Large model-based role dialogue method, agent, device and storage medium

Abstract

The present disclosure provides a large model-based role dialogue method, an agent, a device and a storage medium, relates to the technical field of computers, and in particular to the technical field of chat robots, the technical field of large models, and the technical field of agents. The specific implementation scheme is: using a decision-making large model to process perception information and memory information to obtain decision-making information of a dialogue; wherein the perception information is obtained by processing input information of the dialogue based on perception, and the memory information is obtained by processing information of a dialogue initiator; using a dialogue large model to process the decision-making information of the dialogue and role characteristic information to generate reply information of the dialogue. According to the present disclosure, the decision-making large model can be used to generate decision-making information according to the perception information and the memory information, and then the dialogue large model can be used to obtain dialogue replies that can meet the needs of the initiator and are more in line with the characteristics of the role according to the decision-making information and the role information in the memory information.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to the technical field of chat robots, the technical field of large models, and the technical field of agents. BACKGROUND

[0002] A large language model can process various natural language tasks such as text classification, question answering, and dialogue. The capability of the large language model is increasingly improved, and the multi-turn dialogue capability based on the large language model can realize role dialogue. The large language model has basic dialogue capability and role deduction capability, but the roles are relatively fixed, and the chat content replied by the roles is not particularly satisfactory. SUMMARY

[0003] The present disclosure provides a role dialogue method based on a large model, an agent, a device, and a storage medium.

[0004] According to an aspect of the present disclosure, a role dialogue method based on a large model is provided, comprising:

[0005] processing, by a decision large model, the perception information and the memory information to obtain decision information of the dialogue, wherein the perception information is obtained based on input information of the dialogue after perception processing, and the memory information is obtained based on information processing of a dialogue initiator;

[0006] processing, by a dialogue large model, the decision information of the dialogue and the role characteristic information to generate reply information of the dialogue.

[0007] According to another aspect of the present disclosure, an agent is provided, comprising:

[0008] a decision system configured to process, by a decision large model, the perception information and the memory information to obtain decision information of the dialogue, wherein the perception information is obtained based on input information of the dialogue after perception processing, and the memory information is obtained based on information processing of a dialogue initiator;

[0009] a dialogue system configured to process, by a dialogue large model, the decision information of the dialogue and the role characteristic information to generate reply information of the dialogue.

[0010] According to another aspect of the present disclosure, an electronic device is provided, comprising:

[0011] at least one processor; and

[0012] a memory in communication with the at least one processor; wherein

[0013] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any of the embodiments of the present disclosure.

[0014] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method according to any of the embodiments of the present disclosure.

[0015] According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program which, when executed by a processor, implements the method according to any of the embodiments of the present disclosure.

[0016] According to the present disclosure, a decision large model can be used to generate decision information according to perception information and memory information, and then a dialogue large model can be used to obtain a dialogue reply that can meet the needs of the initiator and is more in line with the characteristics of the role according to the decision information and the role information in the memory information.

[0017] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are used to better understand the present scheme and do not limit the present disclosure. Among them:

[0019] Figure 1 is a flowchart of a role dialogue method based on a large model according to an embodiment of the present disclosure;

[0020] Figure 2 is a flowchart of a role dialogue method based on a large model according to another embodiment of the present disclosure;

[0021] Figure 3 is a flowchart of a role dialogue method based on a large model according to another embodiment of the present disclosure;

[0022] Figure 4 is a schematic diagram of an artificial intelligence agent according to an embodiment of the present disclosure;

[0023] Figure 5 is a specific example diagram of a decision flow according to an embodiment of the present disclosure;

[0024] Figure 6 is a structural schematic diagram of an agent according to an embodiment of the present disclosure;

[0025] Figure 7 is a structural schematic diagram of an agent according to another embodiment of the present disclosure;

[0026] Figure 8 is a block diagram of an electronic device for implementing an embodiment of the present disclosure. DETAILED DESCRIPTION

[0027] Exemplary embodiments of the present disclosure are described herein with reference to the accompanying drawings, which are included to provide a thorough understanding of embodiments of the present disclosure by a person of ordinary skill in the art, and should not be construed as limiting the present disclosure to particular embodiments. Thus, it will be apparent to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Also, descriptions of well-known functions and constructions are omitted for clarity and conciseness.

[0028] Based on prompt engineering, thought chain alignment and other methods, the role consistency of large models can be improved, and the basic dialogue ability and role performance ability of the model can be enhanced. Some early role dialogues are based on preset storylines, and the interaction nodes and dialogue content are set by the producer in advance, and the user does not really interact with the system.

[0029] Examples of existing technical solutions for role dialogues include:

[0030] (1) Role-playing products based on language models, which use prompt engineering to make language models perform role-playing. This technical solution, because large language models have read many stories during training, and the 'intelligent' behavior exhibited by larger language models. Users find that the model can often exhibit certain role-playing and dialogue capabilities under such prompts. However, the use of such prompts is highly dependent on the memory of the large language model. If the large language model's memory of the role itself is vague, it cannot mimic a specific role. The use of prompts to limit the definition of role-related knowledge is also vague and cannot effectively prevent the occurrence of large language model hallucination. Using prompts to limit the dialogue style of chatbots will be greatly influenced by the language model. Adjusting the prompts may alleviate this problem, but each specific role requires very fine-tuned prompts. These shortcomings obviously limit the use of this role-playing chatbot.

