WO2021174871A1

WO2021174871A1 - Data query method and system, computer device, and storage medium

Info

Publication number: WO2021174871A1
Application number: PCT/CN2020/123092
Authority: WO
Inventors: 王昊; 张乐情; 罗水权; 刘剑; 李果夫
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-09-01
Filing date: 2020-10-23
Publication date: 2021-09-10
Also published as: CN112035645A; CN112035645B

Abstract

The present application relates to the field of big data, and provides a data query method, comprising: receiving a first request carrying a first original statement; classifying the first original statement to generate a sequence annotation result; acquiring a preset self-attention matrix parameter and, on the basis of said self-attention matrix parameter and of the sequence annotation result, generating a plurality of word attention matrices; generating, on the basis of the plurality of word attention matrices, a plurality of intention stream parsing results; querying said intention stream parsing results by means of a knowledge graph library so as to obtain query results. By means of relating multiple-intent words and non-intent words through intent extraction and a self-attention mechanism, the present application effectively resolves the problem of inaccurate query results when multiple intentions are present in an original query statement, thus enhancing query efficiency.

Description

Data query method, system, computer equipment and storage medium

This application claims the priority of the Chinese patent application filed with the Chinese Patent Office on September 1, 2020 with the application number 202010905604.6 and the invention title "Data query method and system", the entire content of which is incorporated into this application by reference.

Technical field

The embodiments of the present application relate to the field of big data technology, and in particular, to a data query method, system, computer device, and storage medium.

Background technique

At present, automatic retrieval based on knowledge graphs based on search engines generally uses entity recognition replacement and text matching. First, named entity recognition is used to find entity words for replacement. For example, when a user queries "what is Yao Ming's height", Then it finds [姚明], replaces [姚明] with [Celebrity], finds [身高], and replaces [身高] with [data], that is, the user sentence is transformed into "What is the [data] of [Celebrity]", Then the sentence is matched with the sample sentence in the back-end database through a text matching algorithm, and it is successfully matched to the intention of "[Celebrity]’s [information]". Finally, the person named "Yao Ming" is queried in the knowledge graph database and the data is "Height". The attribute value can then be returned to the user with the answer.

However, the inventor realizes that the above-mentioned intent identification search query method requires manual maintenance of a large number of sample sentences. For queries with multiple intents, and when the query does not match the template of the sample sentence, the correct query answer cannot be returned. . The query efficiency and accuracy rate are low through the above-mentioned intent recognition search query method.

Summary of the invention

In view of this, the embodiments of the present application provide a data query method, system, computer equipment, and computer readable storage medium, which are used to solve the problem of low query efficiency and low accuracy through the above-mentioned entity recognition and text matching search query. problem.

The embodiments of this application solve the above technical problems through the following technical solutions:

A data query method, including:

Receiving a first request, the first request carrying a first original sentence;

Classify the first original sentence, and generate a sequence labeling result;

Acquiring preset self-attention matrix parameters, and generating multiple word attention matrices based on the self-attention matrix parameters and sequence labeling results;

Generating multiple intent stream analysis results based on the multiple word attention matrices;

Query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.

In order to achieve the foregoing objective, an embodiment of the present application further provides a data query system, including:

A receiving module, configured to receive a first request, the first request carrying a first original sentence;

The sequence labeling module is used to classify the first original sentence and generate sequence labeling results;

The first generation module is configured to obtain preset self-attention matrix parameters, and generate multiple word attention matrices based on the self-attention matrix parameters and sequence labeling results;

The second generation module is configured to generate multiple intent stream analysis results based on the multiple word attention matrices;

The query module is used to query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.

In order to achieve the foregoing objective, an embodiment of the present application further provides a computer device. The computer device includes a memory, a processor, and a computer program stored on the memory and running on the processor, and the processor executes the Perform the following steps in the computer program:

Classify the first original sentence, and generate a sequence labeling result;

In order to achieve the foregoing objective, the embodiments of the present application also provide a computer-readable storage medium having a computer program stored in the computer-readable storage medium, and the computer program may be executed by at least one processor, so that the at least A processor performs the following steps:

Classify the first original sentence, and generate a sequence labeling result;

The data query method, system, computer equipment, and computer-readable storage medium provided by the embodiments of the present application realize the extraction of intentions and the attention mechanism to associate multiple intentional words and non-intentional words through sequence labeling, which effectively solves the multiplicity of original query sentences. The problem of inaccurate query results in the case of intent improves query efficiency.

