Intelligent toy control method and system based on multi-round dialogues

Abstract

The invention discloses a toy intelligent control method and system based on multi-round dialogues, and relates to the field of intelligent control. A toy intelligent control system based on multi-round dialogues comprises a dialogue management module, an information estimation module, a dialogue analysis module, an intelligent control module and a feedback adjustment module. According to the method, the intention label and the dialogue action label of the user are obtained through the dialogue analysis model, the dialogue analysis model combines the time decay function with the multi-head self-attention network, the global context is considered, meanwhile, the recent context is focused on, the model can better utilize the historical context, and the user experience is improved. And context information with different attention degrees is screened out for different characters by using character-level attention, so that the recognition capability of the model is improved, and effective semantic recognition on multiple rounds of dialogues is realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control, and in particular to a toy intelligent control method and system based on multi-round dialogue. Background Art

[0002] With the rapid development of artificial intelligence and natural language processing technologies, smart toys have gradually become an important tool for children's entertainment and education. These toys can not only interact with children in a simple way, but also provide personalized experiences through learning and adaptation.

[0003] However, most smart toys on the market still have some significant limitations, mainly in the following aspects: Single-round dialogue limitation. Currently, many smart toys can only handle single-round dialogue, that is, they can only respond to a simple command at a time. This approach lacks contextual understanding, making it impossible for toys to conduct continuous, multi-round interactions. For example, a user may ask a series of related questions to a toy, but the toy cannot remember the content of the previous question, resulting in poor coherence and naturalness of the conversation. Insufficient semantic understanding. Existing smart toys still have a lot of room for improvement in semantic understanding. Many toys can only recognize simple commands and keywords, but cannot understand complex sentence structures and multi-level semantics. This limits the application scenarios and intelligence level of toys in actual interactions. Lack of personalization and adaptability. Most current smart toys lack the ability to adjust interaction strategies based on user historical dialogues and behaviors, and cannot provide personalized services based on user interests and needs. Children have different interests and learning needs at different stages of growth. If toys cannot adapt to these changes, it will be difficult to continue to attract and educate users. Insufficient feedback mechanism. Many smart toys do not have an effective feedback mechanism to evaluate and improve their performance after interacting with users. The lack of feedback learning makes it difficult for the toy to optimize its dialogue strategies and interaction methods, making it difficult to continuously improve the user experience. Summary of the invention

[0004] The toy intelligent control method based on multi-round dialogue includes:

[0005] S1. Get the dialogue information input by voice or text when the user uses the toy. If the dialogue information is input in the form of voice, proceed to step S2 for voice information estimation;

[0006] S2. extracting acoustic features of the conversation information input in the form of voice, and sending the acoustic features to the voice information prediction model for calculation to obtain the user's estimated information;

[0007] S3. Send the dialogue information to the dialogue parsing model for intent classification and dialogue action recognition to obtain intent labels and dialogue action labels;

[0008] S4. Match the action instructions or dialogue response content through the intention label, dialogue action label and the user's estimated information, and generate corresponding control instructions;

[0009] S5. The toy performs corresponding actions or voice responses according to the generated control instructions, obtains user feedback information, and adjusts the control instructions according to the feedback information.

[0010] As a preferred technical solution of the present invention, the dialogue parsing model includes an encoder layer, a time-decayed multi-head self-attention network layer and a decoder layer;

[0011] The encoder layer includes a history conversation encoder and a user current conversation encoder. The history conversation encoder and the user current conversation encoder are two different bidirectional long short-term memory networks. The history conversation encoder is used to encode the historical turn sentences in the conversation information to obtain the history conversation vector. The user current conversation encoder is used to encode the sentence currently sent by the user in the conversation information to obtain the user current conversation vector.

[0012] The time-decayed multi-head self-attention network layer contains a multi-layer attention network, which is composed of 3 layers of time-decayed multi-head self-attention networks stacked together. It is used to receive the output vectors from the historical dialogue encoder and the user's current dialogue encoder, concatenate the historical dialogue vectors and the user's current dialogue vectors and input them into the time-decayed multi-head self-attention network layer to perform character-level attention interaction calculations and sentence-level attention interaction calculations.

