WO2022068197A1

WO2022068197A1 - Conversation generation method and apparatus, device, and readable storage medium

Info

Publication number: WO2022068197A1
Application number: PCT/CN2021/091292
Authority: WO
Inventors: 李雅峥; 杨海钦; 姚晓远
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-09-30
Filing date: 2021-04-30
Publication date: 2022-04-07
Also published as: CN112199482B; CN112199482A

Abstract

A conversation generation method and apparatus, a device, and a readable storage medium, relating to the technical field of digital treatment. The method comprises: obtaining questioning information, and converting the questioning information into a first query vector by using a preset first gate recurrent unit (GRU) model (S101); according to the first query vector, determining, by using a preset first end-to-end memory network (MemN2N) model, a common sense vector associated with the first query vector, and forming a questioning vector according to the first query vector and the common sense vector (S102); according to the questioning vector, converting the questioning vector into multiple second query vectors by using a preset second GRU model, and sequentially inputting the second query vectors into a preset second MemN2N model to obtain multiple response vectors (S103); and respectively converting the response vectors into response words, and combining all the response words into response information (S104). According to the method, response information can be quickly and accurately formed in a teleconsultation conversation, such that user experience is improved.

Description

A dialogue generation method, apparatus, device and readable storage medium

CROSS-REFERENCE TO RELATED APPLICATIONS

This application declares that it enjoys the priority of the Chinese patent application with the application number 202011059826.7 and the title of "A Method, Apparatus, Equipment and Readable Storage Medium for Dialog Generation" filed on September 30, 2020, and the overall content of the Chinese patent application Incorporated herein by reference.

technical field

The present application relates to the field of digital medical technology, and in particular, to a dialog generation method, apparatus, device, and readable storage medium.

Background technique

With the continuous development of artificial intelligence, more and more human-machine dialogues are applied to various scenarios; for example, in manual customer service scenarios, by identifying the question information input by the user, the response information corresponding to the question information is formed, thereby reducing manpower However, the inventor realizes that if the traditional open-domain human-machine dialogue system lacks understanding of the background knowledge and relevant common sense information of the user's question, and only starts from the dialogue data, it will produce a general answer that lacks effective information, and may Will have an impact on the readability of the reply message. In addition, how to quickly and accurately form reply information according to the user's question information has become a technical problem that those skilled in the art need to solve urgently.

SUMMARY OF THE INVENTION

The purpose of the present application is to provide a dialogue generation method, apparatus, device and readable storage medium, which can quickly and accurately form reply information in a remote consultation dialogue and improve user experience.

According to one aspect of the present application, there is provided a dialog generation method, the method comprising:

Obtain question information, and utilize the preset first gate recursive unit GRU model to convert the question information into a first query vector;

According to the first query vector, use the preset first end-to-end memory network MemN2N model to determine the common sense vector associated with the first query vector, and form a question according to the first query vector and the common sense vector vector;

According to the question vector, the question vector is converted into a plurality of second query vectors by using the preset second gate recursive unit GRU model, and each second query vector is sequentially input into the preset second end-to-end memory network MemN2N model to get multiple reply vectors;

Each reply vector is converted into reply words separately, and all reply words are combined into reply information.

In order to achieve the above purpose, the present application also provides a dialogue generation device, the device comprising:

an acquisition module for acquiring question information, and converting the question information into a first query vector by using a preset first gate recursive unit GRU model;

The questioning module is configured to use the preset first end-to-end memory network MemN2N model according to the first query vector to determine the common sense vector associated with the first query vector, and to determine the common sense vector associated with the first query vector according to the first query vector and all Describe the common sense vector to form the question vector;

The answering module is configured to convert the question vector into a plurality of second query vectors according to the question vector using a preset second gate recursive unit GRU model, and input each second query vector to the preset second terminal in turn end-to-end memory network MemN2N model to get multiple reply vectors;

The conversion module is used to convert each reply vector into reply words respectively, and combine all reply words into reply information.