The following describes the application in detail with reference to the accompanying drawings and specific embodiments, but it is not intended to limit the application.

Description of the drawings

Fig. 1 is a flow chart of the steps of the data query method according to the first embodiment of the application;

2 is a flow chart of the steps of converting the first original sentence into a vector sequence in the data query method according to the first embodiment of the application;

FIG. 3 is a flowchart of the steps of generating sequence labeling results in the data query method according to the first embodiment of the application; FIG.

4 is a flowchart of the steps of generating a plurality of word attention matrices in the data query method according to the first embodiment of the application;

FIG. 5 is a flowchart of steps for generating multiple intent stream analysis results in the data query method according to the first embodiment of the application; FIG.

6 is a flow chart of the steps of querying through the knowledge graph database in the data query method of the first embodiment of the application;

FIG. 7 is a flowchart of the steps of training a sequence labeling task model in the data query method according to the first embodiment of the application;

FIG. 8 is a schematic diagram of program modules of the data query system according to the second embodiment of the application;

FIG. 9 is a schematic diagram of the hardware structure of a computer device according to the third embodiment of the application.

Detailed ways

In order to make the purpose, technical solutions, and advantages of this application clearer and clearer, the following further describes the application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the application, and are not used to limit the application. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

The technical solutions between the various embodiments can be combined with each other, but they must be based on what can be achieved by a person of ordinary skill in the art. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that such a combination of technical solutions does not exist. It is not within the scope of protection required by this application.

Example one

Please refer to FIG. 1, which shows a flowchart of the steps of the data query method according to an embodiment of the present application. It can be understood that the flowchart in this method embodiment is not used to limit the order of execution of the steps. The following is an exemplary description with computer equipment as the main body of execution, and the details are as follows:

As shown in Figure 1, the data query method may include steps S100 to S500, wherein:

Step S100: Receive a first request, where the first request carries a first original sentence.

In an exemplary embodiment, the first request is used to request a search query for the first original sentence. After receiving the first request, the first original sentence in the first request is obtained, and the first original sentence is converted into a vector sequence.

In an exemplary embodiment, as shown in FIG. 2, the step S100 of receiving the first request, the first request carrying the first original sentence includes:

Step S101: Map the first original sentence into a first identification sequence and a second identification sequence.

Step S102: Map the first identification sequence to a first word vector sequence, and map the second identification sequence to a first position vector sequence.

In an exemplary embodiment, a word list of commonly used words is preset (for example, 2w words commonly used in a dictionary, each word corresponds to a digital ID in the dictionary), and the first identification sequence refers to the first original sentence The numeric ID corresponding to each word in the dictionary.

The first original sentence takes "the name price of the bond issued by Hainan Airlines and the name of the bond issued by its holding parent company" as an example. For this sentence, there are only 29 words, and 29 words are mapped to the digital ID, namely [12, 23, 13,42,56,...], and add 0 to extend the length of the above-mentioned insufficient number of IDs to 512, where 512 is the default maximum sentence length in the model, which can also be changed according to actual needs. The maximum length of a sentence is 512 as an example. If the length of a sentence is greater than 512, the excess part of the digital ID corresponding to the first original sentence will be truncated. That is, the first identification sequence of the first original sentence is [12, 23, 13, 42, 56, 0, 0,..., 0]. The above-mentioned first identification sequence with a length of 512 is mapped to a first word vector sequence of 512*768 (the parameter is bert default and can be changed) through the embeding (word embedding) layer, that is, the first identification sequence is mapped to a distributed first word Vector sequence.

Further, the second identification sequence is the position ID number of each word in the first original sentence in the first original sentence, and the first position vector is a 512*768 position vector code.

In an exemplary embodiment, according to the following formula:

Mapped to the first position vector, where p represents the position ID number, which corresponds to [0-511] according to the position, 2i and 2i+1 correspond to the even and odd positions of p respectively, and dpos here refers to the dimension of the position vector, that is, 768 dimensions .

Step S200: Classify the first original sentence, and generate a sequence labeling result.

