CN-121981146-A - Man-machine interaction method and related device, electronic equipment and storage medium

CN121981146ACN 121981146 ACN121981146 ACN 121981146ACN-121981146-A

Abstract

The application discloses a man-machine interaction method, a related device, electronic equipment and a storage medium, wherein the man-machine interaction method comprises the steps of carrying out data retrieval based on current input of a target object to obtain historical memory related to the current input in historical interaction between the target object and an virtual image, carrying out intention abstraction based on the current input, the historical memory and the current relation state between the target object and the virtual image to obtain interaction intention, carrying out desensitization extraction based on the historical memory to obtain a historical abstract, and sending the interaction intention and the historical abstract to cloud equipment, wherein the cloud equipment generates a reply draft based on the interaction intention and the historical abstract, and carrying out style migration based on the reply draft from the cloud equipment to obtain personalized reply of the virtual image. According to the scheme, personalized man-machine interaction can be realized on the premise of protecting data privacy.

Inventors

ZHANG LINFANG
LONG MINGKANG
LIU KUN
JIAN SHENGQI
WU JIANGZHAO

Assignees

合肥智能语音创新发展有限公司

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (15)

1. The man-machine interaction method is characterized by being applied to local equipment and comprising the following steps of: Performing data retrieval based on the current input of a target object to obtain a history memory related to the current input in history interaction between the target object and the avatar; Performing intention abstraction based on the current input, the history memory and the current relation state between the target object and the virtual image to obtain interaction intention, and performing desensitization extraction based on the history memory to obtain a history abstract; The interaction intention and the history abstract are sent to cloud equipment, wherein the cloud equipment generates a reply draft based on the interaction intention and the history abstract; And performing style migration based on the reply draft from the cloud device to obtain personalized reply of the virtual image.
2. The method of claim 1, wherein the method further comprises, prior to the retrieving data based on the current input of the target object to obtain a history memory associated with the current input in the history interactions between the target object and the avatar: Detecting the current quantized values of all psychological parameters based on the virtual image to obtain detection results, wherein the detection results comprise whether psychological parameters in a starvation state exist or not; Responding to the detection result to represent that psychological parameters in a starvation state do not exist, and keeping a passive response mode to wait for the input of the target object; And responding to the detection result to represent that psychological parameters in a deficient state exist, selecting the psychological parameters in the deficient state as target parameters, generating a complement task corresponding to the requirement of the target parameters based on the target parameters, and generating a target message of the avatar for actively interacting with the target object based on the current quantized values of the psychological parameters and the complement task.
3. The method according to claim 2, wherein the detecting is performed on the basis of the current quantized values of the respective psychological parameters, respectively, and before the detecting results are obtained, the method further comprises: Acquiring first quantized values of the virtual images at the psychological parameters at the moment previous to the current moment, and acquiring external events related to interaction of the virtual images; respectively attenuating the first quantized values of each psychological parameter based on the time interval between the current moment and the previous moment to obtain second quantized values of each psychological parameter; adjusting a second quantized value of a psychological parameter related to the external event based on the event type of the external event to obtain a third quantized value corresponding to the psychological parameter; The current quantized value of the psychological parameter irrelevant to the external event is a second quantized value corresponding to the psychological parameter, and the current quantized value of the psychological parameter relevant to the external event is a third quantized value corresponding to the psychological parameter.
4. The method of claim 2, wherein the generating the avatar for actively interacting with the target object based on the current quantized values of the respective psychological parameters and the complement tasks comprises: weighting based on the current quantized values of the psychological parameters to obtain the current relation state between the target object and the virtual image; and generating a target message for completing the complement task based on the relation state and personality parameters of the avatar.
5. The method of claim 1, wherein the retrieving data based on the current input of the target object to obtain a history memory associated with the current input in the history interactions between the target object and the avatar comprises: Inquiring a vector database based on the semantic features of the current input to obtain a historical event matched with the current input as a target event, wherein the historical database contains feature vectors of each historical event in the historical interaction; Taking a target node representing the target event in an interaction memory map as an initial point, and performing multi-jump traversal in the interaction memory map to obtain an event causal chain of the target event, wherein the interaction memory map comprises a plurality of nodes and connecting edges among different nodes, and the interaction memory map is constructed based on the history interaction; and carrying out format conversion based on the event cause and effect chain to obtain a history memory described in natural language.
