CN115617997A - A dialog state tracking method, device, equipment and medium - Google Patents

Abstract

The present application discloses a dialogue state tracking method, device, equipment and medium, which relate to the field of computers, including: using the first BiGRU neural network to encode the historical dialogue to obtain the first encoding result, and based on the attention and the domain vector to encode the first encoding result Perform feature extraction to obtain the first extraction result containing at least one field; use the second BiGRU neural network to encode the first extraction result to obtain the second encoding result, and perform feature extraction on the second encoding result based on attention and slot word vectors to obtain the second The second extraction result: input the second extraction result, domain-slot word vector and vocabulary vector into the decoder to obtain the decoded vector, calculate the predicted probability distribution of the slot value based on the decoded vector, and calculate the slot value under the domain based on the predicted probability distribution. For bit filling, this application solves the long-term dependency problem of historical dialogues in multiple domains based on the BiGRU neural network, and realizes the information sharing of different domain-slot pairs by processing domains and slots separately and paying attention in turn.

Description

A dialog state tracking method, device, equipment and medium

technical field

The present invention relates to the field of computer technology, in particular to a dialog state tracking method, device, equipment and medium.

Background technique

With the continuous development of artificial intelligence, man-machine dialogue technology has been widely used in various forms such as personal assistants, customer service robots and voice control systems in navigation, entertainment and communication. Dialogue systems can be divided into two categories: task-based and non-task-based. Regarding task-based dialogue systems, there are currently two main methods, namely the pipeline method and the end-to-end method. The pipeline method divides the entire human-machine dialogue process into automatic speech recognition (Automatic SpeechRecognition, ASR), natural language processing (Natural Language Understanding, NLU), dialogue management (Dialogue Management, DM), natural language generation (Natural Language Generation, NLG) and Speech synthesis (Text To Speech, TTS) five modules. Among them, the dialogue management module is equivalent to the brain of the dialogue system, which plays a very important role in the entire dialogue system and determines the corresponding actions to be taken by the system, including questioning, clarification and confirmation, etc., and its main subtasks are dialogue state tracking (Dialogue State Tracking, DST) and dialogue policy generation (Dialogue Policy, DP). The pipeline method can model and optimize each part separately, but it is easy to cause error transmission and accumulation in a multi-round dialogue system, making the overall performance worse. The end-to-end approach reduces the number of modules, simplifies the model structure, and facilitates global optimization. In a task-based dialogue system, it is difficult to fully and clearly express the user's needs in a single-talk dialogue. Therefore, users need to conduct multiple rounds of dialogue to gradually express their needs.

One of the existing technologies is the early dialogue tracking technology, which mainly utilizes human rule-based methods. Dialogue management based on finite state machine, slot filling and information state update all need to manually define a large number of rules. The method based on manual formulating rules needs to formulate rules according to task scenarios, but it cannot guarantee that all possible situations and dialogue rules can be exhaustively enumerated. In general, the dialog state tracking method based on rule templates is only suitable for simple tasks, not for complex tasks and cannot reuse rules after task changes.

In view of the limitations of rule-based methods, scholars mostly use data-driven methods for model research, resulting in generative methods and discriminative methods, which is the second prior art. It has the following disadvantages: (1) The characteristics of historical dialogue information are not fully extracted. In deep learning, Recurrent Neural Network (RNN) can be used to model time series and extract historical dialogue information features. However, RNN has the problem of gradient disappearance and gradient explosion. Although Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) can solve the problem of long sequence dependence through the gating mechanism, but They only have forward context information, ignoring backward context information; (2) In multi-domain scenarios, domain-slot value pair information is not fully shared.

For this reason, how to solve the long-term dependency problem of historical dialogues in multi-domain scenarios without ignoring the forward and backward context information, and realize information sharing between different domain-slot pairs is an urgent problem in this field.

Contents of the invention

In view of this, the object of the present invention is to provide a dialog state tracking method, device, device and medium, which can solve the long-term dependency problem of historical dialogs in multi-field scenarios without ignoring the forward and backward context information, and realize different The specific scheme for information sharing between domain-slot pairs is as follows:

In the first aspect, the present application discloses a dialogue state tracking method, including:

Use the first BiGRU neural network to encode the historical dialogue records to obtain the first encoding result, and perform domain-level feature extraction on the first encoding result based on the attention mechanism and the preset domain vector, and obtain a domain-level feature extraction including at least one domain. The first feature extraction result;

Use the second BiGRU neural network to encode the first feature extraction result to obtain the second encoding result, and perform slot-level feature extraction on the second encoding result based on the attention mechanism and the preset slot word vector , to obtain the second feature extraction result including slots;

Input the second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector to the decoder to obtain a decoded vector, and calculate the slot value to be filled based on the decoded vector The predicted probability distribution of the value of the slot to be filled is then based on the predicted probability distribution of the value of the slot to be filled to fill the slot under the at least one field, so as to realize the dialogue state tracking; wherein, the preset field-slot word The vector is spliced by the preset domain vector and the preset slot word vector.