In an exemplary embodiment, the first original sentence may be classified by a sequence labeling task model, where the sequence labeling task model is a model based on a bert model. Specifically, the sequence labeling task model includes an encoder and a decoder, where the encoder is an encoder with bert as the backbone, sentence features are extracted, and the hidden layer dimension is set to 768. The encoder output is 512*768 dimensions. There are 6 encoders, and each encoder is connected by a transformer (transformation neural network) block with the same structure. Among them, each transformer block is composed of a self-attention layer, a feedforward neural network layer, and a normalization layer. . The first word vector sequence and the first position sequence are added by position and input into the encoder of the sequence labeling task model, and a 512*768 vector set can be output through multiple encoders.

Further, the encoder is followed by the decoder. Each decoder includes: a self-attention layer in the decoder, an encoding-decoding attention layer, and a forward propagation neural network layer. The decoder also includes 512 fully connected layers with an output dimension of 9. Through full connection, the 512*768-dimensional vector set finally output by the encoder is reduced to 512*9 dimensions. 512 corresponds to each position, and 9 corresponds to each type of label, where each type of label is preset. Through the last decoder, each position outputs a 9-dimensional vector, and each dimension corresponds to a label, which is added to form a real number vector, and then the real number vector is classified through the softmax (normalization) layer to obtain the highest probability Position, output sequence labeling result.

In an exemplary embodiment, as shown in FIG. 3, the step S200 of classifying the first original sentence and generating a sequence labeling result includes:

In step S201, the first word vector sequence and the first position vector sequence are added into the sequence labeling task model by position addition.

Step S202: A plurality of encoders based on the sequence labeling task model perform encoding on the first word vector sequence and the first position vector sequence, and output a vector set.

Step S203: Input the vector set into multiple decoders of the sequence labeling task model, perform decoding and dimensionality reduction processing on the vector set based on the multiple decoders, and output a target vector.

Step S204: Perform a linear transformation on the target vector by a classifier, and output a sequence labeling result.

In an exemplary embodiment, taking "the name price of the bond issued by Hainan Airlines and the name of the bond issued by its holding parent company" as an example, the output sequence labeling result is:

[E_B, E_I, E_I, E_I, A_B, A_I, A_I, A_I, A_I, O, T_B, T_I, T_B, T_I, O, O, R_B, A_B, A_I, A_I, A_I, A_I, A_B, A_I, A_I , A_I, A_I, T_B, T_I]. Among them, _B represents the first character of a certain word, _I represents the characters in other positions of a certain word, E represents entity entity, A represents attribute, T represents intent word target, and R represents replace word. The result of the sequence labeling can also be parsed into an easy-to-understand form, namely {entity word (knowledge graph node): ["Hainan Airlines"], attribute word (knowledge graph edge): ["issued bonds", "holding parent company" ,"The bond of the hairstyle"], the pronoun: ["其"], the intent word: ["name", "price", "name"]}.

Step S300: Obtain preset self-attention matrix parameters, and generate multiple word attention matrices based on the self-attention matrix parameters and the sequence labeling result.

In an exemplary embodiment, the self-attention matrix can be extracted from the transformer block of the first layer in the encoder, and the self-attention matrix is a symmetric 512*512 matrix that can be used to represent characters. Correlation with words.

In an exemplary embodiment, as shown in FIG. 4, the step S300 of acquiring preset self-attention matrix parameters and generating multiple word attention matrices based on the self-attention matrix parameters and the sequence labeling result may also be It further includes:

Step S301: Obtain the self-attention matrix parameters from the first layer encoder of the encoder.

Step S302: Generate a self-attention matrix for each word based on the self-attention matrix parameters.

Specifically, the self-attention matrix of each word is the intermediate data in the sequence labeling task model.

Further, the similarity formula can be used

Among them, the vectors Q, K, and V are obtained by multiplying the input vector matrix of the sequence labeling task model and the self-attention matrix parameters. F is the self-attention matrix of the word, Q is the query sample vector, K is the key vector, KT is the transposition of the key vector, V is the value vector, and dk is the dimension of the input vector.

Step S303: Obtain multiple target words and a self-attention matrix of each character in each target word based on the sequence labeling result.

Step S304: Based on the self-attention matrix of each character in each target word, the word attention matrix of the corresponding target word is calculated.

Specifically, the sequence labeling is combined with the self-attention matrix to construct a word attention matrix of each word and other words to express the relevance between words. The attention matrix of each word is the sum of squared attention corresponding to each word

Add and average, where i represents the word at the i-th position, j represents the word at the j-th position, aij represents the self-attention matrix between the i-th word and the j-th word, and m refers to the target The number of characters in a word, n refers to the number of characters in a word in the first original sentence.

