CN110858226A - Dialogue management method and device - Google Patents

Abstract

The invention discloses a conversation management method and device, and relates to the technical field of computers. One embodiment of the method comprises: processing user input data to be processed according to the natural language understanding model to obtain characteristic information of the user input data to be processed; determining historical record information corresponding to the user input data to be processed, and determining the current conversation state according to the characteristic information and the historical record information; and generating response data corresponding to the user input data to be processed based on a neural network algorithm according to the current conversation state. According to the embodiment, the characteristic information of the user input data is acquired by using the natural language understanding model, then the current conversation state is determined according to the historical record information, and then the response data is generated, so that the design and maintenance cost of a large number of regular programs is reduced, the accuracy of recognizing the user input data is improved, and the user experience is improved.

Description

Dialogue management method and device

technical field

The present invention relates to the field of computer technology, and in particular, to a dialog management method and device.

Background technique

In recent years, with the continuous innovation and progress of artificial intelligence, the way users buy and sell goods, provide or receive services, and engage in other business activities has gradually shifted to the way of human-computer interaction. In the interaction process, it is necessary to determine the user's intention according to the user's request through the dialog management method, and then feedback the corresponding response result to the user. Therefore, the dialogue management method plays a crucial role in the process of human-computer interaction, and its quality directly affects the efficiency of human-computer interaction and the user's experience.

The prior art dialog management method is a dialog management method based on vocabulary template matching, which is specifically explained as follows: pre-writing program codes that match sentence patterns, using the pre-written program codes to perform matching processing on user input requests, and then according to the matching As a result, the sentence that the robot needs to reply is obtained from the background database. For example, if the user says "I want to buy a mobile phone", the machine will ask back "What brand of mobile phone do you want?".

In the process of realizing the present invention, the inventor found that there are at least the following problems in the prior art: 1. In the prior art dialog management method, the programmer needs to develop more regular words to match the possible input information of the user. Adding new words, the backend server needs to refactor the program, and the project management is not flexible enough; 2. In the existing dialogue management method, because the program cannot match all the sentences the user wants to say in advance, it will lead to real online There is a situation where the business cannot be matched.

SUMMARY OF THE INVENTION

In view of this, the embodiments of the present invention provide a dialog management method and device, which can reduce the design and maintenance costs of a large number of regular programs, improve the accuracy of identifying user input data, and improve user experience.

To achieve the above object, according to an aspect of the embodiments of the present invention, a dialog management method is provided.

A dialog management method according to an embodiment of the present invention includes: processing user input data to be processed according to a natural language understanding model, acquiring feature information of the user input data to be processed; determining a history record corresponding to the user input data to be processed information, determine the current dialogue state according to the feature information and the history record information; and generate response data corresponding to the user input data to be processed based on the current dialogue state based on the neural network algorithm.

Optionally, before analyzing the user input data to be processed according to the natural language understanding model, the method further includes: acquiring a first sample set, the first sample set including at least one user input sample data; The sample data input by the user is subjected to labeling processing, and the feature information of the sample data input by the user is obtained to generate a second sample set, and the second sample set includes the feature information of the sample data input by the user; using the first sample and the second sample set to construct a training sample set; the training sample set is trained to obtain the natural language understanding model, the input of the natural language understanding model is user input data, and the output is user input data characteristic information.

Optionally, the historical record information includes: user historical input data, feature information of the user historical input data, and a dialog state of the user historical input data, the user historical input data is the pre-processing before the user input data to be processed. Let n rounds of user input data, n be an integer not less than zero.

Optionally, determining the current dialogue state according to the feature information and the historical record information includes: judging the user input data to be processed and the user historical input data according to the overlapping degree principle, the scene classification principle or the model correlation principle. Whether it is related; when the user input data to be processed is related to the user historical input data, update the current dialog state according to the dialog state of the user historical input data and the user input data to be processed; When the user input data to be processed is not related to the historical user input data, a new dialog state is generated according to the user input data to be processed, and the new dialog state is determined as the current dialog state.

Optionally, the feature information includes slot value information; and judging whether the user input data to be processed and the user historical input data are related according to the overlapping degree principle includes: calculating the slot value information of the user input data to be processed and The overlap degree of the slot value information of the user historical input data is determined, and whether the overlap degree is greater than a preset overlap degree is determined. If so, it is confirmed that the user input data to be processed is related to the user historical input data.

Optionally, the feature information includes scene information; and judging whether the user input data to be processed and the user historical input data are related according to the scene classification principle includes: judging the scene information of the user input data to be processed and the user input data to be processed. Whether the scene information of the user's historical input data is the same, if so, consider that the scene of the user's input data to be processed is the same as the scene of the user's historical input data, and confirm that the user's input data to be processed and the user's historical input data are the same. related.

Optionally, judging whether the to-be-processed user input data and the user historical input data are related according to the model correlation principle includes: comparing the to-be-processed user input data and the user historical input data according to a pre-built context-dependent model. A correlation analysis is performed, and if the correlation analysis result is related, it is confirmed that the user input data to be processed is related to the user historical input data.

Optionally, before the correlation analysis is performed on the to-be-processed user input data and the user historical input data according to a pre-built context-dependent model, the method further includes: constructing a context-dependent model based on a dynamic memory network algorithm.

Optionally, after judging whether the user input data to be processed and the user historical input data are related according to the overlapping degree principle, the scene classification principle or the model correlation principle, the method further includes: calculating using a neural network gate function. The gate weight of the user input data to be processed; according to the gate weight, calculate the correlation value between the user input data to be processed and the user historical input data; if the correlation value is greater than the preset correlation threshold, Then, it is confirmed that the user input data to be processed is related to the user historical input data.

To achieve the above object, according to another aspect of the embodiments of the present invention, an object management apparatus is provided.

A dialogue management device according to an embodiment of the present invention includes: an acquisition module, configured to process user input data to be processed according to a natural language understanding model, and acquire feature information of the to-be-processed user input data; a determination module, configured to determine all The historical record information corresponding to the user input data to be processed, according to the feature information and the historical record information, to determine the current dialogue state; the generating module is used for, according to the current dialogue state, based on the neural network algorithm. Process the response data corresponding to the user input data.

Optionally, the obtaining module is further configured to: obtain a first sample set, where the first sample set includes at least one user input sample data; perform labeling processing on the user input sample data, and obtain the user input feature information of the sample data to generate a second sample set, the second sample set includes the feature information of the user input sample data; constructing a training sample set by using the first sample set and the second sample set; The training sample set is trained to obtain the natural language understanding model, where the input of the natural language understanding model is user input data, and the output is feature information of the user input data.

Optionally, the historical record information includes: user historical input data, feature information of the user historical input data, and a dialog state of the user historical input data, the user historical input data is the pre-processing before the user input data to be processed. Let n rounds of user input data, n be an integer not less than zero.