[0013] The decoder layer includes an intent classifier, a dialogue action recognition classifier and a slot filling decoder. The intent classifier is a fully connected neural network with a softmax output head, which is used to receive the user's current dialogue vector and the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer to perform intent classification tasks; the dialogue action recognition classifier is a fully connected neural network with a sigmoid output head, which is used to receive the user's current dialogue vector and the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer to perform dialogue action recognition tasks; the slot filling decoder is a unidirectional long short-term memory network, which is used to receive the original encoding of each character of the user's current dialogue vector and the original encoding of each character of the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer, the output of the intent classifier and the dialogue action recognition classifier, and perform slot filling tasks.

[0014] As a preferred technical solution of the present invention, the historical conversation encoder is used to encode the historical turn sentences in the conversation information to obtain the historical conversation vector, and the user current conversation encoder is used to encode the sentence currently sent by the user in the conversation information to obtain the user current conversation vector. The specific steps are:

[0015] A1. Get the initial dialogue information data, record the whole dialogue of the dialogue information data as D, D = {R1, R2, ..., Rt}, t represents the dialogue round, record the historical dialogue as H, H = {R1, R2, ..., Rt} t-1}, R t is the current user dialogue sentence, R t ={w t1 ,w t2 ,…,w tn}, n represents R t The number of characters in ;

[0016] A2. Use the Glove distributed word vector to initialize each conversation sentence, and use the Embedding layer of the bidirectional long short-term memory network to vectorize the words of the historical conversation to obtain each sentence R of the historical conversation H i = {v i1 , v i2 , …, v in} (i represents the historical dialogue of the i-th round, n represents the number of sentence characters), where v in Represents the vector representation of the nth character of the i-th sentence, using the bidirectional long short-term memory network model A to R i Encode and obtain the high-dimensional encoding representation of each historical dialogue sentence:

[0017]

[0018] Where mi represents the sentence vector of the concatenation of the output of the hidden unit of the forward long short-term memory network and the reverse long short-term memory network at the last moment, represents the embedding function, mi∈R d , d represents the dimension of the mi vector, and all historical dialogue vectors M = {m1,m2,…,m t-1}∈R (t-1)*d ;

[0019] A3. Get the user's current conversation R after the word vectorization representation t = {v t1 ,v t2 ,…,v tn}, using the bidirectional long short-term memory network model B to R t Encode and obtain the encoding representation of the sentence currently sent by the user and each character in the high-dimensional space:

[0020]

[0021] where h lastRepresents the sentence vector of the concatenation of the output of the hidden unit of the forward LSTM network and the reverse LSTM network at the last moment, ho = {h1,h2,…,h n}, represents the concatenated character vector of the output of each hidden unit of the forward LSTM network and the reverse LSTM network at each time step, where h i Represents the output of the i-th time step, hi∈R d , get the user's current conversation vector h last ∈R d The character vector of the current conversation with the user ho={h1,h2,…,h n}∈R n*d .

[0022] As a preferred technical solution of the present invention, character-level attention interaction calculation and sentence-level attention interaction calculation are performed, and the specific steps are:

[0023] B1. The user's current conversation vector h last and historical dialogue vector M = {m1,m2,…,m t-1} to obtain the full conversation vector M R ={h last ,m1,m2,…,m t-1}∈R t*d , M R Use the position encoding function to calculate the position encoding PE (M R ), add the position encoding to the full conversation vector to obtain the reconstructed full conversation vector M R* =M R +PE(M R );

[0024] B2. Reconstruct the full conversation vector M R* Use the time-decayed multi-head self-attention network for calculation to obtain the final sentence-level context weighted sum;

[0025] B3. Concatenate the character vector ho of the user's current conversation with the historical conversation vector M to obtain the full conversation character vector M o ={m o1 , m o2 ,…,m on}=[{h1,m1,m2,…,m t-1}, {h2,m1,m2,…,m t-1}, …, {h n ,m1,m2,…,m t-1}, M o Use the position encoding function to calculate the position encoding PE (M o ), add the position code to the full dialogue character vector to obtain the reconstructed full dialogue character vector Mo* =M o +PE(M o );

[0026] B4. Reconstruct the full dialogue character vector M o* The time-decayed multi-head self-attention network is used for calculation to obtain the final character-level context weighted sum.

