CN116453549A

CN116453549A - AI dialogue method based on virtual digital character and online virtual digital system

Info

Publication number: CN116453549A
Application number: CN202310496728.7A
Authority: CN
Inventors: 赵黄婷; 陈媛
Original assignee: Guangxi Muzhe Technology Co ltd
Current assignee: Zhao Huangting
Priority date: 2023-05-05
Filing date: 2023-05-05
Publication date: 2023-07-18

Abstract

The embodiment of the application provides an AI dialogue method and an online virtual digitizing system based on a virtual digital character, which are characterized in that through analyzing first user positive emotion tendency content of the virtual digital character, after obtaining first dialogue attention weight of an AI dialogue event corresponding to the first user positive emotion tendency content, updating the first dialogue attention weight through dialogue evaluation data, and under the condition that user dialogue emotion figures of the virtual digital character are determined, the influence of dialogue evaluation data on the user dialogue effect of the virtual digital character is further combined, so that more accurate online interactive user recommendation is executed on the virtual digital character, and the AI dialogue matching degree of subsequent online interactive users is improved.

Description

AI dialogue method based on virtual digital character and online virtual digital system

Technical Field

The application relates to the technical field of virtual scenes and AI, in particular to an AI dialogue method based on virtual digital characters and an online virtual digital system.

Background

The virtual Digital person (Digital Human/Meta Human) is a Digital figure which is created by using a Digital technology and is close to a Human figure, external input information can be automatically read, analyzed and identified through an intelligent system, a subsequent output text of the Digital person is decided according to an analysis result, a virtual Digital figure model is driven to generate corresponding voice and action so as to enable the Digital person to interact with a user, and recommendation of virtual Digital figures of other related users is carried out according to interaction effects, namely dialogue evaluation data of the virtual Digital figures, so that applicability of the virtual Digital figures is improved. However, in the related art, through the tests of the inventor, if online interactive user recommendation is performed based on only dialogue evaluation data of the virtual digital character, a good AI dialogue matching degree cannot be achieved.

Disclosure of Invention

In order to at least overcome the above-mentioned shortcomings in the prior art, an object of the present application is to provide an AI conversation method based on virtual digital characters and an online virtual digitizing system.

In a first aspect, the present application provides an AI conversation method based on a virtual digital character, applied to an online virtual digitizing system, the method comprising:

acquiring virtual character dialogue voices of a target user in an online virtual digital scene initiating a plurality of AI dialogue events of an AI dialogue aiming at a virtual digital character, and carrying out dialogue emotion analysis on the virtual character dialogue voices to acquire first user positive emotion tendency content of the virtual digital character;

determining a content feature vector corresponding to the first user positive emotion tendencies content, and determining an attention parameter value of the first user positive emotion tendencies content based on the content feature vector, wherein the content feature vector at least comprises a content semantic feature vector and a content context feature vector, and the attention parameter value represents a user dialogue emotion portrait of the virtual digital character;

determining an emotion value corresponding to the first user active emotion tendency content based on the emotion attribute of the first user active emotion tendency content, and obtaining a first dialogue attention weight of an AI dialogue event corresponding to the first user active emotion tendency content based on the emotion value and the attention parameter value;

Acquiring dialogue evaluation data of an AI dialogue event corresponding to the first user positive emotion tendency content, analyzing the dialogue evaluation data to update the first dialogue attention weight based on the dialogue evaluation data and acquiring an updated second dialogue attention weight;

and based on the second dialogue attention weight, performing online interactive user recommendation on the virtual digital character, wherein the online interactive user recommendation is associated with the target user.

In a possible implementation manner of the first aspect, the step of analyzing the dialogue evaluation data to update the first dialogue attention weight based on the dialogue evaluation data and obtain an updated second dialogue attention weight includes:

acquiring a virtual character dialogue voice set of the virtual digital character, and extracting a prior virtual character dialogue voice associated with a time node corresponding to the virtual character dialogue voice from the virtual character dialogue voice set;

performing dialogue emotion analysis on the dialogue voice of the prior virtual character to obtain second user positive emotion tendency content corresponding to the virtual digital character;

comparing the first user positive emotion tendencies content with the second user positive emotion tendencies content, and determining positive emotion tendencies conversion parameters between the first user positive emotion tendencies content and the second user positive emotion tendencies content;

Determining a dialogue evaluation weight value corresponding to the dialogue evaluation data, and determining an update index value corresponding to the dialogue evaluation data based on a fusion parameter value between the dialogue evaluation weight value and the positive emotion tendency conversion parameter;

and updating the first dialogue attention weight based on the updating index value to obtain an updated second dialogue attention weight, wherein the updating direction and the updating amplitude are in positive association with the updating index value and the dialogue evaluation weight value.

In a possible implementation manner of the first aspect, the dialogue evaluation data includes at least scene effect evaluation data, emotion understanding evaluation data, and emotion expression evaluation data, and the step of determining a dialogue evaluation weight value corresponding to the dialogue evaluation data includes:

determining a plurality of scene effect evaluation score ranges and a plurality of emotion understanding evaluation score ranges corresponding to emotion attributes of the first user active emotion tendency content, wherein each scene effect evaluation score range corresponds to a first dialogue evaluation weight value, and each emotion understanding evaluation score range corresponds to a second dialogue evaluation weight value;

based on the scene effect evaluation data and the emotion understanding evaluation data, respectively determining a first dialogue evaluation weight value and a second dialogue evaluation weight value corresponding to the first user positive emotion tendency content;

Acquiring an evaluation knowledge graph of emotion expression evaluation data of an AI dialogue event where the first user positive emotion tendency content is located, and determining a plurality of significant emotion activity data in the AI dialogue event based on the evaluation knowledge graph;

determining a third dialogue evaluation weight value corresponding to the first user positive emotion tendency content based on the plurality of significant emotion activity data;

respectively determining fusion coefficients corresponding to the first dialogue evaluation weight value, the second dialogue evaluation weight value and the third dialogue evaluation weight value, and fusing the first dialogue evaluation weight value, the second dialogue evaluation weight value and the third dialogue evaluation weight value based on the fusion coefficients to obtain dialogue evaluation weight values corresponding to the dialogue evaluation data;

based on the plurality of significant emotion activity data, determining a third dialogue evaluation weight value corresponding to the first user positive emotion tendency content, including:

performing unit splitting on the evaluation knowledge graph, and extracting a plurality of target unit evaluation knowledge graphs associated with the significant emotion activity data from the unit evaluation knowledge graphs obtained after unit splitting aiming at each significant emotion activity data;

Determining map nodes corresponding to the significant emotion activity data in the multiple target unit evaluation knowledge maps respectively, and sequencing the map nodes based on descending order of the crossing number of the map nodes to obtain target map nodes of the significant emotion activity data; the target map node represents the node of the maximum crossing number of the significant emotion activity data;

mapping the evaluation knowledge graph to a dialogue importance space, and determining dialogue importance parameters of the target graph according to the dialogue importance space;

and determining a third dialogue evaluation weight value corresponding to the first user positive emotion tendency content based on the dialogue importance parameter.

In a possible implementation manner of the first aspect, the step of performing dialogue emotion analysis on the virtual character dialogue speech to obtain first user positive emotion tendencies content of the virtual digital character includes:

performing dialogue emotion encoding on the virtual character dialogue speech, and determining an active emotion mapping mean value of each dialogue interactive statement of the virtual character dialogue speech after dialogue emotion encoding so as to determine formatted interactive data and target interactive data to be analyzed in the virtual character dialogue speech based on the active emotion mapping mean value, wherein the active emotion mapping mean value of the formatted interactive data is smaller than the active emotion mapping mean value of the target interactive data;

Respectively determining corresponding associated dialogue interactive sentences in a front Wen Yugou set and a rear Wen Yugou set of each dialogue interactive sentence in the target interactive data, and positive emotion mapping standard deviation values and positive emotion mapping average values between corresponding positive emotion mapping values of the associated dialogue interactive sentences;

determining a significance trend value corresponding to the dialogue interactive statement based on the ratio between the positive emotion mapping standard deviation value and the positive emotion mapping mean value;

and comparing the significance trend value corresponding to each dialogue interactive sentence with a set trend value, and clustering the dialogue interactive sentences in the target interactive data based on a comparison result to obtain first user positive emotion trend content of the virtual digital character, wherein the significance trend value corresponding to the first user positive emotion trend content is larger than the set trend value.