Optionally, the determining module is further configured to: according to the overlapping degree principle, the scene classification principle or the model correlation principle, determine whether the user input data to be processed and the user historical input data are related; When the input data is related to the user's historical input data, the current dialog state is updated according to the dialog state of the user's historical input data and the user input data to be processed; when the user input data to be processed is related to the user input data When the historical input data is irrelevant, a new dialog state is generated according to the user input data to be processed, and the new dialog state is determined as the current dialog state.

Optionally, the feature information includes slot value information; and the determining module is further configured to: calculate the degree of overlap between the slot value information of the user input data to be processed and the slot value information of the user historical input data, and Determine whether the overlap degree is greater than a preset overlap degree, and if so, confirm that the user input data to be processed is related to the user historical input data.

Optionally, the feature information includes scene information; and the determining module is further configured to: determine whether the scene information of the user input data to be processed is the same as the scene information of the user historical input data, and if so, it is considered that the scene information is the same. The scenario of the user input data to be processed is the same as the scenario of the user historical input data, and it is confirmed that the user input data to be processed is related to the user historical input data.

Optionally, the determining module is further configured to: perform a correlation analysis on the user input data to be processed and the user historical input data according to a pre-built context-dependent model, and if the correlation analysis result is relevant, confirm all the relevant data. The user input data to be processed is related to the user historical input data.

Optionally, the determining module is further configured to: construct a context-dependent model based on a dynamic memory network algorithm.

Optionally, the determining module is further configured to: calculate the gate weight of the user input data to be processed by using a neural network gate function; calculate the user input data to be processed and the user historical input data according to the gate weight If the correlation value is greater than the preset correlation threshold, it is confirmed that the user input data to be processed is related to the user historical input data.

To achieve the above object, according to yet another aspect of the embodiments of the present invention, an electronic device is provided.

An electronic device according to an embodiment of the present invention includes: one or more processors; and a storage device for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more programs The processor implements the dialog management method of the embodiment of the present invention.

To achieve the above object, according to yet another aspect of the embodiments of the present invention, a computer-readable medium is provided.

A computer-readable medium of an embodiment of the present invention stores a computer program thereon, and when the program is executed by a processor, the dialog management method of the embodiment of the present invention is implemented.

An embodiment of the above invention has the following advantages or beneficial effects: the natural language understanding model can be used to obtain the characteristic information of the user input data, and then the current dialogue state can be determined according to the historical record information, and then the response data can be generated, so that a large number of regular programs can be reduced. reduce the design and maintenance costs, improve the accuracy of identifying user input data, and improve user experience; in the embodiment of the present invention, the training sample set composed of the first sample set and the second sample set is trained to obtain a natural language understanding model, As a result, a natural language understanding model can be constructed by using massive sample set data, which improves the accuracy of identifying user input data; in the embodiment of the present invention, the user input to be processed is judged from multiple perspectives: the overlapping degree principle, the scene classification principle or the model correlation principle. The correlation between the data and the user's historical input data can improve the accuracy of the predicted context relationship and further improve the user experience; The relevance of the context can be predicted from the perspective of the keywords of the input data; in the embodiment of the present invention, the relevance of the user input data to be processed and the user historical input data is judged according to the scene information, so that the context can be predicted according to the scene of the input data. Correlation; in the embodiment of the present invention, the correlation between the user input data to be processed and the user historical input data is judged according to a pre-built context correlation model, so that the context correlation can be predicted with the help of a model generated by massive data; this In the embodiment of the invention, the correlation value between the user input data to be processed and the user historical input data is also calculated according to the neural network gate function, so that the accuracy of predicting the contextual correlation can be improved.

Further effects of the above non-conventional alternatives will be described below in conjunction with specific embodiments.

Description of drawings

The accompanying drawings are used for better understanding of the present invention and do not constitute an improper limitation of the present invention. in:

1 is a schematic diagram of the main steps of a dialog management method according to an embodiment of the present invention;

2 is a schematic diagram of slot value information under a mobile phone category according to an embodiment of the present invention;

3 is a schematic diagram of a format for labeling slot value information according to an embodiment of the present invention;

4 is a schematic diagram of the main flow of a dialog management method according to a referenced embodiment of the present invention;

Fig. 5 is the overall structure diagram that realizes the dialogue management method of the present invention;

6 is a schematic diagram of the main modules of the dialog management apparatus according to an embodiment of the present invention;

FIG. 7 is an exemplary system architecture diagram to which an embodiment of the present invention may be applied;

FIG. 8 is a schematic structural diagram of a computer system suitable for implementing a terminal device or a server according to an embodiment of the present invention.

Detailed ways

Exemplary embodiments of the present invention are described below with reference to the accompanying drawings, which include various details of the embodiments of the present invention to facilitate understanding and should be considered as exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the invention. Also, descriptions of well-known functions and constructions are omitted from the following description for clarity and conciseness.

FIG. 1 is a schematic diagram of main steps of a dialog management method according to an embodiment of the present invention. As an embodiment of the present invention, as shown in FIG. 1 , the main steps of the dialog management method according to the embodiment of the present invention may include:

Step S101: Process the user input data to be processed according to the natural language understanding model, and acquire feature information of the user input data to be processed.

In the process of human-computer interaction, it is carried out in the form of a sentence of user input data and a sentence of machine response data. The user input data may be text data directly input by the user, or may be voice data input by the user. When the user input data is voice data, it is necessary to first identify the voice data with voice recognition software to obtain its corresponding text data. The response data refers to the data corresponding to the user input data fed back by the machine to the user, which may be response text data or response voice data. The machine in the present invention may be an intelligent robot, an intelligent navigation customer service on a shopping application, or an intelligent customer service software of a computer, which is not limited in the present invention.

A sentence of user input data and its corresponding sentence of response data constitute a round of dialogue data. In the process of human-computer interaction, single or multiple rounds of dialogue data can be included. In the present invention, the user input data in each round of dialogue data can be regarded as the user input data to be processed. In this step, the user input data to be processed is used as the input of the pre-built natural language understanding model, and then the feature information of the user input data to be processed is obtained. For example, the user inputs "I want to buy a computer", and the acquired characteristic information is "purchase" and "computer".

As another embodiment of the present invention, before analyzing the user input data to be processed according to the natural language understanding model in step S101, the dialog management method may further include: constructing a natural language understanding model in advance. Among them, building a natural language understanding model can specifically include:

Step S1011: Obtain a first sample set, where the first sample set may include at least one user input sample data;

Step S1012: Perform labeling processing on the sample data input by the user, and obtain feature information of the sample data input by the user, so as to generate a second sample set, wherein the second sample set may include the feature information of the sample data input by the user;

Step S1013: using the first sample set and the second sample set to construct a training sample set;

Step S1014: Train the training sample set to obtain a natural language understanding model, the natural language understanding model inputs the user input data, and outputs the feature information of the user input data.