[0031] (2) Use the rich knowledge of the large model to fine-tune the role dialogue capability into the model. This technical solution. Because the current large model is more in the form of an assistant answering questions, the logical thinking ability in complex situations is not very satisfactory, and the dialogue often lacks practical meaning, and the user needs to find topics. According to the theory that the generalization ability of the large model is essentially the compression of the training data, after the model is trained on dialogue data, it learns the response language of role dialogue, but it does not learn the motivation and thinking way in the conversation, and it is easy to give replies that do not conform to the scene, and it cannot have depth and temperature like real human conversation.

[0032] In summary, it is necessary to migrate the dialogue style of the model without degrading the basic ability and improve the realism of the dialogue experience of the dialogue model. In order to improve the role-playing ability of the model, in order to improve the dialogue time of the user and the model, the embodiment of the disclosure designs a role dialogue model based on the paradigm of an artificial intelligence (AI) agent (also known as an agent), adds a functional module to the large language model, and reconstructs the thinking link of the dialogue process. Let the large model be able to dialogue like a "person".

[0033] Figure 1 is a flowchart of a role dialogue method based on a large model according to an embodiment of the disclosure. The method can include:

[0034] S101, using a decision-making large model to process perception information and memory information to obtain decision-making information of the dialogue; wherein the perception information is obtained based on the input information of the dialogue after perception processing, and the memory information is obtained based on the information processing of the dialogue initiator;

[0035] S102, using a dialogue large model to process the decision-making information of the dialogue and the role characteristic information to generate reply information of the dialogue.

[0036] In the embodiment of the disclosure, the role dialogue method based on the large model can be executed by the agent, and the agent can also be called an agent (Agent) or a chat robot, etc. The large model can include a large language model (LLM).

[0037] In some examples, the agent can include a perception system, a memory system, a decision-making system, and a dialogue system. The perception system can include a perception large model, which can be trained from an initial large model using training samples of the perception large model. The input information of the dialogue is input into the perception system of the perception large model to obtain the perception information. The input information of the dialogue can include one or more information such as text, voice, image, video, etc. input by the dialogue initiator. The dialogue initiator can include a user and / or a device used by the user, etc.

[0038] In the embodiments of the present disclosure, the memory system can save one or more of the memory information of the object, the memory information of the role, the dialogue context, etc. The name, identity, etc. of the dialogue initiator is input into the memory system, and the memory information of the dialogue initiator can be found in the memory system. The input information and / or the perception information of the dialogue is input into the memory system, and the dialogue context corresponding to the input information and / or the perception information of the dialogue can be found in the memory system. The name, identity, etc. of the role is input into the memory system, and the memory information (which can be referred to as role characteristic information) of the role can be found in the memory system.

[0039] In the embodiments of the present disclosure, the decision system can include a decision large model, which can be trained from an initial large model using training samples of the decision large model. The perception information output by the perception large model and the memory information output by the memory system are input into the decision system, and the decision information of the dialogue can be obtained based on the perception information and the memory information by using the decision large model of the decision system.

[0040] In the embodiments of the present disclosure, the dialogue system can include a dialogue large model, which can be trained from an initial large model using training samples of the dialogue large model. The decision information of the dialogue output by the decision large model and the memory information of the role output by the memory system are input into the dialogue system, and the reply information to the input information of the dialogue can be generated by using the dialogue large model of the dialogue system.

[0041] According to the embodiments of the present disclosure, the decision information can be generated from the perception information and the memory information by using the decision large model, and then the dialogue reply that can meet the needs of the initiator and is more in line with the characteristics of the role can be obtained from the decision information and the role information in the memory information by using the dialogue large model.

[0042] In an implementation manner, the perception information includes one or more of the emotion information, the intention information and the dialogue state information of the dialogue initiator.

[0043] In the embodiments of the present disclosure, the emotion information can represent the emotional state corresponding to the dialogue initiator, such as the states of depression, excitement, calmness, etc. The emotion information can also be referred to as the sentiment information. The intention information can represent the expectation, tendency or desire of the dialogue initiator about the dialogue content, such as the expectation of the topic of the chat related to the weather, the expectation of understanding the content related to the animation, etc. The dialogue state information can be analyzed based on the emotion information and the intention information. The dialogue state information can represent the current state of the dialogue, such as the state of the beginning of the dialogue, the state of the end of the dialogue, the state of the need to switch the topic, etc.

[0044] According to the embodiments of the present disclosure, the needs of the user can be better understood according to the emotion information and the intention information of the dialogue initiator, and more accurate dialogue state information is obtained, so that the large model can make more reasonable decisions based on the emotion information, the intention information and the dialogue state information, and obtain a reply that meets the needs of the user.

[0045] In an embodiment, the memory information includes long-term memory and short-term memory, the long-term memory includes one or more of the object characteristics, the interaction history, and the role characteristic information; and the short-term memory includes one or more of the dialogue context, the emotion information, and the intention information.