In a possible implementation manner of the first aspect, after determining the first user positive emotion tendencies content of the virtual digital person, the method further includes:

business clauses are carried out on the first user positive emotion tendency content, a first user positive emotion mapping statement sequence after the clauses are obtained, and the first user positive emotion mapping statement sequence is composed of a plurality of user positive emotion mapping statements;

For each user positive emotion mapping statement, traversing the first user positive emotion mapping statement sequence in sequence by taking the user positive emotion mapping statement as a starting point to obtain an associated user positive emotion mapping statement which is associated with the user positive emotion mapping statement and has dialogue behaviors;

connecting the associated user positive emotion mapping sentences to obtain a plurality of associated positive emotion tendency contents corresponding to the first user positive emotion tendency contents;

determining the number of dialogue interactive sentences in the plurality of associated positive emotion tendencies and the number of standard dialogue interactive sentences corresponding to the plurality of associated positive emotion tendencies respectively, and comparing the number of dialogue interactive sentences corresponding to each associated positive emotion tendencies with the number of standard dialogue interactive sentences to obtain whether the number of standard dialogue interactive sentences is larger than the number of dialogue interactive sentences;

and if the number of the standard dialogue interactive sentences is larger than the number of the dialogue interactive sentences, removing the associated positive emotion tendencies from the first user positive emotion tendencies.

In a possible implementation manner of the first aspect, before determining the fusion coefficients corresponding to the first dialogue evaluation weight value, the second dialogue evaluation weight value, and the third dialogue evaluation weight value, the method further includes:

Determining an interaction problem label where the virtual digital character is currently located and label weight information corresponding to the dialogue evaluation data under the interaction problem label;

and respectively adjusting fusion coefficients corresponding to the first dialogue evaluation weight value, the second dialogue evaluation weight value and the third dialogue evaluation weight value based on the label weight information.

In a possible implementation manner of the first aspect, the step of determining an attention parameter value of the first user positive emotion tendencies content based on the content feature vector includes:

taking the content feature vector as a training sample and the attention parameter value as a training label to construct a Monte Carlo neural network;

and loading the content characteristic vector into the Monte Carlo neural network, and determining the attention parameter value of the first user positive emotion tendency content.

The method of claim 4, wherein the step of dialog emotion encoding the avatar dialogue speech and determining a positive emotion mapping mean value for each dialog interaction statement of the avatar dialogue speech after dialog emotion encoding comprises:

Loading the virtual character dialogue voice to a dialogue emotion analysis network to perform dialogue emotion analysis on the virtual character dialogue voice through the dialogue emotion analysis network, and determining a dialogue emotion analysis result of the virtual character dialogue voice, wherein the dialogue emotion analysis result comprises a dialogue emotion label of each dialogue interaction statement and a corresponding label mapping value;

performing positive emotion feature analysis on dialogue emotion analysis results of the virtual character dialogue voices to obtain positive emotion mapping average values of each dialogue interactive statement;

the training step of the dialogue emotion analysis network comprises the following steps:

acquiring a first number of virtual character dialogue supervised samples and a second number of virtual character dialogue unsupervised samples, and loading the first number of virtual character dialogue supervised samples and the second number of virtual character dialogue unsupervised samples to a dialogue emotion analysis network; the first number of virtual character dialogue supervision samples respectively carry dialogue emotion annotation information of the contained virtual character dialogue voices; the virtual character dialogue voices carried by the first number of virtual character dialogue supervised samples and the virtual character dialogue voices carried by the second number of virtual character dialogue unsupervised samples belong to the same virtual digital character scene;

Determining first dialogue emotion analysis information of virtual character dialogue voices contained in each virtual character dialogue supervision sample in the dialogue emotion analysis network, and acquiring an associated virtual character dialogue voice sample of each virtual character dialogue supervision sample from a dialogue interaction knowledge resource pool; the dialogue interactive knowledge resource pool comprises the first number of virtual character dialogue supervised samples and the second number of virtual character dialogue unsupervised samples; the dialogue voice sample of the associated virtual character of each virtual character dialogue supervision sample does not carry dialogue emotion marking information carried by the dialogue supervision sample of the belonging virtual character;

determining a first learning effect index based on a sample association value between each virtual character dialogue supervision sample and an associated virtual character dialogue voice sample, and determining a second learning effect index based on first dialogue emotion analysis information corresponding to each virtual character dialogue supervision sample and carried dialogue emotion marking information;

adjusting weight configuration information of the dialogue emotion analysis network according to the first learning effect index and the second learning effect index to generate a target dialogue emotion analysis network; the target dialogue emotion analysis network is used for performing dialogue emotion analysis on virtual character dialogue voices belonging to the virtual digital character scene;

The obtaining the associated avatar dialogue voice sample of each avatar dialogue supervision sample from the dialogue interaction knowledge resource pool comprises the following steps:

generating dialogue semantic characterization information of each virtual character dialogue supervision sample and dialogue semantic characterization information of each virtual character dialogue unsupervised sample in the dialogue emotion analysis network;

generating a dialogue semantic association array based on the dialogue semantic characterization information of each virtual character dialogue supervision sample and the dialogue semantic characterization information of each virtual character dialogue non-supervision sample;

acquiring sample association values between each virtual character dialogue supervision sample and dialogue interaction voices in the dialogue interaction knowledge resource pool from the dialogue semantic association array; determining an associated virtual character dialogue speech sample of each virtual character dialogue supervision sample from the dialogue interaction knowledge resource pool based on sample association values between each virtual character dialogue supervision sample and dialogue interaction speech in the dialogue interaction knowledge resource pool;

the first number of virtual character conversation monitoring samples comprises virtual character conversation monitoring samples i, wherein i is a positive integer not greater than the second number;

The determining, based on the sample association value between each virtual character dialogue supervision sample and dialogue interaction voice in the dialogue interaction knowledge resource pool, an associated virtual character dialogue voice sample of each virtual character dialogue supervision sample from the dialogue interaction knowledge resource pool includes:

cleaning a dialogue sample carrying the same dialogue emotion priori information with the virtual character dialogue supervision sample i from the dialogue interaction knowledge resource pool to generate a candidate dialogue interaction knowledge resource pool;

according to descending order arrangement information of sample association values between the virtual character dialogue supervision sample i and each dialogue interaction voice in the candidate dialogue interaction knowledge resource pool, sequencing each dialogue interaction voice to generate a candidate dialogue interaction knowledge resource pool;

acquiring the number k of the related virtual character dialogue voice samples, and determining the first k dialogue interactive voices in the candidate dialogue interactive knowledge resource pool as the related virtual character dialogue voice samples of the virtual character dialogue supervision sample i; k is a positive integer less than the sum of the second number and the first number;

the generating a dialogue semantic association array based on the dialogue semantic characterization information of each virtual character dialogue supervision sample and the dialogue semantic characterization information of each virtual character dialogue unsupervised sample comprises the following steps:

Obtaining regularized conversion characterization information obtained after regularized conversion of dialogue semantic characterization information of each virtual character dialogue supervision sample, and obtaining regularized conversion characterization information obtained after regularized conversion of dialogue semantic characterization information of each virtual character dialogue non-supervision sample;

acquiring a first characterization knowledge extraction vector array comprising regularized conversion characterization information of each virtual character dialogue supervised sample and a second characterization knowledge extraction vector array comprising regularized conversion characterization information of each virtual character dialogue unsupervised sample;

fusing the first characterization knowledge extraction vector array and the second characterization knowledge extraction vector array to generate a target characterization knowledge extraction vector array;

acquiring a Wen Yuyi vector array before and after attachment of the target characterization knowledge extraction vector array, and determining fusion information of the first characterization knowledge extraction vector array and the Wen Yuyi vector array before and after attachment as the dialogue semantic association array.

