CN115114397A

CN115114397A - Annuity information updating method, device, electronic device, storage medium, and program

Info

Publication number: CN115114397A
Application number: CN202210499645.9A
Authority: CN
Inventors: 闵际达; 郭瑞; 郭计雄; 申世豪
Original assignee: Taikang Insurance Group Co Ltd; Taikang Pension Insurance Co Ltd
Current assignee: Taikang Insurance Group Co Ltd; Taikang Pension Insurance Co Ltd
Priority date: 2022-05-09
Filing date: 2022-05-09
Publication date: 2022-09-27
Anticipated expiration: 2042-05-09
Also published as: CN115114397B

Abstract

The application provides a method, a device, an electronic device, a storage medium and a program for updating annuity information, and belongs to the technical field of computers. The method is applied to an annuity system and comprises the following steps: acquiring information to be updated for updating the record information of the target object; inputting the dependency syntax analysis result and the keyword co-occurrence feature of the information to be updated into a machine information understanding model to obtain a main semantic feature of the information to be updated; comparing the main semantic features of the information to be updated with the main semantic features of the recorded information to obtain main semantic feature similarity between the recorded information and the information to be updated; when the main semantic similarity is larger than the similarity threshold, inputting the semantic word vector of the information to be updated into the attention model to obtain the attention score of the information to be updated; and when the attention score meets the score requirement, updating the record information based on the information to be updated. The effectiveness and integrity of the updated recording information can be effectively improved.

Description

Annuity information updating method, annuity information updating device, electronic device, storage medium, and program

Technical Field

The present application relates to the field of computer technologies, and in particular, to a method, an apparatus, an electronic device, a storage medium, and a program for updating annuity information.

Background

At present, more and more enterprise systems record the relevant information and the client information of the enterprise, and update the enterprise information and the client information at regular time to ensure the real-time performance of the information. When the business department provides service for the client, the business department can analyze the service according to the stored client information or enterprise information, thereby providing the service which is more in line with the requirement of the client.

For example, a business or professional annuity system may record business information and customer information to provide an annuity service that is tailored to the customer or business's actual situation based on the recorded business information and customer information. Thus, in order to ensure the quality of service, it is desirable to have business information and customer information as complete and efficient as possible. The annuity system in the related art will regularly update the enterprise information and the customer information to ensure the real-time performance of the recorded information, and usually the latest entered information is adopted to cover the original information to realize information update, that is, the system will store the enterprise information and the customer information with the latest entered time by default. Although timely updating of information can be achieved in this way, validity of the updated information is difficult to guarantee, for example, address information newly retained by a certain enterprise is xx city, but address information retained before is xx city xx road, obviously the address of the enterprise is not changed, but obviously the address information retained before is more complete, and due to a mechanism for storing information entered nearby in the related art, the enterprise information and the client information stored in the annuity system have missing or invalid records.

Disclosure of Invention

The application provides an annuity information updating method and device, electronic equipment, a storage medium and a program.

An annuity information updating method applied to an annuity system comprises the following steps:

acquiring information to be updated for updating the record information of the target object;

inputting the dependency syntax analysis result and the keyword co-occurrence feature of the information to be updated into a machine information understanding model to obtain a main semantic feature of the information to be updated;

comparing the main semantic features of the information to be updated with the main semantic features of the recorded information to obtain main semantic feature similarity between the recorded information and the information to be updated, wherein the main semantic features of the recorded information are obtained in advance through the machine information understanding model;

when the main body semantic similarity is larger than a similarity threshold value, inputting the semantic word vector of the information to be updated into an attention model to obtain an attention score of the information to be updated;

and when the attention score meets the score requirement, updating the record information based on the information to be updated.

Optionally, the machine information understanding model includes at least: an input layer, a coding layer and a matching layer; the step of inputting the dependency syntax analysis result of the information to be updated and the keyword co-occurrence feature into a machine information understanding model to obtain the main semantic feature of the information to be updated comprises the following steps:

acquiring corresponding problem information according to the information type of the information to be updated;

inputting the question information and the information to be updated into the input layer, wherein the input layer is used for respectively extracting a question dependence syntactic analysis result and a question keyword co-occurrence characteristic of the question information, and an information syntactic analysis result and an information keyword co-occurrence characteristic of the information to be updated;

inputting the question dependency syntax analysis result, the question keyword co-occurrence feature, the information syntax analysis result and the information keyword co-occurrence feature into a coding layer respectively, wherein the coding layer is used for fusing the question dependency syntax analysis result, the question keyword co-occurrence feature, the information syntax analysis result and the information keyword co-occurrence feature into a fused semantic feature;

and inputting the fused semantic features into the matching layer, wherein the matching layer is used for acquiring main semantic features of answer information in the fused semantic features.