[0046] The memory system can include long-term memory and short-term memory, wherein the long-term memory can be used to store the interaction history of the agent and the user, the object characteristics of the user, and the role knowledge base. The short-term memory can be used to store the emotion information, the intention information, and the dialogue state output by the perception system.

[0047] In the embodiments of the present disclosure, the memory system can use the long-term memory and the short-term memory to store the information needed by the agent during the dialogue. The long-term memory can store the characteristics of the object with which the agent has a dialogue. The long-term memory can also store the interaction history of the object and the agent, such as the historical input information and the historical reply information of the dialogue. The long-term memory can also store role characteristic information, such as the knowledge base of the role (including the knowledge information of different roles), the personality of the role, the character of the role, and the language habits of the role. The short-term memory can store the context information of the dialogue. According to the identifier of the current dialogue, the dialogue content of the previous times of the current dialogue and the dialogue content of the previous topics can be found. The short-term memory can also store the perception information output by the perception system, such as the emotion information and the intention information. The long-term memory and the short-term memory of the same object can have a corresponding relationship. For example, based on the identifier of an object, the characteristics, the interaction history, the role characteristic information, the dialogue context, the emotion information, and the intention information of the object can be found.

[0048] In the embodiments of the present disclosure, the memory system can use a preset prompt word template to splice one or more of the context information, the emotion information, and the intention information, to obtain a short-term memory prompt word. One or more of the dialogue context, the emotion information, and the intention information can be stored in the short-term memory according to the form of the short-term memory prompt word, to obtain the short-term memory of the memory system. Further, the memory system can input the short-term memory prompt word of a certain object into the decision system to participate in decision-making.

[0049] According to the embodiments of the present disclosure, the long-term memory and the short-term memory related to the dialogue can be stored in the memory system, and a dialogue that meets the needs of the user and conforms to the role is generated for the decision system based on the rich and diverse memory information.

[0050] Figure 2 is a flowchart of a role dialogue method based on a large model according to another embodiment of the disclosure, which can include one or more features of the role dialogue method based on a large model described above. In an implementation, S101 processes the perception information and the memory information using the decision-making large model to obtain decision-making information of the dialogue, including:

[0051] S201, using a decision-making large model, performing a thinking process based on the dialogue state information to obtain a thinking result;

[0052] S202, based on the thinking result and one or more of the dialogue context, the emotion information, the intention information, the object feature, and the interaction history, performing an action process to obtain an action direction;

[0053] S203, based on the action direction, performing an observation process to obtain an observation result;

[0054] S204, based on the action direction and / or the observation result, obtaining decision-making information of the dialogue.

[0055] For example, the role thinking logic includes one or more of the thinking result, the action result, and the observation result; the decision-making information can include one or more of the content decision-making information, the motivation decision-making information, the emotion decision-making information, and the tone decision-making information.

[0056] In an embodiment of the disclosure, the decision-making system can include a decision-making large model, which can generate decision-making information of the dialogue based on a thinking chain. The thinking chain of the decision-making system can include meta-motivation, action mode, and thinking mode. The meta-motivation of the thinking chain can be to continue the dialogue; the action mode of the thinking chain can adopt a reason + action mode; and the thinking mode of the thinking chain can adopt a thinking-action-observation mode.

[0057] In an embodiment of the disclosure, the decision-making large model can make decisions based on the long-term memory and the short-term memory sent by the memory system and the perception information output by the perception system to obtain decision-making information of the dialogue. The decision-making information of the dialogue can also be referred to as an optimal execution strategy, such as whether to change the topic, a response strategy, a dialogue motivation, a dialogue emotion, a dialogue tone, and a reply content.

[0058] In an embodiment of the disclosure, the decision-making system can first perform a thinking process in a thinking-action-observation mode according to the dialogue state information output by the perception system. For example, the decision-making large model can perform a thinking process based on the dialogue state information to determine whether to continue the current topic or to change a new topic. After the thinking process is performed, a thinking result can be obtained, such as continuing the current topic or changing a new topic.

[0059] In the embodiments of the present disclosure, the decision system can execute the action process in the thinking-action-observation mode according to the thinking result, the dialogue context, the emotion information, the intention information, the object feature, the interaction history, etc., to determine the action direction to be taken, which can also be the response strategy. For example, the thinking result is to continue the current topic, the emotion information is that the mood is not good, and the action direction can include that the role needs to be more positive and enthusiastic. For another example, the thinking result is to change a topic, the interaction history includes that the topic related to game A has been asked before, and the action direction can include the topic related to game A.

[0060] In the embodiments of the present disclosure, the decision system can execute the observation process in the thinking-action-observation mode according to the action direction to obtain the observation result. The observation result can include the reaction that the decision system may encounter after taking the action based on the action direction, the operation details that need to be paid attention to, etc. For example, after changing the topic related to game A, it is observed whether the dialogue initiator has a further response to the topic related to game A.