For example, in a possible implementation manner of the first aspect, the determining the first learning effect index based on the sample association value between each virtual character dialogue supervision sample and the associated virtual character dialogue speech sample includes:

Based on the sample association value between each virtual character dialogue supervised sample and the associated virtual character dialogue voice sample, respectively determining the average sample association value between each virtual character dialogue supervised sample and the associated virtual character dialogue voice sample;

generating a sample association value array based on the average sample association value between each virtual character dialogue supervision sample and the associated virtual character dialogue voice sample;

and determining a variance function of the sample association value array as the first learning effect index.

For example, in a possible implementation manner of the first aspect, the adjusting the weight configuration information of the dialog emotion analysis network according to the first learning effect index and the second learning effect index to generate the target dialog emotion analysis network includes:

weighting calculation is carried out on the first learning effect index and the second learning effect index, and a final learning effect index is generated;

adjusting weight configuration information of the dialogue emotion analysis network according to the final learning effect index;

and when analysis determines that the weight configuration information of the dialogue emotion analysis network is not changed any more, determining the dialogue emotion analysis network as the target dialogue emotion analysis network.

For example, in a possible implementation manner of the first aspect, the method further includes:

loading the first number of virtual character dialogue supervision samples into a basic dialogue emotion analysis network;

determining second dialogue emotion analysis information of the virtual character dialogue voices contained in each virtual character dialogue supervision sample in the basic dialogue emotion analysis network;

and adjusting weight configuration information of the basic dialogue emotion analysis network based on second dialogue emotion analysis information corresponding to each virtual character dialogue supervision sample and dialogue emotion marking information carried by each virtual character dialogue supervision sample, and generating the dialogue emotion analysis network.

For example, in a possible implementation manner of the first aspect, the determining, in the dialog emotion analysis network, first dialog emotion analysis information of a virtual character dialog voice included in each virtual character dialog supervision sample includes:

generating dialogue semantic characterization information of each virtual character dialogue supervision sample in the dialogue emotion analysis network;

performing regularization conversion on dialogue semantic characterization information of each virtual character dialogue supervision sample to generate regularized conversion characterization information of each virtual character dialogue supervision sample;

Determining first dialogue emotion analysis information of each virtual character dialogue supervision sample according to the regularized conversion characterization information of each virtual character dialogue supervision sample;

the first number of virtual character conversation monitoring samples comprises virtual character conversation monitoring samples i, wherein i is a positive integer not greater than the second number; the weight configuration information of the dialogue emotion analysis network comprises weight configuration information of FCL units; the first number of virtual character dialogue supervision samples carry s dialogue emotion priori information, one dialogue emotion priori information corresponds to one dialogue emotion annotation information, and s is a positive integer not greater than the second number;

the determining the first dialogue emotion analysis information of each virtual character dialogue supervision sample according to the regularized conversion characterization information of each virtual character dialogue supervision sample comprises the following steps:

performing regularized conversion on the weight configuration information of the FCL unit to generate conversion weight configuration information of the FCL unit;

determining the mapping confidence coefficient of each dialog emotion annotation information in s dialog emotion annotation information of the virtual character dialog voice contained in the virtual character dialog supervision sample according to the regularized conversion characterization information and the conversion weight configuration information of the virtual character dialog supervision sample;

And determining the mapping confidence degree of the virtual character dialogue voice contained in the virtual character dialogue supervision sample i as the dialogue emotion marking information of each type of dialogue emotion marking information as first dialogue emotion analysis information of the virtual character dialogue supervision sample i.

In a second aspect, embodiments of the present application also provide an online virtual digitizing system comprising a processor and a machine readable storage medium having stored therein a computer program loaded and executed in conjunction with the processor to implement the virtual digital character based AI conversation method of the first aspect above.

In a third aspect, embodiments of the present application provide a computer-readable storage medium storing computer-executable instructions for, when executed by a processor, implementing the virtual digital character-based AI conversation method of the above first aspect.

In a fourth aspect, embodiments of the present application provide a computer program product comprising a computer program or computer executable instructions which, when executed by a processor, implement the above AI conversation method of the first aspect, based on virtual digital characters.

The embodiment of the application has at least the following beneficial effects:

according to the method and the device for processing the user dialogue emotion tendencies, the first dialogue attention weight of the AI dialogue event corresponding to the first user positive emotion tendencies is obtained through analysis of the first user positive emotion tendencies of the virtual digital characters, and then the first dialogue attention weight is updated through dialogue evaluation data, so that the influence of the dialogue evaluation data on the user dialogue effect of the virtual digital characters can be further combined under the condition that the user dialogue emotion portraits of the virtual digital characters are determined, and further accurate online interaction user recommendation is performed on the virtual digital characters, and the AI dialogue matching degree of subsequent online interaction users is improved.

Drawings

Fig. 1 is a flow chart of an AI dialogue method based on a virtual digital character according to an embodiment of the present application.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the present application more apparent, the present application will be described in further detail with reference to the accompanying drawings, and the described embodiments should not be construed as limiting the present application, and all other embodiments obtained by those skilled in the art without making any inventive effort are within the scope of the present application.

In the following description, reference is made to "some embodiments" which describe a subset of all possible embodiments, but it is to be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other on a non-conflicting basis.

In the following description, references to the term "first/second" are merely to distinguish similar virtual character conversational voices and do not represent a particular ordering for objects, it being understood that the "first/second" may be interchanged with a particular order or sequence, as allowed, to enable embodiments of the present application described herein to be implemented in an order other than illustrated or described herein.

Unless defined otherwise, all technical and scientific terms used in the examples of this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the embodiments of the application is for the purpose of describing the embodiments of the application only and is not intended to be limiting of the application.

Step S101: and acquiring virtual character dialogue voices of a target user initiating a plurality of AI dialogue events of the AI dialogue aiming at the virtual digital character, and carrying out dialogue emotion analysis on the virtual character dialogue voices to acquire first user positive emotion tendency content of the virtual digital character.

Virtual digital characters are digitized character images created using digital technology that approximate human images, which can establish AI dialogues with target users to perform dialog interactions for online services (e.g., online medical services, online e-commerce services, online industrial services, online educational services, etc.). In the dialogue interaction process, the virtual character dialogue voice corresponding to the virtual digital character and the target user under the current AI dialogue event can be collected. After virtual character dialogue voices of a plurality of AI dialogue events of the AI dialogue initiated by the target user aiming at the virtual digital character are acquired, dialogue emotion analysis is needed for the virtual character dialogue voices, so that whether positive emotion tendencies (such as happiness, satisfaction, excitement and the like) exist in the current virtual digital character or not and the first user positive emotion tendency content where the positive emotion tendencies exist are determined.

In some exemplary design ideas, the virtual character dialogue speech of the virtual digital character is divided into formatted interactive data and target interactive data, the emotion tendencies of the formatted interactive data and the target interactive data are greatly different, and the corresponding positive emotion mapping values are also greatly different, so that the target interactive data can be extracted from the virtual character dialogue speech through the positive emotion mapping values of different dialogue interactive sentences.

The positive emotion mapping mean value of each dialogue interactive statement of the virtual character dialogue voice after dialogue emotion encoding can be determined, and further formatted interactive data and target interactive data to be analyzed in the virtual character dialogue voice are determined based on the positive emotion mapping mean value. The positive emotion mapping mean value of the formatted interactive data is smaller than that of the target interactive data.

After determining the corresponding associated dialogue interactive sentences in the front and rear Wen Yugou sets of each dialogue interactive sentence in the target interactive data, determining the positive emotion mapping value corresponding to each associated dialogue interactive sentence, and calculating the positive emotion mapping mean value between the positive emotion mapping values and the positive emotion mapping standard difference value between the positive emotion mapping values. After the positive emotion mapping standard deviation value and the positive emotion mapping mean value are obtained, determining a significance trend value C corresponding to the dialogue interactive statement based on the ratio between the positive emotion mapping standard deviation value and the positive emotion mapping mean value. And comparing the corresponding significance trend value with the set trend value for each dialogue interactive statement so as to obtain a corresponding comparison result. Based on the comparison result, the dialogue interactive sentences in the target interactive data of the virtual digital characters can be clustered, so that the first user positive emotion tendency content is obtained.