The user input data in the first sample set may be online real user data. Of course, after the real user data is obtained in the present invention, these real data need to be cleaned, and words without information content in the real data are removed, for example, "Haha" and "Hey", etc., the blacklist data should also be cleaned, for example, "Voice is forbidden on the Internet" and "Profanity". The present invention performs simple labeling and manual labeling on the cleaned real user data, and then obtains characteristic information of the real user data. After a long period of statistics and accumulation, online real user data belongs to massive data, thus improving the accuracy of the model.

In this embodiment of the present invention, the feature information may include scene information and slot value information. Context information refers to determining the user's intent based on user data. For example, if the user enters "I want to buy a computer" through the intelligent customer service, it can be determined that the user's purpose is to buy, and the intelligent customer service conducts pre-sales consultation; if the user enters "I want to return the previously ordered product" through the intelligent customer service, It can be determined that the user's purpose is to return goods, and after-sales return consultation is carried out through intelligent customer service. The slot value information refers to the keyword information in the data, and the slot value information is divided into the slot value information inform_slot in the user input data and the slot value information request_slot desired by the machine. For example, if a user enters "I want to buy a Xiaomi phone and want a bigger screen", it is necessary to mark "Xiaomi" as the brand word in the inform_slot, and "mobile phone" as the product word in the inform_slot. The machine only knows that the Xiaomi phone cannot make a recommendation immediately. If you need to know the price, color, memory and other slot value information request_slot, you can make better product recommendations. 2 is a schematic diagram of slot value information under a mobile phone category according to an embodiment of the present invention, where the slot value information may be in the form of a key-value pair, for example, the key of the slot value information is a brand word and the value is Huawei.

The simple labeling of the cleaned real user data above means that the program preliminarily labels the real user data according to the rules formulated in advance. Optimizing annotation refers to sending the simply annotated data to the annotator, who manually annotates, revises the wrong one, and retains the correct one, and generates the characteristic information of the real user data. Considering that the present invention involves multiple rounds of conversations, the slot value information that needs to be marked includes not only product words, brand words, modifiers, query range, channel number, etc., but also attribute slot values, such as product size, color, etc. , usage, etc., you may also need to mark the dynamic slot value, that is, the slot value that may change, such as the memory size under the mobile phone category, and the size under the clothing category. FIG. 3 is a schematic diagram of a format of labeling slot value information according to an embodiment of the present invention.

Step S102: Determine the history record information corresponding to the user input data to be processed, and determine the current dialog state according to the feature information and the history record data information. The dialogue state is the dialogue state obtained by the machine according to the user's input data during the entire multi-round conversation process. For example, the user's first round of input data is "I want to buy a computer". According to the user's first round of input data, the machine obtains the dialog status as "The user needs to buy a computer", and the response data is "May I ask what brand of computer do you need?" , and then the user's second round of input data is "brand A", and then the machine obtains the current dialogue state according to the user's second round of input data and the previous dialogue state is "the user needs to buy a computer with brand A".

In this embodiment of the present invention, the historical record information may include: user historical input data, feature information of the user historical input data, and dialog status of the user historical input data. Wherein, the historical user input data is the user input data of n preset rounds before the user input data to be processed, and n is an integer not less than zero. If n is zero, the history information is set to null. The historical record information is used to let the machine have memory knowledge. For example, in the above example, the user's second round of input data is the user's input data to be processed, then the machine needs to remember what the user's first round of input data is, so that in the subsequent human-computer interaction In , the machine remembers that the user's purpose is to buy a computer, and will consult the specific information about the computer the user needs, such as brand, price, and size. The purpose of recording the feature information of the first round of input data is to use the feature information to determine the relevance of the context, which will be described in detail below, and will not be explained here. The purpose of recording the dialog state of the user's historical input data is to determine whether the dialog state of the user input data to be processed has changed, which will be described in detail below and will not be explained in detail here.

As yet another embodiment of the present invention, determining the current dialog state according to the feature information and the history record information in step S102 may include:

S1021: According to the overlapping degree principle, the scene classification principle or the model correlation principle, determine whether the user input data to be processed is related to the user historical input data;

S1022: when the user input data to be processed is related to the user historical input data, update the current dialogue state according to the dialog state of the user historical input data and the user input data to be processed;

S1023: When the user input data to be processed and the user historical input data are not related, generate a new dialog state according to the user input data to be processed, and determine the new dialog state as the current dialog state.

In the present invention, at least one of the three options of the overlapping degree principle, the scene classification principle or the model correlation principle is used to determine whether the user's historical input data and the user's input data to be processed are related, and if so, the previous dialogue state is updated. , otherwise, start a new dialog state. For example, the user's input data in the first round is "I want to buy a computer", and the machine determines the dialog state according to the input data in the first round as "The user needs to buy a computer", and then sends the user "May I ask what brand of computer do you want?" Then the user's input data in the second round is "I want to buy a mobile phone", the machine judges that the user's input data in the first round and the input data in the second round are irrelevant, therefore, the generated dialogue state is "user Need to buy a mobile phone", and send the response data of "What brand of mobile phone do you need" to the user, then the user's input data in the third round is "brand B", and the machine determines the dialog state according to the input data in the first round as "" The user needs to buy a computer", the machine judges that the user's second round of input data and the third round of input data are related, and the update dialog state is "the user needs to buy a mobile phone of brand B".

It should be noted that the historical user input data in the present invention is the user input data of n preset rounds before the user input data to be processed, and n is an integer not less than zero. Therefore, if the user input data to be processed is the fifth round user input data, and the value of n is 3, then the method for judging whether the fifth round user input data and user historical input data are related is specifically explained as:

First, determine whether the 5th round of user input data and the 4th round of user input data are relevant, if relevant, update the current dialogue state according to the dialog state corresponding to the 4th round of user input data and the 5th round of user input data;

When the 5th round of user input data and the 4th round of user input data are not related, judge whether the 5th round of user input data and the 3rd round of user input data are related, if relevant, according to the dialog state corresponding to the 3rd round of user input data and the fifth round of user input data to update the current dialog state;

When the 5th round of user input data and the 3rd round of user input data are not related, judge whether the 5th round of user input data and the 2nd round of user input data are related, if relevant, according to the dialog state corresponding to the 3rd round of user input data and the fifth round of user input data to update the current dialog state;

When the user input data in the fifth round is not related to the user input data in the second round, the current dialog state is generated directly according to the user input data in the fifth round.

In the embodiment of the present invention, after judging whether the user input data to be processed and the user historical input data are related in the above step S1021, the dialog management method may further include: using a neural network gate function to calculate the gate weight of the user input data to be processed; , calculate the correlation value between the user input data to be processed and the user historical input data; if the correlation value is greater than the preset correlation threshold, it is confirmed that the user input data to be processed and the user historical input data are related. In the present invention, after judging that the user input data to be processed is related to the user historical input data, the correlation value between the user input data to be processed and the user historical input data can be calculated, and the specific correlation degree can be obtained according to the correlation value. When the value is a certain value, it is considered that the user input data to be processed is related to the user historical input data, which can further ensure the accuracy of the context correlation prediction.

