Intelligent interaction method and system based on AIGC

Abstract

The invention relates to the technical field of intelligent interaction, in particular to an intelligent interaction method and system based on AIGC. According to the intelligent interaction method and system based on AIGC, an input text interacted for the first time is split into interaction vocabularies, then term description depth levels of the interaction vocabularies are analyzed, and contribution differences between the interaction vocabularies and effective interaction vocabularies are analyzed; the interaction depth of the interaction vocabularies is obtained, then an initial sampling weight is set, pertinent activation of expert parameters is carried out, and an output text of first interaction is obtained; and then according to the input text of the secondary interaction, the change tendency and change degree of the interaction demand of the user are analyzed to obtain the adjustment sampling weight of the secondary interaction, the activated expert parameters are subjected to targeted suppression, and the deformed text of the secondary interaction is output. According to the method, the matching degree of the output deep knowledge and the actual interaction stage of the user can be effectively improved, and the response duration of interaction is optimized under the condition of ensuring that the AIGC interaction content is more professional, so that the immersion feeling is maintained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent interaction, in particular to an intelligent interaction method and system based on AIGC. BACKGROUND

[0002] AIGC is an interactive method for outputting artificial intelligence generated content, involving multidimensional technology implementation and user experience optimization. The core is to realize human-computer interaction in an intelligent, personalized and natural way. The actual scene content of the interaction is used as the activation evaluation of the corresponding expert parameters of the model. The output result is corresponded to the user input interaction request through the activation parameter ratio and the professional degree scale, so as to optimize the actual application experience of the interaction.

[0003] Under the change requirement of the field agent, the accurate evaluation of the subdivided field, and the cross-field combination generated by the user's interactive demand, the actual interactive input request text of the user needs to be evaluated to understand the actual demand of the user. However, in the actual process, due to the user's misunderstanding of the search content or the lack of input information, the AIGC cannot accurately understand the actual demand of the user, resulting in a difference between the output interactive text and the subjective intention of the user.

[0004] In the prior art, the expert parameters are activated by matching the input context of the keywords input by the user during the interaction. However, the understanding span generated by the user in the initial interaction will cause the output of deep knowledge to be unmatched with the actual interactive stage of the user, thereby causing the waste of activation of expert parameters. SUMMARY

[0005] In order to solve the above technical problems, the purpose of the present application is to provide an intelligent interaction method and system based on AIGC.

[0006] According to the first aspect of the embodiment of the present application, an intelligent interaction method based on AIGC is provided, and the technical solution is as follows: Based on the AIGC platform, the input text of the first interaction of the user is obtained, and the input text is split into interactive vocabulary; The depth level of the term description of the interactive vocabulary is analyzed to obtain the depth contribution of each interactive vocabulary; According to the depth contribution, the effective interactive vocabulary is determined, and the contribution difference between the interactive vocabulary and the effective interactive vocabulary is analyzed to obtain the interactive depth of the interactive vocabulary; According to the interactive depth, the initial sampling weight is set, the expert parameters are activated in a targeted manner, and the output text of the first interaction is obtained; If the user interacts again based on the output text, the change tendency and change degree of the user's interactive demand are analyzed according to the input text of the reinteraction, an adjustment sampling weight of the reinteraction is obtained, the activated expert parameters are inhibited in a targeted manner, and a deformation text of the reinteraction is output.

[0007] In some embodiments of the present application, the depth level of the term description of the interaction vocabulary is analyzed to obtain the depth contribution of each interaction vocabulary, including: In the domain knowledge graph established in the AIGC platform, the corresponding interaction vocabulary same as the interaction vocabulary is located, the distance of the corresponding interaction vocabulary to the domain name main node of the domain knowledge graph is extracted, and the depth level of the term description corresponding to the interaction vocabulary is obtained; For each interaction vocabulary, the corresponding maximum depth level of the term description in the domain knowledge graph to which it belongs is extracted, and the depth contribution of each interaction vocabulary is obtained in combination with the depth level of the term description corresponding to the interaction vocabulary.

[0008] In some embodiments of the present application, the effective interaction vocabulary is determined according to the depth contribution, including: The median of the depth contribution of all interaction vocabularies is obtained; The interaction vocabulary corresponding to the depth contribution higher than the median is determined as the effective interaction vocabulary.

[0009] In some embodiments of the present application, the contribution difference between the interaction vocabulary and the effective interaction vocabulary is analyzed to obtain the interaction depth of the interaction vocabulary, including: The difference value of the depth contribution between the effective interaction vocabulary and the interaction vocabulary is calculated to obtain the contribution difference between the interaction vocabulary and the effective interaction vocabulary; According to the order of the input text, all the interaction vocabularies are sorted to obtain an interaction vocabulary sequence; In the interaction vocabulary sequence, the distance between the interaction vocabulary and the effective interaction vocabulary is calculated to obtain a distance parameter; The contribution difference and the distance parameter are combined to obtain the interaction depth of the interaction vocabulary.