6. The method of claim 5, wherein the type of connection edge comprises at least one of cause, occur, similar to, participate in; and/or the plurality of nodes comprises at least one of event nodes, emotion nodes and entity nodes; and/or the vector database and the interaction memory map are stored in the local equipment; and/or after the current interaction is finished, updating the vector database and the interaction memory map based on the current interaction.
7. The method of claim 1, wherein the personalized reply is obtained by performing a style migration on the reply draft by a style migration model running on the local device, and the step of obtaining the training dataset of the style migration model comprises: Extracting and obtaining the current adjustment amplitude of the target object in each preference dimension based on the behavior data of the target object on the interaction with the virtual image; Incrementally updating the current value of the preference dimension based on the current adjustment amplitude of the target object in the preference dimension to obtain a new current value of the target object in the preference dimension; Determining a scaling factor between corpora of different styles based on the new current values of the target object in each preference dimension; And carrying out data screening based on the proportionality coefficients among the corpora of different styles to obtain the training data set of the style migration model for the fine tuning.
8. The method of claim 7, wherein each of the preference dimensions includes at least one of intimate acceptance, formal preference, redundancy preference, humor style preference, proactive acceptance, conflict tolerance; And/or the behavior data comprises at least one of interaction behavior of the target object on the personalized reply, editing record of actively modifying reply content of the target object, personality parameters set by the target object for the avatar, duration of dialogue between the target object and the avatar, emotion polarity when the target object replies, and reuse frequency of vocabulary expression of the target object.
9. The method of any of claims 1-8, wherein performing the intent abstraction also results in a desired emotion to the avatar, a privacy level of a current conversation, and the desired emotion, privacy level, and interaction intent are sent as middle tier instructions with the history summary to the cloud device for the cloud device to generate the reply draft based on the middle tier instructions and the history summary; And/or the cloud device runs a large language model for generating the reply draft.
10. The man-machine interaction method is characterized by being applied to cloud equipment and comprising the following steps of: Receiving an interaction intention and a history abstract, wherein the interaction intention and the history abstract are sent by local equipment of a target object, the local equipment performs data retrieval based on current input of the target object to obtain a history memory related to the current input in history interaction between the target object and the avatar, performs intention abstraction based on the current input, the history memory and the current relation state between the target object and the avatar to obtain the interaction intention, and performs desensitization extraction based on the history memory to obtain the history abstract; generating a reply draft based on the interaction intention and the history abstract; and sending the reply draft to the local equipment, wherein the local equipment carries out style migration based on the reply draft to obtain personalized reply of the virtual image.
11. A human-machine interaction device, characterized by being applied to a local apparatus, comprising: the data retrieval module is used for retrieving data based on the current input of the target object to obtain a history memory related to the current input in the history interaction between the target object and the virtual image; the abstract extraction module is used for carrying out intention abstraction based on the current input, the history memory and the current relation state between the target object and the virtual image to obtain interaction intention, and carrying out desensitization extraction based on the history memory to obtain a history abstract; The information sending module is used for sending the interaction intention and the history abstract to cloud equipment, wherein the cloud equipment generates a reply draft based on the interaction intention and the history abstract; And the style migration module is used for performing style migration based on the reply draft from the cloud device to obtain personalized replies of the virtual image.
12. A human-computer interaction device, characterized in that, be applied to high in the clouds equipment, include: the information receiving module is used for receiving interaction intention and history abstract, wherein the interaction intention and the history abstract are sent by local equipment of a target object, the local equipment performs data retrieval based on current input of the target object to obtain history memory related to the current input in history interaction between the target object and the virtual image, performs intention abstraction based on the current input, the history memory and the current relation state between the target object and the virtual image to obtain interaction intention, and performs desensitization extraction based on the history memory to obtain history abstract; the reply generation module is used for generating a reply draft based on the interaction intention and the history abstract; And the information sending module is used for sending the reply draft to the local equipment, wherein the local equipment carries out style migration based on the reply draft to obtain personalized reply of the virtual image.
13. An electronic device comprising at least a memory and a processor coupled to each other, the memory having at least program instructions stored therein, the processor being configured to execute the program instructions to implement the human-machine interaction method of any one of claims 1-10.
14. A man-machine interaction system, comprising a local device and a cloud device, wherein the local device and the cloud device are in communication connection with each other, the local device is configured to implement the man-machine interaction method according to any one of claims 1 to 9, and the cloud device is configured to implement the man-machine interaction method according to claim 10.
15. A computer readable storage medium, characterized in that program instructions executable by a processor for implementing the human-machine interaction method of any one of claims 1 to 10 are stored.