Step S400: Generate multiple intent stream analysis results based on the multiple word attention matrices.

In an exemplary embodiment, as shown in FIG. 5, the step S400 of generating multiple intent stream analysis results based on the multiple word attention matrices may further include:

Step S401: Take an average value from the word attention matrix of a plurality of target words, and define the average value as a threshold.

In step S402, the word attention matrix of a plurality of the target words is compared with the threshold to generate a comparison result.

Step S403, based on the comparison result: determine that the target word whose word attention matrix is greater than the threshold is an intention-related word.

Step S404: Generate multiple intent stream analysis results based on the sequence labeling result and the intent related words.

In an exemplary embodiment, when the sequence labeling result contains a pronoun, the pronoun needs to be replaced with the word with the highest word attention matrix. Following the example above, replace the pronouns in the multi-intent stream analysis result with the word with the highest attention. Here, replace "其" with "Hainan Airlines".

In an exemplary embodiment, for the three intentions (two types) {1_name, 2_price, 3_name} obtained in the first original sentence of the above example, find the three words that satisfy the attention For words with a score greater than M, the result of multi-intent flow analysis is obtained {1_Name: [Hainan Airlines, bonds issued], 2_Price: [Hainan Airlines, bonds issued], 3_Name: [Hainan Airlines, controlling parent Company, bonds issued]}.

Step S500: Query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.

In an exemplary embodiment, when the knowledge graph table/library is a small knowledge graph library/table, the query result can be obtained directly through the query of the knowledge graph table/library, and the query result is returned to the user end. Compared with the traditional search that can only return related webpages, the embodiments of the present application are based on the knowledge graph and can directly return the results required by the user, which improves the user experience.

In an exemplary embodiment, the knowledge graph library is stored in a blockchain. Exemplarily, as shown in FIG. 6, when the knowledge graph table/library is a massive knowledge graph table/library, query the multiple intent stream analysis results through the knowledge graph library to obtain the query result S500 It can further include:

Step S501: Assemble the multiple intention flow analysis results into corresponding second sentences.

Step S502: Identify the second sentence, and generate an intent type corresponding to the second sentence.

Step S503, based on the intent type of the second sentence and the second sentence, query in the corresponding knowledge graph database to obtain the query result.

Specifically, multiple intent stream analysis results can be spliced directly with a separator to form multiple corresponding second sentences, such as {1_name: [海南航空, issued bonds], 2_price: [海南航空, Bond issued], 3_ name: [其, holding parent company, bond issued]}, spliced into three sentences, for example, using &&& as a separator, the three sentences are the name of the bond issued by Hainan Airlines &&&, Hainan The price of bonds issued by aviation &&&, and the bonds issued by Hainan Airlines (replaced by it) &&& holding parent company &&&).

Specifically, the multi-intent stream recognition results are assembled into a sentence and sent to the Textcnn (text classification) model for classification. For example, the intents are divided into ten categories: 1. Query the bond name, 2. Query the bond price, 3...., Among them, the intention categories 1 and 2 correspond to the knowledge graph bond table. The Textcnn model is an existing model, so I won’t repeat it here.

In an exemplary embodiment, as shown in FIG. 7, the method further includes:

Step S601: Obtain a plurality of sample original sentences and sample label labels corresponding to the plurality of sample original sentences.

Step S602: Map the multiple sample original sentences into a first sample identification sequence and a second sample identification sequence.

Step S603: Map the first sample identification sequence to a first sample word vector sequence, and map the second sample identification sequence to a first sample position vector.

Step S604: Input the first sample word vector sequence and the first sample position vector to the sequence labeling task model, and output a sample sequence labeling result through the sequence labeling task model.

Step S605: Calculate a cross-entropy loss value based on the sample labeling label and the sample sequence labeling result, and adjust the sequence labeling task model based on the cross-entropy loss value to obtain an optimized sequence labeling task model.

Specifically, the manually constructed sample is labeled as y _label , and the sample sequence output by the model is labeled as y _predict , by minimizing the cross-entropy loss function

Just label the task model in the training sequence.