[0010] In some embodiments of the present application, the contribution difference and the distance parameter are combined to obtain the interaction depth of the interaction vocabulary, including: The contribution difference and the distance parameter are combined to obtain the partial depth contribution of all the effective interaction vocabularies; All the effective interaction vocabularies in the input text are traversed, the sum of the partial depth contributions is calculated, and the interaction depth of the interaction vocabulary is obtained in combination with the total number of the interaction vocabularies.

[0011] In some embodiments of the application, according to the interaction depth, the initial sampling weight is set, the expert parameter is activated, and the output text of the first interaction is obtained, including: In the domain knowledge graph established in the AIGC platform, the domain name main node corresponding to all the interaction words is obtained, and the expert sub-network corresponding to the domain name main node is activated; The interaction depth is normalized by softsign to obtain a normalized interaction depth; The initial sampling weight is set to the normalized interaction depth, and the output string of all activated expert sub-networks is taken as the result of the normalized interaction depth, to obtain the output text of the first interaction.

[0012] In some embodiments of the application, according to the input text of the second interaction, the change tendency of the user's interaction demand is analyzed, including: According to the input text of the second interaction, the second effective interaction word of the second interaction is extracted, and the domain name main node corresponding to the second effective interaction word is obtained; The total number of occurrences of each domain name main node corresponding to the second effective interaction word in all interactions is counted, and the domain name main node corresponding to each second effective interaction word is obtained. The domain emphasis evaluation of the domain name main node is combined with the total number of interactions to obtain the domain emphasis evaluation of the domain name main node corresponding to each second effective interaction word; The number of occurrences of each domain name main node corresponding to the second effective interaction word in the second interaction is counted, and the number of inquiries of the domain name main node corresponding to each second effective interaction word is obtained. The total number of all interaction words in the second interaction is combined to obtain the number of inquiries of the domain name main node corresponding to each second effective interaction word; The domain emphasis evaluation and the number of inquiries are combined to obtain the change tendency of the user's interaction demand.

[0013] In some embodiments of the application, according to the input text of the second interaction, the change degree of the user's interaction demand is analyzed, including: The partial depth contribution of the second effective interaction word is obtained, and the domain name main node corresponding to the second effective interaction word is obtained; According to the partial depth contribution of the second effective interaction word and the domain name main node, the interaction vector of the second interaction is obtained; According to the input text of the last interaction, the interaction vector of the last interaction is obtained; The similarity between the interaction vectors corresponding to the second interaction and the last interaction is analyzed to obtain the change degree of the user's interaction demand in the second interaction and the last interaction.

[0014] According to a second aspect of the embodiments of the present application, an AIGC-based intelligent interaction system is provided, comprising a memory, a processor and a display, wherein: The memory is configured to store program codes. The processor is configured to read the program codes stored in the memory and execute the method according to the first aspect of the embodiments of the present application. The display is configured to display the input text of the user and the output text after processing the input text of the user.

[0015] In some embodiments of the present application, the processor comprises: An interactive vocabulary acquisition module is configured to acquire the input text of the first interaction of the user based on the AIGC platform, and split the input text into interactive vocabularies. An interactive depth analysis module is configured to analyze the term description depth level of the interactive vocabularies to obtain the depth contribution of each interactive vocabulary, and determine the effective interactive vocabulary according to the depth contribution, and analyze the contribution difference between the interactive vocabularies and the effective interactive vocabulary to obtain the interactive depth of the interactive vocabularies. An output text acquisition module is configured to set an initial sampling weight according to the interactive depth, and perform targeted activation of expert parameters to obtain the output text of the first interaction. An output text optimization module is configured to, if the user performs a second interaction based on the output text, analyze the change tendency and change degree of the interactive demand of the user according to the input text of the second interaction to obtain an adjusted sampling weight of the second interaction, perform targeted inhibition on the activated expert parameters, and output a transformed text of the second interaction.