Optionally, the coding layer comprises at least: a multi-headed self-attention network and a fully-connected feedforward network;

the inputting the question dependency syntactic analysis result, the question keyword co-occurrence feature, the information syntactic analysis result and the information keyword co-occurrence feature into the coding layer respectively comprises:

inputting the question dependency syntactic analysis result, the question keyword co-occurrence feature, the information syntactic analysis result and the information keyword co-occurrence feature into the multi-head self-attention network, wherein the multi-head self-attention network is used for performing projection through multiple linear transformations so as to splice the question dependency syntactic analysis result, the question keyword co-occurrence feature, the information syntactic analysis result and the information keyword co-occurrence feature to obtain a spliced semantic feature,

inputting the spliced semantic features to the fully-connected feed-forward network through a residual error network, wherein the fully-connected feed-forward network is used for extracting fused semantic features from the spliced semantic features;

performing post-processing on the fusion semantic features through a regularization layer, and then optionally outputting, wherein the attention model includes: an embedding layer, an encoding layer, an attention layer and an output layer;

the inputting the semantic word vector of the information to be updated into an attention model to obtain the attention score of the information to be updated includes:

inputting the information to be updated into the embedding layer, wherein the embedding layer is used for inquiring the sememe in the information to be updated in a preset sememe library and marking the sememe in the information to be updated;

inputting the information to be updated marked with the sememe into the coding layer, wherein the coding layer is used for extracting sememe word vectors from the information to be updated;

inputting the semantic word vector into the attention layer, wherein the attention layer is used for acquiring a similarity score between the noticed semantic word vector and an attention object through an attention function, and carrying out weighted summation on the semantic word vector and the similarity score based on a preset weight to obtain an attention word vector;

inputting the attention word vector into an output layer, wherein the output layer is used for determining answer information corresponding to a preset question in the attention word vector and an attention score of the answer information.

Optionally, the attention layer comprises at least: a first attention layer and a second attention layer;

the inputting the semantic word vector to the attention layer comprises:

inputting the semantic word vector to the first attention layer, wherein the first attention layer is used for extracting a first attention word vector from low-depth features of the semantic word vector;

and recoding the first attention word vector and inputting the recoded first attention word vector into the second attention layer, wherein the second attention layer is used for extracting a second attention word vector from the high-depth features of the semantic word vector to serve as a final output attention word vector.

Optionally, after the comparing the main semantic features of the information to be updated with the main semantic features of the recorded information to obtain the main semantic feature similarity between the recorded information and the information to be updated, the method further includes:

and updating the record information based on the information to be updated when the similarity of the main semantic features is smaller than or equal to a similarity threshold.

Optionally, when the attention score meets a score requirement, the record information is updated based on the information to be updated, and the method further includes:

and replacing partial information which belongs to the same information type as the partial information to be updated in the recorded information by utilizing the partial information to be updated with the highest attention score in the information to be updated.

Some embodiments of the present application provide an annuity information updating apparatus, which is applied to an annuity system, and the apparatus includes:

the acquisition module is configured to acquire information to be updated for updating the record information of the target object;

the processing module is configured to input the dependency syntax analysis result and the keyword co-occurrence feature of the information to be updated into a machine information understanding model to obtain a main semantic feature of the information to be updated;

an updating module configured to update the record information based on the information to be updated when the attention score meets a score requirement.

Optionally, the machine information understanding model includes at least: an input layer, a coding layer and a matching layer; the processing module further configured to:

respectively inputting the problem dependency syntax analysis result, the problem keyword co-occurrence feature, the information syntax analysis result and the information keyword co-occurrence feature into a coding layer, wherein the coding layer is used for fusing the problem dependency syntax analysis result, the problem keyword co-occurrence feature, the information syntax analysis result and the information keyword co-occurrence feature into a fused semantic feature;

Optionally, the coding layer includes at least: a multi-headed self-attention network and a fully-connected feedforward network;

the processing module further configured to:

inputting the information to be updated marked with the sememe to the coding layer, wherein the coding layer is used for extracting sememe word vectors from the information to be updated;

inputting the semantic word vector to the attention layer, wherein the attention layer is used for acquiring a similarity score between the noticed semantic word vector and an attention object through an attention function, and performing weighted summation on the semantic word vector and the similarity score based on a preset weight to obtain an attention word vector;

the processing module further configured to:

and re-encoding the first attention word vector and inputting the re-encoded first attention word vector to the second attention layer, wherein the second attention layer is used for extracting a second attention word vector from the high-depth features of the semantic word vector to serve as the finally output attention word vector.

Optionally, the update module is further configured to:

and when the similarity of the main semantic features is smaller than or equal to a similarity threshold, updating the recorded information based on the information to be updated.

Optionally, the update module is further configured to:

Some embodiments of the present application provide a computing processing device comprising:

a memory having computer readable code stored therein;

one or more processors which, when executed by the computer readable code, perform the annuity information update method described above.