In order to achieve the above object, the present application also provides a computer device, the computer device specifically includes: a memory, a processor, and a computer program stored in the memory and running on the processor, the processor executes the computer program. The following steps are implemented when the computer program is described:

Obtain question information, and utilize the preset first gate recursive unit GRU model to convert the question information into the first query vector;

In order to achieve the above purpose, the present application also provides a computer-readable storage medium on which a computer program is stored, and the computer program implements the following steps when executed by a processor:

The dialogue generation method, device, device and readable storage medium provided by this application combine the MemN2N architecture of the end-to-end memory network with the GRU network to find out the common sense information related to it according to the question information, and comprehensively consider the question information and the common sense information The reply message is determined. In the process of encoding the question information into the question vector, the form of GRU+MemN2N is used to encode the question information. For the question information input by the user, the GRU network is used to replace the EmbeddingB in the MemN2N network, and the final hidden layer state of the GRU network is used. It is input into the MemN2N network as a query vector. In the process of decoding the question vector into answer information, the form of GRU+MemN2N is also used to generate the answer information. The application can quickly and accurately form reply information in the remote consultation dialogue, and improve user experience.

Description of drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustrating preferred embodiments only and are not to be considered limiting of the application. Also, the same components are denoted by the same reference numerals throughout the drawings. In the attached image:

FIG. 1 is an optional schematic flowchart of a dialog generation method provided in Embodiment 1;

FIG. 2 is a schematic diagram of an optional composition structure of the dialogue generation device provided in Embodiment 2;

FIG. 3 is a schematic diagram of an optional hardware architecture of the computer device provided in the third embodiment.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

Example 1

An embodiment of the present application provides a dialog generation method, as shown in FIG. 1 , the method specifically includes the following steps:

Step S101: Obtain question information, and use a preset first GRU (Gate Recurrent Unit, gate recursive unit) model to convert the question information into a first query vector.

Specifically, step S101 includes:

Step A1: performing word segmentation processing on the question information, and forming a word sequence with multiple keywords obtained after the word segmentation processing; wherein, the word sequence includes N keywords;

Step A2: For a target keyword in the word sequence, according to the hidden influence factor of the keyword located before the target keyword in the word sequence to the target keyword, use the first gate a recursive unit GRU model, which calculates the hidden influence factor of the target keyword to the keyword located after the target keyword in the word sequence;

Step A3: Use the hidden influence factor calculated according to the last keyword in the word sequence as the first query vector u ₁ corresponding to the question information.

Step S102: According to the first query vector, use a preset first MemN2N (End-to-end Memory Networks, end-to-end memory network) model to determine the common sense vector associated with the first query vector, and according to The first query vector and the common sense vector form a question vector.

Specifically, step S102 includes:

Step B1: In the first loop hop of the first end-to-end memory network MemN2N model, calculate the first query vector u ₁ and the ith common sense head vector in the preset common sense head group respectively. correlation value p _{i of x i} _;

Wherein, p _i =Softmax((u ₁ ) ^T * _xi ), and T is a transposition function.

Step B2: Calculate the question sub-vector a ₁ of the first cycle according to the correlation value p _i of the ith common sense head vector x _i and the ith common sense tail vector _yi in the preset common sense tail group;

where a ₁ =∑ _i p _i y _i .

Step B3: adding the first query vector u ₁ and the question vector a ₁ to obtain the first query vector u ₂ of the second cycle;

Step B4: Repeat steps B1 to B3 until the question sub-vector a _M of the Mth cycle is calculated;

Step B5: Use the question sub-vector a _M of the Mth cycle as the question vector.

Further, the method also includes:

Step C1: obtaining a common sense information base; wherein, the common sense information base includes a plurality of common sense information represented in the form of knowledge triples, and the common sense information includes: a head, a relation part, and a tail;

Taking "a cat is an animal" as an example, the knowledge triple form represents bits (h: cat, r: belongs to, t: animal), where h represents the head, t represents the tail, and r represents the difference between the head and the tail. Relations Department.

Step C2: Convert the head in each common sense information into a common sense head vector by presetting the first hidden layer matrix EmbeddingA, thereby forming a common sense head group;

Step C3: By presetting the second hidden layer matrix EmbeddingC, the tail in each common sense information is converted into a common sense tail vector, thereby forming a common sense tail group;

Step C4: Establish a correspondence between the common sense head vector and the common sense tail vector according to the relationship part in each common sense information.

In the process of encoding the Encoder, the process of encoding the question information into a question vector, this embodiment uses the form of GRU+MemN2N to encode the question information, and for the question information input by the user, the GRU network is used to replace the EmbeddingB in the MemN2N network, and the The final hidden layer state of the GRU network is input into the MemN2N network as a query vector. The entire MemN2N network is superimposed by multiple hops. In each hop, the correlation between the query vector and each common sense information in the Memory is calculated separately. In this embodiment, the Encoder is implemented by using GRU+MemN2N, and on the premise that the complete question information is extracted by using the GRU, common sense information that is highly correlated with the entire question information can continue to be added, avoiding the search for a single entity word. information bias. In addition, the common sense information of Memory is calculated in the form of weighted sum, which avoids selecting a single knowledge triplet as compensation information, making the acquired common sense information more comprehensive.