Description

Man-machine interaction method and related device, electronic equipment and storage medium Technical Field The present application relates to the field of man-machine interaction technologies, and in particular, to a man-machine interaction method, a related apparatus, an electronic device, and a storage medium. Background With the rapid development of large language model technology, AI companion products are becoming an emerging field of man-machine interaction. Such products are intended to provide emotional chaperones, psychological support, personalized interactive experiences, etc. for the user. Currently, the prior art is generally implemented by deploying a cloud large model, but for safety compliance, after the cloud large model is trained by RLHF (Reinforcement Learning from Human Feedback, human feedback reinforcement learning), when facing a request related to intimacy, emotion depth expression and specific subculture context, the cloud large model actively refuses or gives a reply of too neutral and passenger air, that is, the interaction mode is biased towards homogeneity, surfacing and generalization, so that personalized interaction service is difficult to provide. In view of this, how to implement personalized man-machine interaction on the premise of protecting data privacy becomes a problem to be solved urgently. Disclosure of Invention The application mainly solves the technical problem of providing a man-machine interaction method, a related device, electronic equipment and a storage medium, and realizing personalized man-machine interaction on the premise of protecting data privacy. In order to solve the technical problems, the first aspect of the application provides a man-machine interaction method which is applied to local equipment and comprises the steps of carrying out data retrieval based on current input of a target object to obtain historical memory related to the current input in historical interaction between the target object and an avatar, carrying out intention abstraction based on the current input, the historical memory and the current relation state between the target object and the avatar to obtain interaction intention, carrying out desensitization extraction based on the historical memory to obtain a historical abstract, and sending the interaction intention and the historical abstract to cloud equipment, wherein the cloud equipment generates a reply draft based on the interaction intention and the historical abstract, and carrying out style migration based on the reply draft from the cloud equipment to obtain personalized reply of the avatar. The second aspect of the application provides a man-machine interaction method applied to cloud equipment, which comprises the steps of receiving interaction intention and history abstract, wherein the interaction intention and the history abstract are sent by local equipment of a target object, the local equipment performs data retrieval based on current input of the target object to obtain history memory related to the current input in history interaction between the target object and an avatar, performs intention abstraction based on the current input, the history memory and the current relation state between the target object and the avatar to obtain interaction intention, performs desensitization extraction based on the history memory to obtain history abstract, generates a reply draft based on the interaction intention and the history abstract, and sends the reply draft to the local equipment, wherein the local equipment performs style migration based on the reply draft to obtain personalized reply of the avatar. In order to solve the technical problems, the third aspect of the application provides a man-machine interaction device which is applied to local equipment and comprises a data retrieval module, an abstract extraction module, an information sending module and a style migration module, wherein the data retrieval module is used for retrieving data based on the current input of a target object to obtain a history memory related to the current input in the history interaction between the target object and an avatar, the abstract extraction module is used for carrying out intention abstraction based on the current input, the history memory and the current relation state between the target object and the avatar to obtain an interaction intention and carrying out desensitization extraction based on the history memory to obtain a history abstract, the information sending module is used for sending the interaction intention and the history abstract to cloud equipment, the cloud equipment generates a reply draft based on the interaction intention and the history abstract, and the style migration module is used for carrying out style migration based on the reply draft from the cloud equipment to obtain personalized reply of the avatar. In order to solve the technical problems, the fourth aspect of the application provides