As another embodiment of the present invention, the feature information may include: slot value information. Therefore, judging whether the user input data to be processed and the user historical input data are related according to the overlapping degree principle in the above step S1021 may include: calculating the degree of overlap between the slot value information of the user input data to be processed and the slot value information of the user historical input data, And it is judged whether the overlap degree is greater than the preset overlap degree, and if so, it is confirmed that the user input data to be processed is related to the user historical input data. For example, the user's input data in the first round is "I want to buy a computer", then the slot value information of the first round of input data is "computer", and the user's input data in the second round is "I want to buy a computer of brand A" ", the slot value information of the second round of input data is "computer" and "brand A", it is determined that the overlap of the slot value information of the user's first round of input data and the user's second round of input data is very high, Therefore, it is considered that the user's first round input data and the user's second round input data are related.

As another embodiment of the present invention, the feature information may include: scene information. Therefore, judging whether the user input data to be processed and the user historical input data are related according to the scene classification principle in the above step S1021 may include: judging whether the scene information of the user input data to be processed and the scene information of the user historical input data are the same, and if so, then It is considered that the scenario of the user input data to be processed is the same as the scenario of the user historical input data, and it is confirmed that the user input data to be processed is related to the user historical input data. For example, the user's input data in the first round is "I want to buy a computer", then the scene information of the first round of input data is "buy a computer", and the user's input data in the second round is "I want to buy a mobile phone", then the first round of input data is "I want to buy a mobile phone". The scene information of the 2 rounds of input data is "purchasing a mobile phone", then it is considered that the scene information of the user's 1st round of input data and the user's 2nd round of input data are different. The 2 rounds of input data are irrelevant.

As yet another embodiment of the present invention, determining whether the user input data to be processed and the user history input data are related according to the model correlation principle in the above step S1021 may include: the user input data to be processed and the user history input data to be processed according to a pre-built context-dependent model Correlation analysis is performed on the input data. If the correlation analysis result is related, it is confirmed that the user input data to be processed is related to the user historical input data.

In the embodiment of the present invention, before the correlation analysis is performed on the user input data to be processed and the user historical input data according to the pre-built context-dependent model, the dialog management method may further include: constructing a context-dependent model based on a dynamic memory network algorithm. The dynamic memory network algorithm in the present invention includes four sub-modules, namely: an input text sub-module, a question sub-module, a situational memory network sub-module and an answer generation sub-module. The question sub-module is the user input data to be processed in the present invention, the episodic memory network sub-module uses a deep learning neural network to construct the relationship between the pending user input data and the user's historical input data, and the answer generation sub-module is to predict the pending user input data. The relationship between the input data and the user's historical input data, if there is a relationship, output yes, if not, output no.

The input text sub-module is to convert Chinese phrases into distributed vector numbers recognized by the computer. The input information can be voice data or text data. If it is voice data, it is necessary to first convert the voice data into text with the help of a voice recognition model. The data is then processed using natural language techniques. In the present invention, the input sequence T ₁ is composed of characters w ₁ , w ₂ , . . . , w _T and so on. The input sequence of the present invention is encoded by a cyclic neural network (Recurrent Neural Network, abbreviated as RNN) (it may be, but not limited to, RNN encoding, or other encoding forms, which are not limited in the present invention), and at each time t, the RNN network Both update the hidden state h _t =RNN(L[w _t ],h _t-1 ), where L is the encoding matrix, and w _t is the character index of the t-th Chinese character in the input sequence. If the user input text is a single short sentence, directly enter the RNN network and output the hidden state; if the user input text is a relatively long sequence, the invention will extract the slot value, and then store the slot value in the list, and the vector of the hidden state is used as output. The present invention can also select GRU (Gated Recurrent Unit, which is a variant of cyclic neural network, with relatively simple structure, composed of input gate, forget gate and output gate function) with lower computational complexity for coding. Assuming that at time t, the input text sequence is x _t and the hidden state is h _t , the structure of the GRU is defined as follows:

z _t =σ(W ^(z) x _t +U ^(z) h _t-1 +b ^(z) )

r _t =σ(W ^(r) x _t +U ^(r) h _t-1 +b ^(r) )

是自定义的一个算子，W是神经网路输入层的权重，U是隐藏层的权重，b是常量，W^(z)、W^(r)、

U^(z)、U^(r)、

和n这7个均为超参数，综合可以概括为h_t＝GRU(x_t,h_t-1)。in

is a user-defined operator, W is the weight of the input layer of the neural network, U is the weight of the hidden layer, b is a constant, W ^(z) , W ^(r) ,

U ^(z) , U ^(r) ,

and n are all hyperparameters, and the synthesis can be summarized as h _t =GRU(x _t ,h _t-1 ).

其中L表示字符短语编码矩阵，

表示问题中第t个字符或词语的索引。Like the input text sequence, the question sub-module is also input into the model in the form of characters, and is encoded by the recurrent neural network RNN. The question T _Q consists of a series of strings and is encoded as a mathematical vector,

where L represents the character phrase encoding matrix,

Represents the index of the t-th character or word in the question.

其中c_t是输入文本，m^i-1是记忆网络，q是问题，在本发明中指当前用户输入信息，门函数达到预先设定阈值或者满足最大迭代次数，网络就收敛，停止迭代。The episodic memory network sub-module is used to update the episodic memory inside the input text sub-module. The core component is iterative memory through the attention mechanism. In each iteration, the comprehensive vector c of the attention mechanism takes into account the current problem q and the previous Memory ^mi-1 once to generate the current scenario e ⁱ , and then the scenario vector e ⁱ and memory vector ^mi will be used. The formula for updating the scenario memory network is mi =GRU(e ⁱ ,m ^{i -1} ), The initial vector of GRU is the problem itself m ⁰ =q. With the construction of the network, various episodic memories may be required. For example, in the first sentence of the intelligent assistant, the user said "I browsed the mobile phone of brand B yesterday", and the second sentence said "But I want to buy a mobile phone with brand C today", these two sentences can be regarded as multiple scenarios. The user's current input information "I want to buy a mobile phone with brand C with a larger screen" is equivalent to a problem. Dynamic Memory Network It is necessary to identify whether the current information is related to the previous input according to the user's input of the first two sentences. The present invention uses the gate function as the attention mechanism, and the calculation formula is

Where c _t is the input text, m ^i-1 is the memory network, q is the problem, in the present invention it refers to the current user input information, the gate function reaches the preset threshold or satisfies the maximum number of iterations, the network converges and stops the iteration.