Optionally, before said utilizing the first BiGRU neural network to encode the historical dialogue record, before obtaining the first encoding result, it also includes:

Acquiring several rounds of dialogue records, and splicing the several rounds of dialogue records to obtain the historical dialogue records.

Optionally, after inputting the second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector into the decoder, and obtaining the decoded vector, it further includes:

mapping the decoded vector to a target probability distribution based on a dialogue state tracking strategy;

When the target probability distribution is the first probability distribution, it means that the user has not mentioned the slots under the at least one field; The slots below remain indifferent, and when the target probability distribution is the third probability distribution, it means that the user has mentioned the slots under the at least one domain.

Optionally, the calculating the predicted probability distribution of the slot value to be filled based on the decoded vector includes:

When the target probability distribution is the third probability distribution, then calculate the probability distribution of slot values in the preset vocabulary dictionary and the probability distribution of slot values in the historical dialogue record based on the decoded vector Probability distributions;

The predicted probability distribution of the slot values to be filled is calculated according to the probability distribution of the slot values in the preset vocabulary dictionary and the probability distribution of the slot values in the historical dialogue records.

Optionally, the formula for calculating the probability distribution of slot values in the preset vocabulary dictionary and the probability distribution of slot values in the historical dialogue record based on the decoded vector is:

Wherein, V represents the preset vocabulary dictionary vector, H _decode represents the decoded vector, h represents the first encoding result, and p ^vocab represents the probability of a slot value in the preset vocabulary dictionary distribution, and p ^history represents the probability distribution of the slot values in the historical dialogue records.

Optionally, according to the probability distribution of the slot values in the preset vocabulary dictionary and the probability distribution of the slot values in the historical dialogue record, the formula for calculating the predicted probability distribution of the slot values to be filled for:

p ^value = p ^gen × p ^vocab + (1-p ^gen ) × p ^history ;

Wherein, p ^gen represents the weight of generating the value of the slot to be filled from the preset vocabulary dictionary.

Optionally, the second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector are input to the decoder to obtain a decoded vector, including:

The second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector are input to the decoder constructed by the third BiGRU neural network to obtain the decoded vector.

In a second aspect, the present application discloses a dialogue state tracking device, including:

The first encoding module is used to encode the historical dialogue record by using the first BiGRU neural network to obtain the first encoding result;

A domain-level feature extraction module, configured to perform domain-level feature extraction on the first encoding result based on an attention mechanism and a preset domain vector, to obtain a first feature extraction result including at least one domain;

The second encoding module is used to encode the first feature extraction result by using the second BiGRU neural network to obtain the second encoding result;

The feature extraction module at the slot level is used to perform feature extraction at the slot level on the second encoding result based on the attention mechanism and the preset slot word vector, to obtain a second feature extraction result including the slot;

A decoding module, configured to input the second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector to the decoder to obtain a decoded vector;

A dialog state tracking module, configured to calculate a predicted probability distribution of the slot values to be filled based on the decoded vector, and then perform an operation on the slots under the at least one domain based on the predicted probability distribution of the slot values to be filled. filling, so as to realize dialogue state tracking; wherein, the preset domain-slot word vector is spliced by the preset domain vector and the preset slot word vector.

In a third aspect, the present application discloses an electronic device, comprising:

memory for storing computer programs;

A processor is configured to execute the computer program to implement the dialog state tracking method disclosed above.

In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned dialog state tracking method disclosed above is implemented.