[0016] Compared with the prior art, the AIGC-based intelligent interaction method and system provided by the present application has the following beneficial effects: The application firstly provides basic data for subsequent analysis by splitting the input text into interactive words for model analysis; then the depth contribution of each interactive word is obtained by analyzing the term description depth level of the interactive word, and the interaction depth of the interactive word is obtained by further analyzing the contribution difference between the interactive word and the effective interactive word, that is, the input result of the current interactive word can be accurately reflected by analyzing the semantic contribution of the interactive word, so as to realize the evaluation of the output process; then the sampling ratio of the activation parameter is determined according to the interaction depth of the sentence, and the output text is obtained; then the output content is re-evaluated after the user interaction level is deepened by analyzing the change of the interaction demand embodied in the re-input text compared with the previous input text, the activated expert parameter is inhibited, and the transformed text for re-interaction is output; the technical scheme provided by the application effectively improves the matching degree of the output depth knowledge and the actual interaction stage of the user, ensures that the AIGC interaction content is more professional, and optimizes the interaction experience. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 A basic flow diagram of an AIGC-based intelligent interaction method provided by an embodiment of the present application; Figure 2 A basic composition diagram of an AIGC-based intelligent interaction system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the AIGC-based intelligent interaction method and system according to the present application are described in detail below in combination with the drawings and preferred embodiments, the specific implementation, structure, features and effects thereof. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The use of the terms "at least" and "one or more of should be taken as encompassing an open-ended range of meanings, such that, for example, the phrase "at least two" should be taken to include at least two, at least three, at least four, at least five, at least six, at least seven, at least eight, at least nine, at least ten, at least eleven, at least twelve, at least thirteen, at least fourteen, at least fifteen, at least sixteen, at least seventeen, at least eighteen, at least nineteen, at least twenty, at least twenty-one, at least twenty-two, at least twenty-three, at least twenty-four, at least twenty-five, at least twenty-six, at least twenty-seven, at least twenty-eight, at least twenty-nine, at least thirty, at least thirty-one, at least thirty-two, at least thirty-three, at least thirty-four, at least thirty-five, at least thirty-six, at least thirty-seven, at least thirty-eight, at least thirty-nine, at least forty, and so forth, unless otherwise limited. The use of the terms "at least one of should be taken as encompassing an open-ended range of meanings, such that, for example, the phrase "at least one of A and B" should be taken to include A, B, or both A and B. The use of the terms "one or more of should be taken as encompassing an open-ended range of meanings, such that, for example, the phrase "one or more of A and B" should be taken to include A, B, or both A and B. The use of the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a circuit, item, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such circuit, item, or apparatus.

[0021] The specific scheme of the intelligent interaction method based on AIGC provided by the present application is specifically described below in combination with the accompanying drawings.

[0022] Referring to Figure 1 , a basic flow of an intelligent interaction method based on AIGC provided by an embodiment of the present application is shown.

[0023] As Figure 1 shown, the intelligent interaction method based on AIGC provided by an embodiment of the present application specifically comprises: S100: Based on an AIGC platform, input text of first interaction of a user is acquired, and the input text is split into interactive vocabulary.

[0024] The AIGC platform is visually displayed through a graphical user interface (GUI), and has an interactive module for the user. The input text of first interaction of the user is extracted based on the graphical user interface of the AIGC platform. The input text is preprocessed, such as text cleaning and word segmentation. Specifically, abnormal words in the input text of first interaction of the user are replaced using a stop word replacement dictionary, to ensure that all words of the input text belong to the technical field range of model training; special symbols that are not characters are removed; the input text is split and segmented by the accurate mode of the jieba word segmentation algorithm, to split the input text into interactive vocabulary.

[0025] Up to now, the acquisition and preprocessing of the input text of first interaction of the user are completed, and the interactive vocabulary available for model analysis is obtained.

[0026] S200: The depth level of the term description of the interactive vocabulary is analyzed, to obtain the depth contribution of each interactive vocabulary.

[0027] A generative pre-trained transformer (GPT) is an advanced artificial intelligence language model that understands and generates natural language text through deep learning techniques, particularly the Transformer architecture. GPT is pre-trained on a large amount of text data to learn the patterns and structures of language, enabling it to predict and generate coherent and meaningful text content. GPT models can be widely applied to various natural language processing tasks such as text generation, dialogue systems, and automatic summarization. In embodiments of the present invention, a pre-trained generative pre-trained transformer model is used as the basis for processing input text.

[0028] When users interact, they usually ask questions about a certain professional knowledge point to get a description of the key technology. Therefore, during the interaction process in a certain professional field, the user evaluates the output process by evaluating the term input level of professional knowledge points. If the user is very familiar with a certain technical field, the input terms are more professional, i.e., the input professional terms belong to the deep vocabulary of one or more technical fields.

[0029] Based on the above analysis, in some embodiments of the present invention, the depth contribution of each interaction vocabulary is obtained by analyzing the term description depth level of the interaction vocabulary. Further, it includes: First, the same corresponding interaction vocabulary as the interaction vocabulary is located in the domain knowledge graph established in the AIGC platform, and the distance of the corresponding interaction vocabulary to the domain name main node of the domain knowledge graph is extracted to obtain the term description depth level corresponding to the interaction vocabulary. The specific implementation is: using the node name and relationship name fields of the domain knowledge graph established in the AIGC platform to filter the split interaction vocabulary, locate the same corresponding interaction vocabulary as the interaction vocabulary in the domain knowledge graph; the domain knowledge graph is a tree structure, the farther the distance from the domain name main node, the deeper, therefore, the distance of the corresponding interaction vocabulary of the interaction vocabulary to the domain name main node of the domain knowledge graph is extracted , i.e., the term description depth level corresponding to the interaction vocabulary is , for example: vehicle (domain name main node) - driving mechanism - steering unit - steering wheel (current interaction vocabulary ), the distance between the steering wheel and the vehicle is 3, so the depth level of the current interaction vocabulary (steering wheel) is 3.