In an exemplary embodiment, when the sample data is too small and the training is insufficient, the method further includes: additionally constructing a loss function

Auxiliary training. Where I _ij is the indicator function,

x _i and x _j are the word vectors of two words that are transformed by a fully connected layer, which is the intermediate result of the first transformer block, x=W*word vector sequence+b, W and b are both trainable The model parameter, the dimension is 512*768, and d is the hidden layer dimension. In this example, d is always equal to 768. The additionally constructed loss function is added to the aforementioned minimized cross-entropy loss function L to obtain the final loss function L _total =L+L _extra , and the sequence labeling task model is trained according to the value of the minimized cross-entropy loss function.

This application uses sequence labeling to realize the extraction of intentions and the attention mechanism to associate multi-intentional words and non-intentional words, query the analysis of information flow, and can accurately identify multi-intention query sentences, improve accuracy, and improve query efficiency; and Based on the knowledge graph, the results required by the user can be directly returned, which improves the user experience; this application uses the sequence tagging task model to extract the intent, and does not need to maintain a large number of sentence style templates, saving labor costs; and by adding additional attention loss The function guarantees that the accuracy of the model is still sufficient for effective application even with a small training set sample.

Example two

Please continue to refer to FIG. 8, which shows a schematic diagram of program modules of the data query system of the present application. In this embodiment, the data query system 20 may include or be divided into one or more program modules, and the one or more program modules are stored in a storage medium and executed by one or more processors to complete the present invention. Apply, and realize the above-mentioned data query method. The program module referred to in the embodiments of the present application refers to a series of computer program instruction segments capable of completing specific functions, and is more suitable for describing the execution process of the data query system 20 in the storage medium than the program itself. The following description will specifically introduce the functions of each program module in this embodiment:

The receiving module 700 is configured to receive a first request, where the first request carries a first original sentence.

The sequence labeling module 710 is configured to classify the first original sentence and generate a sequence labeling result.

The first generating module 720 is configured to obtain preset self-attention matrix parameters, and generate multiple word attention matrices based on the self-attention matrix parameters and the sequence labeling result.

The second generating module 730 is configured to generate multiple intent stream analysis results based on the multiple word attention matrices.

The query module 740 is configured to query the multiple intent stream analysis results through the knowledge graph library to obtain the query results.

In an exemplary embodiment, the receiving module 700 is further configured to: map the first original sentence into a first identification sequence and a second identification sequence; and map the first identification sequence into a first word vector sequence , Mapping the second identification sequence to a first position vector sequence.

In an exemplary embodiment, the sequence labeling module 710 is further configured to: add the first word vector sequence and the first position vector sequence to the sequence labeling task model by position addition; and label the task based on the sequence The multiple encoders of the model perform encoding on the first word vector sequence and the first position vector sequence, and output a vector set; input the vector set into multiple decoders of the sequence labeling task model, based on the multiple The decoder performs decoding and dimensionality reduction processing on the vector set, and outputs a target vector; performs a linear transformation on the target vector through a classifier, and outputs a sequence labeling result.

In an exemplary embodiment, the first generation module 720 is further configured to: obtain the self-attention matrix parameters from the first-layer encoder of the encoder; and generate each parameter based on the self-attention matrix parameters. Self-attention matrix of a word; based on the sequence labeling results to obtain multiple target words and the self-attention matrix of each word in each target word; based on the self-attention matrix of each word in each target word, calculate Get the word attention matrix of the corresponding target word.

In an exemplary embodiment, the second generating module 730 is further configured to: take an average value from the word attention matrix of a plurality of the target words, and define the average value as a threshold; The word attention matrix of the word attention matrix is compared with the threshold value to generate a comparison result; based on the comparison result: the target word whose word attention matrix is greater than the threshold value is determined to be an intent related word; based on the sequence labeling result and Intent related words generate multiple intent stream analysis results.

In an exemplary embodiment, the knowledge graph library is stored in the blockchain, and the query module 740 is further configured to: assemble the multiple intent stream analysis results into corresponding second sentences; and identify the The second sentence generates the intent type corresponding to the second sentence; based on the intent type of the second sentence and the second sentence, query in the corresponding knowledge graph database to obtain the query result.

In an exemplary embodiment, the system further includes a training module 750, and the training module 750 is further configured to: obtain a plurality of sample original sentences and sample labeling labels corresponding to the plurality of sample original sentences; A sample of original sentences is mapped to a first sample identification sequence and a second sample identification sequence; the first sample identification sequence is mapped to a first sample word vector sequence, and the second sample identification sequence is mapped to the first sample This position vector; the first sample word vector sequence and the first sample position vector are input to the sequence labeling task model, and the sample sequence labeling result is output through the sequence labeling task model; label and sample sequence are labelled by the sample The labeling result calculates a cross-entropy loss value, and adjusts the sequence labeling task model based on the cross-entropy loss value to obtain an optimized sequence labeling task model.