It can be seen that the present application proposes a dialogue state tracking method, including: using the first BiGRU neural network to encode the historical dialogue record, obtaining the first encoding result, and encoding the first encoding result based on the attention mechanism and the preset domain vector As a result, the feature extraction at the domain level is carried out to obtain the first feature extraction result including at least one domain; the second BiGRU neural network is used to encode the first feature extraction result to obtain the second encoding result, and based on the attention mechanism and the prediction The set slot word vector performs feature extraction at the slot level on the second encoding result to obtain a second feature extraction result including the slot; the second feature extraction result, the preset domain-slot word vector And the preset vocabulary dictionary vector is input to the decoder to obtain the decoded vector, and calculate the predicted probability distribution of the slot value to be filled based on the decoded vector, and then based on the predicted probability of the slot value to be filled Distributing to fill slots under the at least one domain so as to realize dialogue state tracking; wherein, the preset domain-slot word vector consists of the preset domain vector and the preset slot word Vector concatenation. In summary, since the BiGRU (bidirectional gated recurrent unit) neural network can extract context information using the forward and backward gated recurrent unit (GRU) structures, this application does not ignore the forward and backward context information based on the BiGRU neural network. Under the premise of solving the long-term dependency problem of historical dialogues in multi-domain scenarios; in addition, the traditional dialogue state tracking is to perform overall feature extraction on domain-slot pairs. In this way, the information between different slots in the same domain It is irrelevant, and the same slot information in different fields is also irrelevant, but this application separates the fields and slots and performs attention calculations on the fields and slots in turn, so that the difference between the extracted different field-slot pairs information can be shared.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention, and those skilled in the art can also obtain other drawings according to the provided drawings without creative work.

FIG. 1 is a flow chart of a dialog state tracking method disclosed in the present application;

FIG. 2 is a flow chart of a specific dialog state tracking method disclosed in the present application;

FIG. 3 is a schematic structural diagram of a dialog state tracking device disclosed in the present application;

FIG. 4 is a structural diagram of an electronic device disclosed in the present application.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Generative methods and discriminative methods have the following disadvantages: (1) The characteristics of historical dialogue information are not fully extracted. In deep learning, recurrent neural networks can be used to model time series and extract historical dialogue information features. However, RNN has the problem of gradient disappearance and gradient explosion. Although the long-short-term memory network and the gated recurrent unit can solve the problem of long-term sequence dependence through the gating mechanism, they only have forward context information and ignore the backward context. (2) In multi-domain scenarios, domain-slot pair information is not fully shared.

For this reason, the embodiment of the present application proposes a dialogue state tracking scheme, which can solve the long-term dependency problem of historical dialogues in multi-domain scenarios without ignoring the forward and backward context information, and realize the communication between different domain-slot pairs. Information Sharing.

The embodiment of the present application discloses a dialog state tracking method, as shown in Fig. 1, the method includes:

Step S11: Use the first BiGRU neural network to encode the historical dialogue records to obtain the first encoding result, and perform domain-level feature extraction on the first encoding result based on the attention mechanism and the preset domain vector to obtain at least A domain's first feature extraction results.

It should be pointed out that in the process of multiple rounds of dialogue tasks, there will often be meaningless dialogues and topic shifts, which will lead to unclear domain discrimination in multi-domain scenarios and mutual influence between contexts. Capturing historical dialogue information in different domains, so that it can better focus on information in its own domain without being interfered by information in other domains. For the problem of information interference in other domains, since the attention mechanism can ignore the distance of elements in the input sequence and directly capture the key information in the task, this application can better focus on the information in its own domain without interference from information in other domains. The specific process is as follows:

需要指出的是，此处编码是为了将领域词编码为领域向量。进一步的，对编码后的对话记录和编码后得到的领域向量进行注意力计算，从而完成领域特征提取。所述对编码后的对话记录和编码后得到的领域向量进行计算的过程，通过公式可以表现为以下形式：In this embodiment, it is first necessary to obtain several rounds of dialogue records, and splice the several rounds of dialogue records to obtain the historical dialogue records. Specifically, multiple rounds of dialogue D _T of T rounds of dialogue are sequentially spliced into x={u ₁ , s ₁ , u ₂ , s ₂ ,... u _T , s _T }, where u _t represents the User utterance, s _t represents the system utterance at time t. Further, use the first BiGRU neural network to encode x={u ₁ , s ₁ , u ₂ , s ₂ ,... u _T , s _T } to obtain the first encoding result, expressed as h={h ₁ ,h ₂ ,...h _T }, the first encoding result is the hidden layer output of the BiGRU neural network. It should be pointed out that the encoding here is for the subsequent extraction of domain-level features. Among them, the BiGRU neural network can use the forward and backward GRU structure to extract context information, and fully capture the semantic information contained in the dialogue sentence. After encoding the spliced dialog records, encode the domain words in the preset domain word database to obtain the preset domain vector, specifically, for a certain specific domain word in the preset domain word database domain word d _i , which can be coded as