[0030] ​​​The interaction description mode of the term description depth level higher interaction vocabulary is analyzed, that is, the term description depth level higher interaction vocabulary represents the depth and specialization of the technology, and the interaction range of the user is more narrow, but the description mode matched with the term description depth level higher interaction vocabulary is more specialized than the term description depth level lower interaction vocabulary, that is, the term range level close to it is affected, that is, the interaction demand of the term description depth level higher interaction vocabulary needs to be exhibited with the similar level of the remaining interaction vocabulary. For example, "CKF and UKF are both first and second moments of filtering distribution through a set of sampling points with weights after the conversion of a nonlinear system", wherein the nonlinear system, filtering distribution, and second moment vocabulary level is 7. Therefore, if the input text of the first interaction of the user has more term description depth levels close to and higher term description depth levels, the first interaction depth rating of the user is higher, so that the activated expert parameter type is more accurate.

[0031] Therefore, in some embodiments of the application, for each interaction vocabulary, the corresponding maximum term description depth level in the domain knowledge graph to which it belongs is extracted, and the depth contribution of each interaction vocabulary is obtained by combining the term description depth level corresponding to the interaction vocabulary. Specific implementation is as follows: For the interaction vocabulary , the corresponding maximum term description depth level in the domain knowledge graph to which it belongs is extracted (every domain knowledge graph has its corresponding deepest level number); then the ratio of the term description depth level corresponding to the interaction vocabulary to the maximum term description depth level is calculated, which is:

[0032] In the formula, represents the ratio of the term description depth level corresponding to the interaction vocabulary to the maximum term description depth level; represents the term description depth level corresponding to the interaction vocabulary ; represents the maximum term description depth level corresponding to the interaction vocabulary in the domain knowledge graph to which it belongs.

[0033] And the average value of the ratio of the maximum term description depth level to the distance of all interaction vocabularies corresponding to the first interaction input text to the domain name main node of the domain knowledge graph is calculated, which is:

[0034] wherein, represents the average of the ratio of the maximum term description depth level to the term description depth level corresponding to the first interaction input text, respectively; represents the interaction vocabulary belonging to the domain knowledge graph corresponding to the maximum term description depth level in the domain knowledge graph represents the first interaction vocabulary corresponding to the first interaction input text. distance to the domain name main node of the domain knowledge graph If the same corresponding interaction vocabulary as the interaction vocabulary cannot be located in the domain knowledge graph , there is no , i.e. value is 0);

[0035] represents the average of the term description depth level in the current domain knowledge graph , which shows the user's interactive request in the domain knowledge graph belonging to the domain's tendency, i.e. whether the user asks professional questions in the current domain.

[0036] Finally, the ratio of the depth contribution of the interaction vocabulary to the average of the ratio of the corresponding value of the domain knowledge graph is obtained, and the depth contribution of the interaction vocabulary is:

[0037] wherein, represents the depth contribution of the interaction vocabulary , and the remaining symbol meanings are the same as above.

[0038] The depth contribution of each interaction vocabulary is obtained by traversing the interaction vocabulary corresponding to the first interaction input text.

[0039] S300: According to the depth contribution, determine the effective interaction vocabulary, and analyze the contribution difference between the interaction vocabulary and the effective interaction vocabulary, to obtain the interactive depth of the interaction vocabulary.

[0040] According to the order of the input text, all the interaction words are sorted to form an interaction word sequence, the depth contribution of the interaction words is placed according to the order of the interaction word sequence, and the description of the words is expressed as: the greater the distance difference between the interaction words of two adjacent high-depth contribution positions, the higher the contribution of other interaction words close to them, and the stronger the description ability of the current interaction word to the interaction words of high-depth contribution; and further, the number of interaction words contained in the current interaction word sequence is counted to realize accurate evaluation of the interaction depth of the input text.

[0041] Based on the above analysis, in some embodiments of the present application, according to the depth contribution, the effective interaction words are determined, and the contribution difference between the interaction words and the effective interaction words is analyzed to obtain the interaction depth of the interaction words. Wherein: According to the depth contribution, the effective interaction words are determined, including: obtaining the median of the depth contribution of all interaction words; the interaction words corresponding to the depth contribution higher than the median are determined as the effective interaction words.

[0042] The contribution difference between the interaction words and the effective interaction words is analyzed to obtain the interaction depth of the interaction words, including: calculating the difference value of the depth contribution between the effective interaction words and the interaction words to obtain the contribution difference between the interaction words and the effective interaction words; according to the order of the input text, all the interaction words are sorted to obtain an interaction word sequence; in the interaction word sequence, the distance between the interaction words and the effective interaction words is calculated to obtain a distance parameter; the contribution difference, the distance parameter and are combined to obtain the interaction depth of the interaction words.

[0043] Wherein, the contribution difference, the distance parameter and are combined to obtain the interaction depth of the interaction words, which further includes: first, the contribution difference and the distance parameter are combined to obtain the partial depth contribution of each effective interaction word, and the partial depth contribution calculation formula of the effective interaction word is constructed. The partial depth contribution calculation formula of the effective interaction word is:

[0044] In the formula, represents the partial depth contribution of the effective interaction word ; represents the depth contribution of the effective interaction word ; represents the depth contribution of the non-effective interaction word closest to the position of the effective interaction word , if there is more than one non-effective interaction word closest to the position, the non-effective interaction word closest to the position (i.e. with the smallest value) is taken . ; represents the distance between the valid interactive vocabulary and the non-valid interactive vocabulary. .