[0061] In the embodiments of the present disclosure, the decision system can generate the decision information including one or more of the content decision information, the motivation decision information, the emotion decision information and the tone decision information according to the action direction and / or the observation result. The content decision information can include the text, voice, image or video content of the expected reply. For example, “Is the mood not good? Do you need to open game A to relax”. The emotion decision information can include the emotion of the role and / or the object. For example, the object's low emotion corresponds to the role's enthusiastic emotion. The tone decision information can include the tone of the role's reply. For example, a soft expected reply is adopted for the object's low emotion.

[0062] According to the embodiments of the present disclosure, through the thinking process, the action process and the observation process, one or more of the dialogue state information, the dialogue context, the emotion information, the intention information, the object feature and the interaction history can be used to generate the strategy of the dialogue reply that meets the object's demand and the role's characteristics, which is conducive to generating more lively, more flexible, more object-demand-meeting and more role-characteristic replies in the dialogue system.

[0063] Figure 3 is a flowchart of a role dialogue method based on a large model according to another embodiment of the present disclosure. The method can include one or more features of the role dialogue method based on a large model described above. In an implementation, S102 processes the decision information of the dialogue and the role feature information using the dialogue large model to generate the reply information of the dialogue, including:

[0064] S301, input the decision information of the dialogue output by the decision large model and the target role feature information obtained from the memory system into the dialogue large model to generate the reply information meeting the characteristic information of the target role and the dialogue state information.

[0065] In the embodiments of the present disclosure, the dialogue system can receive decision information from the decision system and target role characteristic information from the memory system, input the information into the dialogue large model, and generate reply information conforming to the target role characteristic information and the current dialogue scene. The current dialogue scene can be determined based on the dialogue state information. For example, if the dialogue state information is an initial dialogue state, the current dialogue scene is a start scene. If the dialogue state information includes a topic switching state, the current dialogue scene is a topic switching scene.

[0066] In the embodiments of the present disclosure, before the dialogue with the intelligent agent, the role of the intelligent agent can be preselected through the dialogue initiation object, or the role of the intelligent agent can be determined based on input information analysis. The role of the intelligent agent can include an animation role, a film and television role, a literature role, a historical role, a science and technology role, etc. The memory system can prestore long-term memory and / or short-term memory of various roles. According to the role of the intelligent agent, the target role characteristic information can be retrieved in the long-term memory. After retrieving the target role characteristic information, the memory system can send the target role characteristic information to the dialogue system.

[0067] According to the embodiments of the present disclosure, the dialogue large model can be used to generate more vivid, flexible, object-demand-conforming, and role-characteristic reply according to the decision information conforming to the object demand and the target role characteristic information.

[0068] In an implementation, the training sample of the decision large model includes one or more of the following:

[0069] perception sample information;

[0070] memory sample information;

[0071] role thinking logic to be processed and / or decision information to be processed obtained by the large language model based on the perception sample information and / or the memory sample information;

[0072] labeled role thinking logic obtained by labeling the role thinking logic to be processed;

[0073] labeled decision information obtained by labeling the decision information to be processed.

[0074] In the embodiments of the present disclosure, the large language model can be used to reversely deduce the role thinking logic to be processed based on the perception sample information and / or the memory sample information according to the thinking chain, such as deducing an observation result, deducing an action direction, and deducing a thinking result. Then, the deduced observation result, the deduced action direction, and the deduced thinking result are labeled by using machine labeling or manual labeling. The large language model can further reversely deduce the decision information to be processed based on the deduced observation result, the deduced action direction, and the deduced thinking result, such as one or more of deduced content decision information, deduced motivation decision information, deduced emotion decision information, and deduced tone decision information.

[0075] In the training process of the decision large model, the perception sample information and / or the memory sample information in the training sample can be used as the input feature of the model, and the labeled role thinking logic and / or the labeled decision information in the training sample can be used as the expected input feature of the model. The initial decision large model is fine-tuned to obtain the trained decision large model.

[0076] According to the embodiments of the present disclosure, the decision large model trained by using the training sample of the decision large model can obtain decision information more consistent with the perception information and the memory information, and further obtain reply information more consistent with the role characteristics and the object requirements.

[0077] In an implementation, the training sample of the dialogue large model includes: obtaining role characteristic sample information and role dialogue sample information based on a role dialogue data set, the role characteristic sample information including one or more of role knowledge, role background, role personality, role character, and role language habit.

[0078] In the embodiments of the present disclosure, the role dialogue data set can be understood as a role data source. For example, the role dialogue data set can include literary works, film and television works, historical documents, etc. The role data source is processed by using a large model, and role characteristic sample information such as role knowledge, role background, role personality, role character, and role language habit can be extracted from the role data source. These role characteristic sample information can be stored in a memory system or stored separately in a training sample set. The role knowledge can include the content that the role can understand. For example, the knowledge of a physicist role includes physical knowledge. The knowledge of a doctor role includes medical knowledge. For another example, the role background can include the life background, the learning background, the work background, etc. of the role. For another example, the role personality can include introversion and extroversion. The role character can include bravery, intelligence, wit, loyalty, etc. The role personality and the role character can have the same or similar meanings. The role personality and / or the role character can also be referred to as the role individuality. For another example, the role language habit can include the habitual language of the role, vividness, directness, stiffness, etc.