The answer generation submodule is used to generate answers through vector computation. The present invention can build another GRU network to generate answers. The last state m ^TM of the memory network is used as the input of the GRU network. At each moment, the input also includes the problem vector q, the hidden state a _t-1 of the previous layer and the previous prediction output y _t-1 , the calculation formula is y _t =soft max(W ^(a) at ), at = GRU([y _{t -1} , _q ], at _-1 ), the loss function of the present invention is cross entropy, and the binary classification is "yes" or "no" predictions, denoting contextually "relevant" or "irrelevant".

In the present invention, the parameters of the context-dependent model constructed based on the dynamic memory network algorithm are set as follows:

(1) The number of cells of the cyclic neural network, expressed in English as recurrent_cell_size, can be set to 128 in the present invention;

(2) In-depth vector dimension of text characters, expressed as D in English, and can be set to 50 in the present invention;

(3) The learning rate of the optimization algorithm in deep learning, expressed in English as Learning_rate, can be set to 0.005 in the present invention;

(4) The random neuron loss rate of the input layer, expressed as Input_p in English, can be set to 0.5 in the present invention;

(5) The random neuron loss rate of the output layer, expressed in English as Output_p, can be set to 0.5 in the present invention;

(6) The amount of data for training one batch at a time, expressed in English as Batch_size, can be set to 128 in the present invention;

(7) The memory segment in the episodic memory network, expressed in English as Passes, can be set to 4 in the present invention;

(8) The hidden layer network size of forward propagation, expressed in English as Ff_hidden_size, can be set to 256 in the present invention;

(9) The weight automatic decay rate, expressed in English as Weight_decay, can be set to 0.00000001 in the present invention;

(10) The number of questions that the network can train each time, expressed in English as Training_iterations_count, which can be set to 400,000 in the present invention;

(11) How many iterations are displayed in the verification set, which is expressed as Display_step in English, and can be set to 100 in the present invention;

(12) Session context, expressed in English as Context;

(13) Sentence end marker, expressed in English as Input_sentence_endings;

(14) The gate function of the neural network, expressed in English as Input_gru.

Step S103: According to the current dialogue state, the response data corresponding to the user input data to be processed is generated based on the neural network algorithm. For example, if the user's input data is "I want to buy a mobile phone with brand A", the brand slot value is "A", and the product slot value is "mobile phone". To make a recommendation immediately, you need to know the price, color, memory and other slot values. Therefore, the machine will generate the response data "May I ask what the price of the mobile phone is", and the user will continue to enter "I want to buy the price of the phone" when the user obtains the response data from the machine mobile phones in the range of 1,000 yuan to 2,000 yuan", then the machine will obtain the dialog status as "the user wants to buy a mobile phone with brand A and price range between 1,000 yuan and 2,000 yuan", and will then ask for other slot value information , and then recommend the mobile phone to the user. If the machine finds that there is no mobile phone suitable for the user, it will recommend other similar mobile phones to the user.

FIG. 4 is a schematic diagram of the main flow of a dialog management method according to a referenced embodiment of the present invention. As shown in Figure 4, the main flow of the dialogue management method may include:

Step S401: constructing a natural language understanding model, the constructed natural language understanding model inputs the user input data, and outputs the scene information and slot value information of the user input data;

Step S402: Process the user input data to be processed according to the natural language understanding model, and obtain scene information and slot value information of the user input data to be processed;

Step S403: Determine the historical record information corresponding to the user input data to be processed. The determined historical record information may include: the user historical input data, the feature information of the user historical input data, and the dialog state of the user historical input data, wherein the user historical input data The data is the user input data of the preset n rounds before the user input data to be processed, and n is an integer not less than zero;

Step S404: Calculate the degree of overlap between the slot value information of the user input data to be processed and the slot value information of the user's historical input data;

Step S405: determine whether the degree of overlap is greater than the preset degree of overlap, if so, go to step S409, otherwise, go to step S411;

Step S406: Determine whether the scene information of the user input data to be processed is the same as the scene information of the user historical input data, if so, go to step S409, otherwise, go to step S411;

Step S407: constructing a context-dependent model based on the dynamic memory network algorithm, and performing a correlation analysis on the user input data to be processed and the user's historical input data according to the pre-built context-dependent model;

Step S408: According to the correlation analysis result, determine whether the scene information of the user input data to be processed is related to the user historical input data, if so, go to step S409, otherwise, go to step S411;

Step S409: Calculate the correlation value between the user input data to be processed and the user's historical input data, and determine whether the calculated correlation value is greater than the preset correlation threshold, if so, go to step S410, otherwise, go to step S411;

Step S410: Update the current dialog state according to the dialog state of the user's historical input data and the user input data to be processed;

Step S411: generating a new dialogue state according to the user input data to be processed, and determining the new dialogue state as the current dialogue state;

Step S412: According to the current dialogue state, the response data corresponding to the user input data to be processed is generated based on the neural network algorithm.

It should be noted that the construction of the natural language understanding model in the above step S401 has been specifically explained in the above (steps S1011 to S1014 ), and will not be repeated here. In addition, the above also explains in detail how to calculate the correlation value between the user input data to be processed and the user historical input data, so step S409 will not be repeated. It should also be noted that in the above steps, as long as the user input data to be processed and the user historical input data meet the requirements in terms of the overlap of slot values, the same scene information, or the results of correlation analysis, the user input data to be processed and the user input data can be confirmed. The user's historical input data is relevant, and then step S409 is executed. In practical applications, all three aspects can be set to meet the requirements to confirm that the user input data to be processed and the user historical input data are related, or the method of setting weights for these three aspects can be used. Only when the conditions are met (for example, greater than a certain value), it can be confirmed that the user input data to be processed and the user historical input data are related. Of course, other methods can also be used, which are not limited in the present invention.

For the convenience of understanding, taking the "machine" as an "intelligent assistant" as an example, the overall framework for implementing the dialog management method of the present invention is described in detail, and the overall architecture diagram for implementing the dialog management method of the present invention is obtained as shown in FIG. 5 . As shown in FIG. 5 , the overall framework for implementing the dialogue management method of the present invention may include: an intelligent assistant multi-round dialogue labeling module, an intelligent assistant multi-round dialogue framework module, a dialogue management context relationship extraction module, a dynamic memory network algorithm module, a multi-round dialogue frame module Dialogue effect module and online test cut traffic module.

Among them, the intelligent assistant multi-round dialogue labeling module includes three units: voice assistant log table, dialogue cleaning and labeling slot value design. The voice assistant log table unit is used to extract real online user data, let the annotator mark it, obtain the marked corpus, and train the real data and the marked expected input model to generate a natural speech understanding model. Conversation cleaning The user cleans the user's input data and the assistant's response data. Annotation slot values are designed to define which slot value information needs to be obtained.