[0045] Then, the sum of the partial depth contributions of all valid interactive vocabularies in the input text is calculated, and the interaction depth of the interactive vocabulary is obtained by combining the total number of interactive vocabularies:

[0046] In the formula, represents the interaction depth of the interactive vocabulary corresponding to the first interaction; represents the partial depth contribution of the valid interactive vocabulary . represents the total number of valid interactive vocabularies corresponding to the first interaction input text; . represents the total number of all interactive vocabularies corresponding to the first interaction input text.

[0047] The greater the value, the longer the input text of the first interaction of the user, and therefore the depth contribution is weighted to reflect the dilution of the depth contribution in long sentences, and the interaction depth evaluation of the current input interactive vocabulary sequence can more accurately reflect the input result of the current interactive vocabulary sequence.

[0048] So far, the semantic contribution analysis of the interactive vocabulary corresponding to the first interaction input text has been completed.

[0049] S400: According to the interaction depth, set the initial sampling weight, and perform targeted activation of the expert parameters to obtain the output text of the first interaction.

[0050] After obtaining the interaction depth evaluation result of the interactive vocabulary corresponding to the first interaction, the expert parameters are activated. The expert parameters refer to the parameters of each subnetwork included in the total model, i.e., the result of training using knowledge parameters in a specific field, so the purpose achieved is the targeted activation of the input parameters to the subnetwork, thereby determining the proportion of different professional knowledge in the output result content. Therefore, according to the interaction depth of the sentence, the sampling proportion of the activated parameters is determined, and the output content is further re-evaluated after deepening the user interaction level.

[0051] Based on the above analysis, in some embodiments of the present application, according to the interaction depth, the initial sampling weight is set, the expert parameters are activated, and the output text of the first interaction is obtained. Specifically, it includes: In the domain knowledge graph established in the AIGC platform, the domain name main node corresponding to all interactive words is obtained, and the expert sub-network corresponding to the domain name main node is activated (the domain name main node is the same as the expert sub-network name).

[0052] The interaction depth is obtained by setting the initial sampling weight as the normalized interaction depth. The initial sampling weight is set as the normalized interaction depth, and the output string of all activated expert sub-networks is obtained by taking the front initial sampling weight (the total weight is set as the sum of the initial sampling weights of all expert sub-networks to avoid weight overflow caused by setting the fixed weight as 1). S500: If the user interacts again based on the output text, the change tendency and change degree of the user's interaction demand are analyzed according to the input text of the second interaction, the adjustment sampling weight of the second interaction is obtained, the activated expert parameters are inhibited, and the deformed text of the second interaction is output.

[0053] After the user obtains the output text, the user will have a new focus on a certain field in the next interaction process, so the change of the output result is evaluated according to the change of the user's depth contribution to the interaction of a certain field.

[0054] Therefore, if the user interacts again based on the output text, the change tendency and change degree of the user's interaction demand are analyzed according to the input text of the second interaction, the adjustment sampling weight of the second interaction is obtained, the activated expert parameters are inhibited, and the deformed text of the second interaction is output.

[0055] According to the input text of the second interaction, the change tendency of the user's interaction demand is analyzed, including: First, according to the input text of the second interaction, the second effective interaction word of the second interaction is extracted, and the domain name main node corresponding to the second effective interaction word is obtained. Among them, according to the input text of the second interaction, the second effective interaction word of the second interaction is extracted, and the specific method is steps S100 to S300, that is, based on the AIGC platform, the second input text of the user's second interaction is obtained, and the second input text is split into second interaction words; the depth contribution of each second interaction word is obtained by analyzing the term description depth level of the second interaction word; and the second effective interaction word corresponding to the second input text is determined according to the depth contribution. For details, see steps S100 to S300, which will not be repeated here. Then, if the same corresponding second interaction word as the second effective interaction word can be located in the domain knowledge graph ​​​The domain name master node in the field name master node is the domain name master node corresponding to the re-interactive vocabulary.

[0056] Then, the re-interactive vocabulary between the historical number of interactions is compared, that is, the way of evaluating the inquiry, the number and type of the field surface faced after the interaction represent the more real interaction demand of the user, for example, the professional level of the interaction gradually deepens, then the rest of the re-interactive vocabulary reflects that the domain name master node gradually decreases, and the concentrated descriptive vocabulary is concentrated in a small number of domain name master nodes, which reflects that the interaction field gradually narrows and gradually deepens. Therefore, it is necessary to evaluate the repetition of the domain name master node first.

[0057] Based on the above analysis, in some embodiments of the present application, by counting the total number of times of the domain name master node corresponding to each re-interactive vocabulary appearing in all interactions, and combining the total number of interactions, the domain focus evaluation of the domain name master node corresponding to each re-interactive vocabulary is obtained as:

[0058] In the formula, the domain focus evaluation of the domain name master node corresponding to the re-interactive vocabulary of the re-interaction (the first interaction) is represented. the total number of times of the domain name master node corresponding to the re-interactive vocabulary appearing in all interactions is represented. the total number of interactions of the user up to the present is represented.