Step S103: According to the question vector, use the preset second gate recursive unit GRU model to convert the question vector into a plurality of second query vectors, and input each second query vector into the preset second end-to-end in sequence. memory network in the MemN2N model to obtain multiple reply vectors.

Specifically, step S103 includes:

Step D1: Use the question vector as the hidden influence factor h ₀ of the first layer, and input the preset starting character vector s ₀ into the second gate recursive unit GRU model to obtain the output vector s ₁ and transfer to The hidden influence factor h ₁ of the second layer;

where (s ₁ , h ₁ )=GRU(s ₀ , h ₀ ).

Step D2: Inputting the output vector s ₁ as a second query vector into the second end-to-end memory network MemN2N model to obtain a reply vector r ₁ ;

Further, step D2 includes:

Step D21: In the first loop hop of the second end-to-end memory network MemN2N model, calculate the second query vector s ₁ and the i-th reply header vector in the preset reply header group respectively. the correlation value p _{i of k i} _;

Wherein, p _i =Softmax((s ₁ ) ^T k _i ), T is the transpose function;

Step D22: Calculate the reply sub-vector o ₁ of the first cycle according to the correlation value p _i of the _ith reply head vector ki and the _ith reply tail vector li in the preset reply tail group;

Wherein, o ₁ =∑ _i p _i l _i ;

Step D23: adding the second query vector s ₁ and the reply sub-vector o ₁ of the first loop to obtain the second query vector s ₂ of the second loop hop;

Step D24: Repeat steps D21 to D23 until the question sub-vector o _N of the Nth loop hop is calculated;

Step D25: Use the question sub-vector o _N of the Nth cycle as the reply vector r ₁ .

Further, the method also includes:

Step E1: obtaining a reply information base; wherein, the reply information base includes a plurality of reply information represented in the form of knowledge triples, and the reply information includes: a head, a relation part and a tail;

Step E2: Converting the header in each reply message into a reply header vector through a preset conversion and embedding the TransE algorithm, thereby forming a reply header group;

Step E3: Converting the tail in each reply message into a reply tail vector through a preset transformation and embedding the TransE algorithm, thereby forming a reply tail group;

Wherein, k=(h,r,t)=MLP(TransE(h,r,t));

k _i =h; l _i =t.

Step E4: Establish a corresponding relationship between the reply head vector and the reply tail vector according to the relationship part in each reply information.

Step D3: Re-input the output vector s ₁ and the hidden influence factor h ₁ of the second layer into the second gate recursive unit GRU model to obtain the output vector s ₂ and the hidden influence factor passed to the third layer h ₂ , and re-input the output vector s ₂ into the second end-to-end memory network MemN2N model to obtain a reply vector r ₂ , and so on until the output of the second gate recursive unit GRU model The vector is the default end character vector.

Step S104: Convert each reply vector into reply words respectively, and combine all reply words into reply information.

Specifically, step S104 includes:

The reply word _wi corresponding to the reply vector _ri is obtained according to the following formula:

P(r _i = _wi )=softmax(Wr _i );

Wherein, W is a preset matrix containing multiple reply words, and the word with the largest P value in the calculated matrix W is used as the reply word _wi corresponding to _ri .

In the process of decoding the Decoder, the process of decoding the question vector into the reply information, the form of GRU+MemN2N is also used to generate the reply information; the initial hidden layer state of the GRU network is the output of the Encoder part. For Memory, different from the Encoder part, the TransE algorithm is used to complete the encoding of knowledge triples, instead of Embedding A and Embedding C in the MemoryN2N model. In addition, unlike the Encoder, which uses the output of the last moment of the GRU network as the input of MemN2N, the Decoder part uses each hidden state of the GRU as the query vector query of MemN2N.

In this embodiment, the implementation of the Decoder part avoids the distinction between entity words and common words when generating a reply, so that all reply words can be obtained according to the vocabulary. In addition, this patent uses the idea of Kay Value Memory Network to distinguish the similarity calculation part of Memory and query from the weighted and output part, so that the query is closer to the head entity in the knowledge triplet, and the output is closer to the knowledge The tail entities in the triplet are closer, reducing the repetition rate of the model to generate responses and questions.