It should be pointed out that the encoding here is to encode domain words into domain vectors. Further, attention calculation is performed on the encoded dialogue record and the encoded domain vector, so as to complete domain feature extraction. The process of calculating the encoded dialogue record and the encoded domain vector can be expressed in the following form through the formula:

Among them, y ⁱ represents the correlation between the i-th field in the preset field word vector and the historical dialogue record; y ⁱ _softmax represents the normalized y ⁱ , that is, the weight of each field i corresponding to the historical dialogue record ; y ⁱ _context represents the history vector h with domain weight; the history vector is the context vector, that is, the first encoding result obtained after encoding the historical dialogue records; T indicates transposition.

含有整个历史对话记录的上下文信息，因此，本申请使用

作为特定领域d_i的表示，然后，利用注意力机制计算上下文，通过赋予权重的方式提取神经网络输出的隐含表示，本实施例中，所述第一特征提取结果也即yⁱ _context。because

Contains the context information of the entire historical conversation record, therefore, this application uses

As the representation of the specific field d _i , then use the attention mechanism to calculate the context, and extract the hidden representation output by the neural network by assigning weights. In this embodiment, the first feature extraction result is also y ⁱ _context .

Step S12: Use the second BiGRU neural network to encode the first feature extraction result to obtain the second encoding result, and perform slot level analysis on the second encoding result based on the attention mechanism and the preset slot word vector feature extraction to obtain the second feature extraction result including slots.

需要指出的是，此处编码是为了将槽位词编码为槽位词向量，进一步的，对第二编码结果以及编码后得到的槽位词向量进行注意力计算，从而完成槽位词特征提取。所述对第二编码结果以及编码后得打的槽位词向量进行注意力计算的过程，通过公式可以表现为以下形式：In this embodiment, after obtaining the first feature extraction result with domain key information, the second BiGRU neural network is used to encode the first feature extraction result to obtain the second encoding result, denoted as h', it needs to be pointed out Interestingly, the encoding here is for the subsequent extraction of features at the slot level. After encoding the first feature extraction result, encode the slot word in the preset slot word database to obtain the preset slot word vector, specifically, for the preset slot word database A specific slot word s _j in , can be coded as

It should be pointed out that the encoding here is to encode the slot word into a slot word vector, and further, perform attention calculation on the second encoding result and the slot word vector obtained after encoding, so as to complete the slot word feature extraction . The process of calculating attention to the second encoding result and the slot word vectors to be played after encoding can be expressed in the following form through the formula:

Among them, z ^j represents the correlation between the jth field in the preset slot word vector and the historical dialogue record, and z ^j _softmax represents the normalized z ^j , that is, each slot word corresponds to the historical dialogue record weight; z ^j _context represents the history vector h with slot word weights.

含有整个历史对话信息的上下文信息，因此，本申请使用

作为特定槽位s_j的表示。然后，利用注意力机制计算上下文，通过赋予权重的方式提取神经网络输出的隐含表示，实施例中，所述第二特征提取结果也即z^j _context。如此一来，本申请能够将历史对话信息中关于槽位的信息突出表达。See how domain features are handled, due to

Contains the context information of the entire historical dialogue information, therefore, this application uses

as a representation of a specific slot s _j . Then, the attention mechanism is used to calculate the context, and the hidden representation output by the neural network is extracted by assigning weights. In the embodiment, the second feature extraction result is also z ^j _context . In this way, the present application can highlight the information about slots in the historical dialogue information.