[0079] The role data source is processed using a large model, and role dialogue sample information can be extracted from the role data source. The role dialogue sample information can include dialogues of a role appearing in literary works, film and television works, historical documents, and the like. For example, a conversation involving a role R in a literary work includes a question Q1 and a reply A1, a question Q2 and a reply A2, and a question Q3 and a reply A3. The role feature sample information and the questions in the role dialogue sample information can be used as input features of a model, and the replies in the role dialogue sample information can be used as output features of the model. The initial dialogue large model is fine-tuned to obtain a trained dialogue large model. The trained dialogue large model has the ability to play different roles and reply to dialogues.

[0080] According to the embodiments of the present disclosure, the dialogue large model trained using the training samples of the dialogue large model can generate reply information that is more consistent with the role characteristics and the needs of the object, and thus the reply information is more realistic and more flexible.

[0081] In an implementation, the training samples of the perception large model include one or more of the following:

[0082] input sample information;

[0083] to-be-processed perception information obtained by the large language model based on the perception sample information;

[0084] annotated perception information obtained by annotating the to-be-processed perception information.

[0085] In the embodiments of the present disclosure, text, voice, image, or video information can be input into a large language model for reverse deduction to obtain to-be-processed perception information such as deduced emotional information, deduced intention information, and deduced dialogue state. Then, the to-be-processed perception information such as the deduced emotional information, the deduced intention information, and the deduced dialogue state is annotated using machine annotation or manual annotation to obtain annotated perception information such as annotated emotional information, annotated intention information, and annotated dialogue state.

[0086] In the training process, the input sample information can be used as input features of an initial perception large model, and the annotated perception information can be used as output features of the initial perception large model. The initial perception large model is fine-tuned to obtain a trained perception large model.

[0087] According to the embodiments of the present disclosure, the perception large model trained using the training samples of the perception large model can obtain reply information of a dialogue that is consistent with the emotions and intentions of an object based on perception information that is consistent with the emotions and intentions of the object.

[0088] In order to let the dialogue initiator truly experience the role dialogue experience as human dialogue, the embodiments of the present disclosure can construct an artificial intelligence agent (AI agent) based on the dialogue capability of a large language model, with an independent perception system, memory system and decision system as an aid. When performing role dialogue, the perception system can analyze the current dialogue state according to the perceived emotions and desires of the dialogue initiator, and the decision system can plan the dialogue development direction according to the results of the perception system and the role characteristic information, dialogue initiator portrait and interaction history in the memory system. In this way, not only the understanding of the dialogue initiator by the model is improved, but also the diversity of the model's reply is greatly improved, making the dialogue development have infinite possibilities.

[0089] The schematic diagram of the artificial intelligence agent (AI agent) of the embodiments of the present disclosure is shown in Figure 4 The artificial intelligence agent can include four parts, including a perception system, a memory system, a decision system and a dialogue model (or called dialogue system). The artificial intelligence agent can perform role dialogue according to the following steps: the first step is that the perception system perceives the emotions, intentions and dialogue states of the dialogue initiator according to the dialogue initiator input and the current dialogue; the second step is that the decision system analyzes and decides the optimal execution strategy (also called decision information) based on the results of the perception system and the object characteristics, interaction history (interaction history), role knowledge base and other data in the memory system; the third step is that the dialogue model gives a reply with clear meaning, distinct role style and the ability to promote the dialogue to continue forward under the guidance of the optimal execution strategy of the decision system based on the dialogue capability of the large model and the role playing ability learned in the fine-tuning process. Different trained large models or deep learning models will be used to realize the four parts. Specifically as follows:

[0090] I. Perception system: The perception system is mainly responsible for processing and understanding the input of the dialogue initiator. It perceives the situation of the dialogue initiator according to the input of the dialogue initiator and the current dialogue history, including the emotions, intentions and dialogue states of the dialogue initiator. The meaning of the perception system is equivalent to observing the facial expressions in real dialogue. This ability is very important, which determines whether the dialogue initiator has the desire to continue the dialogue with the model. Since the emotions and desires are relatively shallow semantics, we can use a smaller large model here to save computing cost and reduce response time.

[0091] II. Memory System: Memory system is an important part of the Agent solution. Memory mainly solves the problem of knowledge transfer. It can make the Agent have long-term and short-term memory, and make the Agent more intelligent. Long-term memory mainly includes the characteristics of the dialogue initiator, the interaction history, and the role-related knowledge graph, etc. Short-term memory mainly includes the context dialogue of the dialogue initiator, the desire (intention) of the dialogue initiator, and the emotion (emotion) and other knowledge that are strongly related to the current dialogue. The long-term memory part takes the vector database as the main implementation form. Because of the embedded storage, it can be more beneficial to compress information and efficiently retrieve under the premise of large-scale data. Short-term memory generally uses prompt as a transfer medium. After the memory system obtains it, it is delivered to the decision system through the prompt. At the same time, these information can be used in the decision system to help the system make better decisions.