The intelligent assistant multi-round dialogue framework module is the overall framework logic abstraction of multi-round conversation, including: natural language understanding module, dialogue management sub-module and natural language generation sub-module. Among them, the natural speech understanding module is used to identify the user's intention (ie, identify the user's scene information), and extract slot value information such as product words, brand words, and modifier words. The natural speech understanding module can be subdivided into a scene intent classification model and a user natural language understanding model. The role of the scene intent classification model is to classify the user's input sentences in the intelligent assistant and determine which business scenario it is; the role of the user's natural language understanding model is to analyze the user's input sentences in the intelligent assistant, identify brand words, Slot value information such as product words and modifiers. The dialog management sub-module is used to identify the current dialog state of the user and determine whether the dialog state needs to be switched. The dialog management sub-module can be subdivided into multi-round dialog units, knowledge graph units and dialog state maintenance units. The multi-round dialogue unit is the interaction between the intelligent assistant and the user; the knowledge graph unit allows the intelligent assistant to have memorized knowledge. For example, if the user says "the capital of China", the intelligent assistant knows that it is Beijing; the dialogue state maintenance unit is the entire multi-round dialogue process. , the unit that makes the conversation work. The natural language generation module is used by the intelligent assistant to generate response data.

The dialogue management context extraction module is used to manage the context, identify whether the user's current input still belongs to the previous round of dialogue, and determine which of the extracted product words and brand words need to be remembered and which can be forgotten. The dialogue management context relationship extraction module is divided into three levels, namely, the overlap of slot values, the overlap of intentions and the dynamic memory network algorithm. These three levels predict the relationship of context from three different angles.

The dynamic memory network algorithm module is used to manage the context of multiple rounds of dialogue. If the user enters the intelligent assistant, it will be matched with the previous sentence by the dynamic memory network. If the network features match, it will be predicted as context-related, and it will not be closed. In the current dialogue, if the network features do not match, it will be predicted that the context is irrelevant, the current dialogue will be closed, and the next dialogue state will be performed. In this module, the above information input by the user and the current information input by the user are input into the dynamic network with episodic memory, and the prediction result is obtained to determine whether the context is related.

The multi-round dialogue effect module includes two parts: algorithm evaluation indicators and business evaluation indicators, which are used to evaluate the accuracy of dialogue management algorithms. The online test traffic cut module is used to test the intelligent assistant, test various user input data, and obtain the online model test results before the intelligent assistant goes online, so as to verify the effect of the dialogue management method, find problems and improve according to the test results. , so as to ensure the efficiency and reliability of on-line testing.

According to the technical solution for dialogue management in the embodiment of the present invention, it can be seen that the natural language understanding model can be used to obtain the characteristic information of the user input data, and then the current dialogue state can be determined according to the historical record information, and then the response data can be generated, so that a large number of regular expressions can be reduced. Program design and maintenance costs, improve the accuracy of identifying user input data, and improve user experience; in the embodiment of the present invention, the training sample set composed of the first sample set and the second sample set is trained to obtain a natural language understanding model , so that a natural language understanding model can be constructed by using massive sample set data, and the accuracy of identifying user input data is improved; in the embodiment of the present invention, the user to be processed is judged from multiple perspectives: the principle of overlap, the principle of scene classification, or the principle of model correlation The correlation between the input data and the user's historical input data can improve the accuracy of the predicted context relationship and further improve the user experience; Therefore, the relevance of the context can be predicted from the perspective of the keywords of the input data; in the embodiment of the present invention, the relevance of the user input data to be processed and the user historical input data can be judged according to the scene information, so that the context can be predicted according to the scene of the input data. In the embodiment of the present invention, the correlation between the user input data to be processed and the user historical input data is judged according to a pre-built context correlation model, so that the context correlation can be predicted with the help of a model generated by massive data; In the embodiment of the present invention, the correlation value between the user input data to be processed and the user historical input data is also calculated according to the neural network gate function, so that the accuracy of predicting the contextual correlation can be improved.

FIG. 6 is a schematic diagram of main modules of a dialog management apparatus according to an embodiment of the present invention. As shown in FIG. 6 , the dialog management apparatus 600 according to the embodiment of the present invention mainly includes the following modules: an acquisition module 601 , a determination module 602 , and a generation module 603 .

The obtaining module 601 may be configured to process the user input data to be processed according to the natural language understanding model, and obtain feature information of the user input data to be processed. The determining module 602 may be configured to determine the history record information corresponding to the user input data to be processed, and determine the current dialog state according to the feature information and the history record information. The generating module 603 may be configured to generate response data corresponding to the user input data to be processed based on the neural network algorithm according to the current dialogue state.

In this embodiment of the present invention, the obtaining module 601 may be further configured to: obtain a first sample set, where the first sample set includes at least one user-input sample data; perform labeling processing on the user-input sample data, and obtain feature information of the user-input sample data , to generate a second sample set, the second sample set contains the feature information of the user input sample data; use the first sample set and the second sample set to construct a training sample set; train the training sample set to obtain a natural language understanding model , the input of the natural language understanding model is the user input data, and the output is the feature information of the user input data.

In this embodiment of the present invention, the historical record information may include: user historical input data, feature information of the user historical input data, and dialog status of the user historical input data. The historical user input data is the user input data of the preset n rounds before the user input data to be processed, where n is an integer not less than zero.

In this embodiment of the present invention, the determining module 602 may be further configured to: determine whether the user input data to be processed and the user history input data are related according to the overlapping degree principle, the scene classification principle or the model correlation principle; When the input data is related, the current dialog state is updated according to the dialog state of the user's historical input data and the user's input data to be processed; when the user's input data to be processed is not related to the user's historical input data, a new dialog is generated according to the user's input data to be processed. state, and determine the new dialog state as the current dialog state.

In this embodiment of the present invention, the feature information may include: slot value information. The determining module 602 can also be used to: calculate the overlap degree between the slot value information of the user input data to be processed and the slot value information of the user historical input data, and determine whether the overlap degree is greater than the preset overlap degree, and if so, confirm the user input data to be processed. Related to user historical input data.

In this embodiment of the present invention, the feature information may further include: scene information. The determining module 602 can also be used for: judging whether the scene information of the user input data to be processed is the same as the scene information of the user's historical input data, if so, it is considered that the scene of the user input data to be processed is the same as the scene of the user's historical input data, and confirms that the scene information of the user input data to be processed is the same. Processing user input data is related to user historical input data.

In this embodiment of the present invention, the determining module 602 may be further configured to: perform a correlation analysis between the user input data to be processed and the user historical input data according to a pre-built context-dependent model, and if the correlation analysis result is relevant, confirm the user input data to be processed Related to user historical input data.

In this embodiment of the present invention, the determining module 602 may be further configured to: construct a context-dependent model based on a dynamic memory network algorithm.