Illustratively, business clauses are carried out on the first user positive emotion tendencies content, and a first user positive emotion mapping statement sequence after the clauses is obtained. The first user positive emotion mapping statement sequence after clause is composed of a plurality of user positive emotion mapping statements. For each user positive emotion mapping statement, traversing the first user positive emotion mapping statement sequence in sequence by taking the user positive emotion mapping statement as a starting point, and further determining the associated user positive emotion mapping statement associated with the user positive emotion mapping statement with dialogue behaviors. And if the associated user positive emotion mapping statement exists, connecting the associated user positive emotion mapping statements in sequence until all the user positive emotion mapping statements after the first user positive emotion mapping statement sequence are traversed, and obtaining a plurality of associated positive emotion tendency contents corresponding to the first user positive emotion tendency contents. Wherein the associated positive emotion tendencies content may be one or more.

Step S102: determining a content feature vector corresponding to the first user positive emotion tendencies content, and determining an attention parameter value of the first user positive emotion tendencies content based on the content feature vector, wherein the content feature vector at least comprises a content semantic feature vector and a content context feature vector, and the attention parameter value represents a user dialogue emotion portrait of the virtual digital character.

The content feature vectors include at least a content semantic feature vector and a content context feature vector, and then, based on the content feature vector, an attention parameter value for the first user positive emotion tendencies content is determined. The attention parameter value represents the user dialogue emotion figure of the virtual digital character, and the larger the attention parameter value is, the larger the user dialogue emotion satisfaction degree corresponding to the virtual digital character is.

The attention parameter value and the content characteristic vector value are in a linear relation, the content characteristic vector can be used as a training sample, the attention parameter value is used as a training label, and the Monte Carlo neural network is constructed.

After the Monte Carlo neural network is constructed, the content feature vector of the first user positive emotion tendency content is substituted into the Monte Carlo neural network, so that the attention parameter value of the first user positive emotion tendency content can be determined.

Step S103: based on the emotion attribute of the first user positive emotion tendency content, determining the emotion value corresponding to the first user positive emotion tendency content, and based on the emotion value and the attention parameter value, obtaining a first dialogue attention weight corresponding to the AI dialogue event by the first user positive emotion tendency content.

After the first user positive emotion tendency content is identified, the emotion value corresponding to the current first user positive emotion tendency content can be determined from the preset mapping relation based on the emotion attribute of the first user positive emotion tendency content. And further, the emotion value and the attention parameter value corresponding to the first user positive emotion tendencies content are fused, so that the first dialogue attention weight of the AI dialogue event corresponding to the first user positive emotion tendencies content is obtained. The first dialog attention weight represents an impact weight of the first user positive emotion tendencies content on the virtual digital character.

Step S104: and collecting dialogue evaluation data of the first user positive emotion tendency content corresponding to the AI dialogue event, analyzing the dialogue evaluation data to update the first dialogue attention weight based on the dialogue evaluation data and obtaining an updated second dialogue attention weight.

And analyzing the dialogue evaluation data by collecting dialogue evaluation data of the AI dialogue event corresponding to the first user positive emotion tendency content, and further updating the first dialogue attention weight based on the dialogue evaluation data to obtain an updated second dialogue attention weight.

Illustratively, a set of avatar dialogue voices of the avatar is obtained, and a preceding avatar dialogue voice closest to the time of the avatar dialogue voice is extracted from the set of avatar dialogue voices. And carrying out dialogue emotion analysis on the dialogue voice of the prior virtual character, and determining the second user positive emotion tendency content corresponding to the virtual digital character. After the second user positive emotion tendencies content is obtained, the first user positive emotion tendencies content and the second user positive emotion tendencies content are compared, and positive emotion tendencies conversion parameters between the first user positive emotion tendencies content and the second user positive emotion tendencies content are determined. Wherein the positive emotion tendencies conversion parameter indicates the extent of development of the positive emotion tendencies when the target user is in a conversation with the virtual digital character.

Further, after the positive emotion tendency conversion parameter of the positive emotion tendency content of the first user is obtained, a dialogue evaluation weight value corresponding to the dialogue evaluation data is determined, and an update index value corresponding to the dialogue evaluation data is determined based on a fusion parameter value between the dialogue evaluation weight value and the positive emotion tendency conversion parameter.

The dialogue evaluation data at least comprises scene effect evaluation data, emotion understanding evaluation data and emotion expression evaluation data.

The dialogue evaluation weight value is adapted to dialogue evaluation data, a plurality of scene effect evaluation score ranges and a plurality of emotion understanding evaluation score ranges corresponding to emotion attributes of active emotion tendency contents of a first user are required to be determined, each scene effect evaluation score range corresponds to a first dialogue evaluation weight value, and each emotion understanding evaluation score range corresponds to a second dialogue evaluation weight value, so that the first dialogue evaluation weight value and the second dialogue evaluation weight value corresponding to the active emotion tendency contents of the current first user can be determined based on the acquired scene effect evaluation data and emotion understanding evaluation data. The evaluation knowledge graph of emotion expression evaluation data of the AI dialogue event where the positive emotion tendency content of the first user is located can be generated through the dialogue evaluation data, so that a plurality of significant emotion activity data in the AI dialogue event can be determined in the evaluation knowledge graph.

After determining the plurality of significant emotion activity data, a third dialog evaluation weight value corresponding to the first user positive emotion tend content may be determined based on the plurality of significant emotion activity data. And carrying out unit splitting on the evaluation knowledge graph, and extracting a plurality of target unit evaluation knowledge graphs associated with the significant emotion activity data from the unit evaluation knowledge graphs obtained after unit splitting aiming at each significant emotion activity data. The overall evaluation knowledge graph formed by fusing the plurality of target unit evaluation knowledge graphs contains significant emotion activity data. After the target unit evaluation knowledge graph is extracted, the graph nodes corresponding to the significant emotion activity data in the multiple target unit evaluation knowledge graphs can be determined, and the graph nodes are ordered based on descending order of the intersection number of the graph nodes, so that the target graph nodes of the significant emotion activity data are obtained. The map nodes refer to the nodes of the maximum crossing number of the part of the significant emotion activity data in a certain target unit evaluation knowledge map, and the target map nodes represent the nodes of the maximum crossing number of the significant emotion activity data. After determining the target spectrum node, mapping the whole evaluation knowledge spectrum to a dialogue importance space, and determining dialogue importance parameters of the target spectrum according to the dialogue importance space. At this time, a third dialogue evaluation weight value corresponding to the first user positive emotion tendencies content may be determined based on the dialogue importance parameter.

After the first dialogue evaluation weight value, the second dialogue evaluation weight value and the third dialogue evaluation weight value corresponding to the dialogue evaluation data are respectively determined, the fusion coefficient corresponding to each dialogue evaluation weight value can be determined through a preset mapping relation. The preset mapping relation comprises the mapping relation between each dialogue evaluation data and the corresponding fusion coefficient. And based on the fusion coefficient, fusing the first dialogue evaluation weight value, the second dialogue evaluation weight value and the third dialogue evaluation weight value to obtain the dialogue evaluation weight value corresponding to the dialogue evaluation data.

The emotion expression evaluation data, the scene effect evaluation data and the emotion understanding evaluation data are all affected by the interaction problem labels, so that when the influence of the dialogue evaluation data on the virtual digital character can be measured, the factors of the interaction problem labels are considered, and after the interaction problem labels of the current virtual digital character and the label weight information corresponding to the dialogue evaluation data under the current interaction problem labels are determined, fusion coefficients corresponding to the first dialogue evaluation weight value, the second dialogue evaluation weight value and the third dialogue evaluation weight value are respectively adjusted based on the label weight information. The label weight information corresponding to the first dialogue evaluation weight value and the second dialogue evaluation weight value is respectively in direct proportion to scene effect evaluation data and emotion understanding evaluation data; the label weight information corresponding to the third dialogue evaluation weight value is in direct proportion to the significance value of the significance emotion activity data.