[0092] III. Decision System: The decision system takes the sustainable progress of the dialogue as the meta-motive, and solves the problem in the mode of thinking-acting-observing with the method of reason + action. For example, the thinking process can think whether to continue the current topic or find a new topic; the action process can determine the action direction of the role, such as the strategy of the role to respond, the motivation of the role to dialogue, and the emotion of the role to dialogue; for each action direction of the action process, the corresponding observation result can be given in the observation link. The decision system can finally output what the role says with what tone. The decision system analyzes the information such as the emotion, intention, and historical behavior of the dialogue initiator according to the results of the perception system and the data in the memory system, and comprehensively considers and decides the optimal execution strategy in terms of content, motivation, emotion, tone, etc.

[0093] For example, as shown in FIG. 1, in the historical dialogue and the input information, the input information 1 of the object includes “Who are you?” The reply 1 of the role (for example, AI) includes “Don’t you know me? I am Y!” The input information 2 of the object includes “Oh.” The reply 2 of the role includes “Have you seen any new movies recently?” The input information 3 of the object includes “No.” The decision system uses the method of reason + action to execute the process of thinking-acting-observing. Figure 5

[0094] For example, the thinking process can include: whether the role continues the current topic or finds a new topic (thinking):

[0095] 1. The current topic is whether the role asks the object if he has seen a new movie, and the object says no;

[0096] 2. The current topic has no content to promote, and a new topic needs to be found.​

[0097] For example, the action process and the observation process can include: what strategy does the role need to respond (action) with:

[0098] 1. The object feature shows that the object likes playing games, and the role can cut in from the game and ask the object if he / she has played a new game recently;

[0099] Observation: The object does not respond to the topic of the movie, and may not respond to the game;

[0100] 2. The interaction history with the object shows that the object is in a low mood, and the role can cut in from this angle and care about the emotional state of the object;

[0101] Observation: Do not touch the object's sad place;

[0102] 3. The object is currently in a low mood, and the role can be more enthusiastic and positive;

[0103] Observation: Pay attention not to make the object feel stressed;

[0104] For example, the decision information can include: what does the role need to say? What tone does the role need to say?

[0105] Tone: gentle and considerate, enthusiastic;

[0106] Content: "You don't seem to be in a good mood, is it because you've been too tired recently? If you want to talk to someone, I'm always here. If you don't want to talk about it, I can also play games with you."

[0107] Four, dialogue system: the dialogue model is mainly responsible for generating responses to the dialogue initiator according to the optimal execution strategy of the decision system. The implementation of the dialogue model is based on the fine-tuning of large language models, and the ability of the model to learn dialogue and role-playing is learned through training to generate replies with clear meaning, distinctive role style, and the ability to push the dialogue forward.

[0108] The construction of the training samples of the large models of the perception system, the decision system and the dialogue system of the embodiments of the present disclosure, and the construction of the long and short term memory data of the memory system are shown in Table 1, and are as follows:

[0109] Table 1

[0110]

[0111] The following are data examples of several systems:

[0112] I. Perception system: Emotion can include the psychological state of the thought, feeling and performance of the dialogue initiator. The emotion and desire of the dialogue. The present disclosure can input and add annotations for the user. When labeling, the context will be used as prior knowledge, and the dialogue context can be used for labeling. The content of the labeling can include the emotion of the dialogue initiator, the desire of the dialogue initiator and the dialogue state.

[0113] II. Memory system: The characteristics of the dialogue initiator are described. The characteristics of the dialogue initiator are divided into basic characteristics and time characteristics, and the basic characteristics include the characteristics formed by the long-term behavior of the dialogue initiator. The time characteristics can be extracted according to the behavior of the dialogue initiator in the past period of time, and can represent the behavior characteristics of the dialogue initiator in a period of time.

[0114] III. Decision system: The dialogue data is deduced by a large model, the psychological thinking logic when the role answers is obtained, and the annotation conforming to the general thinking logic is performed, and finally modified. Since the thinking process of the dialogue is various, the expression of the thinking paradigm can be specifically limited. After the trial adjustment, the present disclosure enriches the guidance criteria of the large model reverse deduction, so as to construct a more comprehensive and more accurate analysis dimension, and finally the strategy is presented in the form of a thinking chain.

[0115] IV. Dialogue system: The role dialogue data set can be obtained based on novels, scripts, and constructed in combination with high-quality interesting dialogue data obtained from the network. The role dialogue data set is mainly constructed around the following three dimensions, which are the knowledge and background of the role, the personality or character of the role, and the language habit of the role. Each role has its own background, so in the construction of the ChatBot, the present disclosure enables the ChatBot to understand the setting of the corresponding story.

[0116] Personality or character: The personality and character setting of the character is a very important part in animation, film and even game works. The personality and character setting needs to be consistent in the whole work. Some literary works even define the personality setting of the character first when creating, and then perform subsequent writing work. The present disclosure makes the personality and character reacted by the agent consistent with the original setting of the work through training.

[0117] Language habit: Language habit is the most easily imitated by language model, and the language model can imitate output by giving appropriate examples in the text. The dialogue large model of the present disclosure can enable the fans of literary and television works to interact with the product of the present disclosure, and can "reproduce" the classic scenes of the literary and television works, so that the fans of these works can obtain a better experience.

