by that device but do not necessarily have to be part of that device. For example, NVM / storage device 1308 may be accessed via a network through (one or more) input / output devices 1310.

[0133] One or more input / output devices 1310 may provide an interface for device 1300 to communicate with any other suitable device. Input / output devices 1310 may include communication components, pinyin components, sensor components, etc. A network interface 1312 may provide an interface for device 1300 to communicate via one or more networks. Device 1300 may wirelessly communicate with one or more components of a wireless network according to any standard and / or protocol of one or more wireless network standards and / or protocols, such as accessing wireless networks based on communication standards, such as WiFi, 2G, 3G, 4G, 5G, etc., or combinations thereof. In one embodiment, at least one of the processors 1302 may be logically packaged with one or more controllers (e.g., memory controller modules) of control module 1304. In one embodiment, at least one of the processors 1302 may be logically packaged with one or more controllers of control module 1304 to form a system-in-package (SiP). In one embodiment, at least one of the processors 1302 may be logically integrated with one or more controllers of control module 1304 on the same die. In one embodiment, at least one of the processors 1302 may be integrated with the logic of one or more controllers of the control module 1304 on the same die to form a system-on-a-chip (SoC).

[0134] In various embodiments, device 1300 may be, but is not limited to, a server, desktop computing device, or mobile computing device (e.g., laptop computing device, handheld computing device, tablet computer, netbook, etc.). In various embodiments, device 1300 may have more or fewer components and / or different architectures. For example, in some embodiments, device 1300 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including a touchscreen display), a non-volatile memory port, multiple antennas, a graphics chip, an application-specific integrated circuit (ASIC), and a speaker.

[0135] This application provides an electronic device, including: one or more processors; and one or more machine-readable media having instructions stored thereon, which, when executed by the one or more processors, cause the electronic device to perform one or more methods as described in this application.

[0136] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0137] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0138] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable information processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable information processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable information processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable information processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. Although preferred embodiments of the embodiments of this application have been described, those skilled in the art, once they learn the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of this application.

[0139] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element. The training method, matching degree acquisition method, matching method, and apparatus of the matching model provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application; at the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for training a matching model, characterized in that, The method includes: Acquire at least one first training data and an encoding network to be trained, wherein the first training data includes event information of a first sample event in a first domain; The parameters of the encoding network to be trained are adjusted using at least one first training data to obtain the encoding network of the first domain; Acquire at least one second training data and create a matching model that includes an encoding network of the first domain; the second training data includes: event information of the second sample event in the first domain, object information of the sample object in the first domain, and the annotation matching degree between the second sample event in the first domain and the sample object in the first domain; The parameters in the matching model, which includes the encoding network of the first domain, are trained using at least one second training data until the parameters in the matching model, which includes the encoding network of the first domain, converge, thereby obtaining the first matching model. Acquire at least one third training data and create a matching model that includes the first matching model; the third training data includes: event information of sample events in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the labeled matching degree between sample events in the target second domain and sample objects in the target second domain; The parameters of the matching model, which includes the first matching model, are adjusted using at least one third training data to obtain the second matching model.

2. The method according to claim 1, characterized in that, The matching model, which includes the encoding network of the first domain, includes: the encoding network of the first domain and the matching network; The encoding network of the first domain is used to encode the event information of the second sample event in the first domain to obtain the event features of the second sample event in the first domain, and to encode the object information of the sample object in the first domain to obtain the object features of the sample object in the first domain. The matching network is used to obtain the matching degree between the second sample event in the first domain and the sample object in the first domain based on the event characteristics of the second sample event in the first domain and the object characteristics of the sample object in the first domain.

3. The method according to claim 2, characterized in that, The matching network includes a feature processing layer, a feature concatenation layer, and a prediction layer; The feature operation layer is used to perform cross-operations on the event features of the second sample event in the first domain and the object features of the sample object in the first domain to obtain the operation features; The feature concatenation layer is used to concatenate the operational features, the event features of the second sample events in the first domain, and the object features of the sample objects in the first domain to obtain the concatenated features. The prediction layer is used to predict the matching degree between the second sample event in the first domain and the sample object in the first domain based on the splicing features.

4. The method according to claim 1, characterized in that, The matching model, including the first matching model, includes: an auxiliary feature extraction network, a first-domain encoding network, and a matching network; The auxiliary feature extraction network is used to extract auxiliary information associated with the event information of sample events in the target second domain and the object information of sample objects in the target second domain, and to encode the auxiliary information to obtain auxiliary features. The auxiliary information includes: keywords in the event information of sample events in the target second domain, the character overlap rate between the event information of sample events in the target second domain and the object information of sample objects in the target second domain, the word overlap rate between the event information of sample events in the target second domain and the object information of sample objects in the target second domain, and the hyper-vocabulary and hypo-vocabulary of the event information of sample events in the target second domain. The encoding network of the first domain is used to encode the event information of the sample events in the target second domain to obtain the event features of the sample events in the target second domain, and to encode the object information of the sample objects in the target second domain to obtain the object features of the sample objects in the first domain. The matching network is used to obtain the matching degree between sample events and sample objects in the second target domain based on auxiliary features, event features of sample events in the second target domain, and object features of sample objects in the second target domain.

5. The method according to claim 4, characterized in that, The matching network includes a feature processing layer, a feature concatenation layer, and a prediction layer; The feature operation layer is used to perform cross-operations on the event features of sample events in the second domain of the target and the object features of sample objects in the second domain of the target to obtain the operation features; The feature concatenation layer is used to concatenate the operational features, the event features of sample events in the second domain of the target, the object features of sample objects in the second domain of the target, and auxiliary features to obtain the concatenated features; The prediction layer is used to predict the matching degree between sample events in the second domain of the target and sample objects in the second domain of the target, based on the splicing features.