Some embodiments of the present application provide a computer program comprising computer readable code which, when run on a computing processing device, causes the computing processing device to perform an annuity information update method as described above.

Some embodiments of the present application provide a non-transitory computer readable medium in which the annuity information update method as described above is stored.

According to the annuity information updating method, the device, the electronic equipment, the storage medium and the program, in the recorded information updating process, after information with unobvious information change is identified by combining a machine information understanding model with dependency syntax analysis and keyword co-occurrence, the recorded information is updated by extracting the content with the attention score meeting the requirement from the information to be updated through the attention model with the sense original characteristics, and the integrity and the effectiveness of the recorded information of a target object are guaranteed.

The foregoing description is only an overview of the technical solutions of the present application, and the present application can be implemented according to the content of the description in order to make the technical means of the present application more clearly understood, and the following detailed description of the present application is given in order to make the above and other objects, features, and advantages of the present application more clearly understandable.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative efforts.

Fig. 1 schematically illustrates a flowchart of an annuity information updating method provided in some embodiments of the present application;

FIG. 2 schematically illustrates a flow chart of another annuity information update method provided by some embodiments of the present application;

FIG. 3 schematically illustrates one of the principles of another annuity information update method provided by some embodiments of the present application;

FIG. 4 schematically illustrates a second schematic diagram of another annuity information updating method provided by some embodiments of the present application;

FIG. 5 schematically illustrates a third schematic diagram of another annuity information update method provided by some embodiments of the present application;

FIG. 6 is a schematic flow chart diagram illustrating yet another annuity information update method provided in some embodiments of the present application;

FIG. 7 schematically illustrates one of the principles of yet another annuity information update method provided by some embodiments of the present application;

FIG. 8 schematically illustrates a second schematic diagram of another annuity information updating method provided by some embodiments of the present application;

FIG. 9 is a schematic diagram illustrating an annual fund information updating apparatus according to some embodiments of the present disclosure;

FIG. 10 schematically illustrates a block diagram of a computing processing device for performing a method according to some embodiments of the present application;

FIG. 11 schematically illustrates a memory unit for holding or carrying program code implementing methods according to some embodiments of the present application.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are some embodiments of the present application, but not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

An annuity system of an enterprise in the related art records enterprise information and customer information to provide an annuity service adapted to a customer or an actual situation of the enterprise based on the recorded enterprise information and customer information. Thus, in order to ensure the quality of service, it is desirable to have business information and customer information as complete and efficient as possible. The annuity system in the related art will regularly update the enterprise information and the client information to ensure the real-time performance of the recorded information, and usually update the information by using the latest entered information to cover the original information, that is, the system will store the enterprise information and the client information with the latest entered time by default. Although timely information updating can be achieved in this manner, validity of updated information is difficult to guarantee, for example, address information newly retained by a certain enterprise is xx city, but address information retained before is xx road in xx city, obviously, the address of the enterprise is not changed, but obviously, address information retained before is more complete, and due to a mechanism for storing information recorded nearby in the related art, the enterprise information and client information stored in the annuity system have missing or invalid records.

Therefore, the annuity information updating method is provided to solve the problem that information is missing or invalid due to a mechanism of storing information input nearby when an enterprise annuity system updates the information as far as possible, so that the integrity and the effectiveness of the information in the enterprise annuity system are guaranteed.

Fig. 1 schematically shows a flow chart of an annuity information updating method provided by the present application, which is applied to an annuity system, and the method includes:

step 101, obtaining information to be updated for updating the record information of the target object.

It should be noted that the execution subject of the present application may be a server, a terminal device, or an electronic device that is provided by an enterprise or a professional annuity system and has data processing, data storage, and data transmission, and is used to store and update record information stored in the enterprise annuity system.

In the embodiment of the present application, the target object is a subject object for which information recording is required, and the information recorded in the recording information for the target object, for example, if the target object is a business, a business address, a business name, a business asset, a business corporate law, and the like can be recorded as the recording information, and if the target object is an individual, a name, an address, a contact address, and the like can be recorded as the recording information. The information to be updated refers to information that is obtained latest and has information types that are repeated with information content of the recorded information after the recorded information has been stored for the target object, for example, when the enterprise information is recorded by the enterprise annuity system, the enterprise a has obtained the enterprise information 1 from the information source such as the network and the like once in Y month in X year and recorded as the recorded information, and then has obtained the enterprise information 2 of the enterprise a again from the information source such as the network and the like in Y month in X +1 year, at this time, the enterprise information 2 and the previously recorded enterprise information 1 are both enterprise information of the enterprise a, it is necessary to determine whether to update the enterprise information 1 to the enterprise information 2, and the enterprise information 2 is the information to be updated, which is only an exemplary description here, and may be specifically set according to actual needs, and is not limited here.