In this embodiment of the present invention, the determining module 602 may be further configured to: calculate the gate weight of the user input data to be processed by using the neural network gate function; calculate the correlation value between the user input data to be processed and the user historical input data according to the gate weight; If the degree value is greater than the preset correlation threshold, it is confirmed that the user input data to be processed is related to the user historical input data

It can be seen from the above description that the natural language understanding model can be used to obtain the characteristic information of the user input data, and then the current dialogue state can be determined according to the historical record information, and then the response data can be generated, which can reduce the design and maintenance costs of a large number of regular programs and improve the Identify the accuracy of user input data and improve user experience; in the embodiment of the present invention, a training sample set composed of a first sample set and a second sample set is trained to obtain a natural language understanding model, so that a massive sample set can be used The data constructs a natural language understanding model, which improves the accuracy of identifying user input data; in the embodiment of the present invention, the user input data to be processed and the user historical input data are judged from multiple perspectives: the overlapping degree principle, the scene classification principle or the model correlation principle. Correlation, thereby improving the accuracy of predicting the context relationship and further improving the user experience; in the embodiment of the present invention, the degree of overlap between the user input data to be processed and the user historical input data is judged according to the slot value information, so that the keywords of the input data can be determined from the input data. From the perspective of the input data, the correlation of the context is predicted; in the embodiment of the present invention, the correlation between the user input data to be processed and the historical input data of the user is judged according to the scene information, so that the correlation of the context can be predicted according to the scene of the input data; the embodiment of the present invention According to the pre-built context-related model, the correlation between the user input data to be processed and the user's historical input data is judged, so that the context correlation can be predicted with the help of the model generated by massive data; The network gate function calculates the correlation value between the user input data to be processed and the user's historical input data, so that the accuracy of predicting contextual correlation can be improved.

FIG. 7 shows an exemplary system architecture 700 to which a dialog management method or a dialog management apparatus according to an embodiment of the present invention may be applied.

As shown in FIG. 7 , the system architecture 700 may include terminal devices 701 , 702 , and 703 , a network 704 and a server 705 . The network 704 is the medium used to provide the communication link between the terminal devices 701 , 702 , 703 and the server 705 . Network 704 may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others.

The user can use the terminal devices 701, 702, 703 to interact with the server 705 through the network 704 to receive or send messages and the like. Various communication client applications may be installed on the terminal devices 701 , 702 and 703 , such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only examples).

The terminal devices 701, 702, 703 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

The server 705 may be a server that provides various services, such as a background management server that provides support for shopping websites browsed by the terminal devices 701 , 702 , and 703 (just an example). The background management server can analyze and process the received product information query request and other data, and feed back the processing results (such as target push information, product information—just an example) to the terminal device.

It should be noted that the dialog management method provided by the embodiment of the present invention is generally executed by the server 705 , and accordingly, the dialog management apparatus is generally set in the server 705 .

It should be understood that the numbers of terminal devices, networks and servers in FIG. 7 are only illustrative. There can be any number of terminal devices, networks and servers according to implementation needs.

Referring next to FIG. 8 , it shows a schematic structural diagram of a computer system 800 suitable for implementing a terminal device according to an embodiment of the present invention. The terminal device shown in FIG. 8 is only an example, and should not impose any limitations on the functions and scope of use of the embodiments of the present invention.

As shown in FIG. 8, a computer system 800 includes a central processing unit (CPU) 801, which can be loaded into a random access memory (RAM) 803 according to a program stored in a read only memory (ROM) 802 or a program from a storage section 808 Instead, various appropriate actions and processes are performed. In the RAM 803, various programs and data necessary for the operation of the system 800 are also stored. The CPU 801 , the ROM 802 , and the RAM 803 are connected to each other through a bus 804 . An input/output (I/O) interface 805 is also connected to bus 804 .

The following components are connected to the I/O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc. ; and a communication section 809 including a network interface card such as a LAN card, a modem, and the like. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I/O interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is mounted on the drive 810 as needed so that a computer program read therefrom is installed into the storage section 808 as needed.

In particular, the processes described above with reference to the flowcharts may be implemented as computer software programs in accordance with the disclosed embodiments of the present invention. For example, embodiments disclosed herein include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart. In such an embodiment, the computer program may be downloaded and installed from the network via the communication portion 809, and/or installed from the removable medium 811. When the computer program is executed by the central processing unit (CPU) 801, the above-described functions defined in the system of the present invention are performed.

It should be noted that the computer-readable medium shown in the present invention may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or a combination of any of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), fiber optics, portable compact disk read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, however, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device . Program code embodied on a computer readable medium may be transmitted using any suitable medium including, but not limited to, wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logical functions for implementing the specified functions executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It is also noted that each block of the block diagrams or flowchart illustrations, and combinations of blocks in the block diagrams or flowchart illustrations, can be implemented in special purpose hardware-based systems that perform the specified functions or operations, or can be implemented using A combination of dedicated hardware and computer instructions is implemented.

The modules involved in the embodiments of the present invention may be implemented in a software manner, and may also be implemented in a hardware manner. The described modules can also be provided in the processor, for example, it can be described as: a processor includes an acquisition module, a determination module and a generation module. Among them, the names of these modules do not constitute a limitation of the module itself under certain circumstances. For example, the acquisition unit can also be described as "processing the user input data to be processed according to the natural language understanding model, and obtaining the user input data to be processed. module for feature information".

As another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiments; or may exist alone without being assembled into the device. The above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by a device, the device includes: processing the user input data to be processed according to the natural language understanding model, and obtaining the user input data to be processed. feature information of the data; determine the historical record information corresponding to the user input data to be processed, and determine the current dialog state according to the feature information and historical record information; generate response data corresponding to the user input data to be processed based on the current dialog state based on a neural network algorithm .

According to the technical solution of the embodiment of the present invention, the natural language understanding model can be used to obtain the characteristic information of the user input data, and then the current dialogue state can be determined according to the historical record information, and then the response data can be generated, thereby reducing the design and maintenance costs of a large number of regular programs. , improve the accuracy of identifying user input data and improve user experience; in the embodiment of the present invention, the training sample set composed of the first sample set and the second sample set is trained to obtain a natural language understanding model, so that a massive The sample set data builds a natural language understanding model, which improves the accuracy of identifying user input data; in the embodiment of the present invention, the user input data to be processed and the user historical input are judged from multiple perspectives: the principle of overlap, the principle of scene classification, or the principle of model correlation Therefore, the accuracy of predicting the context relationship can be improved, and the user experience can be further improved; in the embodiment of the present invention, the degree of overlap between the user input data to be processed and the user historical input data is determined according to the slot value information, so that the input data can be obtained from the input data. From the perspective of keywords, the correlation of the context is predicted; in the embodiment of the present invention, the correlation between the user input data to be processed and the historical input data of the user is judged according to the scene information, so that the correlation of the context can be predicted according to the scene of the input data; the present invention In the embodiment, the correlation between the user input data to be processed and the user's historical input data is judged according to a pre-built context-related model, so that the context correlation can be predicted with the help of a model generated by massive data; Calculate the correlation value between the user input data to be processed and the user historical input data according to the neural network gate function, so as to improve the accuracy of predicting contextual correlation