It should be pointed out that in the dialogue state tracking, the traditional operation is to encode domain-slot pairs as a whole. In this way, different slots in the same domain have nothing to do with each other, and the same slots in different domains have nothing to do with each other. Also irrelevant. But in the case of multi-domain dialogue, all domain-slot pairs are not completely irrelevant. For example, location slot information will appear in the taxi and meal order domains. Furthermore, the data size between different domains in a multi-domain data set is often not completely balanced, and the amount of data in some domains may be small, resulting in insufficient model training, that is, traditional operations do not fully share domain-slot pairs. Compared with the traditional processing method based on domain-slot pairs, this application separates domains and slots, and performs attention calculations on domains and slots in turn, so as to realize the difference between different domain-slot pairs. information sharing. For example, assuming that domain A and domain B exist, the slot values corresponding to a certain slot in domain A are A1 and A2, and the slot values corresponding to the same slot in domain B are B1 and B2. The traditional operation is to Domain-slot pairs are encoded as a whole, that is, A-A1, A-A2, B-B1, and B-B2. The result of the encoding is that there is no relationship between the same slots under domain A and domain B, but in multi-domain dialogue In the case of , all field-slot pairs are not completely irrelevant. Therefore, this application separates the fields and slots, and performs attention calculations on the fields and slots in turn, that is, pays attention to the fields first. Operation, extract domain A and domain B, and then perform attention calculation on the slots, and then determine that the slot values to be filled may be A1, A2, B1, B2. At this time, the slot values corresponding to the slots under the domain are not only It may include A1 and A2, and may also include B1 and B2. In this way, information sharing among different field-slot pairs is realized.

Step S13: Input the second feature extraction result, the preset domain-slot word vector and the preset vocabulary dictionary vector into the decoder to obtain the decoded vector, and calculate the slot to be filled based on the decoded vector The predicted probability distribution of values, and then fill the slots under the at least one domain based on the predicted probability distribution of the slot values to be filled, so as to realize dialogue state tracking; wherein, the preset domain-slot The word vector is concatenated by the preset domain vector and the preset slot word vector.

Specifically, the second feature extraction result, the preset domain-slot word vector and the preset vocabulary dictionary vector are input to the decoder constructed by the third BiGRU neural network to obtain the decoded vector.

In this embodiment, after inputting the second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector into the decoder to obtain the decoded vector, it also includes: The decoded vector is mapped to a target probability distribution; when the target probability distribution is the first probability distribution, it means that the user has not mentioned the slots under the at least one domain; when the target probability distribution is the second probability distribution If the target probability distribution is the third probability distribution, it means that the user has mentioned the slots in the at least one field. The dialogue state tracking strategy is specifically the current mainstream DST (Dialogue State Tracking, dialogue state tracking) update strategy, that is, the classifier maps the decoded vector into the probability distribution of NONE, DONTCARE and MENTIONED, wherein when the probability distribution of mapping to NONE distribution (the first probability distribution), it means that the user has not mentioned the slots under the at least one domain; The slot is in an indifferent state, and when mapped to a MENTIONED probability distribution (the third probability distribution), it means that the user has mentioned the slot under the at least one domain.

In this embodiment, in the historical dialogue records, since the user's expression may be ambiguous, it is impossible to directly calculate the value of the slot to be filled according to the user's expression. Therefore, this application introduces a preset vocabulary dictionary, and the preset The vocabulary dictionary includes several slot values. Further, because the vocabulary dictionary cannot include all slot values, this application calculates the probability distribution of the slot values to be filled based on the vocabulary dictionary and historical dialogue records. . Specifically, when mapping to the third probability distribution, the probability distribution of slot values in the preset vocabulary dictionary and the probability distribution of slot values in the historical dialogue records are calculated based on the decoded vector ; Calculate the predicted probability distribution of the slot values to be filled according to the probability distribution of the slot values in the preset vocabulary dictionary and the probability distribution of the slot values in the historical dialogue records. Wherein, the formula for calculating the probability distribution of the slot value in the preset vocabulary dictionary and the probability distribution of the slot value in the historical dialogue record based on the decoded vector is:

Wherein, V represents the preset vocabulary dictionary vector, H _decode represents the decoded vector, h represents the first encoding result, and p ^vocab represents the probability of a slot value in the preset vocabulary dictionary distribution, and p ^history represents the probability distribution of the slot values in the historical dialogue records.

The formula for calculating the predicted probability distribution of the slot value to be filled according to the probability distribution of the slot value in the preset vocabulary dictionary and the probability distribution of the slot value in the historical dialogue record is:

p ^value = p ^gen × p ^vocab + (1-p ^gen ) × p ^history ;

Wherein, p ^gen represents the weight of generating the value of the slot to be filled from the preset vocabulary dictionary.

In this way, the present application determines the corresponding slot value according to the probability distribution of the slot values to be filled, and uses the slot value to fill slots in at least one domain.