In practical applications, the enterprise annuity system records the client information as record information by taking the client as a target object, and can perform periodic updating, such as updating once per week or once per month, and the like, wherein the updating frequency can be flexibly configured according to actual requirements. In the updating process, the latest client information can be acquired by requesting the client to re-input or reading and re-acquiring from various data sources to serve as the information to be updated so as to update the recorded information, but because the information to be updated may have information loss and other problems relative to the recorded information, if the recorded information is replaced by referring to the information to be updated, the problem that the updated recorded information has information loss and invalid recording may occur, and therefore, the integrity and accuracy of the information to be updated need to be further judged before the recorded information is updated in the embodiment of the present application.

And 102, inputting the dependency syntax analysis result of the information to be updated and the keyword co-occurrence characteristics into a machine information understanding model to obtain the main semantic characteristics of the information to be updated.

In the embodiment of the application, the dependency parsing result is a result of parsing based on a dependency parsing mechanism, according to the parsing, only one component in one sentence in the dependency parsing is independent, other components of the sentence are all dependent on a certain component, any one component cannot depend on two or more components, and other components on the left side and the right side of the center component are not related to each other.

The keyword co-occurrence feature is a feature extracted from information to be updated through a keyword co-occurrence mechanism, and it can be understood that the associated feature of the question and the answer is the keyword, and the keyword with the question appears near the position of the answer in the article. For example, in the problem of "XX road XX enterprise in rich district of beijing city", the keyword sequence of the whole problem is [ beijing city, rich district, XX road, XX enterprise ], the first keyword sequence is the most critical word, wherein 'beijing city' is the most critical word, it can be considered that if the city is not beijing city, the information is basically excluded, the later the more the critical is, the weaker the more the later the critical is, the threshold is set to 3, that is, the first 3 keywords can be taken as co-occurrence keyword features.

Optionally, the machine information understanding model includes at least: input layer, coding layer and matching layer, referring to fig. 2, the step 102 may include:

step 1021, obtaining corresponding question information according to the information type of the information to be updated.

Step 1022, inputting the question information and the information to be updated to the input layer, where the input layer is configured to extract a question dependency syntax analysis result and a question keyword co-occurrence feature of the question information, and an information syntax analysis result and an information keyword co-occurrence feature of the information to be updated, respectively.

And 1023, inputting the question dependency syntax analysis result, the question keyword co-occurrence feature, the information syntax analysis result and the information keyword co-occurrence feature into a coding layer, wherein the coding layer is used for fusing the question dependency syntax analysis result, the question keyword co-occurrence feature, the information syntax analysis result and the information keyword co-occurrence feature into a fused semantic feature.

Step 1024, inputting the fused semantic features into the matching layer, where the matching layer is used to obtain main semantic features of answer information in the fused semantic features.

In the embodiment of steps 1021 to 1024 of the present application, referring to fig. 3, where PrtNet (PionterNetworks) is a matching layer, Merge is a fusion layer, Self-attention is a Self-attention feature, CNN (Convolutional Neural Network) is a Convolutional layer, Encoder is an Encoder, begin (bidirectional Encoder registration from transformations) is a pre-trained language characterization model, Question is problem information, Context is text information, i.e., information to be updated in the embodiment of the present application, start is a start node of a main semantic feature, and end is an end node of the main semantic feature.

The machine information understanding model can be obtained by fusing the dependency syntactic analysis result and the keyword co-occurrence characteristics into a pre-trained language representation model, the semantic characteristic extraction model can be a Bert model of residual training, the pre-trained model can greatly improve the effect of a natural language processing task, the pre-trained model can be directly applied to the current task, the defect of insufficient training corpus is overcome, and the convergence rate of the model can be increased. The pre-training model is combined with dependency syntactic analysis information (dependency) and keyword co-occurrence characteristics (keyword co-occurence), and mainly comprises an input layer, a coding layer and a matching layer. The input layer vectorizes the questions and the information to be updated and extracts features, the coding layer fuses text information and the features, the matching layer searches answers corresponding to the questions in the information to be updated and outputs answer intervals in the information to be updated, and the questions are preset for the information to be updated of different data types.

Optionally, the coding layer comprises at least: the multi-head self-attention network and the fully-connected feedforward network, the step 1023, may include:

a1, inputting the question dependency syntactic analysis result, the question keyword co-occurrence feature, the information syntactic analysis result and the information keyword co-occurrence feature into the multi-head self-attention network, wherein the multi-head self-attention network is used for projecting through multiple linear transformations so as to splice the question dependency syntactic analysis result, the question keyword co-occurrence feature, the information syntactic analysis result and the information keyword co-occurrence feature to obtain a spliced semantic feature,

a2, inputting the splicing semantic features to the fully-connected feed-forward network through a residual error network, wherein the fully-connected feed-forward network is used for extracting fusion semantic features from the splicing semantic features;

and A3, performing post-processing on the fusion semantic features through a regularization layer, and then outputting.