[0027] As a preferred technical solution of the present invention, the time-decayed multi-head self-attention score is calculated by an adaptive time attention function composed of three time decay functions: a linear time decay function, a convex time decay function, and a concave time decay function;

[0028] As a preferred technical solution of the present invention, speech information estimation specifically includes extracting acoustic features of dialogue information, including pitch, timbre, power spectrum of sound and linear prediction parameters; sending the acoustic features into a speech information estimation model for calculation, and outputting the user's age estimation label, gender estimation label and emotion estimation label;

[0029] As a preferred technical solution of the present invention, the action instructions or dialogue response content are matched through the intention label, the dialogue action label and the user's estimated information, and the corresponding control instructions are generated. The specific steps are as follows:

[0030] C1. Match the standard control instructions consistent with the intention requirements in the control instruction library through the intention tag, obtain the first matching result and send it to C2 to match the behavior requirements;

[0031] C2. Match the standard control instruction with the behavior requirement consistent with the dialog action label in the first matching result to obtain the second matching result. If the estimated information of the user is not received, the standard control instruction with the largest number in the second matching result is output as the control instruction. Otherwise, the second matching result is sent to C3 to match the personalized requirement;

[0032] C3. In the second matching result, a standard control instruction with a personalized requirement consistent with the estimated information of the user is matched, and the standard control instruction with the largest number in the third matching result is output as the control instruction.

[0033] As a preferred technical solution of the present invention, the user's feedback information is obtained, and the control instructions are adjusted according to the feedback information. The specific steps are:

[0034] D1. After the toy starts outputting control instructions, it monitors the user's conversation and feeds the user's conversation into the conversation parsing model to obtain the user's feedback intention label and conversation action label;

[0035] D2. Monitor the intention tags and dialogue action tags of the feedback for abnormalities, and search the intention tags and dialogue action tags of the feedback with the intention tags and dialogue action tags of the control instructions generated by the toy. If there is a tag comparison that is consistent with that in the feedback processing library, the standard control instructions in the feedback processing library are output to adjust the toy;

[0036] D3. Collect the number of abnormal situations in abnormal monitoring, and add weights to the label comparison in the feedback processing library according to the number of abnormal situations.

[0037] The toy intelligent control system based on multi-round dialogue includes:

[0038] A dialogue management module, used to obtain dialogue information input by the user through voice or text when using the toy, and store and manage the dialogue information;

[0039] The information prediction module is used to extract the acoustic features of the conversation information input in the form of voice, send the acoustic features to the voice information prediction model for calculation, and obtain the user's estimated information;

[0040] The dialogue parsing module is used to feed the dialogue information into the dialogue parsing model for intent classification and dialogue action recognition to obtain intent labels and dialogue action labels;

[0041] Intelligent control module, which is used to match action instructions or dialogue response content through intent labels, dialogue action labels and user's estimated information, and generate corresponding control instructions;

[0042] The feedback adjustment module is used to obtain user feedback information and adjust the control instructions according to the feedback information.

[0043] The present invention has the following advantages:

[0044] 1. The present invention obtains the user's intention label and dialogue action label through a dialogue parsing model. The dialogue parsing model combines the time decay function with a multi-head self-attention network, which not only considers the global context but also focuses on the recent context, so that the model can better utilize the historical context, and uses character-level attention to filter out context information of different attention levels for different characters, thereby improving the recognition ability of the model and realizing effective semantic recognition of multi-round dialogues.