Embodiment 2

An embodiment of the present application provides a dialogue generation device, as shown in FIG. 2 , the device specifically includes the following components:

The obtaining module 201 is used for obtaining question information, and utilizes the preset first gate recursive unit GRU model to convert the question information into a first query vector;

The questioning module 202 is configured to use a preset first end-to-end memory network MemN2N model according to the first query vector to determine a common sense vector associated with the first query vector, and to determine the common sense vector associated with the first query vector according to the first query vector and the common sense vector forms a questioning vector;

The answering module 203 is configured to convert the question vector into a plurality of second query vectors according to the question vector using a preset second gate recursive unit GRU model, and input each second query vector into a preset second query vector in sequence. end-to-end memory network MemN2N model to get multiple reply vectors;

The conversion module 204 is configured to convert each reply vector into reply words respectively, and combine all reply words into reply information.

Specifically, the acquisition module 201 is used for:

The question information is subjected to word segmentation processing, and a plurality of keywords obtained after the word segmentation processing is formed into a word sequence; for a target keyword in the word sequence, according to the word sequence located before the target keyword The hidden influence factor of the keyword passed to the target keyword, using the first gate recursive unit GRU model to calculate the target keyword passed to the key word sequence located after the target keyword The hidden influence factor of the word; the hidden influence factor calculated according to the last keyword in the word sequence is used as the first query vector u ₁ corresponding to the question information.

Further, the questioning module 202 is specifically used for:

In the first cycle of the first end-to-end memory network MemN2N model, the correlation between the first query vector u ₁ and the ith common sense head vector x _i in the preset common sense head group is calculated respectively. degree value p _i ; according to the correlation degree value p _i of the ith common sense head vector x _i and the ith common sense tail vector y _i in the preset common sense tail group, calculate the question sub-vector a ₁ of the first cycle ; Add the first query vector u ₁ and the question vector a ₁ to obtain the first query vector u 2 of the second cycle; Recalculate the first query vector u ₂ of the second cycle according to the first query vector u ₂ of the second cycle The question sub-vector a ₂ of 2 cycles and the first query vector u ₃ of the third cycle, and so on, until the question sub-vector a _M of the M-th cycle is calculated; The vector a _M serves as the question vector.

Further, the device also includes:

The processing module is used to obtain a common sense information base; wherein, the common sense information base includes a plurality of common sense information represented in the form of knowledge triples, and the common sense information includes: a head, a relationship part and a tail; A hidden layer matrix converts the head in each common sense information into a common sense head vector, thereby forming a common sense head group; the second hidden layer matrix is preset to convert the tail in each common sense information into a common sense tail vector , so as to form a common sense tail group; establish the corresponding relationship between the common sense head vector and the common sense tail vector according to the relationship part in each common sense information.

Further, the reply module 203 is specifically used for:

The question vector is used as the hidden influence factor h ₀ of the first layer, and the preset starting character vector s ₀ is input into the second gate recursive unit GRU model to obtain the output vector s ₁ and pass to the second layer. _The hidden impact factor _h ₁ _of The hidden influence factor h ₁ of the second layer is re-input into the second gate recursive unit GRU model to obtain the output vector s ₂ and the hidden influence factor h ₂ passed to the third layer, and the output vector s ₂ is re-input Input into the second end-to-end memory network MemN2N model to obtain a reply vector r ₂ , and so on, until the output vector of the second gate recursive unit GRU model is the preset end character vector.

Further, when the reply module 203 implements the step of inputting the output vector s ₁ as the second query vector into the second end-to-end memory network MemN2N model to obtain the reply vector r ₁ , it specifically includes:

In the first cycle of the second end-to-end memory network MemN2N model, the correlation between the second query vector s ₁ and the _ith reply header vector ki in the preset reply header group is calculated respectively degree value p _i ; according to the correlation degree value p _i of the i-th reply head vector _ki and the _i -th reply tail vector li in the preset reply tail group, calculate the reply sub-vector o ₁ of the first cycle ; Add the second query vector s ₁ and the reply sub-vector o ₁ of the first cycle to obtain the second query vector s ₂ of the second cycle; According to the second query vector s ₂ of the second cycle Recalculate the reply sub-vector o ₂ of the 2nd cycle and the second query vector s ₃ of the 3rd cycle, and so on, until the reply sub-vector o _N of the Nth cycle is calculated; The circular reply sub-vector o _N is taken as the reply vector r ₁ .