[0059] In addition, the number of times of the domain name master node corresponding to each re-interactive vocabulary appearing in the re-interaction is counted, and the inquiry quantity parameter of the domain name master node corresponding to each re-interactive vocabulary is obtained by combining the total number of all re-interactive vocabularies of the re-interaction as:

[0060] In the formula, the inquiry quantity parameter of the domain name master node corresponding to the re-interactive vocabulary of the re-interaction (the first interaction) is represented. the number of times of the domain name master node corresponding to the re-interactive vocabulary appearing in the re-interaction (the first interaction) is represented. the total number of all re-interactive vocabularies corresponding to the re-interaction input text is represented.

[0061] ​​​​​In the case of in-depth inquiry, the level of inquiry is deeper, so the inquiry method includes more detailed content in a single field, indicating that the user's inquiry is more professional, and therefore more in-depth terminology in the same field is needed for description.

[0062] Finally, combining the domain-specific evaluation and inquiry quantity parameters, the changing trend of user interaction needs is as follows:

[0063] In the formula, Indicates further interaction (the first) The user's interaction needs under the second interaction (for the domain name master node) The tendency of change; Indicates further interaction (the first) The domain name master node corresponding to the next valid interaction term (secondary interaction). The field focuses on evaluation; Indicates further interaction (the first) The domain name master node corresponding to the next valid interaction term (secondary interaction). The number of queries parameter.

[0064] Based on the input text from subsequent interactions, analyze the degree of change in the user's interaction needs, including: First, obtain the partial depth contribution of the re-effective interaction terms and the domain name master nodes corresponding to the re-effective interaction terms. The specific method for obtaining the partial depth contribution of the re-effective interaction terms is the same as in step S300, that is, combining the contribution difference and distance parameters to obtain the partial depth contribution of all re-effective interaction terms. Detailed steps are described in step S300 and will not be repeated here. Then, if in the domain knowledge graph... If the corresponding repeated interaction words that are the same as the words used in the repeated effective interaction can be located, then the knowledge graph of that domain can be established. The domain name master node in the context is the domain name master node corresponding to the valid interaction term. .

[0065] Then, based on the partial depth contribution of the effective interaction vocabulary and the domain name master node, the comparison point is set as... ,in Indicates the domain name master node, Indicates a word that represents another valid interaction. The partial depth contribution; the interaction vector of the second interaction is obtained by arranging the comparison points in the order of the second interaction words of the re-input text, denoted as .

[0066] Similarly, based on the input text of the previous interaction (the previous interaction related to the next interaction level), the interaction vector of the previous interaction is obtained, denoted as... ; Finally, by analyzing the similarity between the interaction vectors corresponding to the second interaction and the previous interaction, we can obtain the degree of change in the user's interaction needs between the second interaction and the previous interaction:

[0067] In the formula, This indicates that the user is interacting again (the first time). The first interaction) and the previous interaction (the second interaction) The degree of change in interaction needs (multiple interactions); The interaction vector representing the previous interaction related to the current interaction level. Interaction vectors with subsequent interactions Cosine similarity between them.

[0068] The current query method is determined by the query order among interaction vectors. Specifically, the similarity between interaction vectors is used as the background activation parameter for the current query, taking into account the query result deviation (similarity between interaction vectors). This is further combined with the domain name master node. The tendency of change, that is, combining the domain name master node The domain focuses on evaluating and querying quantity parameters to obtain the domain name master node. The sampling weights are adjusted in the subsequent interaction as follows:

[0069] In the formula, Indicates the domain name master node In the next interaction (the first) Adjusting sampling weights under (secondary interaction); Indicates further interaction (the first) The user's interaction needs under the second interaction (for the domain name master node) The tendency of change; This indicates that the user is interacting again (the first time). The first interaction) and the previous interaction (the second interaction) The degree of change in interaction needs (multiple interactions).

[0070] By analyzing the two proportional values ​​( and The sum of 1 and -1 is used to complete the mapping from logarithmic values ​​to 1, thereby ensuring that the activation level of expert parameters is similar when the questioning method does not change much, and stabilizing the output of the response results in the interactive process.

[0071] Similarly, the main nodes of each domain name are obtained in the subsequent interaction (the first... adjust the sampling weight under the second interaction. According to the adjusted sampling weight, the adjusted sampling weight is set, the activated expert parameters are inhibited in a targeted manner, and the deformed text of the second interaction is output. The specific operation method is the same as step S400, that is, the adjusted sampling weight is obtained by performing softsign normalization processing on the interaction depth ; the initial sampling weight is set to the normalized interaction depth , and the first normalized interaction depth of the output string of all activated expert subnetworks is taken (the total weight is set to the sum of the initial sampling weights of all expert subnetworks, so as to avoid weight overflow caused by the fixed weight being 1), and the deformed text of the second interaction is output.

[0072] Finally, the output interaction result is visualized and displayed. Specifically: After the interactive sentence is generated, real-time streaming output is performed: Server-Sent Events (SSE) is used to support word-by-word / sentence-by-sentence text pushing, and the text is sent to the memory of the client.