[0045] 2. The present invention estimates the user's voice information, obtains the user's estimated information, combines the user's estimated information with the intention tag and the dialogue action tag to match the toy control command, thereby improving the personalization of the toy control;

[0046] 3. The present invention monitors the intention labels and dialogue action labels of user feedback for abnormalities, adjusts the control instructions of the toy in real time according to the monitoring results, and adds weights according to the number of abnormal situations, thereby achieving effective feedback and continuous optimization of toy control. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a schematic diagram of the structure of a toy intelligent control system based on multi-round dialogues adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0049] Embodiment 1, a toy intelligent control method based on multi-round dialogue, includes:

[0050] S1. Get the dialogue information input by voice or text when the user uses the toy. If the dialogue information is input in the form of voice, proceed to step S2 for voice information estimation;

[0051] S2. extracting acoustic features of the conversation information input in the form of voice, and sending the acoustic features to the voice information prediction model for calculation to obtain the user's estimated information;

[0052] Extract the acoustic features of the conversation information, including pitch, timbre, power spectrum of the sound, and linear prediction parameters; send the acoustic features to the voice information prediction model for calculation, and output the user's age estimation label, gender estimation label, and emotion estimation label;

[0053] The voice information prediction model is built based on a neural network model and is trained using collected voice data. The output of the model is the user's prediction information, which can be used for personalized intelligent control of toys.

[0054] S3. Send the dialogue information to the dialogue parsing model for intent classification and dialogue action recognition to obtain intent labels and dialogue action labels;

[0055] S4. Match the action instructions or dialogue response content through the intention label, dialogue action label and the user's estimated information, and generate corresponding control instructions;

[0056] In step S4, the action instruction or the dialogue response content is matched through the intention label, the dialogue action label and the user's estimated information, and the corresponding control instruction is generated. The specific steps are as follows:

[0057] C1. Match the standard control instructions consistent with the intention requirements in the control instruction library through the intention tag, obtain the first matching result and send it to C2 to match the behavior requirements;

[0058] C2. Match the standard control instruction with the behavior requirement consistent with the dialog action label in the first matching result to obtain the second matching result. If the estimated information of the user is not received, the standard control instruction with the largest number in the second matching result is output as the control instruction. Otherwise, the second matching result is sent to C3 to match the personalized requirement;

[0059] C3. In the second matching result, a standard control instruction with a personalized requirement consistent with the estimated information of the user is matched, and the standard control instruction with the largest number in the third matching result is output as the control instruction.

[0060] S5. The toy performs corresponding actions or voice responses according to the generated control instructions, obtains user feedback information, and adjusts the control instructions according to the feedback information.

[0061] In step S5, feedback information from the user is obtained, and the control instructions are adjusted according to the feedback information. The specific steps are as follows:

[0062] D1. After the toy starts outputting control instructions, it monitors the user's conversation and feeds the user's conversation into the conversation parsing model to obtain the user's feedback intention label and conversation action label;

[0063] D2. Monitor the intention tags and dialogue action tags of the feedback for abnormalities, and search the intention tags and dialogue action tags of the feedback with the intention tags and dialogue action tags of the control instructions generated by the toy. If there is a tag comparison that is consistent with that in the feedback processing library, the standard control instructions in the feedback processing library are output to adjust the toy;

[0064] D3. Collect the number of abnormal situations in abnormal monitoring, and add weights to the label comparison in the feedback processing library according to the number of abnormal situations.

[0065] The dialogue parsing model consists of an encoder layer, a time-decayed multi-head self-attention network layer, and a decoder layer;

[0066] The encoder layer includes a history conversation encoder and a user current conversation encoder. The history conversation encoder and the user current conversation encoder are two different bidirectional long short-term memory networks. The history conversation encoder is used to encode the historical turn sentences in the conversation information to obtain the history conversation vector. The user current conversation encoder is used to encode the sentence currently sent by the user in the conversation information to obtain the user current conversation vector.

[0067] The time-decayed multi-head self-attention network layer contains a multi-layer attention network, which is composed of 3 layers of time-decayed multi-head self-attention networks stacked together. It is used to receive the output vectors from the historical dialogue encoder and the user's current dialogue encoder, concatenate the historical dialogue vectors and the user's current dialogue vectors and input them into the time-decayed multi-head self-attention network layer to perform character-level attention interaction calculations and sentence-level attention interaction calculations.