Further, the processing module is also used for:

Obtain a reply information base; wherein, the reply information base includes a plurality of reply information represented in the form of knowledge triples, and the reply information includes: a head, a relation part and a tail; The head in each reply message is converted into a reply head vector, thereby forming a reply head group; the tail in each reply message is converted into a reply tail vector by the preset transformation and embedded TransE algorithm, thereby forming a reply tail group; according to each reply tail group; The relation part in each reply information establishes the correspondence between the reply head vector and the reply tail vector.

Embodiment 3

This embodiment also provides a computer device, such as a smart phone, a tablet computer, a notebook computer, a desktop computer, a rack server, a blade server, a tower server or a cabinet server (including independent servers, or A server cluster composed of multiple servers), etc. As shown in FIG. 3 , the computer device 30 in this embodiment at least includes but is not limited to: a memory 301 and a processor 302 that can be communicatively connected to each other through a system bus. It should be noted that FIG. 3 only shows the computer device 30 having components 301-302, but it should be understood that implementation of all of the illustrated components is not required, and more or fewer components may be implemented instead.

In this embodiment, the memory 301 (that is, a readable storage medium) includes a flash memory, a hard disk, a multimedia card, a card-type memory (eg, 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 Disk, Optical Disk, etc. In some embodiments, the memory 301 may be an internal storage unit of the computer device 30 , such as a hard disk or a memory of the computer device 30 . In other embodiments, the memory 301 may also be an external storage device of the computer device 30, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash memory card (Flash Card), etc. Of course, the memory 301 may also include both the internal storage unit of the computer device 30 and its external storage device. In this embodiment, the memory 301 is generally used to store the operating system and various application software installed on the computer device 30 . In addition, the memory 301 can also be used to temporarily store various types of data that have been output or will be output.

In some embodiments, the processor 302 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 302 is typically used to control the overall operation of the computer device 30 .

Specifically, in this embodiment, the processor 302 is configured to execute the program of the dialogue generation method stored in the processor 302, and the following steps are implemented when the program of the dialogue generation method is executed:

For the specific embodiment process of the above method steps, reference may be made to the first embodiment, which will not be repeated in this embodiment.

Embodiment 4

This embodiment also provides a computer-readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), only Read-Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Programmable Read-Only Memory (PROM), Magnetic Memory, Magnetic Disk, Optical Disc, Server, App Store, etc., the computer-readable storage medium It can be non-volatile or volatile, and a computer program is stored thereon, and when the computer program is executed by the processor, the following method steps are implemented:

It should be noted that, herein, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation.

The above are only the preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present application, or directly or indirectly applied in other related technical fields , are similarly included within the scope of patent protection of this application.

Claims

A dialogue generation method, wherein the method comprises:

Obtain question information, and utilize the preset first gate recursive unit GRU model to convert the question information into a first query vector;

According to the first query vector, use the preset first end-to-end memory network MemN2N model to determine the common sense vector associated with the first query vector, and form a question according to the first query vector and the common sense vector vector;

According to the question vector, the question vector is converted into a plurality of second query vectors by using the preset second gate recursive unit GRU model, and each second query vector is sequentially input into the preset second end-to-end memory network MemN2N model to get multiple reply vectors;

Each reply vector is converted into reply words separately, and all reply words are combined into reply information.
The dialogue generation method according to claim 1, wherein the acquiring question information and converting the question information into a first query vector by using a preset first gate recursive unit GRU model, comprising:

Perform word segmentation processing on the question information, and form a word sequence with a plurality of keywords obtained after the word segmentation processing;

For a target keyword in the word sequence, according to the hidden influence factor of the keyword located before the target keyword in the word sequence transmitted to the target keyword, the first recursive unit GRU is used. The model calculates the hidden influence factor of the target keyword transmitted to the keyword located after the target keyword in the word sequence;

The hidden influence factor calculated according to the last keyword in the word sequence is used as the first query vector u 1 corresponding to the question information.
The dialogue generation method according to claim 2, wherein the common sense vector associated with the first query vector is determined by using a preset first end-to-end memory network MemN2N model according to the first query vector, and form a question vector according to the first query vector and the common sense vector, including:

In the first cycle of the first end-to-end memory network MemN2N model, the correlation between the first query vector u 1 and the ith common sense head vector x i in the preset common sense head group is calculated respectively. degree value p i ;

According to the correlation value pi of the ith common sense head vector x i and the ith common sense tail vector yi in the preset common sense tail group, calculate the question sub-vector a 1 of the first cycle;

adding the first query vector u 1 and the question vector a 1 to obtain the first query vector u 2 of the second cycle;

According to the first query vector u 2 of the second cycle, the question sub-vector a 2 of the second cycle and the first query vector u 3 of the third cycle are recalculated, and so on until the M-th cycle is calculated. cyclic question sub-vector a M ;