[0073] The message history area is displayed as: a) User input and model reply are displayed on the left and right (similar to a chat interface); b) User message (right side): light background + right alignment; c) Model reply (left side): dark background + left alignment + fixed maximum width.

[0074] The input control area is arranged as: a text input box (supporting multiple lines), a send button, and an additional function button (such as clearing history, setting).

[0075] A status indicator is used to indicate whether the output has been completed.

[0076] Loading animation (such as progress bar, rotating icon), error prompt area.

[0077] Based on the same inventive concept as the above method, the embodiment also provides an intelligent interaction system based on AIGC.

[0078] Please refer to Figure 2 , which shows the basic composition of an intelligent interaction system based on AIGC provided by an embodiment of the present application.

[0079] As shown in Figure 2 , an intelligent interaction system based on AIGC includes a memory 10, a processor 20, and a display 30, wherein: The memory 10 is used to store program codes.

[0080] The processor 20 is configured to read the program code stored in the memory 10, and perform the AIGC platform-based operation to obtain input text of first interaction of a user, and split the input text into interaction words; analyze a term description depth level of the interaction words to obtain depth contribution of each interaction word; determine effective interaction words according to the depth contribution, and analyze contribution difference between the interaction words and the effective interaction words to obtain interaction depth of the interaction words; set initial sampling weight according to the interaction depth, perform targeted activation of expert parameters to obtain output text of the first interaction; if the user performs second interaction based on the output text, analyze variation tendency and variation degree of interaction demand of the user according to input text of the second interaction to obtain adjustment sampling weight of the second interaction, perform targeted inhibition on the activated expert parameters, and output deformed text of the second interaction.

[0081] The display 30 is configured to display the input text of the user and display the output text after processing the input text of the user.

[0082] Further, the processor 20 includes an interaction word obtaining module 21, an interaction depth analyzing module 22, an output text obtaining module 23 and an output text optimizing module 24. Wherein: The interaction word obtaining module 21 is configured to obtain input text of first interaction of a user based on the AIGC platform, and split the input text into interaction words; The interaction depth analyzing module 22 is configured to analyze a term description depth level of the interaction words to obtain depth contribution of each interaction word, and determine effective interaction words according to the depth contribution, and analyze contribution difference between the interaction words and the effective interaction words to obtain interaction depth of the interaction words; The output text obtaining module 23 is configured to set initial sampling weight according to the interaction depth, perform targeted activation of expert parameters to obtain output text of the first interaction; The output text optimizing module 24 is configured to, if the user performs second interaction based on the output text, analyze variation tendency and variation degree of interaction demand of the user according to input text of the second interaction to obtain adjustment sampling weight of the second interaction, perform targeted inhibition on the activated expert parameters, and output deformed text of the second interaction.

[0083] The display 30 includes a message history display area, an input control area, a state indicator and an animation loading area. Wherein: The message history display area is arranged as: a) User input and model reply are displayed on the left and right (similar to a chat interface); b) User message (right side): light background + right alignment; c) Model reply (left side): dark background + left alignment + fixed maximum width.

[0084] The input control area layout is: text input box (supporting multiple lines), send button, additional function button (such as clearing history, setting).

[0085] A status indicator is used to indicate whether the output has been completed.

[0086] The animation loading area includes a progress bar, a rotating icon, and an error prompt sub-area.

[0087] It should be noted that the above-mentioned embodiment order of the application is only for description, and does not represent the pros and cons of the embodiment. The process depicted in the drawing does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0088] Various embodiments in the specification are described in a progressive manner, and the same or similar parts between various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments.