After determining the update index value, updating the first dialogue attention weight based on the update index value, and performing fusion calculation on the update index value and the first dialogue attention weight to obtain an operation result which is the updated second dialogue attention weight. Wherein, the update index value is greater than 1, and the update direction and amplitude, the update index value and the dialogue evaluation weight value are in positive association relation.

Step S105: based on the second dialog attention weight, online interactive user recommendation is made to the virtual digital character in association with the target user presence.

The second dialog attention weight may be used to characterize the degree of matching of the virtual digital character to the target user for dialog interactions, i.e., the greater the second dialog attention weight, the greater the degree of matching of the virtual digital character to the target user for dialog interactions. Thus, the virtual digital persona may be recommended to an online interactive user that is associated with the target user. By way of example, an online interactive user associated with the target user presence may refer to a user representation matching an online interactive user with the target user.

Based on the above steps, in the embodiment of the present application, by analyzing the first user positive emotion tendency content of the virtual digital character, after obtaining the first dialogue attention weight of the AI dialogue event corresponding to the first user positive emotion tendency content, updating the first dialogue attention weight through dialogue evaluation data, so that under the condition of determining the user dialogue emotion portrait of the virtual digital character, the influence of dialogue evaluation data on the user dialogue effect of the virtual digital character can be further combined, thereby executing more accurate online interactive user recommendation on the virtual digital character, and further improving the AI dialogue matching degree of the subsequent online interactive user.

In an exemplary design concept, the step of performing dialogue emotion encoding on the virtual character dialogue speech and determining an active emotion mapping mean value of each dialogue interactive statement of the virtual character dialogue speech after dialogue emotion encoding includes: loading the virtual character dialogue voice to a dialogue emotion analysis network to perform dialogue emotion analysis on the virtual character dialogue voice through the dialogue emotion analysis network, and determining a dialogue emotion analysis result of the virtual character dialogue voice, wherein the dialogue emotion analysis result comprises a dialogue emotion label of each dialogue interaction statement and a corresponding label mapping value; performing positive emotion feature analysis on dialogue emotion analysis results of the virtual character dialogue voices to obtain positive emotion mapping average values of each dialogue interactive statement;

the embodiment of the application provides an artificial intelligence-based virtual character dialogue emotion analysis method, which comprises the following steps.

Step W101, a first number of virtual character dialogue supervised samples and a second number of virtual character dialogue unsupervised samples are obtained, and the first number of virtual character dialogue supervised samples and the second number of virtual character dialogue unsupervised samples are loaded to a dialogue emotion analysis network; the second number and the first number are both positive integers; the first number of virtual character dialogue supervision samples respectively carry dialogue emotion annotation information of the contained virtual character dialogue voices; the virtual character dialogue voices carried by the first number of virtual character dialogue supervision samples and the virtual character dialogue voices carried by the second number of virtual character dialogue non-supervision samples belong to the same virtual digital character scene;

The online virtual digital system can obtain a first number of virtual character dialogue supervision samples and a second number of virtual character dialogue non-supervision samples, and the specific values of the second number and the first number can be flexibly determined based on actual requirements.

Wherein the avatar dialogue supervised sample and the avatar dialogue unsupervised sample comprise avatar dialogue voices of the same avatar scene.

Thus, the avatar dialogue supervision sample may be a dialogue sample to which dialogue emotion label information of the included avatar dialogue speech is added, and one avatar dialogue supervision sample may include one avatar dialogue sample, and the dialogue emotion label information carried by one avatar dialogue supervision sample indicates actual dialogue emotion information of the avatar dialogue speech included in the avatar dialogue supervision sample.

The virtual character conversation non-supervision sample may be any collected conversation interactive voice, and in general, conversation emotion information of virtual character conversation voice contained in the virtual character conversation non-supervision sample is different from conversation emotion information of virtual character conversation voice contained in the virtual character conversation supervision sample, wherein preferably conversation emotion information of virtual character conversation voice contained in the virtual character conversation non-supervision sample is completely different from conversation emotion information of virtual character conversation voice contained in the virtual character conversation supervision sample. The virtual character dialogue unsupervised sample is a dialogue sample to which dialogue emotion prior information is not added.

Therefore, the online virtual digitizing system can load the obtained first number of virtual character dialogue supervised samples and second number of virtual character dialogue unsupervised samples to the dialogue emotion analysis network, and adjust the network weight configuration information of the dialogue emotion analysis network to obtain a target dialogue emotion analysis network, wherein the target dialogue emotion analysis network is used for dialogue emotion analysis of virtual character dialogue voices of the virtual digital character scene, and please refer to the following description.

The network weight configuration information adjustment may have two branches, including a first branch network weight configuration information adjustment and a second branch network weight configuration information adjustment. The above-mentioned dialog emotion analysis network may be obtained by performing network weight configuration information adjustment on the basic dialog emotion analysis network through a virtual character dialog supervision sample, and the process of performing network weight configuration information adjustment on the basic dialog emotion analysis network through the virtual character dialog supervision sample to obtain the dialog emotion analysis network may be referred to as a process of adjusting network weight configuration information of the first branch. The process of performing network weight configuration information adjustment on the dialog emotion analysis network through the virtual character dialog supervision sample and the virtual character dialog unsupervised sample together to obtain the target dialog emotion analysis network may be referred to as a second-branch network weight configuration information adjustment process, which is specifically described in the embodiments of the present application.

Here, a process of training the basic dialog emotion analysis network to obtain the dialog emotion analysis network is described:

the online virtual digitizing system may load the first number of virtual character dialogue supervised samples to a basic dialogue emotion analysis network, and further generate dialogue semantic representation information of each virtual character dialogue supervised sample through the basic dialogue emotion analysis network, and further the basic dialogue emotion analysis network may respectively determine dialogue emotion analysis information of virtual character dialogue voices included in each virtual character dialogue supervised sample based on the generated dialogue semantic representation information of each virtual character dialogue supervised sample, and may refer to the dialogue emotion analysis information as second dialogue emotion analysis information.

The first number of virtual character dialogue supervision samples can carry s dialogue emotion priori information, one dialogue emotion priori information corresponds to one dialogue emotion annotation information, and s is a positive integer not greater than the second number.

The basic dialog emotion analysis network can identify the mapping confidence of each virtual character dialog supervision sample in the loaded first number of virtual character dialog supervision samples for each dialog emotion annotation information through the FCL unit.

Therefore, the second dialog emotion analysis information may be a mapping confidence level that virtual character dialog voices in a virtual character dialog supervision sample obtained by the basic dialog emotion analysis network decision are each of s dialog emotion label information, and a determined mapping confidence level corresponds to one dialog emotion label information between one virtual character dialog supervision sample.

Furthermore, the online virtual digitizing system can calculate and obtain a second learning effect index of the basic dialogue emotion analysis network through the second dialogue emotion analysis information obtained by decision and the dialogue emotion marking information carried by each virtual character dialogue supervision sample, wherein the second learning effect index characterizes the difference between the dialogue emotion analysis result (such as the second dialogue emotion analysis information) of the basic dialogue emotion analysis network and the actual dialogue emotion information (such as the dialogue emotion indicated by the dialogue emotion marking information carried by the virtual character dialogue supervision sample). Therefore, the weight configuration information of the basic dialogue emotion analysis network can be adjusted through the second learning effect index, and the adjustment direction is the best learning effect index.

Thus, the dialogue emotion analysis network can be obtained through basic dialogue emotion analysis network training, and the online virtual digitizing system can load the obtained first number of virtual character dialogue supervised samples and second number of virtual character dialogue unsupervised samples to the dialogue emotion analysis network.