[0068] The decoder layer includes an intent classifier, a dialogue action recognition classifier and a slot filling decoder. The intent classifier is a fully connected neural network with a softmax output head, which is used to receive the user's current dialogue vector and the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer to perform intent classification tasks; the dialogue action recognition classifier is a fully connected neural network with a sigmoid output head, which is used to receive the user's current dialogue vector and the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer to perform dialogue action recognition tasks; the slot filling decoder is a unidirectional long short-term memory network, which is used to receive the original encoding of each character of the user's current dialogue vector and the original encoding of each character of the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer, the output of the intent classifier and the dialogue action recognition classifier, and perform slot filling tasks.

[0069] Get the historical conversation vector and the user's current conversation vector. The specific steps are:

[0070] A1. Obtain the initial dialogue information data, and record the entire dialogue round of dialogue information data as D, where D = {R1, R2, ..., R t}, t represents the turn of the conversation, and the historical conversation is recorded as H, H = {R1, R2, ..., R t-1}, R t is the current user dialogue sentence, R t ={w t1 ,w t2 ,…,w tn}, n represents R t The number of characters in ;

[0071] A2. Use the Glove distributed word vector to initialize each conversation sentence, and use the Embedding layer of the bidirectional long short-term memory network to vectorize the words of the historical conversation to obtain each sentence R of the historical conversation H i = {v i1 , v i2 , …, v in} (i represents the historical dialogue of the i-th round, n represents the number of sentence characters), where v in Represents the vector representation of the nth character of the i-th sentence, using the bidirectional long short-term memory network model A to R iEncode and obtain the high-dimensional encoding representation of each historical dialogue sentence:

[0072]

[0073] Where mi represents the sentence vector of the concatenation of the output of the hidden unit of the forward long short-term memory network and the reverse long short-term memory network at the last moment, represents the embedding function, mi∈R d , d represents the dimension of the mi vector, and all historical dialogue vectors M = {m1,m2,…,m t-1}∈R (t-1)*d ;

[0074] A3. Get the user's current conversation R after the word vectorization representation t = {v t1 ,v t2 ,…,v tn}, using the bidirectional long short-term memory network model B to R t Encode and obtain the encoding representation of the sentence currently sent by the user and each character in the high-dimensional space:

[0075]

[0076] where h last Represents the sentence vector of the concatenation of the output of the hidden unit of the forward LSTM network and the reverse LSTM network at the last moment, ho = {h1,h2,…,h n}, represents the concatenated character vector of the output of each hidden unit of the forward LSTM network and the reverse LSTM network at each time step, where h i Represents the output of the i-th time step, hi∈R d , get the user's current conversation vector h last ∈R d The character vector of the current conversation with the user ho={h1,h2,…,h n}∈R n*d .

[0077] The character-level attention interaction calculation and sentence-level attention interaction calculation are as follows:

[0078] B1. The user's current conversation vector h last and historical dialogue vector M = {m1,m2,…,m t-1} to obtain the full conversation vector M R ={h last ,m1,m2,…,m t-1}∈R t*d , M RUse the position encoding function to calculate the position encoding PE (M R ), add the position encoding to the full conversation vector to obtain the reconstructed full conversation vector M R* =M R +PE(M R );

[0079] B2. Reconstruct the full conversation vector M R* Use the time-decayed multi-head self-attention network for calculation to obtain the final sentence-level context weighted sum;

[0080] B3. Concatenate the character vector ho of the user's current conversation with the historical conversation vector M to obtain the full conversation character vector M o ={m o1 , m o2 ,…,m on}=[{h1,m1,m2,…,m t-1}, {h2,m1,m2,…,m t-1}, …, {h n ,m1,m2,…,m t-1}, M o Use the position encoding function to calculate the position encoding PE (M o ), add the position code to the full dialogue character vector to obtain the reconstructed full dialogue character vector M o* =M o +PE(M o );

[0081] B4. Reconstruct the full dialogue character vector M o* The time-decayed multi-head self-attention network is used for calculation to obtain the final character-level context weighted sum.