[0118] The large model of the perception system, decision system, and dialogue system of the embodiments of the present disclosure can train the model with constructed data to improve the emotion and desire perception ability of the perception system, planning ability, dialogue model role dialogue ability, etc. The memory system can store the information required by the agent when performing role dialogue. The specific training method of each model and the use method of the memory system are as follows:

[0119] I. Perception system: The perception large model can be obtained by fine-tuning a small-scale large model, for example, a 7B to 14B large language model, and limiting the output of the perception large model to formatted output emotions, desires, and dialogue states using prompts.

[0120] II. Memory system: By recommending knowledge graph reasoning algorithms and vector representation methods of graph models, a large-scale dialogue initiator portrait heterogeneous graph data system is constructed, and the similarity is calculated using graph vector representation to match the personalized content of a single dialogue initiator. The large model extracts key information from the dialogue history of the dialogue initiator, and constructs a retrieval knowledge base according to the importance and time sequence of the key information. When replying, the model constructs a more appropriate reply based on the dialogue history and the current dialogue topic, combines the large model and the knowledge graph technology, and applies it to the role dialogue. Breakthroughs are achieved in the long-term memory and speciality of the model to improve the personalization degree of the role dialogue. The dialogue model adopts the way of predicting the next reply, which can predict the next reply based on the existing multi-round historical dialogue.

[0121] III. Dialogue system: The dialogue large model can be fine-tuned based on a large language model, and the multi-round dialogue ability is more emphasized on the basis of general knowledge ability. Based on the large language model, the model is trained by self-focused conversion (Self-Focused Transformer, SFT) and differential policy optimization (Differentiable Policy Optimization, DPO), continuously optimizing the dialogue large model ability, enhancing the dialogue ability of the dialogue large model in the role's background, personality setting, dialogue interaction, emotional quotient, etc., and improving the generalization and reply quality of the dialogue model.

[0122] The role dialogue large model needs to have some basic capabilities, including that the reply content conforms to the scene, conforms to the character setting, and provides similar emotional value. In role dialogue, individuality is crucial. Each role should exhibit unique personality traits and language styles, allowing the dialogue initiator to clearly distinguish between different roles, thereby increasing the interest of the dialogue. At the same time, emotional expression is also indispensable. The dialogue should truly exhibit the unique emotions of the characters, such as love, hate, and betrayal, etc. By allowing the dialogue initiator to deeply feel the emotional fluctuations of the characters during the dialogue, it can promote the dialogue initiator to have empathy. In order to maintain the authenticity and natural flow of the dialogue, the dialogue content should conform to the character setting and the situation, and be consistent with the interests and desires of the dialogue initiator, promoting natural and smooth communication and interaction between the dialogue initiator and the model, and avoiding the model's reply being too rigid or unrealistic. In addition, the dialogue large model can use more vivid language, combined with some interesting and gong expressions, to increase the interest of the dialogue initiator and make the dialogue more lively and interesting.

[0123] Four, decision system: the decision large model can be trained or fine-tuned with the same or similar base model as the dialogue large model. The embodiments of the present disclosure can provide rich dialogue scene corpus training, which helps the decision large model to understand the dialogue paradigm and infer the emotional expression of the dialogue party.

[0124] The embodiments of the present disclosure provide a perception system, a memory system, and a decision system for dialogue large models based on large model fine-tuning by constructing an artificial intelligence agent (AI Agent), thereby significantly improving the shortcomings of dialogue models such as poor logic, lack of knowledge, and insufficient dialogue analysis capabilities, and can bring great help and advantages to products. The role dialogue realized by the present disclosure can bring a more lively and realistic dialogue experience to the dialogue initiator. The perception system, memory system, and decision system possessed by the artificial intelligence agent (AI Agent) can make its own dialogue large model more intelligent and flexible, thereby better understanding the needs of the dialogue initiator, remembering the preferences of the dialogue initiator, and making more reasonable replies and decisions. This personalized dialogue exchange method not only increases the participation and interactivity of the dialogue initiator, but also greatly improves the attractiveness and competitiveness of the product.

[0125] The role dialogue realized by the present disclosure will provide the dialogue initiator with a more intelligent and personalized document editing experience, and also open up broader development space for the product, creating more value and convenience for the dialogue initiator.

[0126] Figure 6 is a structural schematic diagram of an intelligent agent according to an embodiment of the present disclosure. In an implementation manner, the intelligent agent can include:

[0127] The decision system 601 is configured to process the perception information and the memory information using a decision large model to obtain decision information of the dialogue; wherein the perception information is obtained by processing the input information of the dialogue based on perception, and the memory information is obtained by processing information of a dialogue initiator;

[0128] The dialogue system 602 is configured to process the decision information of the dialogue and the role characteristic information using a dialogue large model to generate reply information of the dialogue.

[0129] Figure 7 FIG. 6 is a structural schematic diagram of an intelligent agent according to another embodiment of the present disclosure, which can include one or more features of the intelligent agent described above. In an implementation, the intelligent agent further includes:

[0130] The perception system 603 is configured to obtain the perception information by processing the input information of the dialogue based on perception, and the perception information includes one or more of emotional information, intention information, and dialogue state information of a dialogue initiator.