Step W102, determining first dialogue emotion analysis information of virtual character dialogue voices contained in each virtual character dialogue supervision sample in a dialogue emotion analysis network, and acquiring an associated virtual character dialogue voice sample of each virtual character dialogue supervision sample from a dialogue interaction knowledge resource pool; the dialogue interactive knowledge resource pool comprises a first number of virtual character dialogue supervision samples and a second number of virtual character dialogue unsupervised samples; the dialogue voice sample of the associated virtual character of each virtual character dialogue supervision sample does not carry dialogue emotion annotation information carried by the virtual character dialogue supervision sample;

the online virtual digitizing system may determine, in the dialog emotion analysis network, first dialog emotion analysis information for the avatar dialog voices contained in each avatar dialog supervision sample in the first number of avatar dialog supervision samples. Similarly, if the first number of virtual character dialogue supervision samples carries s dialogue emotion priori information, that is, s dialogue emotion labeling information is indicated, the dialogue emotion analysis network can also identify and obtain mapping confidence that virtual character dialogue voices in the virtual character dialogue supervision samples are each dialogue emotion labeling information. Therefore, the first dialogue emotion analysis information is the mapping confidence degree of each dialogue emotion marking information of the virtual character dialogue voice in the virtual character dialogue supervision sample which is decided by the dialogue emotion analysis network.

For example, the dialog emotion analysis network may generate dialog semantic characterization information for each virtual character dialog supervision sample, that is, dialog sample features of the virtual character dialog supervision sample extracted by the dialog emotion analysis network. The dialogue emotion analysis network can also perform regularized conversion on dialogue semantic representation information of each virtual character dialogue supervision sample, namely, the dialogue semantic representation information of the virtual character dialogue supervision sample can be regularized converted into a specific feature labeling interval, the dialogue semantic representation information of each virtual character dialogue supervision sample after regularized conversion is generated, and the dialogue semantic representation information of each virtual character dialogue supervision sample after regularized conversion can be called as regularized conversion representation information. Furthermore, the dialogue emotion analysis network can obtain the mapping confidence degree of the virtual character dialogue voice in each virtual character dialogue supervision sample as each dialogue emotion marking information through the generated regularized conversion characterization information of each virtual character dialogue supervision sample, namely, the first dialogue emotion analysis information of each virtual character dialogue supervision sample is obtained through the decision.

The weight configuration information of the dialog emotion analysis network may further include weight configuration information of the FCL unit, where the weight configuration information is an array, may be recorded as a weight configuration information array w, and in each training of the dialog emotion analysis network, the weight configuration information array w may also be subjected to regularized conversion, and the regularized converted weight configuration information array w may be referred to as conversion weight configuration information, and then the dialog emotion analysis network may determine the first dialog emotion analysis information of each virtual character dialog supervision sample through the conversion weight configuration information.

The weight configuration information of the basic dialogue emotion analysis network also comprises weight configuration information of the FCL unit, and the basic dialogue emotion analysis network can also make decisions through dialogue semantic characterization information after regularized conversion of each virtual character dialogue supervision sample and weight configuration information after regularized conversion of the FCL unit when deciding the second dialogue emotion analysis information of each virtual character dialogue supervision sample each time.

Therefore, the dialogue emotion analysis network can determine the mapping confidence degree of the virtual character dialogue voice in each virtual character dialogue supervision sample for each dialogue emotion marking information through the regularized conversion characterization information of each virtual character dialogue supervision sample and the weight configuration information after the regularized conversion of the FCL unit. Therefore, the mapping confidence degree of the virtual character dialogue voice in the virtual character dialogue supervision sample i for each dialogue emotion marking information can be decided, namely, the first dialogue emotion analysis information of the virtual character dialogue supervision sample is decided.

In some embodiments, the specific scheme of the second dialogue emotion analysis information of the basic dialogue emotion analysis network decision virtual character dialogue supervision sample is the same as the specific scheme of the first dialogue emotion analysis information of the dialogue emotion analysis network decision virtual character dialogue supervision sample, and is different from the regularized conversion characterization information and the conversion weight configuration information of the substitution operation.

The second java dialogue emotion analysis network may also generate dialogue semantic representation information of each virtual character dialogue unsupervised sample in the second number of virtual character dialogue unsupervised samples, and perform regularized conversion on the dialogue semantic representation information of each virtual character dialogue unsupervised sample, and may also obtain regularized conversion representation information of each virtual character dialogue unsupervised sample.

The sequence formed by the first number of virtual character dialogue supervised samples and the second number of virtual character dialogue unsupervised samples can be called a dialogue interaction knowledge resource pool, and the dialogue interaction knowledge resource pool comprises the first number of virtual character dialogue supervised samples and the second number of virtual character dialogue unsupervised samples. And the dialogue emotion analysis network can generate a sample association value between the virtual character dialogue supervision sample and each dialogue interaction voice in the dialogue interaction knowledge resource pool through regularized conversion representation information of each virtual character dialogue supervision sample and dialogue semantic representation information of each virtual character dialogue non-supervision sample. And further, through a sample association value between the virtual character dialogue supervision sample and each dialogue interaction voice in the dialogue interaction knowledge resource pool, an associated virtual character dialogue voice sample of each virtual character dialogue supervision sample can be obtained from the dialogue interaction knowledge resource pool, and one virtual character dialogue supervision sample can have one or more associated virtual character dialogue voice samples.

The virtual character dialogue supervision sample and the associated virtual character dialogue voice sample do not carry the same dialogue emotion marking information, that is, the virtual character dialogue voice in the virtual character dialogue supervision sample and the virtual character dialogue voice in the associated virtual character dialogue voice sample of the virtual character dialogue supervision sample need the dialogue emotion information of the virtual character dialogue voice which belongs to different types.

The dialog interactive knowledge resource pool may not include the first number of avatar dialog supervision samples, but only include the second number of avatar dialog unsupervised samples. The first number of virtual character dialogue supervised samples are added into the dialogue interactive knowledge resource pool to participate in the acquisition process of the associated virtual character dialogue voice samples of each virtual character dialogue supervised sample, so that training data can be expanded, the training data can be conveniently acquired from all virtual character dialogue supervised samples and virtual character dialogue unsupervised samples when the associated virtual character dialogue voice samples of each virtual character dialogue supervised sample are acquired, and the selection range of the associated virtual character dialogue voice samples of each virtual character dialogue supervised sample is expanded.

Step W103, determining a first learning effect index based on a sample association value between each virtual character dialogue supervision sample and an associated virtual character dialogue voice sample, and determining a second learning effect index based on first dialogue emotion analysis information corresponding to each virtual character dialogue supervision sample and carried dialogue emotion marking information;

the online virtual digitization system may calculate a first learning effect index by a sample association value (which may be referred to as a sample association value) between each avatar dialogue supervision sample and its associated avatar dialogue speech sample. The first learning effect index characterizes a difference in dialogue emotion information of the virtual character dialogue speech for the virtual character dialogue speech in each virtual character dialogue supervision sample by the dialogue emotion analysis network.

The online virtual digitizing system can calculate and obtain a second learning effect index of the dialogue emotion analysis network through the first dialogue emotion analysis information of each virtual character dialogue supervision sample and the dialogue emotion marking information carried by each virtual character dialogue supervision sample, wherein the second learning effect index characterizes the difference between the dialogue emotion analysis information (such as the first dialogue emotion analysis information) identified by the dialogue emotion analysis network and the dialogue emotion information of the virtual character dialogue speech indicated by the dialogue emotion marking information carried by the virtual character dialogue supervision sample.

The online virtual digitizing system can adjust the weight configuration information of the dialogue emotion analysis network through the second learning effect index and the first learning effect index of the dialogue emotion analysis network together so as to obtain the target dialogue emotion analysis network, and please refer to the following description.

Step W104, adjusting weight configuration information of the dialogue emotion analysis network according to the first learning effect index and the second learning effect index to generate a target dialogue emotion analysis network; the target dialogue emotion analysis network is used for performing dialogue emotion analysis on virtual character dialogue voices belonging to the virtual digital character scene;

the online virtual digitizing system may perform a weighted calculation on the obtained first learning effect index and the obtained second learning effect index, and may refer to the weighted calculation value as a final learning effect index. The online virtual digitizing system can adjust the weight configuration information of the dialogue emotion analysis network through the final learning effect index, namely, adjust the weight configuration information of the dialogue emotion analysis network, so that the final learning effect index is maximized.

Further embodiments of the training steps of the above dialog emotion analysis network are provided below.