[0082] For the intent classification and dialogue action recognition tasks, both of which belong to text classification tasks, we only need to classify the entire sentence into one or several categories, that is, calculate the context weight score for the entire sentence; for the slot filling task, we need to classify each character of the sentence. The context on which each character depends may be different, so we need to calculate the context weight score for each character separately;

[0083] Attention score calculation: The time-decayed multi-head self-attention score is calculated by an adaptive time attention function consisting of three time decay functions: linear time decay function, convex time decay function, and concave time decay function;

[0084] Input MR* into the fully connected layers K, Q and V respectively to obtain MR*(Q)=MR*×WQ, MR*(K)=MR*×WK and MR*(V)=MR*×WV (WQ, WK and Wv∈Rd*Ha, Ha represents the size of the hidden layer dimension). Divide MR*(Q), MR*(K) and MR*(V) into h heads for actual calculation. Let H j (Q), H j (K) and H j (V) represents the jth head after multi-head division, and the dimension size of each head is H i ∈R t×(Ha / h );

[0085] Calculate the time difference d(R i ), d(R i ) represents the historical dialogue R of round i i To the user's current round of dialogue R t Distance, historical dialogue i The current conversation round R of the user t The farther away, the time difference d(R i ) is larger, for each head j of the multi-head attention, let D j Represents the distance definition set of the entire historical round of dialogue, that is, D j ={d j (R1), d j (R2),…,d j (R t-1 ))};

[0086] Get the linear time decay function V j1 , convex time decay function V j2 and the concave time decay function V j3 , where is the linear time decay function;

[0087]

[0088] k j and b j is a trainable parameter, h represents the number of heads in the multi-head attention network;

[0089] The adaptive time decay function is obtained by weighted summing of these three time decay functions. Will The time weight is normalized in the input softmax function to obtain a universal adaptive time attention score The original attention score is calculated as:

[0090]

[0091] Among them, z represents the hidden layer dimension size of each head, and the final time-decayed multi-head self-attention of the j-th head is, is the balance parameter between temporal attention and raw attention;

[0092] The obtained attention vector is passed through a layer of residual connection and layer normalization, a layer of feedforward neural network and a final layer of residual connection and layer normalization to obtain the final sentence-level context weighted sum, which is subsequently used for intent classification and dialogue action recognition tasks;

[0093] The context weighted sum of the sentence-level context and h last Performing residual splicing can effectively prevent the gradient vanishing problem during model training;

[0094] Since there is a strong correlation between the three tasks of intent classification, dialogue action recognition, and slot filling, a joint modeling training model is adopted, using categorical cross entropy loss in intent classification and slot filling tasks, and binary cross entropy loss in dialogue action recognition task;

[0095] Example 2, a toy intelligent control system based on multi-round dialogue, see Figure 1 As shown, including:

[0096] A dialogue management module, used to obtain dialogue information input by the user through voice or text when using the toy, and store and manage the dialogue information;

[0097] The information prediction module is used to extract the acoustic features of the conversation information input in the form of voice, send the acoustic features to the voice information prediction model for calculation, and obtain the user's estimated information;

[0098] The dialogue parsing module is used to feed the dialogue information into the dialogue parsing model for intent classification and dialogue action recognition to obtain intent labels and dialogue action labels;

[0099] Intelligent control module, which is used to match action instructions or dialogue response content through intent labels, dialogue action labels and user's estimated information, and generate corresponding control instructions;

[0100] The feedback adjustment module is used to obtain user feedback information and adjust the control instructions according to the feedback information.

[0101] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A toy intelligent control method based on multi-round dialogue, characterized in that: include: S1. Get the dialogue information input by voice or text when the user uses the toy. If the dialogue information is input in the form of voice, proceed to step S2 for voice information estimation; S2. extracting acoustic features of the conversation information input in the form of voice, and sending the acoustic features to the voice information prediction model for calculation to obtain the user's estimated information; S3. Send the dialogue information to the dialogue parsing model for intent classification and dialogue action recognition to obtain intent labels and dialogue action labels; S4. Match action instructions or dialogue response content through intent labels, dialogue action labels and user's estimated information, and generate corresponding control instructions; S5. The toy performs corresponding actions or voice responses according to the generated control instructions, obtains user feedback information, and adjusts the control instructions according to the feedback information.