In the embodiment of the present application, refer to fig. 4, where Add is a residual network layer and Norm is a regularization layer. The coding layer at this stage is used for calculating the context perception of the problem and the information to be updated, and in the process of reading the information, the coding layer can use a self-attention model to fuse article semantics and problem semantics. And traversing the text of the information to be updated to obtain an area needing important attention, namely an attention focus, and then putting more attention resources into the area to obtain more detailed information of the target needing attention. Entering regularization layer normalization, the coding layer has two sublayers: the first sub-layer is multi-head self-attention, the second sub-layer is a fully-connected feedforward network, the two sub-layers are connected through a residual error network structure, and then a regularization layer is connected. The multi-head attention is projected through multiple linear transformations, and then different features are spliced together to form final word vector representation.

Referring to fig. 5, where start is a start node of the main body semantic features of the output answer information, end is an end node of the main body semantic features of the output answer information, PrtNet is a matching layer, Merge is a fusion layer, Self-attribute is an attention feature, Hb. And outputting the dependency syntax analysis result and the keyword co-occurrence characteristics for the convolutional layer. According to the embodiment of the application, the idea of pointer-network (pointer network) is used as a reference for a matching layer, context analysis is carried out on information to be updated, answer probability p which is matched with a problem in fineness to be updated and belongs to [0,1] is calculated, and the probability determines a start node start and an end node end.

The multi-level attention model in the embodiment of the application is based on an LSTM (Long Short-Term Memory neural network) encoder, and meanwhile, a global attention mechanism is carried out, so that the encoding process of the model is improved, and the model is richer and more accurate. The multi-level attention is to perform attention matching on the problem and the article at different depths, the vector semantic information characteristics of different levels are different, the attention between different levels of semantics can be acquired through the multi-level attention, and single-level attention deviation errors are reduced.

The machine information understanding model improved through dependency syntactic analysis and keyword co-occurrence is based on a pre-trained semantic representation model, the pre-trained semantic representation model can greatly improve the effect of natural language processing tasks, the pre-trained model can be directly applied to the current tasks, the defect of insufficient training corpora is overcome, and the convergence rate of the model can be accelerated. In reading long text, especially long text, the core keywords are easily covered by extraneous words due to excessive attention to some extraneous words, and noise interference of these extraneous words plus long distance correlation span between words of the long text makes the model obscure attention, resulting in error propagation and deviation of results. The dependency syntactic analysis can reflect the modification relation among all components of the sentence, the central word after the dependency syntactic analysis is independent of the physical position of the components of the sentence, the association information among long-distance words is obtained, the most important association characteristic of the question and the article is the key word in the question, the predicted answer span type corresponds to the answer type given by the question, and the correct answer is in the occurrence range of the key word of the question. The dependency syntactic characteristics and the keyword co-occurrence characteristics are added into the model, so that the attention of the model to key keywords is improved, and the model is richer and more accurate.

Step 103, comparing the main semantic features of the information to be updated with the main semantic features of the recorded information to obtain the main semantic feature similarity between the recorded information and the information to be updated, wherein the main semantic features of the recorded information are obtained in advance through the machine information understanding model.

In the embodiment of the present application, since the record information is information that has been acquired in advance, if the record information is also updated by the annuity information updating method in the embodiment of the present application when the record information is acquired, the main semantic features are acquired in advance through the processing of the machine information understanding model, and the main semantic features of the record information can be stored and then directly acquired in the information updating, the record information does not need to be input to the machine information understanding model for processing, and the data processing amount required in the information updating process is reduced.

The main semantic feature similarity obtained by comparing the main semantic information of the information to be updated with the recorded information can represent whether the content between the information to be updated and the recorded information is changed, and for example, a similarity threshold of 90%, 95% or the like is set as a reference, that is, if the main semantic feature similarity is greater than the similarity threshold, the content of the information to be updated is considered to be unchanged compared with the content of the recorded information, and if the main semantic feature similarity is less than or equal to the similarity threshold, the content of the information to be updated is considered to be changed compared with the content of the recorded information. For the information to be updated whose content has not changed, that is, all the content has not changed, the department content in the information to be updated may still be changed compared with the record information, for example, for the enterprise information, only the enterprise legal person in the enterprise information may have changed, and at this time, it is necessary to determine the partial information that actually changes from the information to be updated to update the record information.

And 104, when the semantic similarity of the main body is greater than a similarity threshold, inputting the semantic word vector of the information to be updated into an attention model to obtain an attention score of the information to be updated.

In the embodiment of the application, the information to be updated is read and understood through the multilayer assistant model fused with the sense original information, semantic information features of different depths are obtained, and which information can be used for updating the recorded information is judged according to the attention scores of all parts of information in the information to be updated.