Step W201, based on the dialogue semantic characterization information of each virtual character dialogue supervision sample and the dialogue semantic characterization information of each virtual character dialogue unsupervised sample, generating a dialogue semantic association array;

after obtaining the dialogue semantic representation information of each virtual character dialogue supervised sample and each virtual character dialogue unsupervised sample in the first number of virtual character dialogue supervised samples, the dialogue emotion analysis network can also obtain the regularized conversion representation information of each virtual character dialogue supervised sample after normalization and the regularized conversion representation information of each virtual character dialogue unsupervised sample after regularized conversion. The specific scheme for obtaining the regularized conversion characterization information of the virtual character dialogue non-supervision sample based on the dialogue semantic characterization information of the virtual character dialogue non-supervision sample is the same as the specific scheme for obtaining the regularized conversion characterization information of the virtual character dialogue supervision sample based on the dialogue semantic characterization information of the virtual character dialogue supervision sample.

The method for obtaining regularized conversion characterization information of the virtual character dialogue supervised samples by the dialogue emotion analysis network is that a characterization knowledge extraction vector array of a first number of virtual character dialogue supervised samples can be generated, the characterization knowledge extraction vector array of the first number of virtual character dialogue supervised samples can be called as a first characterization knowledge extraction vector array, the first characterization knowledge extraction vector array comprises regularized conversion characterization information of each virtual character dialogue supervised sample, and one row in the first characterization knowledge extraction vector array is the regularized conversion characterization information of one virtual character dialogue supervised sample. Similarly, the conversation emotion analysis network may obtain regularized conversion characterization information of the virtual character conversation unsupervised samples by generating a characterization knowledge extraction vector array of a second number of virtual character conversation unsupervised samples, and the characterization knowledge extraction vector array of the second number of virtual character conversation unsupervised samples may be referred to as a second characterization knowledge extraction vector array, where the second characterization knowledge extraction vector array includes regularized conversion characterization information of each virtual character conversation unsupervised sample, and one line in the second characterization knowledge extraction vector array is the regularized conversion characterization information of one virtual character conversation unsupervised sample.

Therefore, the online virtual digitizing system can fuse the first characterization knowledge extraction vector array and the second characterization knowledge extraction vector array to obtain the target characterization knowledge extraction vector array, wherein the target characterization knowledge extraction vector array comprises the first characterization knowledge extraction vector array and the second characterization knowledge extraction vector array.

The online virtual digitizing system may also obtain an attached front-to-back Wen Yuyi vector array of the target representation knowledge extraction vector array, where the dimension of the attached front-to-back Wen Yuyi vector array is d (second number+first number). The online virtual digitizing system may obtain the fusion information between the first token knowledge extraction vector array and the Wen Yuyi vector arrays before and after the attachment of the target token knowledge extraction vector array, where the fusion information is also an array, and the array may be referred to as a dialogue semantic association array, where the dimension of the dialogue semantic association array is a second number (second number+first number), which represents a second number of rows and a second number+first number of columns. One row in the dialogue semantic association array may correspond to one virtual character dialogue supervise sample, and each element in the one row is a sample association value between the corresponding virtual character dialogue supervise sample and each dialogue interaction voice in the second number+the second number of dialogue interaction voices included in the dialogue interaction knowledge resource pool.

Step W202, determining an associated virtual character dialogue voice sample of each virtual character dialogue supervision sample from a dialogue interaction knowledge resource pool based on the dialogue semantic association array;

in the present application, the specific scheme for obtaining the related virtual character dialogue voice sample of each virtual character dialogue supervision sample from the dialogue interaction knowledge resource pool based on the dialogue semantic association array is the same, and here, the related virtual character dialogue voice sample of the virtual character dialogue supervision sample i is taken as an example for illustration.

Optionally, since the dialogue emotion information of the virtual character dialogue voices included in the virtual character dialogue unsupervised sample of the second number of virtual character dialogue voices is generally different from the dialogue emotion information of the virtual character dialogue voices included in the virtual character dialogue supervised sample of the first number of virtual character dialogue voices, the dialogue sample carrying the same dialogue emotion priori information as the virtual character dialogue supervised sample i can be cleaned out from the dialogue interaction knowledge resource pool to generate a candidate dialogue interaction knowledge resource pool, so that the candidate dialogue interaction knowledge resource pool can be considered to include the dialogue samples which do not carry the dialogue emotion marking information carried by the virtual character dialogue supervised sample i in the dialogue interaction knowledge resource pool.

Assuming that k associated virtual character dialogue voice samples of each virtual character dialogue supervision sample need to be acquired, k is a positive integer less than or equal to the second number+the first number, k may be referred to as the number of associated virtual character dialogue voice samples, and the specific value of k may be determined based on the actual application scenario. Therefore, the online virtual digitizing system can acquire sample association values between the virtual character dialogue supervision samples i and each dialogue interaction voice in the candidate dialogue interaction knowledge resource pool from the dialogue semantic association array, and can take k dialogue interaction voices with the largest sample association values between the candidate dialogue interaction knowledge resource pool and the virtual character dialogue supervision samples i as associated virtual character dialogue voice samples of the virtual character dialogue supervision samples i.

For example, each dialogue interactive voice in the candidate dialogue interactive knowledge resource pool may be ranked according to descending order ranking information based on a sample association value between each dialogue interactive voice in the candidate dialogue interactive knowledge resource pool and the virtual character dialogue supervision sample i, and a dialogue interactive knowledge resource pool obtained by ranking the dialogue samples in the candidate dialogue interactive knowledge resource pool may be referred to as a candidate dialogue interactive knowledge resource pool. Thus, the online virtual digitizing system may take the top k dialogue interactive voices in the candidate dialogue interactive knowledge resource pool as the associated avatar dialogue voice samples of the avatar dialogue supervision samples i.

In the second number of virtual character dialogue unsupervised samples, if the dialogue emotion information of the virtual character dialogue voice included in the first number of virtual character dialogue supervised samples is the same as the dialogue emotion information of the virtual character dialogue voice included in the first number of virtual character dialogue supervised samples, then the candidate dialogue interaction knowledge resource pool can be continuously cleaned, that is, t dialogue interaction voices with the maximum sample correlation value between the candidate dialogue interaction knowledge resource pool and the virtual character dialogue supervisory samples i are likely to be the same as the dialogue emotion information of the virtual character dialogue voice included in the virtual character supervisory samples i, so as to ensure that the dialogue emotion information of the virtual character dialogue voice included in the obtained virtual character dialogue supervisory samples i is as different from the dialogue emotion information of the virtual character dialogue voice included in the virtual character dialogue supervisory samples i.

Therefore, k dialogue interactive voices with the largest sample association value with the virtual character dialogue supervision sample i in the candidate dialogue interactive knowledge resource pool for cleaning t dialogue interactive voices can be used as the associated virtual character dialogue voice samples of the virtual character dialogue supervision sample i. In this way, the case where the dialogue emotion information of the avatar dialogue voice included in the associated avatar dialogue voice sample of the obtained avatar dialogue supervision sample i is the same as the dialogue emotion information of the avatar dialogue voice included in the avatar dialogue supervision sample i can be well avoided.

The online virtual digitization system may obtain the associated avatar dialogue voice sample for each avatar dialogue supervisory sample through the same specific scheme as the above-described associated avatar dialogue voice sample for obtaining avatar dialogue supervisory sample i.

Step W203, calculating a first learning effect index based on a sample association value between each virtual character dialogue supervision sample and an associated virtual character dialogue speech sample to which the virtual character dialogue supervision sample belongs;

in this application, since one avatar dialogue supervision sample may have a plurality of associated avatar dialogue voice samples, the online virtual digitizing system may further obtain an average value of sample association values between each avatar dialogue supervision sample and a plurality of associated avatar dialogue voice samples thereof, and the average value may be referred to as an average sample association value. For example, if the associated virtual character speech samples of the virtual character speech monitor sample i include the associated virtual character speech sample 1, the associated virtual character speech sample 2, and the associated virtual character speech sample 3, and the sample association value between the virtual character speech monitor sample i and its associated virtual character speech sample 1 is 0.2, the sample association value between the virtual character speech monitor sample i and its associated virtual character speech sample 2 is 0.4, and the sample association value between the virtual character speech monitor sample i and its associated virtual character speech sample 3 is 0.6, the average sample association value corresponding to the virtual character speech monitor sample i is (0.2+0.4+0.6)/3, that is, equal to 0.4.