The question sub-vector a M of the M-th cycle is used as the question vector.
The dialogue generation method of claim 3, wherein the method further comprises:

Obtaining a common sense information base; wherein, the common sense information base includes a plurality of common sense information represented in the form of knowledge triples, and the common sense information includes: a head, a relation part, and a tail;

By presetting the first hidden layer matrix, the head in each common sense information is converted into a common sense head vector, thereby forming a common sense head group;

By presetting the second hidden layer matrix, the tail in each common sense information is converted into a common sense tail vector, thereby forming a common sense tail group;

The corresponding relationship between the common sense head vector and the common sense tail vector is established according to the relationship part in each common sense information.
The dialogue generation method according to claim 1, wherein, according to the question vector, the question vector is converted into a plurality of second query vectors by using a preset second gate recursive unit GRU model, and each second query vector is converted into a plurality of second query vectors. The query vectors are sequentially input into the preset second end-to-end memory network MemN2N model to obtain multiple reply vectors, including:

The question vector is used as the hidden influence factor h 0 of the first layer, and the preset starting character vector s 0 is input into the second gate recursive unit GRU model to obtain the output vector s 1 and pass to the second layer. The hidden impact factor h 1 of ;

inputting the output vector s 1 into the second end-to-end memory network MemN2N model as a second query vector to obtain a reply vector r 1 ;

Re-input the output vector s 1 and the hidden influence factor h 1 of the second layer into the second gate recursive unit GRU model to obtain the output vector s 2 and the hidden influence factor h 2 passed to the third layer, and re-input the output vector s 2 into the second end-to-end memory network MemN2N model to obtain a reply vector r 2 , and so on, until the output vector of the second gate recursive unit GRU model is the Let the ending character vector.
The dialogue generation method according to claim 5, wherein the inputting the output vector s 1 as a second query vector into the second end-to-end memory network MemN2N model to obtain a reply vector r 1 , comprising: :

In the first cycle of the second end-to-end memory network MemN2N model, the correlation between the second query vector s 1 and the ith reply header vector ki in the preset reply header group is calculated respectively degree value p i ;

Calculate the reply sub-vector o 1 of the first cycle according to the correlation value p i of the ith reply head vector ki and the i th reply tail vector li in the preset reply tail group;

adding the second query vector s 1 and the reply sub-vector o 1 of the first cycle to obtain the second query vector s 2 of the second cycle;

The reply sub-vector o 2 of the second cycle and the second query vector s 3 of the third cycle are recalculated according to the second query vector s 2 of the second cycle, and so on until the Nth cycle is calculated. cyclic reply subvector o N ;

Take the reply sub-vector o N of the Nth cycle as the reply vector r 1 .
The dialogue generation method of claim 6, wherein the method further comprises:

obtaining a reply information base; wherein the reply information base includes a plurality of reply information represented in the form of knowledge triples, and the reply information includes: a head, a relation part and a tail;

The head in each reply message is converted into a reply head vector by a preset conversion and embedded TransE algorithm, thereby forming a reply head group;

The tail in each reply message is converted into a reply tail vector by the preset transformation and embedded TransE algorithm, thereby forming a reply tail group;

The correspondence between the reply head vector and the reply tail vector is established according to the relation part in each reply information.
A dialogue generation device, wherein the device comprises:

an acquisition module for acquiring question information, and converting the question information into a first query vector by using a preset first gate recursive unit GRU model;

The questioning module is configured to use the preset first end-to-end memory network MemN2N model according to the first query vector to determine the common sense vector associated with the first query vector, and to determine the common sense vector associated with the first query vector according to the first query vector and all Describe the common sense vector to form the question vector;

The answering module is configured to convert the question vector into a plurality of second query vectors according to the question vector using a preset second gate recursive unit GRU model, and input each second query vector to the preset second terminal in turn end-to-end memory network MemN2N model to get multiple reply vectors;

The conversion module is used to convert each reply vector into reply words respectively, and combine all reply words into reply information.
A computer device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program :

Obtain question information, and utilize the preset first gate recursive unit GRU model to convert the question information into a first query vector;

According to the first query vector, use the preset first end-to-end memory network MemN2N model to determine the common sense vector associated with the first query vector, and form a question according to the first query vector and the common sense vector vector;

According to the question vector, the question vector is converted into a plurality of second query vectors by using the preset second gate recursive unit GRU model, and each second query vector is sequentially input into the preset second end-to-end memory network MemN2N model to get multiple reply vectors;