Optionally, the attention model comprises: an embedding layer, an encoding layer, an attention layer, and an output layer, referring to fig. 6, the step 104 may include:

step 1041, inputting the information to be updated to the embedding layer, where the embedding layer is configured to query an semantic source in the information to be updated in a preset semantic source library, and mark the semantic source in the information to be updated.

Step 1042, inputting the information to be updated marked with the semantic element to the coding layer, where the coding layer is configured to extract a semantic element word vector from the information to be updated.

Step 1043, inputting the primitive word vector to the attention layer, where the attention layer is configured to obtain a similarity score between the noticed primitive word vector and an attention object through an attention function, and perform weighted summation on the primitive word vector and the similarity score based on a preset weight to obtain an attention word vector.

Step 1044, inputting the attention word vector into an output layer, where the output layer is configured to determine answer information corresponding to a preset question in the attention word vector and an attention score of the answer information.

In an embodiment of the present application, steps 1041 through 1044, refer to fig. 7, wherein Softmax represents an activation function that may map the output of a plurality of neurons into a (0, 1) interval. Wherein the multi-layer attention model mainly comprises an embedding layer, an encoding layer, a multi-layer attention layer and an output layer. The SAT model framework is a combination of each Word meaning as its sememe, performs Word meaning disambiguation according to its context iteration, and learns the sememe, Word meaning and Word representation by extending Skip-Gram in Word2 vec. The SAT model considers both context information and semantic information of words, and the purpose of fusing the semantic information is to make the model easier and better understand the meaning of the words. The method comprises the following steps of firstly analyzing the information of context words so as to better disambiguate the words in the center in a word sense, calculating a weight value of the context information of the words on each semantic of the words by using an attention mechanism by a model, finally weighting the vectors of the word senses, and then calculating the average value of the vectors to represent the words, so that the information of the semantic is fused into the representation learning of the words, and the representation effect of the word vectors is further improved by the model with the fused semantic.

The coding layer uses bidirectional LSTM to extract semantic features of articles and problems, the LSTM adds a gating mechanism and memory cells in a circulating neural network, and the gating mechanism interacts with the states of the memory cells to change information carried by the memory cells. The gate control mechanism in the cyclic neural network adopts a Sigmoid activation function, the introduction of the gate control units effectively relieves the problem of memory loss, and the problem of gradient disappearance of the traditional cyclic neural network is solved.

Specifically, referring to fig. 8, the gating mechanism of the LSTM unit includes a forgetting gate, forget, an input gate, and an output gate. The input gate input get i structure controls which information needs to be updated and stored in the memory cell at the current moment; the structure of a forgetting gate forget gate f controls how much information is forgotten in the memory cell Ct-1 at the previous moment and how much information is reserved to the current moment; the output gate o structure controls how much information in the memory cell needs to be output.

Optionally, the attention layer comprises at least: the first attention layer and the second attention layer, step 1043, may include:

b1, inputting the semantic word vector to the first attention layer, wherein the first attention layer is used for extracting a first attention word vector from low-depth features of the semantic word vector;

and B2, recoding the first attention word vector and inputting the recoded first attention word vector into the second attention layer, wherein the second attention layer is used for extracting a second attention word vector from the high-depth feature of the semantic word vector as a final output attention word vector.

In the embodiment of the application, the multi-level attention mechanism uses the attention mode in the BiDaf model for context interaction. And respectively encoding the vector of the information to be updated and the vector of the problem by respective bidirectional LSTM encoders to obtain respective context semantic information H and U, wherein the output vector of the BilTM layer needs to be input to an attention mechanism layer for further feature extraction. When a model inputs a long text sequence, if the sequence information is encoded as a vector and then attention matching is performed using the encoded vector, the position of the answer is often not found. Assuming that an original sentence contains 10 words, when a Model processes a first word of a question, the Model may have the maximum relationship with the first word of an article and may also be related to the 10 th word, which means that when a matching answer is made, the Model needs to match information of the 1 st step and the 10 th step, although a long and short memory neural network can alleviate the problem to some extent, in actual use, a single-layer Attention mechanism cannot be selected according to similar semantic information, the number of network layers needs to be deepened, the Model can learn different semantic information, and a multi-layer Attention mechanism (Attention Model) is formed based on the theory.

The attention mechanism assigns different attention degrees according to different scores of the attention object on the attention object, and further assigns different attention levels, namely, multiple levels of attention layers, wherein the input of the attention mechanism comprises the attention object (a group of vectors Y) and an attention object (a vector x). The vector x needs to pay attention to each object in the Y, and the distribution of each attention weight value may be different, the magnitude of the attention weight value depends on the score given to the attention object Y by the vector x, and a higher score indicates that more attention needs to be allocated. The scoring mode is to score by using an attention function, namely the similarity of two vectors, such as an inner product function, and finally, the score is normalized by Softmax and is calculated, weighted and summed with the object to be noticed, so as to obtain an attention vector C. The multi-level attention is to perform attention matching on the problem and the article at different depths, the vector semantic information characteristics of different levels are different, the attention between different levels of semantics can be acquired through the multi-level attention, and single-level attention deviation errors are reduced.