It can be seen that the present application proposes a dialogue state tracking method, including: using the first BiGRU neural network to encode the historical dialogue record, obtaining the first encoding result, and encoding the first encoding result based on the attention mechanism and the preset domain vector As a result, the feature extraction at the domain level is carried out to obtain the first feature extraction result including at least one domain; the second BiGRU neural network is used to encode the first feature extraction result to obtain the second encoding result, and based on the attention mechanism and the prediction The set slot word vector performs feature extraction at the slot level on the second encoding result to obtain a second feature extraction result including the slot; the second feature extraction result, the preset domain-slot word vector And the preset vocabulary dictionary vector is input to the decoder to obtain the decoded vector, and calculate the predicted probability distribution of the slot value to be filled based on the decoded vector, and then based on the predicted probability of the slot value to be filled Distributing to fill slots under the at least one domain so as to realize dialogue state tracking; wherein, the preset domain-slot word vector consists of the preset domain vector and the preset slot word Vector concatenation. In summary, since the BiGRU (bidirectional gated recurrent unit) neural network can extract context information using the forward and backward gated recurrent unit (GRU) structures, this application does not ignore the forward and backward context information based on the BiGRU neural network. Under the premise of solving the long-term dependency problem of historical dialogues in multi-domain scenarios; in addition, the traditional dialogue state tracking is to perform overall feature extraction on domain-slot pairs. In this way, the information between different slots in the same domain It is irrelevant, and the same slot information in different fields is also irrelevant, but this application separates the fields and slots and performs attention calculations on the fields and slots in turn, so that the difference between the extracted different field-slot pairs information can be shared.

Fig. 2 is a flow chart of a specific dialog state tracking method disclosed in the present application. Referring to Fig. 2, the present application (1) first uses the first BiGRU neural network to encode the spliced historical dialogs to obtain the first coding result; (2) Perform domain-level feature extraction on the first encoding result based on the attention mechanism and the preset domain vector to obtain the first feature extraction result; (3) Use the second BiGRU neural network structure to encode the first feature extraction result , to obtain the second encoding result; (4) perform slot-level feature extraction on the second encoding result based on the attention mechanism and the preset slot word vector to obtain the second feature extraction result; (5) extract the second feature The result, the preset vocabulary dictionary vector and the field-slot word vector are input to the decoder to obtain the decoded vector, and the decoder adopts the third BiGRU neural network structure; (6) Calculate the slot value in the preset vocabulary dictionary and the probability of the slot value in the historical dialogue record; and calculate the final predicted probability distribution of the slot value based on the probability of the slot value in the preset vocabulary dictionary, the probability of the slot value in the historical dialogue record, and the decoded vector , and determine the corresponding slot value according to the final probability distribution of the slot value, and then complete the filling.

It can be seen that since the bidirectional gating unit can use the forward and backward gating cyclic unit structure to extract context information, this application solves the historical dialogue in multi-domain scenarios based on the BiGRU neural network without ignoring the forward and backward context information. Long-term dependency problem; this application realizes information sharing between different domain-slot pairs by separately processing domains and slots and performing attention operations on domains and slots in turn.

Correspondingly, the embodiment of the present application also discloses a dialog state tracking device, as shown in Fig. 3, the device includes:

The first encoding module 11 is used to encode the historical dialogue record by using the first BiGRU neural network to obtain the first encoding result;

A domain-level feature extraction module 12, configured to perform domain-level feature extraction on the first encoding result based on an attention mechanism and a preset domain vector, to obtain a first feature extraction result including at least one domain;

The second encoding module 13 is used to encode the first feature extraction result by using the second BiGRU neural network to obtain the second encoding result;

The slot-level feature extraction module 14 is used to perform feature extraction at the slot level on the second encoding result based on the attention mechanism and the preset slot word vector, to obtain a second feature extraction result including the slot;

The decoding module 15 is configured to input the second feature extraction result, the preset field-slot word vector and the preset vocabulary dictionary vector to the decoder to obtain the decoded vector;

Dialogue state tracking module 16, configured to calculate the predicted probability distribution of the slot value to be filled based on the decoded vector, and then calculate the slots under the at least one domain based on the predicted probability distribution of the slot value to be filled Filling is performed to realize dialogue state tracking; wherein, the preset domain-slot word vector is spliced by the preset domain vector and the preset slot word vector.

For the more specific working process of each of the above modules, reference may be made to the corresponding content disclosed in the foregoing embodiments, which will not be repeated here.