Each reply vector is converted into reply words separately, and all reply words are combined into reply information.
The computer device according to claim 9, wherein the processor executes the computer program to realize the acquisition of question information, and converts the question information into a first query by using a preset first gate recursive unit GRU model When a vector, include:

Perform word segmentation processing on the question information, and form a word sequence with a plurality of keywords obtained after the word segmentation processing;

For a target keyword in the word sequence, according to the hidden influence factor transmitted to the target keyword by the keyword located before the target keyword in the word sequence, the first recursive unit GRU is used. The model calculates the hidden influence factor of the target keyword to the keyword located after the target keyword in the word sequence;

The hidden influence factor calculated according to the last keyword in the word sequence is used as the first query vector u 1 corresponding to the question information.
The computer device according to claim 10, wherein the processor executes the computer program to realize the determination of the correlation with the predetermined first end-to-end memory network MemN2N model according to the first query vector. When the common sense vector associated with the first query vector is described, and the question vector is formed according to the first query vector and the common sense vector, it includes:

In the first cycle of the first end-to-end memory network MemN2N model, the correlation between the first query vector u 1 and the ith common sense head vector x i in the preset common sense head group is calculated respectively. degree value p i ;

According to the correlation value pi of the ith common sense head vector x i and the ith common sense tail vector yi in the preset common sense tail group, calculate the question sub-vector a 1 of the first cycle;

adding the first query vector u 1 and the question vector a 1 to obtain the first query vector u 2 of the second cycle;

According to the first query vector u 2 of the second cycle, the question sub-vector a 2 of the second cycle and the first query vector u 3 of the third cycle are recalculated, and so on until the M-th cycle is calculated. cyclic question sub-vector a M ;

The question sub-vector a M of the M-th cycle is used as the question vector.
The computer device according to claim 11, wherein the processor further implements the following steps when executing the computer program:

Obtaining a common sense information base; wherein, the common sense information base includes a plurality of common sense information represented in the form of knowledge triples, and the common sense information includes: a head, a relation part, and a tail;

By presetting the first hidden layer matrix, the head in each common sense information is converted into a common sense head vector, thereby forming a common sense head group;

By presetting the second hidden layer matrix, the tail in each common sense information is converted into a common sense tail vector, thereby forming a common sense tail group;

The correspondence between the common sense head vector and the common sense tail vector is established according to the relationship part in each common sense information.
The computer device according to claim 9, wherein the processor executes the computer program to realize the conversion of the question vector into a plurality of The second query vector, and each second query vector is sequentially input into the preset second end-to-end memory network MemN2N model to obtain multiple reply vectors, including:

The question vector is used as the hidden influence factor h 0 of the first layer, and the preset starting character vector s 0 is input into the second gate recursive unit GRU model to obtain the output vector s 1 and pass to the second layer. The hidden impact factor h 1 of ;

inputting the output vector s 1 into the second end-to-end memory network MemN2N model as a second query vector to obtain a reply vector r 1 ;

Re-input the output vector s 1 and the hidden influence factor h 1 of the second layer into the second gate recursive unit GRU model to obtain the output vector s 2 and the hidden influence factor h 2 passed to the third layer, and re-input the output vector s 2 into the second end-to-end memory network MemN2N model to obtain a reply vector r 2 , and so on, until the output vector of the second gate recursive unit GRU model is the Let the ending character vector.
14. The computer device of claim 13, wherein the processor executes the computer program to implement the input of the output vector s 1 as a second query vector to the second end-to-end memory network MemN2N model , to get the reply vector r1 , including:

In the first cycle of the second end-to-end memory network MemN2N model, the correlation between the second query vector s 1 and the ith reply header vector ki in the preset reply header group is calculated respectively degree value p i ;

Calculate the reply sub-vector o 1 of the first cycle according to the correlation value p i of the ith reply head vector ki and the i th reply tail vector li in the preset reply tail group;

adding the second query vector s 1 and the reply sub-vector o 1 of the first cycle to obtain the second query vector s 2 of the second cycle;

The reply sub-vector o 2 of the second cycle and the second query vector s 3 of the third cycle are recalculated according to the second query vector s 2 of the second cycle, and so on until the Nth cycle is calculated. cyclic reply subvector o N ;

Take the reply sub-vector o N of the Nth cycle as the reply vector r 1 .
The computer device of claim 14, wherein the processor further implements the following steps when executing the computer program:

obtaining a reply information base; wherein, the reply information base includes a plurality of reply information represented in the form of knowledge triples, and the reply information includes: a head, a relation part and a tail;