6. A method for obtaining matching degree, characterized in that, The method includes: Obtain event information of the target event in the second domain of the target, as well as object information of the target object in the second domain of the target; Based on the event information of the target event in the second domain, the object information of the target object in the second domain, and the trained second matching model, the matching degree between the target event and the target object is obtained. The second matching model is obtained by adjusting the parameters of a matching model including the first matching model using at least one third training data. The third training data includes: event information of sample events in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the labeled matching degree between sample events and sample objects in the target second domain. The first matching model is obtained by training the parameters of a matching model including the encoding network of the first domain using at least one second training data until the parameters of the matching model including the encoding network of the first domain converge. The second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the labeled matching degree between second sample events and sample objects in the first domain. The encoding network of the first domain is obtained by adjusting the parameters of the encoding network to be trained using at least one first training data, and the first training data includes event information of first sample events in the first domain.

7. A matching method, characterized in that, The method includes: Obtain the event information of the target event, including the title of the target event and the descriptive text of the target event; Retrieve object information for multiple objects to be matched. The object information includes: the object's name and the object's classification structure. Obtain keywords to describe the subject matter of the target event based on the event information; Among multiple objects to be matched, the object information is determined to include the first object containing the keyword; The sorting order of the first object is set to be before the second object among multiple objects to be matched, excluding the first object; Select at least one object that matches the target event from among the multiple objects to be matched, according to the sorting order among the multiple objects to be matched.

8. The method according to claim 7, characterized in that, The method further includes: When there are two or more first objects, for any one first object, based on the event information of the target event and the object information of the first object, obtain the text matching degree between the target event and the first object, obtain the scene matching degree between the target event and the first object, and obtain the keyword matching degree between the target event and the first object; obtain the matching score between the target event and the first object based on the text matching degree, scene matching degree, and keyword matching degree, and sort the two or more first objects in descending order of their matching scores with the target event; And / or, When there are two or more second objects, for any one second object, based on the event information of the target event and the object information of the second object, obtain the text matching degree between the target event and the second object, obtain the scene matching degree between the target event and the second object, and obtain the keyword matching degree between the target event and the second object; obtain the matching score between the target event and the second object based on the text matching degree, scene matching degree, and keyword matching degree, and sort the two or more second objects in descending order of their matching scores with the target event.

9. A training device for a matching model, characterized in that, The device includes: The first acquisition module is used to acquire at least one first training data and an encoding network to be trained, wherein the first training data includes event information of a first sample event in a first domain; The first adjustment module is used to adjust the parameters of the encoding network to be trained using at least one first training data to obtain the encoding network of the first domain. The second acquisition module is used to acquire at least one second training data and create a matching model including a coding network of the first domain; the second training data includes: event information of the second sample event of the first domain, object information of the sample object of the first domain, and the annotation matching degree between the second sample event of the first domain and the sample object of the first domain; A training module is used to train the parameters in a matching model that includes a first-domain encoding network using at least one second training data until the parameters in the matching model that includes the first-domain encoding network converge, thereby obtaining a first matching model. The third acquisition module is used to acquire at least one third training data and create a matching model including the first matching model; the third training data includes: event information of sample events in multiple second domains cascaded in the first domain, object information of sample objects in the target second domain, and the labeled matching degree between sample events in the target second domain and sample objects in the target second domain. The second adjustment module is used to adjust the parameters of the matching model, which includes the first matching model, using at least one third training data, so as to obtain the second matching model.

10. An apparatus for obtaining a matching degree, characterized in that, The device includes: The fourth acquisition module is used to acquire event information of target events in the second domain of the target and object information of target objects in the second domain of the target; The fifth acquisition module is used to acquire the matching degree between the target event and the target object based on the event information of the target event in the second domain, the object information of the target object in the second domain, and the trained second matching model. The second matching model is obtained by adjusting the parameters of a matching model including the first matching model using at least one third training data. The third training data includes: event information of sample events in multiple second domains cascaded from the first domain, object information of sample objects in the target second domain, and the labeled matching degree between sample events and sample objects in the target second domain. The first matching model is obtained by training the parameters of a matching model including the encoding network of the first domain using at least one second training data until the parameters of the matching model including the encoding network of the first domain converge. The second training data includes: event information of second sample events in the first domain, object information of sample objects in the first domain, and the labeled matching degree between second sample events and sample objects in the first domain. The encoding network of the first domain is obtained by adjusting the parameters of the encoding network to be trained using at least one first training data, and the first training data includes event information of first sample events in the first domain.

11. A matching device, characterized in that, The device includes: The sixth acquisition module is used to acquire event information of the target event, including the title of the target event and the description text of the target event; The seventh acquisition module is used to acquire object information of multiple objects to be matched. The object information includes: the name of the object and the classification structure of the object. The eighth acquisition module is used to acquire keywords that describe the theme of the target event based on the event information of the target event; The determining module is used to determine, among multiple objects to be matched, a first object whose object information includes the keyword; The settings module is used to set the sorting order of the first object to be placed before the second object (excluding the first object) among multiple objects to be matched. The selection module is used to select at least one object that matches the target event from among a plurality of objects to be matched, according to the sorting order among the plurality of objects to be matched.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the steps of the method as claimed in any one of claims 1 to 8 when executing the program.

13. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as claimed in any one of claims 1 to 8.

Images