It can be seen that the present application proposes a dialogue state tracking method, including: using the first BiGRU neural network to encode the historical dialogue record, obtaining the first encoding result, and encoding the first encoding result based on the attention mechanism and the preset domain vector As a result, the feature extraction at the domain level is carried out to obtain the first feature extraction result including at least one domain; the second BiGRU neural network is used to encode the first feature extraction result to obtain the second encoding result, and based on the attention mechanism and the prediction The set slot word vector performs feature extraction at the slot level on the second encoding result to obtain a second feature extraction result including the slot; the second feature extraction result, the preset domain-slot word vector And the preset vocabulary dictionary vector is input to the decoder to obtain the decoded vector, and calculate the predicted probability distribution of the slot value to be filled based on the decoded vector, and then based on the predicted probability of the slot value to be filled Distributing to fill slots under the at least one domain so as to realize dialogue state tracking; wherein, the preset domain-slot word vector consists of the preset domain vector and the preset slot word Vector concatenation. In summary, since the BiGRU (bidirectional gated recurrent unit) neural network can extract context information using the forward and backward gated recurrent unit (GRU) structures, this application does not ignore the forward and backward context information based on the BiGRU neural network. Under the premise of solving the long-term dependency problem of historical dialogues in multi-domain scenarios; in addition, the traditional dialogue state tracking is to perform overall feature extraction on domain-slot pairs. In this way, the information between different slots in the same domain It is irrelevant, and the same slot information in different fields is also irrelevant, but this application separates the fields and slots and performs attention calculations on the fields and slots in turn, so that the difference between the extracted different field-slot pairs information can be shared.

Further, the embodiment of the present application also provides an electronic device. Fig. 4 is a structural diagram of an electronic device 20 according to an exemplary embodiment, and the content in the diagram should not be regarded as any limitation on the application scope of the present application.

FIG. 4 is a schematic structural diagram of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21 , at least one memory 22 , a display screen 23 , an input/output interface 24 , a communication interface 25 , a power supply 26 and a communication bus 27 . Wherein, the memory 22 is used to store a computer program, and the computer program is loaded and executed by the processor 21 to implement relevant steps in the dialog state tracking method disclosed in any of the foregoing embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.

In this embodiment, the power supply 26 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 25 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows is applicable Any communication protocol in the technical solution of the present application is not specifically limited here; the input and output interface 24 is used to obtain external input data or output data to the external world, and its specific interface type can be selected according to specific application needs, here Not specifically limited.

In addition, as a resource storage carrier, the memory 22 may be a read-only memory, random access memory, magnetic disk or optical disk, etc., and the resources stored thereon may include computer programs 221, and the storage method may be temporary storage or permanent storage. Wherein, the computer program 221 may further include a computer program capable of completing other specific tasks in addition to the computer program capable of completing the dialog state tracking method performed by the electronic device 20 disclosed in any of the foregoing embodiments.

Furthermore, the embodiment of the present application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the aforementioned dialog state tracking method disclosed above is implemented.

Regarding the specific steps of the method, reference may be made to the corresponding content disclosed in the foregoing embodiments, and details are not repeated here.

Each embodiment in this application is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be disclosed for the embodiments by referring to each other. As far as the device is concerned, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and for relevant parts, please refer to the description of the method part.

Professionals can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software or a combination of the two. In order to clearly illustrate the possible Interchangeability, in the above description, the components and steps of each example have been generally described according to their functions. Whether these functions are executed by hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be regarded as exceeding the scope of the present application.

The steps of the methods or algorithms described in connection with the embodiments disclosed herein may be directly implemented by hardware, software modules executed by a processor, or a combination of both. Software modules can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other Any other known storage medium.

Finally, it should also be noted that in this text, relational terms such as first and second etc. are only used to distinguish one entity or operation from another, and do not necessarily require or imply that these entities or operations, any such actual relationship or order exists. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

A dialogue state tracking method, device, equipment, and storage medium provided by this application have been introduced in detail above. In this paper, specific examples are used to illustrate the principle and implementation of this application. The description of the above embodiments is only for To help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation and application scope. In summary, the content of this specification It should not be construed as a limitation of the application.

Claims (10)

1. A dialog state tracking method, comprising:

encoding the historical dialogue records by using a first BiGRU neural network to obtain a first encoding result, and performing domain-level feature extraction on the first encoding result based on an attention mechanism and a preset domain vector to obtain a first feature extraction result containing at least one domain;

coding the first feature extraction result by using a second BiGRU neural network to obtain a second coding result, and performing slot position level feature extraction on the second coding result based on an attention mechanism and a preset slot position word vector to obtain a second feature extraction result containing slot positions;

inputting the second feature extraction result, a preset field-slot position word vector and a preset word table dictionary vector into a decoder to obtain a decoded vector, calculating the prediction probability distribution of slot position values to be filled based on the decoded vector, and filling the slot positions in at least one field based on the prediction probability distribution of the slot position values to be filled so as to realize dialogue state tracking; the preset field-slot word vector is formed by splicing the preset field vector and the preset slot word vector.