Further embodiments are provided below, including the following steps.

Step W301, a first number of virtual character dialogue supervised samples and a second number of virtual character dialogue unsupervised samples are obtained; the second number and the first number are both positive integers; the first number of virtual character dialogue supervision samples respectively carry dialogue emotion annotation information of the contained virtual character dialogue voices; the virtual character dialogue voices carried by the first number of virtual character dialogue supervision samples and the virtual character dialogue voices carried by the second number of virtual character dialogue non-supervision samples belong to the same virtual digital character scene;

the online virtual digitizing system can obtain a first number of virtual character dialogue supervised samples and a second number of virtual character dialogue unsupervised samples, wherein the first number of virtual character dialogue supervised samples carry dialogue emotion marking information of contained virtual character dialogue voices, the dialogue emotion priori information indicates dialogue emotion information of the virtual character dialogue voices, the second number of virtual character dialogue unsupervised samples do not carry dialogue emotion marking information of contained virtual character dialogue voices, and the first number of virtual character dialogue supervised samples and the second number of virtual character dialogue unsupervised samples are used together as sample dialogue interactive voices for network weight configuration information adjustment

Step W302, loading a first number of virtual character dialogue supervision samples into a basic dialogue emotion analysis network, determining second dialogue emotion analysis information of virtual character dialogue voices contained in each virtual character dialogue supervision sample in the basic dialogue emotion analysis network, and adjusting weight configuration information of the basic dialogue emotion analysis network based on the second dialogue emotion analysis information corresponding to each virtual character dialogue supervision sample and dialogue emotion marking information carried by each virtual character dialogue supervision sample to generate a dialogue emotion analysis network;

step W303, loading a first number of virtual character dialogue supervised samples and a second number of virtual character dialogue unsupervised samples into a dialogue emotion analysis network, determining first dialogue emotion analysis information of virtual character dialogue voices contained in each virtual character dialogue supervised sample in the dialogue emotion analysis network, and acquiring associated virtual character dialogue voice samples of each virtual character dialogue supervised sample from a dialogue interaction knowledge resource pool; the dialogue interactive knowledge resource pool comprises a first number of virtual character dialogue supervision samples and a second number of virtual character dialogue unsupervised samples; the dialogue voice sample of the associated virtual character of each virtual character dialogue supervision sample does not carry dialogue emotion annotation information carried by the virtual character dialogue supervision sample;

Step W304, determining a first learning effect index based on a sample association value between each virtual character dialogue supervision sample and an associated virtual character dialogue voice sample, and determining a second learning effect index based on first dialogue emotion analysis information corresponding to each virtual character dialogue supervision sample and carried dialogue emotion marking information;

step W305, adjusting weight configuration information of the dialogue emotion analysis network according to the first learning effect index and the second learning effect index to generate a target dialogue emotion analysis network; the target dialogue emotion analysis network is used for performing dialogue emotion analysis on virtual character dialogue voices belonging to the virtual digital character scene;

the online virtual digitizing system can obtain a first learning effect index through a sample association value between the virtual character dialogue supervision sample and the associated virtual character dialogue voice sample, and obtain a second learning effect index through first dialogue emotion analysis information corresponding to the virtual character dialogue supervision sample and carried dialogue emotion marking information. Furthermore, the online virtual digitizing system can adjust the weight configuration information of the dialogue emotion analysis network through the first learning effect index and the second learning effect index in the second stage of network weight configuration information adjustment process so as to train the dialogue emotion analysis network to obtain the target dialogue emotion analysis network.

In some design considerations, an online virtual digitizing system, which may be a server, is provided that includes a processor, a memory, an Input/Output interface (I/O) and a communication interface. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface is connected to the system bus through the input/output interface. Wherein the processor of the online virtual digitizing system is operable to provide computing and control capabilities. The memory of the online virtual digital system includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the online virtual digitizing system is used for storing the data related to the method. The model-loaded data/output interface of the online virtual digitizing system is used for exchanging information between the processor and the external device. The communication interface of the online virtual digital system is used for communicating with an external terminal through network connection. The computer program, when executed by a processor, implements an AI dialog method based on a virtual digital character.

In some design considerations, an online virtual digitizing system is provided, which may be a terminal. The online virtual digitizing system includes a processor, a memory, an input/output interface, a communication interface, a display unit, and an input device. The processor, the memory and the input/output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input/output interface. Wherein the processor of the online virtual digitizing system is operable to provide computing and control capabilities. The memory of the online virtual digital system comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The model-loaded data/output interface of the online virtual digitizing system is used for exchanging information between the processor and the external device. The communication interface of the online virtual digital system is used for carrying out wired or wireless communication with an external terminal, and the wireless mode can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program, when executed by a processor, implements an AI dialog method based on a virtual digital character. The display unit of the online virtual digitizing system is used for forming a visually viewable picture.

In some design considerations, an online virtual digitizing system is provided, including a memory and a processor, where the memory stores a computer program, and the processor implements the steps of the method embodiments described above when executing the computer program.

In some design considerations, a computer readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method embodiments described above.

In some design considerations, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of the method embodiments described above.

The above examples only represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the present application. It should be noted that it would be apparent to those skilled in the art that various modifications and improvements could be made without departing from the spirit of the present application, which would be within the scope of the present application. Accordingly, the scope of protection of the present application shall be subject to the appended claims.

Claims

1. An AI conversation method based on virtual digital characters, applied to an online virtual digitizing system, the method comprising:

Acquiring virtual character dialogue voices of a target user in the online virtual digital scene for a plurality of AI dialogue events of an AI dialogue initiated by a virtual digital character, and performing dialogue emotion analysis on the virtual character dialogue voices to acquire first user positive emotion tendency content of the virtual digital character;

2. The AI dialog method based on a virtual digital character of claim 1, wherein analyzing the dialog rating data to update the first dialog focus weight based on the dialog rating data, obtaining an updated second dialog focus weight, comprises:

3. The AI dialogue method based on a virtual digital character according to claim 2, wherein the dialogue evaluation data includes at least scene effect evaluation data, emotion understanding evaluation data, and emotion expression evaluation data, and the step of determining a dialogue evaluation weight value corresponding to the dialogue evaluation data includes:

4. The AI conversation method based on a virtual digital character of claim 1 wherein the step of performing conversation emotion analysis on the virtual character conversation voice to obtain first user positive emotion tendencies content of the virtual digital character comprises:

5. The virtual digital character-based AI conversation method of claim 4 wherein after determining the first user positive emotion tendencies content of the virtual digital character, the method further comprises:

6. The AI dialog method of claim 3 wherein prior to determining the respective fusion coefficients for the first dialog evaluation weight value, the second dialog evaluation weight value, and the third dialog evaluation weight value, the method further comprises:

7. The virtual digital character-based AI conversation method of claim 1 wherein the step of determining an attention parameter value for the first user positive emotion tendencies content based on the content feature vector comprises:

8. The method of claim 4, wherein the step of encoding dialog emotion for the virtual character dialog voice and determining a positive emotion mapping mean value for each dialog interaction sentence for the dialog emotion encoded virtual character dialog voice comprises:

acquiring a Wen Yuyi vector array before and after attachment of the target characterization knowledge extraction vector array, and determining fusion information of the first characterization knowledge extraction vector array and the Wen Yuyi vector array before and after attachment as the dialogue semantic association array;

acquiring sample association values between each virtual character dialogue supervision sample and dialogue interaction voices in the dialogue interaction knowledge resource pool from the dialogue semantic association array;

the first number of virtual character conversation monitoring samples includes virtual character conversation monitoring samples i, i being a positive integer no greater than the second number.

9. A computer program product comprising a computer program or computer executable instructions which, when executed by a processor, implement the virtual digital character based AI dialog method of any of claims 1-8.

10. An online virtual digitization system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the virtual digital character-based AI conversation method of any of claims 1-8.