Example three

Refer to FIG. 9, which is a schematic diagram of the hardware architecture of the computer device according to the third embodiment of the present application. In this embodiment, the computer device 2 is a device that can automatically perform numerical calculation and/or information processing in accordance with pre-set or stored instructions. The computer device 2 may be a rack server, a blade server, a tower server, or a cabinet server (including an independent server or a server cluster composed of multiple servers). As shown in FIG. 9, the computer device 2 at least includes, but is not limited to, a memory 21, a processor 22, a network interface 23, and a data query system 20 that can communicate with each other through a system bus. in:

In this embodiment, the memory 21 includes at least one type of computer-readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory ( RAM), static random access memory (SRAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), programmable read only memory (PROM), magnetic memory, magnetic disks, optical disks, etc. In some embodiments, the memory 21 may be an internal storage unit of the computer device 2, for example, the hard disk or memory of the computer device 2. In other embodiments, the memory 21 may also be an external storage device of the computer device 2, such as a plug-in hard disk, a smart media card (SMC), and a secure digital (Secure Digital, SMC) equipped on the computer device 2. SD) card, flash card (Flash Card), etc. Of course, the memory 21 may also include both the internal storage unit of the computer device 2 and its external storage device. In this embodiment, the memory 21 is generally used to store an operating system and various application software installed in the computer device 2, for example, the program code of the data query system 20 in the foregoing embodiment. In addition, the memory 21 can also be used to temporarily store various types of data that have been output or will be output.

In some embodiments, the processor 22 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 22 is generally used to control the overall operation of the computer device 2. In this embodiment, the processor 22 is used to run the program code or process data stored in the memory 21, for example, to run the data query system 20, so as to implement the data query method of the foregoing embodiment.

The network interface 23 may include a wireless network interface or a wired network interface, and the network interface 23 is generally used to establish a communication connection between the computer device 2 and other electronic devices. For example, the network interface 23 is used to connect the computer device 2 with an external terminal through a network, and establish a data transmission channel and a communication connection between the computer device 2 and the external terminal. The network may be Intranet, Internet, Global System of Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G Network, Bluetooth (Bluetooth), Wi-Fi and other wireless or wired networks.

It should be pointed out that FIG. 9 only shows the computer device 2 with components 20-23, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

In this embodiment, the data query system 20 stored in the memory 21 may also be divided into one or more program modules. The one or more program modules are stored in the memory 21 and are composed of one or more program modules. It is executed by two processors (the processor 22 in this embodiment) to complete the application.

For example, FIG. 8 shows a schematic diagram of program modules for implementing the second embodiment of the data query system 20. In this embodiment, the data-based query system 20 can be divided into a receiving module 700, a sequence labeling module 710, and a first generating module. The module 720, the second generation module 730, and the query moqua 740. Among them, the program module referred to in this application refers to a series of computer program instruction segments that can complete specific functions, and is more suitable than a program to describe the execution process of the data query system 20 in the computer device 2. The specific functions of the program modules 700-740 have been described in detail in the second embodiment, and will not be repeated here.

Example four

This embodiment also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), only Readable memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disks, optical disks, servers, App application malls, etc., on which computer programs are stored, The corresponding function is realized when the program is executed by the processor. The computer-readable storage medium may be non-volatile or volatile. The computer-readable storage medium of this embodiment is used to store the data query system 20, and when executed by a processor, it implements the data query method of the foregoing embodiment.

The serial numbers of the foregoing embodiments of the present application are for description only, and do not represent the superiority or inferiority of the embodiments.

Through the description of the above implementation manners, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。

The above are only the preferred embodiments of the application, and do not limit the scope of the patent for this application. Any equivalent structure or equivalent process transformation made using the content of the description and drawings of the application, or directly or indirectly applied to other related technical fields , The same reason is included in the scope of patent protection of this application.