2. The toy intelligent control method based on multi-round dialogue according to claim 1, characterized in that: The dialogue parsing model consists of an encoder layer, a time-decayed multi-head self-attention network layer, and a decoder layer; The encoder layer includes a history conversation encoder and a user current conversation encoder. The history conversation encoder and the user current conversation encoder are two different bidirectional long short-term memory networks. The history conversation encoder is used to encode the historical turn sentences in the conversation information to obtain the history conversation vector. The user current conversation encoder is used to encode the sentence currently sent by the user in the conversation information to obtain the user current conversation vector. The time-decayed multi-head self-attention network layer contains a multi-layer attention network, which is composed of 3 layers of time-decayed multi-head self-attention networks stacked together. It is used to receive the output vectors from the historical dialogue encoder and the user's current dialogue encoder, concatenate the historical dialogue vectors and the user's current dialogue vectors and input them into the time-decayed multi-head self-attention network layer to perform character-level attention interaction calculations and sentence-level attention interaction calculations. The decoder layer includes an intent classifier, a conversation action recognition classifier, and a slot filling decoder. The intent classifier is a fully connected neural network with a softmax output head. It is used to receive the user's current conversation vector and the user's current conversation vector interactively calculated by the time-decayed multi-head attention network layer to perform intent classification tasks. The dialogue action recognition classifier is a fully connected neural network with a sigmoid output head, which is used to receive the user's current dialogue vector and the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer, and perform the dialogue action recognition task; the slot filling decoder is a unidirectional long short-term memory network, which is used to receive the original encoding of each character of the user's current dialogue vector and the original encoding of each character of the user's current dialogue vector interactively calculated by the time-decayed multi-head attention network layer, the output of the intent classifier and the dialogue action recognition classifier, and perform the slot filling task.

3. The toy intelligent control method based on multi-round dialogue according to claim 2, characterized in that: The historical conversation encoder is used to encode the historical turn sentences in the conversation information to obtain the historical conversation vector. The user current conversation encoder is used to encode the sentence currently sent by the user in the conversation information to obtain the user current conversation vector. The specific steps are as follows: A1. Obtain the initial dialogue information data, and record the entire dialogue round of the dialogue information data as D, D = {R1, R2, ···, R t }, t represents the round of dialogue, and the historical dialogue is recorded as H, H = {R1, R2, ···, R t-1 }, R t is the current user dialogue sentence, R t ={w t1 ,w t2 ,···,w tn }, n represents R t The number of characters in ; A2. Use the Glove distributed word vector to initialize each conversation sentence, and use the Embedding layer of the bidirectional long short-term memory network to vectorize the words of the historical conversation to obtain each sentence R of the historical conversation H i = {v i1 ,v i2 ,···,v in } (i represents the historical dialogue of the i-th round, n represents the number of sentence characters), where v in Represents the vector representation of the nth character of the i-th sentence, using the bidirectional long short-term memory network model A to R i Encode and obtain the high-dimensional encoding representation of each historical dialogue sentence: Where mi represents the sentence vector of the concatenation of the output of the hidden unit of the forward long short-term memory network and the reverse long short-term memory network at the last moment, represents the embedding function, mi∈R d , d represents the dimension of the mi vector, and all historical dialogue vectors are obtained A3. Get the user's current conversation R after the word vectorization representation t = {v t1 ,v t2 ,···,v tn }, using the bidirectional long short-term memory network model B to R t Encode and obtain the encoding representation of the sentence currently sent by the user and each character in the high-dimensional space: where h last Represents the sentence vector of the concatenation of the output of the hidden unit of the forward LSTM network and the reverse LSTM network at the last moment, ho = {h1,h2,···,h n }, represents the concatenated character vector of the output of each hidden unit of the forward LSTM network and the reverse LSTM network at each time step, where h i Represents the output of the i-th time step, hi∈R d , get the user's current conversation vector h last ∈R d Character vector of the current conversation with the user 4. The toy intelligent control method based on multi-round dialogue according to claim 2, characterized in that: Perform character-level attention interaction calculation and sentence-level attention interaction calculation. The specific steps are: B1. The user's current conversation vector h last and historical conversation vector M = {m1,m2,···,m t-1 } to obtain the full conversation vector M R Use the position encoding function to calculate the position encoding PE (M R ), add the position encoding to the full conversation vector to obtain the reconstructed full conversation vector M R* =M R +PE(M R ); B2. Reconstruct the full conversation vector M R* Use the time-decayed multi-head self-attention network for calculation to obtain the final sentence-level context weighted sum; B3. Concatenate the character vector ho of the user's current conversation with the historical conversation vector M to obtain the full conversation character vector M o ={m o1 ,m o2 ,···,m on }=[{h1,m1,m2,···,m t-1 },{h2,m1,m2,···,m t-1 },···,{h n ,m1,m2,···,m t-1 }, M o Use the position encoding function to calculate the position encoding PE (M o ), add the position code to the full dialogue character vector to obtain the reconstructed full dialogue character vector M o* =M o +PE(M o ); B4. Reconstruct the full dialogue character vector M o* The time-decayed multi-head self-attention network is used for calculation to obtain the final character-level context weighted sum.