The above-mentioned specific embodiments do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (20)

1. a dialogue management method, is characterized in that, comprises: Process the user input data to be processed according to the natural language understanding model, and obtain feature information of the user input data to be processed; Determine the historical record information corresponding to the user input data to be processed, and determine the current dialog state according to the feature information and the historical record information; According to the current dialogue state, the response data corresponding to the user input data to be processed is generated based on a neural network algorithm. 2. The method according to claim 1, wherein before analyzing the user input data to be processed according to the natural language understanding model, the method further comprises: obtaining a first sample set, the first sample set including at least one user input sample data; Perform labeling processing on the user-input sample data, and obtain feature information of the user-input sample data to generate a second sample set, where the second sample set includes the feature information of the user-input sample data; Use the first sample set and the second sample set to construct a training sample set; The training sample set is trained to obtain the natural language understanding model, where the input of the natural language understanding model is user input data, and the output is feature information of the user input data. 3. The method according to claim 1, wherein the historical record information comprises: user historical input data, feature information of the user historical input data, and a dialog state of the user historical input data, the user historical input data is the preset n rounds of user input data before the to-be-processed user input data, where n is an integer not less than zero. 4. The method according to claim 3, wherein determining the current dialogue state according to the feature information and the history record information comprises: According to the overlapping degree principle, the scene classification principle or the model correlation principle, determine whether the user input data to be processed and the user historical input data are related; When the to-be-processed user input data is related to the user's historical input data, updating the current dialog state according to the dialog state of the user's historical input data and the to-be-processed user input data; When the user input data to be processed is not related to the user historical input data, a new dialog state is generated according to the user input data to be processed, and the new dialog state is determined as the current dialog state. 5. The method according to claim 4, wherein the feature information comprises slot value information; and judging whether the user input data to be processed and the user historical input data are related according to the overlapping degree principle comprises: Calculate the degree of overlap between the slot value information of the user input data to be processed and the slot value information of the user historical input data, and determine whether the overlap degree is greater than a preset overlap degree, and if so, confirm the pending user input The data is related to the user's historical input data. 6. The method according to claim 4, wherein the feature information comprises scene information; and judging whether the user input data to be processed and the user historical input data are related according to the scene classification principle comprises: Determine whether the scene information of the user input data to be processed is the same as the scene information of the user's historical input data, if so, consider that the scene of the user input data to be processed is the same as the scene of the user's historical input data, and confirm The user input data to be processed is related to the user historical input data. 7. The method according to claim 4, wherein judging whether the user input data to be processed and the user historical input data are related according to the model correlation principle comprises: Correlation analysis is performed on the user input data to be processed and the user historical input data according to a pre-built contextual correlation model, and if the correlation analysis result is related, the user input data to be processed and the user historical input data are confirmed. data related. 8. The method according to claim 7, wherein before the correlation analysis is performed on the user input data to be processed and the user historical input data according to a pre-built context-dependent model, the method further comprises: A context-dependent model is constructed based on the dynamic memory network algorithm. 9. The method according to claim 4, wherein after judging whether the user input data to be processed and the user historical input data are related according to the overlapping degree principle, the scene classification principle or the model correlation principle, the The method also includes: Calculate the gate weight of the user input data to be processed by using the neural network gate function; According to the gate weight, calculate the correlation value between the user input data to be processed and the user historical input data; If the relevancy value is greater than a preset relevancy threshold, it is confirmed that the user input data to be processed is related to the user historical input data. 10. A dialogue management device, comprising: an acquisition module, configured to process the user input data to be processed according to the natural language understanding model, and acquire feature information of the to-be-processed user input data; a determining module, configured to determine the historical record information corresponding to the user input data to be processed, and determine the current dialogue state according to the feature information and the historical record information; A generating module is configured to generate response data corresponding to the user input data to be processed based on a neural network algorithm according to the current dialogue state. 11. The apparatus according to claim 10, wherein the acquisition module is further configured to: obtaining a first sample set, the first sample set including at least one user input sample data; Perform labeling processing on the user-input sample data, and obtain feature information of the user-input sample data to generate a second sample set, where the second sample set includes the feature information of the user-input sample data; Use the first sample set and the second sample set to construct a training sample set; The training sample set is trained to obtain the natural language understanding model, where the input of the natural language understanding model is user input data, and the output is feature information of the user input data. 12. The device according to claim 10, wherein the historical record information comprises: user historical input data, feature information of user historical input data, and a dialog state of user historical input data, the user historical input data is the preset n rounds of user input data before the to-be-processed user input data, where n is an integer not less than zero. 13. The apparatus according to claim 12, wherein the determining module is further configured to: According to the overlapping degree principle, the scene classification principle or the model correlation principle, determine whether the user input data to be processed and the user historical input data are related; When the to-be-processed user input data is related to the user's historical input data, updating the current dialog state according to the dialog state of the user's historical input data and the to-be-processed user input data; When the user input data to be processed is not related to the user historical input data, a new dialog state is generated according to the user input data to be processed, and the new dialog state is determined as the current dialog state. 14. The apparatus according to claim 13, wherein the feature information comprises slot value information; and the determining module is further configured to: Calculate the degree of overlap between the slot value information of the user input data to be processed and the slot value information of the user historical input data, and determine whether the overlap degree is greater than a preset overlap degree, and if so, confirm the pending user input The data is related to the user's historical input data. 15. The apparatus according to claim 13, wherein the feature information comprises scene information; and the determining module is further configured to: Determine whether the scene information of the user input data to be processed is the same as the scene information of the user's historical input data, if so, consider that the scene of the user input data to be processed is the same as the scene of the user's historical input data, and confirm The user input data to be processed is related to the user historical input data. 16. The apparatus according to claim 13, wherein the determining module is further configured to: Correlation analysis is performed on the user input data to be processed and the user historical input data according to a pre-built contextual correlation model, and if the correlation analysis result is related, the user input data to be processed and the user historical input data are confirmed. data related. 17. The apparatus according to claim 16, wherein the determining module is further configured to: construct a context-dependent model based on a dynamic memory network algorithm. 18. The apparatus according to claim 13, wherein the determining module is further configured to: Calculate the gate weight of the user input data to be processed by using the neural network gate function; According to the gate weight, calculate the correlation value between the user input data to be processed and the user historical input data; If the relevancy value is greater than a preset relevancy threshold, it is confirmed that the user input data to be processed is related to the user historical input data. 19. An electronic device, characterized in that, comprising: one or more processors; storage means for storing one or more programs, The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any one of claims 1-9. 20. A computer-readable medium on which a computer program is stored, characterized in that, when the program is executed by a processor, the method according to any one of claims 1-9 is implemented.

Images