Claims

1. An AIGC-based intelligent interaction method, characterized in that, The method comprises: Based on the AIGC platform, the input text of the first interaction of the user is obtained, and the input text is split into interactive vocabulary; Analyze the term description depth level of the interactive vocabulary to obtain the depth contribution of each interactive vocabulary; According to the depth contribution, determine the effective interactive vocabulary, and analyze the contribution difference between the interactive vocabulary and the effective interactive vocabulary to obtain the interactive depth of the interactive vocabulary; According to the interactive depth, set the initial sampling weight, activate the expert parameter, obtain the output text of the first interaction; If the user interacts again based on the output text, according to the input text of the second interaction, analyze the change tendency and change degree of the user's interactive demand, obtain the adjustment sampling weight of the second interaction, and inhibit the activated expert parameter, output the deformation text of the second interaction. 2.The AIGC-based intelligent interaction method of claim 1, wherein, Analyze the term description depth level of the interactive vocabulary to obtain the depth contribution of each interactive vocabulary, including: Locate the same corresponding interactive vocabulary as the interactive vocabulary in the domain knowledge graph established in the AIGC platform, extract the distance from the corresponding interactive vocabulary to the domain name main node of the domain knowledge graph, and obtain the term description depth level corresponding to the interactive vocabulary; For each interactive vocabulary, extract the corresponding maximum term description depth level in the domain knowledge graph to which it belongs, and combine the term description depth level corresponding to the interactive vocabulary to obtain the depth contribution of each interactive vocabulary. 3.The AIGC-based intelligent interaction method of claim 1, wherein, According to the depth contribution, determine the effective interactive vocabulary, including: Obtain the median of the depth contribution of all interactive vocabularies; The interactive vocabulary corresponding to the depth contribution higher than the median is determined as the effective interactive vocabulary. 4.The AIGC-based intelligent interaction method of claim 3, characterized in that, Analyze the contribution difference between the interactive vocabulary and the effective interactive vocabulary to obtain the interactive depth of the interactive vocabulary, including: Calculate the difference value of the depth contribution between the effective interactive vocabulary and the interactive vocabulary to obtain the contribution difference between the interactive vocabulary and the effective interactive vocabulary; According to the order of the input text, sort all the interactive vocabularies to obtain the interactive vocabulary sequence; In the interactive vocabulary sequence, calculate the distance between the interactive vocabulary and the effective interactive vocabulary to obtain the distance parameter; Combine the contribution difference and the distance parameter to obtain the interactive depth of the interactive vocabulary. 5.The AIGC-based intelligent interaction method according to claim 4, characterized in that, Combine the contribution difference and the distance parameter to obtain the interactive depth of the interactive vocabulary, including: Combine the contribution difference and the distance parameter to obtain the partial depth contribution of all the effective interactive vocabularies; Traverse all the effective interactive vocabularies in the input text, calculate the sum of the partial depth contribution, and combine the total number of interactive vocabularies to obtain the interactive depth of the interactive vocabulary. 6.The AIGC-based intelligent interaction method according to claim 2, characterized in that, According to the interactive depth, set the initial sampling weight, activate the expert parameter, obtain the output text of the first interaction, including: In the domain knowledge graph established in the AIGC platform, obtain the domain name main node corresponding to all the interactive vocabularies, and activate the expert subnetwork corresponding to the domain name main node; The interaction depth is normalized by softsign to obtain a normalized interaction depth; The initial sampling weight is set as a percentage of the normalized interaction depth, and the output string of all activated expert sub-networks is taken as a result of the first percentage of the normalized interaction depth to obtain the output text of the first interaction. 7.The AIGC-based intelligent interaction method according to claim 4, characterized in that, According to the input text of the second interaction, the change tendency of the user's interaction demand is analyzed, including: According to the input text of the second interaction, the second effective interaction vocabulary of the second interaction is extracted, and the domain name master node corresponding to the second effective interaction vocabulary is obtained; The total number of appearances of each domain name master node corresponding to the second effective interaction vocabulary in all second interactions is counted, and a domain emphasis evaluation of each domain name master node corresponding to the second effective interaction vocabulary is obtained in combination with the total number of interactions; The number of appearances of each domain name master node corresponding to the second effective interaction vocabulary in the second interaction is counted, and an inquiry quantity parameter of each domain name master node corresponding to the second effective interaction vocabulary is obtained in combination with the total number of interaction vocabularies of the second interaction; The domain emphasis evaluation and the inquiry quantity parameter are combined to obtain the change tendency of the user's interaction demand. 8.The AIGC-based intelligent interaction method of claim 7, characterized in that, According to the input text of the second interaction, the change degree of the user's interaction demand is analyzed, including: The partial depth contribution of the second effective interaction vocabulary is obtained, and the domain name master node corresponding to the second effective interaction vocabulary is obtained; According to the partial depth contribution of the second effective interaction vocabulary and the domain name master node, an interaction vector of the second interaction is obtained; According to the input text of the last interaction, an interaction vector of the last interaction is obtained; The similarity between the interaction vectors corresponding to the second interaction and the last interaction is analyzed to obtain the change degree of the user's interaction demand in the second interaction and the last interaction.

9. An AIGC-based intelligent interaction system, characterized in that, The system comprises a memory, a processor and a display, wherein: The memory is used to store program codes; The processor is used to read the program codes stored in the memory and execute the method according to any one of claims 1 to 8; The display is used to display the input text of the user and display the output text after processing the input text of the user. 10.The AIGC-based intelligent interaction system according to claim 9, characterized in that, The processor comprises: An interaction vocabulary acquisition module is used to acquire the input text of the first interaction of the user based on an AIGC platform, and split the input text into interaction vocabularies; An interaction depth analysis module is used to analyze the term description depth level of the interaction vocabularies to obtain the depth contribution of each interaction vocabulary, and determine effective interaction vocabularies according to the depth contribution, and analyze the contribution difference between the interaction vocabularies and the effective interaction vocabularies to obtain the interaction depth of the interaction vocabularies; An output text acquisition module is used to set an initial sampling weight according to the interaction depth, activate expert parameters in a targeted manner, and obtain the output text of the first interaction. The output text optimization module is configured to, if the user interacts again based on the output text, analyze a change tendency and a change degree of the interactive demand of the user according to input text of the interaction again, obtain an adjustment sampling weight of the interaction again, perform targeted inhibition on the activated expert parameters, and output a transformed text of the interaction again.

Images