5. The toy intelligent control method based on multi-round dialogue according to claim 4, characterized in that: The time-decayed multi-head self-attention score is calculated by an adaptive time-attention function consisting of three time decay functions: linear time decay function, convex time decay function, and concave time decay function.

6. The toy intelligent control method based on multi-round dialogue according to claim 1, characterized in that: Speech information prediction, specifically, extracts the acoustic features of conversation information, including pitch, timbre, power spectrum of sound, and linear prediction parameters; sends the acoustic features to the speech information prediction model for calculation, and outputs the user's age estimation label, gender estimation label, and emotion estimation label.

7. The toy intelligent control method based on multi-round dialogue according to claim 1, characterized in that: Match action instructions or dialogue response content through intent labels, dialogue action labels, and user's estimated information, and generate corresponding control instructions. The specific steps are as follows: C1. Match the standard control instructions consistent with the intention requirements in the control instruction library through the intention tag, obtain the first matching result and send it to C2 to match the behavior requirements; C2. Match the standard control instruction with the behavior requirement consistent with the dialog action label in the first matching result to obtain the second matching result. If the estimated information of the user is not received, the standard control instruction with the largest number in the second matching result is output as the control instruction. Otherwise, the second matching result is sent to C3 to match the personalized requirement; C3. In the second matching result, a standard control instruction with a personalized requirement consistent with the estimated information of the user is matched, and the standard control instruction with the largest number in the third matching result is output as the control instruction.

8. The toy intelligent control method based on multi-round dialogue according to claim 1, characterized in that: Obtain user feedback information and adjust control instructions based on the feedback information. The specific steps are as follows: D1. After the toy starts outputting control instructions, it monitors the user's conversation and feeds the user's conversation into the conversation parsing model to obtain the user's feedback intention label and conversation action label; D2. Monitor the intention tags and dialogue action tags of the feedback for abnormalities, and search the intention tags and dialogue action tags of the feedback with the intention tags and dialogue action tags of the control instructions generated by the toy. If there is a tag comparison that is consistent with that in the feedback processing library, the standard control instructions in the feedback processing library are output to adjust the toy; D3. Collect the number of abnormal situations in abnormal monitoring, and add weights to the label comparison in the feedback processing library according to the number of abnormal situations.

9. Based on multi-round dialogue, the toy intelligent control system is characterized by The system applies the toy intelligent control method based on multi-round dialogue according to any one of claims 1 to 8, including: A dialogue management module, used to obtain dialogue information input by the user through voice or text when using the toy, and store and manage the dialogue information; The information prediction module is used to extract the acoustic features of the conversation information input in the form of voice, send the acoustic features to the voice information prediction model for calculation, and obtain the user's estimated information; The dialogue parsing module is used to feed the dialogue information into the dialogue parsing model for intent classification and dialogue action recognition to obtain intent labels and dialogue action labels; Intelligent control module, which is used to match action instructions or dialogue response content through intent labels, dialogue action labels and user's estimated information, and generate corresponding control instructions; The feedback adjustment module is used to obtain user feedback information and adjust the control instructions according to the feedback information.

Images