The head in each reply message is converted into a reply head vector by the preset conversion and embedded TransE algorithm, thereby forming a reply head group;

The tail in each reply message is converted into a reply tail vector by the preset transformation and embedded TransE algorithm, thereby forming a reply tail group;

The correspondence between the reply head vector and the reply tail vector is established according to the relation part in each reply information.
A computer-readable storage medium on which a computer program is stored, wherein the computer program implements the following steps when executed by a processor:

Obtain question information, and utilize the preset first gate recursive unit GRU model to convert the question information into a first query vector;

According to the first query vector, use the preset first end-to-end memory network MemN2N model to determine the common sense vector associated with the first query vector, and form a question according to the first query vector and the common sense vector vector;

According to the question vector, the question vector is converted into a plurality of second query vectors by using the preset second gate recursive unit GRU model, and each second query vector is sequentially input into the preset second end-to-end memory network MemN2N model to get multiple reply vectors;

Each reply vector is converted into reply words separately, and all reply words are combined into reply information.
The computer-readable storage medium according to claim 16, wherein the computer program is executed by the processor to realize the obtaining of the question information, and use a preset first gate recursive unit GRU model to convert the question information into The first query vector includes:

Perform word segmentation processing on the question information, and form a word sequence with a plurality of keywords obtained after the word segmentation processing;

For a target keyword in the word sequence, according to the hidden influence factor transmitted to the target keyword by the keyword located before the target keyword in the word sequence, the first recursive unit GRU is used. The model calculates the hidden influence factor of the target keyword to the keyword located after the target keyword in the word sequence;

The hidden influence factor calculated according to the last keyword in the word sequence is used as the first query vector u 1 corresponding to the question information.
The computer-readable storage medium according to claim 17, wherein, when the computer program is executed by the processor, the computer program is executed to realize the determination according to the first query vector by using a preset first end-to-end memory network MemN2N model. When generating a common sense vector associated with the first query vector, and forming a question vector according to the first query vector and the common sense vector, including:

In the first cycle of the first end-to-end memory network MemN2N model, the correlation between the first query vector u 1 and the ith common sense head vector x i in the preset common sense head group is calculated respectively. degree value p i ;

According to the correlation value pi of the ith common sense head vector x i and the ith common sense tail vector yi in the preset common sense tail group, calculate the question sub-vector a 1 of the first cycle;

adding the first query vector u 1 and the question vector a 1 to obtain the first query vector u 2 of the second cycle;

According to the first query vector u 2 of the second cycle, the question sub-vector a 2 of the second cycle and the first query vector u 3 of the third cycle are recalculated, and so on until the M-th cycle is calculated. cyclic question sub-vector a M ;

The question sub-vector a M of the M-th cycle is used as the question vector.
The computer-readable storage medium of claim 18, wherein the computer program, when executed by the processor, further implements the following steps:

Obtaining a common sense information base; wherein, the common sense information base includes a plurality of common sense information represented in the form of knowledge triples, and the common sense information includes: a head, a relation part, and a tail;

By presetting the first hidden layer matrix, the head in each common sense information is converted into a common sense head vector, thereby forming a common sense head group;

By presetting the second hidden layer matrix, the tail in each common sense information is converted into a common sense tail vector, thereby forming a common sense tail group;

The corresponding relationship between the common sense head vector and the common sense tail vector is established according to the relationship part in each common sense information.
The computer-readable storage medium according to claim 16, wherein the computer program is executed by the processor to realize the transformation of the question vector according to the question vector using a preset second gate recursive unit GRU model is multiple second query vectors, and each second query vector is sequentially input into the preset second end-to-end memory network MemN2N model to obtain multiple reply vectors, including:

The question vector is used as the hidden influence factor h 0 of the first layer, and the preset starting character vector s 0 is input into the second gate recursive unit GRU model to obtain the output vector s 1 and pass to the second layer. The hidden impact factor h 1 of ;

inputting the output vector s 1 into the second end-to-end memory network MemN2N model as a second query vector to obtain a reply vector r 1 ;

Re-input the output vector s 1 and the hidden influence factor h 1 of the second layer into the second gate recursive unit GRU model to obtain the output vector s 2 and the hidden influence factor h 2 passed to the third layer, and re-input the output vector s 2 into the second end-to-end memory network MemN2N model to obtain a reply vector r 2 , and so on, until the output vector of the second gate recursive unit GRU model is the Let the ending character vector.