For example: if the attention object is the vector representation of the word "Beijing", and the attention object is the vector representation of the word "Beijing city | Torpedo | XX way | XX Enterprise", the attention mechanism result is 0.9 × E ("Beijing City") +0.06 × E ("Torpedo zone") +0.03 × E ("XX way") +0.01 × E ("XX Enterprise").

And 105, updating the record information based on the information to be updated when the attention score meets the score requirement.

In this embodiment of the application, the score requirement may refer to that the attention score is greater than a score threshold, or the attention score is within a preset score range, or the attention score has a maximum value, and may be specifically set according to an actual requirement, which is not limited herein. It can be understood that the larger the attention score value is, the higher the relevance between the information to be updated and the problem information is, so that the attention score can be limited to screen the information to be updated to select the content which needs to participate in the updating of the record information, and the record information is updated by the information to be updated which meets the score requirement.

Optionally, after step 102, the method further comprises: and updating the record information based on the information to be updated when the similarity of the main semantic features is smaller than or equal to a similarity threshold.

In the embodiment of the present application, if the similarity of the main semantic features is less than or equal to the similarity threshold, it can be determined that the information to be updated is changed compared with the recorded information, and the recorded information can be directly replaced by the information to be updated.

Optionally, the step 105 includes: and replacing partial information which belongs to the same information type as the partial information to be updated in the recorded information by utilizing the partial information to be updated with the highest attention score in the information to be updated.

In the embodiment of the present application, the information to be updated may include a plurality of pieces of information to be updated of different information types, for example, the enterprise information may include information types such as a business name, a business address, a business legal person, and the like, and the annuity information updating method provided by the present application may obtain the attention score of the pieces of information to be updated of different information types, so that the pieces of information to be updated of the highest attention score may be used to update the content of the corresponding information type in the recorded information, thereby ensuring the validity of information updating.

After the annuity information updating method provided by the embodiment of the application is applied to an enterprise annuity system, the recording time of the recorded information is not required to be used as a basis for judging whether the information is updated or not, the newly recorded information and the original information are analyzed by combining with a machine information understanding model with dependency syntactic analysis and keyword co-occurrence, the information with unobvious information change is identified, and the information is not simply covered integrally during updating.

For example, using one of the enterprise information classes (enterprise addresses) to request record information as an example, this embodiment sets attention weight according to all levels of an address, the national level is 0.3, the provincial level is 0.3, the city level is 0.3, the district level is 0.04, the road level is 0.02, the street level is 0.02, the house number is 0.01, and the enterprise name is 0.01, this patent will judge whether the address changes according to the model of syntactic analysis and keyword analysis before calculating the score, if the answer probability changes and is less than 95 percent, the address is considered to have changed, all address information before this time node will not remain, if the answer probability does not change or is more than 95 percent, the address semantic does not change, the score is calculated, then, each enterprise address of the same enterprise will be in the above multi-level attention model, and calculating the score according to the weight, wherein the data information with the highest final score and the remaining data information is regarded as the client gold information, namely the record information, and the same steps are used for all the partial classes of the enterprise remaining information to finally obtain all the client gold information.

In the embodiment of the application, in the process of updating the recorded information, after identifying the information with unobvious information change by combining the machine information understanding model with dependency syntactic analysis and keyword co-occurrence, the attention model with the sense original characteristics is fused to extract the content with the attention score meeting the requirement from the information to be updated so as to update the recorded information, so that the integrity and the effectiveness of the recorded information of the target object are ensured.

Fig. 9 schematically shows a structural diagram of an annuity information updating device 20 provided by the present application, which is applied to an annuity system, and the device includes:

an obtaining module 201 configured to obtain information to be updated for updating record information of a target object;

the processing module 201 is configured to input the dependency syntax analysis result of the information to be updated and the keyword co-occurrence feature to a machine information understanding model, so as to obtain a main semantic feature of the information to be updated;

an updating module 203 configured to update the record information based on the information to be updated when the attention score meets a score requirement.