2. The dialog state tracking method of claim 1, wherein before encoding the historical dialog record using the first BiGRU neural network to obtain the first encoding result, the method further comprises:

and acquiring a plurality of turn conversation records, and splicing the turn conversation records to obtain the historical conversation record.

3. The dialog state tracking method according to claim 1, wherein the step of inputting the second feature extraction result, the predetermined field-slot word vector and the predetermined vocabulary dictionary vector to a decoder to obtain a decoded backward vector further comprises:

mapping the decoded vector to a target probability distribution based on a dialog state tracking policy;

when the target probability distribution is a first probability distribution, the slot position in the at least one domain is not referred to by the user, when the target probability distribution is a second probability distribution, the slot position in the at least one domain is not referred to by the user, and when the target probability distribution is a third probability distribution, the slot position in the at least one domain is referred to by the user.

4. The dialog state tracking method of claim 3 wherein said computing a predictive probability distribution of slot values to be filled based on the decoded vector comprises:

when the target probability distribution is the third probability distribution, calculating the probability distribution of slot position values in the preset word list dictionary and the probability distribution of slot position values in the historical dialogue record based on the decoded backward quantity;

and calculating the prediction probability distribution of the slot values to be filled according to the probability distribution of the slot values in the preset vocabulary dictionary and the probability distribution of the slot values in the historical dialogue records.

5. The dialog state tracking method of claim 4,

the formula for calculating the probability distribution of the slot values in the preset vocabulary dictionary and the probability distribution of the slot values in the historical dialogue records based on the decoded backward quantity is as follows:

wherein V represents the predetermined vocabulary dictionary vector, H _decode Representing said decoded vector, h representing said first coding result, p ^vocab Representing a probability distribution, p, of slot values in said predetermined vocabulary dictionary ^history Representing a probability distribution of slot-level values in the historical conversation record.

6. The dialog state tracking method of claim 5,

calculating a formula of the prediction probability distribution of the slot values to be filled according to the probability distribution of the slot values in the preset vocabulary dictionary and the probability distribution of the slot values in the historical dialogue records, wherein the formula comprises the following steps:

p ^value ＝p ^gen ×p ^vocab +(1-p ^gen )×p ^history ；

wherein p is ^gen And representing the weight of the slot value to be filled generated from the preset vocabulary dictionary.

7. The method according to any one of claims 1 to 6, wherein the step of inputting the second feature extraction result, the predetermined field-slot word vector and the predetermined vocabulary dictionary vector into a decoder to obtain a decoded vector comprises:

and inputting the second feature extraction result, the preset field-slot position word vector and the preset word table dictionary vector into a decoder constructed by a third BiGRU neural network to obtain a decoded vector.

8. A dialog state tracking device, comprising:

the first coding module is used for coding the historical dialogue record by utilizing a first BiGRU neural network to obtain a first coding result;

the domain level feature extraction module is used for extracting the features of the domain level of the first coding result based on an attention mechanism and a preset domain vector to obtain a first feature extraction result containing at least one domain;

the second coding module is used for coding the first feature extraction result by utilizing a second BiGRU neural network to obtain a second coding result;

the slot position level feature extraction module is used for extracting the features of the slot position level of the second coding result based on an attention mechanism and a preset slot position word vector to obtain a second feature extraction result containing the slot position;

the decoding module is used for inputting the second feature extraction result, a preset field-slot position word vector and a preset word list dictionary vector into a decoder to obtain a decoded vector;

the conversation state tracking module is used for calculating the prediction probability distribution of the slot position value to be filled based on the decoded vector, and then filling the slot position in at least one field based on the prediction probability distribution of the slot position value to be filled so as to realize conversation state tracking; the preset field-slot position word vector is formed by splicing the preset field vector and the preset slot position word vector.

9. An electronic device, comprising:

a memory for storing a computer program;

a processor for executing the computer program to implement the dialog state tracking method of any of claims 1 to 7.

10. A computer-readable storage medium for storing a computer program; wherein the computer program when executed by a processor implements a dialog state tracking method according to any of claims 1 to 7.

Images