Claims

A data query method, which includes:

Receiving a first request, the first request carrying a first original sentence;

Classify the first original sentence, and generate a sequence labeling result;

Acquiring preset self-attention matrix parameters, and generating multiple word attention matrices based on the self-attention matrix parameters and sequence labeling results;

Generating multiple intent stream analysis results based on the multiple word attention matrices;

Query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.
The data query method according to claim 1, wherein the receiving the first request, the first request carrying the first original sentence comprises:

Mapping the first original sentence into a first identification sequence and a second identification sequence;

The first identification sequence is mapped to a first word vector sequence, and the second identification sequence is mapped to a first position vector sequence.
The data query method according to claim 2, wherein said classifying said first original sentence and generating a sequence labeling result comprises:

Inputting the first word vector sequence and the first position vector sequence into the sequence labeling task model by position addition;

A plurality of encoders based on the sequence labeling task model perform encoding on the first word vector sequence and the first position vector sequence, and output a vector set;

Inputting the vector set into multiple decoders of a sequence labeling task model, and performing decoding and dimensionality reduction processing on the vector set based on the multiple decoders, and outputting a target vector;

The target vector is linearly transformed by the classifier, and the sequence labeling result is output.
The data query method according to claim 3, wherein said acquiring preset self-attention matrix parameters and generating a plurality of word attention matrices based on said self-attention matrix parameters and sequence labeling results comprises:

Acquiring the self-attention matrix parameter from the first layer encoder of the encoder;

Generating a self-attention matrix for each word based on the self-attention matrix parameters;

Obtain the self-attention matrix of multiple target words and each character in each target word based on the sequence labeling result;

Based on the self-attention matrix of each character in each target word, the word attention matrix of the corresponding target word is calculated.
The data query method according to claim 4, wherein said generating a plurality of intent stream parsing results based on the plurality of word attention matrices comprises:

Taking an average value from the word attention matrix of a plurality of the target words, and defining the average value as a threshold;

Comparing word attention matrices of a plurality of the target words with the threshold to generate a comparison result;

Based on the comparison result: determining that the target word whose word attention matrix is greater than the threshold is an intent-related word;

A plurality of intent stream analysis results are generated based on the sequence labeling result and the intent related words.
The data query method according to claim 1, wherein the knowledge graph library is stored in a blockchain, and querying the multiple intent stream analysis results through the knowledge graph library to obtain the query result further comprises:

Assembling the plurality of intent stream parsing results into corresponding second sentences;

Identifying the second sentence, and generating an intent type corresponding to the second sentence;

Based on the intent type of the second sentence and the second sentence, query in the corresponding knowledge graph database to obtain the query result.
The data query method according to claim 1, wherein the method further comprises:

Acquiring a plurality of sample original sentences and sample labeling labels corresponding to the plurality of sample original sentences;

Mapping the multiple sample original sentences into a first sample identification sequence and a second sample identification sequence;

Mapping the first sample identification sequence to a first sample word vector sequence, and mapping the second sample identification sequence to a first sample position vector;

Inputting the first sample word vector sequence and the first sample position vector to the sequence labeling task model, and outputting a sample sequence labeling result through the sequence labeling task model;

The cross-entropy loss value is calculated according to the sample labeling label and the sample sequence labeling result, and the sequence labeling task model is adjusted based on the cross entropy loss value to obtain an optimized sequence labeling task model.
A data query system, which includes:

A receiving module, configured to receive a first request, the first request carrying a first original sentence;

The sequence labeling module is used to classify the first original sentence and generate sequence labeling results;

The first generation module is configured to obtain preset self-attention matrix parameters, and generate multiple word attention matrices based on the self-attention matrix parameters and sequence labeling results;

The second generation module is configured to generate multiple intent stream analysis results based on the multiple word attention matrices;

The query module is used to query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.
A computer device includes a memory, a processor, and a computer program that is stored on the memory and can run on the processor, wherein the processor executes the following steps when the computer program is executed:

Receiving a first request, the first request carrying a first original sentence;

Classify the first original sentence, and generate a sequence labeling result;

Acquiring preset self-attention matrix parameters, and generating multiple word attention matrices based on the self-attention matrix parameters and sequence labeling results;

Generating multiple intent stream analysis results based on the multiple word attention matrices;

Query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.
The computer device according to claim 9, wherein the processor executes the following steps when executing the computer program:

Mapping the first original sentence into a first identification sequence and a second identification sequence;

The first identification sequence is mapped to a first word vector sequence, and the second identification sequence is mapped to a first position vector sequence.
The computer device according to claim 10, wherein the processor executes the following steps when executing the computer program:

Inputting the first word vector sequence and the first position vector sequence into the sequence labeling task model by position addition;

A plurality of encoders based on the sequence labeling task model perform encoding on the first word vector sequence and the first position vector sequence, and output a vector set;

Inputting the vector set into multiple decoders of a sequence labeling task model, and performing decoding and dimensionality reduction processing on the vector set based on the multiple decoders, and outputting a target vector;

The target vector is linearly transformed by the classifier, and the sequence labeling result is output.
The computer device according to claim 11, wherein the processor executes the following steps when executing the computer program:

Acquiring the self-attention matrix parameter from the first layer encoder of the encoder;

Generating a self-attention matrix for each word based on the self-attention matrix parameters;

Obtain the self-attention matrix of multiple target words and each character in each target word based on the sequence labeling result;

Based on the self-attention matrix of each character in each target word, the word attention matrix of the corresponding target word is calculated.
The computer device according to claim 12, wherein the processor executes the following steps when executing the computer program:

Taking an average value from the word attention matrix of a plurality of the target words, and defining the average value as a threshold;

Comparing word attention matrices of a plurality of the target words with the threshold to generate a comparison result;

Based on the comparison result: determining that the target word whose word attention matrix is greater than the threshold is an intent-related word;

A plurality of intent stream analysis results are generated based on the sequence labeling result and the intent related words.
The computer device according to claim 9, wherein the processor executes the following steps when executing the computer program:

Assembling the plurality of intent stream parsing results into corresponding second sentences;

Identifying the second sentence, and generating an intent type corresponding to the second sentence;

Based on the intent type of the second sentence and the second sentence, query in the corresponding knowledge graph database to obtain the query result.
A computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, and the computer program can be executed by at least one processor, so that the at least one processor executes the following steps:

Receiving a first request, the first request carrying a first original sentence;

Classify the first original sentence, and generate a sequence labeling result;

Acquiring preset self-attention matrix parameters, and generating multiple word attention matrices based on the self-attention matrix parameters and sequence labeling results;

Generating multiple intent stream analysis results based on the multiple word attention matrices;

Query the analysis results of the multiple intent streams through the knowledge graph library to obtain the query results.
The computer-readable storage medium according to claim 15, wherein the processor executes the following steps when executing the computer program:

Request to carry the first original sentence includes:

Mapping the first original sentence into a first identification sequence and a second identification sequence;

The first identification sequence is mapped to a first word vector sequence, and the second identification sequence is mapped to a first position vector sequence.
The computer-readable storage medium according to claim 16, wherein the processor executes the following steps when executing the computer program:

Inputting the first word vector sequence and the first position vector sequence into the sequence labeling task model by position addition;

A plurality of encoders based on the sequence labeling task model perform encoding on the first word vector sequence and the first position vector sequence, and output a vector set;

Inputting the vector set into multiple decoders of a sequence labeling task model, and performing decoding and dimensionality reduction processing on the vector set based on the multiple decoders, and outputting a target vector;

The target vector is linearly transformed by the classifier, and the sequence labeling result is output.
The computer-readable storage medium according to claim 17, wherein the processor executes the following steps when executing the computer program:

Acquiring the self-attention matrix parameter from the first layer encoder of the encoder;

Generating a self-attention matrix for each word based on the self-attention matrix parameters;

Obtain the self-attention matrix of multiple target words and each character in each target word based on the sequence labeling result;

Based on the self-attention matrix of each character in each target word, the word attention matrix of the corresponding target word is calculated.
The computer-readable storage medium according to claim 18, wherein the processor executes the following steps when executing the computer program:

Taking an average value from the word attention matrix of a plurality of the target words, and defining the average value as a threshold;

Comparing word attention matrices of a plurality of the target words with the threshold to generate a comparison result;

Based on the comparison result: determining that the target word whose word attention matrix is greater than the threshold is an intent-related word;

A plurality of intent stream analysis results are generated based on the sequence labeling result and the intent related words.
The computer-readable storage medium according to claim 15, wherein the processor executes the following steps when executing the computer program:

Assembling the plurality of intent stream parsing results into corresponding second sentences;

Identifying the second sentence, and generating an intent type corresponding to the second sentence;

Based on the intent type of the second sentence and the second sentence, query in the corresponding knowledge graph database to obtain the query result.