Optionally, the machine information understanding model includes at least: an input layer, a coding layer and a matching layer; the processing module 201 is further configured to:

Optionally, the coding layer comprises at least: a multi-head self-attention network and a fully-connected feed-forward network;

the processing module 201 is further configured to:

the fusion semantic features are post-processed through a regularization layer and then output

Optionally, the attention model comprises: an embedding layer, an encoding layer, an attention layer and an output layer;

the processing module 201 is further configured to:

inputting the semantic word vector to the first attention layer, the first attention layer being configured to extract a first attention word vector from low-depth features of the semantic word vector;

Optionally, the updating module 203 is further configured to:

According to the embodiment of the application, in the process of updating the recorded information, after the information with unobvious information change is identified by combining the dependency syntax analysis and the machine information understanding model with the co-occurrence of the keywords, the content with the attention score meeting the requirement is extracted from the information to be updated by the attention model with the sense original characteristics fused to update the recorded information, and the integrity and the effectiveness of the recorded information of the target object are guaranteed.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

The various component embodiments of the present application may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that a microprocessor or Digital Signal Processor (DSP) may be used in practice to implement some or all of the functionality of some or all of the components in a computing processing device according to embodiments of the application. The present application may also be embodied as apparatus or device programs (e.g., computer programs and computer program products) for performing a portion or all of the methods described herein. Such programs implementing the present application may be stored on a non-transitory computer readable medium or may be in the form of one or more signals. Such a signal may be downloaded from an internet website or provided on a carrier signal or in any other form.

For example, FIG. 10 illustrates a computing processing device in which methods according to the present application may be implemented. The computing processing device conventionally includes a processor 310 and a computer program product or non-transitory computer-readable medium in the form of a memory 320. The memory 320 may be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read only memory), an EPROM, a hard disk, or a ROM. The memory 320 has a storage space 330 for program code 331 for performing any of the method steps of the above-described method. For example, the storage space 330 for the program code may include respective program codes 331 respectively for implementing various steps in the above method. The program code can be read from or written to one or more computer program products. These computer program products comprise a program code carrier such as a hard disk, a Compact Disc (CD), a memory card or a floppy disk. Such a computer program product is typically a portable or fixed storage unit as described with reference to fig. 11. The memory unit may have memory segments, memory spaces, etc. arranged similarly to the memory 320 in the computing processing device of fig. 10. The program code may be compressed, for example, in a suitable form. Typically, the memory unit comprises computer readable code 331', i.e. code that can be read by a processor, such as 310, for example, which when executed by a computing processing device causes the computing processing device to perform the steps of the method described above.

It should be understood that, although the steps in the flowcharts of the figures are shown in order as indicated by the arrows, the steps are not necessarily performed in order as indicated by the arrows. The steps are not performed in the exact order shown and may be performed in other orders unless explicitly stated herein. Moreover, at least a portion of the steps in the flow chart of the figure may include multiple sub-steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, which are not necessarily performed in sequence, but may be performed alternately or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

Reference herein to "one embodiment," "an embodiment," or "one or more embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present application. Moreover, it is noted that instances of the word "in one embodiment" are not necessarily all referring to the same embodiment.

In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The application may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions in the embodiments of the present application.

Claims

1. An annuity information updating method, characterized in that the method comprises:

2. The method of claim 1, wherein the machine information understanding model comprises at least: an input layer, a coding layer and a matching layer; the step of inputting the dependency syntax analysis result of the information to be updated and the keyword co-occurrence feature into a machine information understanding model to obtain the main semantic feature of the information to be updated comprises the following steps:

3. The method according to claim 2, wherein the coding layer comprises at least: a multi-headed self-attention network and a fully-connected feedforward network;

and performing post-processing on the fusion semantic features through a regularization layer, and then outputting.

4. The method of claim 1, wherein the attention model comprises: an embedding layer, an encoding layer, an attention layer and an output layer;

inputting the original word vector to the attention layer, wherein the attention layer is used for acquiring a similarity score between the noticed original word vector and an attention object through an attention function, and performing weighted summation on the original word vector and the similarity score based on a preset weight to obtain an attention word vector;

5. The method of claim 4, wherein the attention layer comprises at least: a first attention layer and a second attention layer;

the inputting the semantic word vector to the attention layer comprises:

6. The method according to claim 1, wherein after comparing the main semantic features of the information to be updated with the main semantic features of the recorded information to obtain main semantic feature similarity between the recorded information and the information to be updated, the method further comprises:

7. The method according to claim 1, wherein the updating the record information based on the information to be updated when the attention score meets a score requirement, the method further comprising:

8. An annuity information updating device applied to an annuity system, the device comprising:

an acquisition module configured to acquire information to be updated for updating the record information of the target object;

when the main semantic similarity is larger than a similarity threshold value, inputting the semantic word vector of the information to be updated into an attention model to obtain an attention score of the information to be updated;

9. A computing processing device, comprising:

a memory having computer readable code stored therein;

one or more processors that, when the computer readable code is executed by the one or more processors, the computing processing device performs the annuity information updating method of any of claims 1-7.

10. A non-transitory computer-readable medium in which a computer program of the annuity information updating method according to any one of claims 1 to 7 is stored.