[0131] In an implementation, the intelligent agent further includes a memory system 604 configured to extract and store memory information, wherein the memory information includes long-term memory and short-term memory, the long-term memory includes one or more of object characteristics, interaction history, and role characteristic information, and the short-term memory includes one or more of dialogue context, emotional information, and intention information.

[0132] In an implementation, the decision system 601 includes:

[0133] The thinking module 701 is configured to use a decision large model to perform a thinking process based on dialogue state information to obtain a thinking result.

[0134] The action module 702 is configured to perform an action process based on the thinking result and one or more of dialogue context, emotional information, intention information, object characteristics, and interaction history to obtain an action direction.

[0135] The observation module 703 is configured to perform an observation process based on the action direction to obtain an observation result.

[0136] The decision module 704 is configured to obtain decision information of the dialogue based on the action direction and / or the observation result.

[0137] In an implementation, the dialogue system 602 is further configured to input the decision information of the dialogue output by the decision large model and target role characteristic information obtained from the memory system into the dialogue large model to generate reply information that conforms to the characteristic information and dialogue state information of the target role.

[0138] In an implementation, the training sample of the decision large model includes one or more of:

[0139] perception sample information;

[0140] memory sample information;

[0141] obtaining, by the large language model, to-be-processed role thinking logic and / or to-be-processed decision information based on the perception sample information and / or the memory sample information;

[0142] annotating the to-be-processed role thinking logic to obtain annotated role thinking logic;

[0143] annotating the to-be-processed decision information to obtain annotated decision information.

[0144] In an embodiment, the training sample of the dialogue large model comprises:

[0145] obtaining role feature sample information and role dialogue sample information based on the role dialogue data set, the role feature sample information comprising one or more of role knowledge, role background, role personality, role character, and role language habit.

[0146] In an embodiment, the training sample of the perception large model comprises one or more of the following:

[0147] input sample information;

[0148] obtaining, by the large language model, to-be-processed perception information based on the perception sample information;

[0149] annotating the to-be-processed perception information to obtain annotated perception information.

[0150] The specific functions and examples of the systems and modules of the agent of the embodiments of the present disclosure are described in the above method embodiments, and will not be described here.

[0151] In the technical solutions of the present disclosure, the acquisition, storage, and application of user personal information comply with relevant laws and regulations, are authorized by the user, and do not violate public order and good customs.

[0152] According to the embodiments of the present disclosure, the present disclosure further provides an electronic device, a readable storage medium, and a computer program product.

[0153] Figure 8A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit the implementations of the present disclosure described and / or claimed in this document.

[0154] As shown, Figure 8 The device 800 includes a computing unit 801 that can perform various appropriate actions and processes in accordance with a computer program stored in a read-only memory (ROM) 802 or a computer program loaded into a random access memory (RAM) 803 from a storage unit 808. Various programs and data required for the operation of the device 800 can also be stored in the RAM 803. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0155] Various components in the device 800 are connected to the I / O interface 805, including an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; the storage unit 808, such as a magnetic disk, a magneto-optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0156] The computing unit 801 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs various methods and processes described above, such as the large model based role dialogue method. For example, in some embodiments, the large model based role dialogue method can be implemented as a computer software program tangibly embodied in a machine readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded onto the RAM 803 and executed by the computing unit 801, one or more steps of the large model based role dialogue method described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform the large model based role dialogue method by any other suitable means, such as by means of firmware.

[0157] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a programmable logic device (PLD), a computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0158] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces a function / operation specified in the flowchart and / or block diagram. The program code can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package, and partially on a remote machine or a server.

[0159] In the context of this disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0160] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0161] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0162] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0163] It should be understood that the various forms of flow shown above can be re-ordered, steps added or removed, etc. For example, the steps recited in the present disclosure can be performed in parallel, in series, in a different order, etc., as long as the desired results of the technology disclosed in the present disclosure are achieved, which is not limited herein.

[0164] The specific embodiments discussed above do not constrain the scope of the present disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and alternatives can be made to the specific embodiments without departing from the principles of the present disclosure. Any such modifications, alternatives, equivalents, and / or alternatives are intended to fall within the scope of the present disclosure.

Claims

1. A large model-based role dialogue method, executed by an agent, the agent comprising a perception system, a memory system, a decision system, and a dialogue system, the perception system having a perception large model, the decision system having a decision large model, and the dialogue system having a dialogue large model, the method comprising: processing, using the decision large model, perception information and memory information to obtain decision information for a dialogue, wherein the perception information is obtained by processing input information for the dialogue based on the perception large model, and the memory information is obtained by searching the memory system based on information of a dialogue initiator; and the perception large model is obtained by fine-tuning an initial perception large model. processing, using the dialogue large model, the decision information for the dialogue and role characteristic information in the memory information to generate reply information for the dialogue, wherein the dialogue large model is obtained by fine-tuning an initial dialogue large model using a question in role characteristic sample information and role dialogue sample information as input features of the model, and using a reply in the role dialogue sample information as output features of the model. The perception information comprises one or more of emotional information, intention information, and dialogue state information of the dialogue initiator.

2. The method of claim 1, wherein, The memory information comprises long-term memory and short-term memory, the long-term memory comprising one or more of object characteristics, interaction history, and role characteristic information, and the short-term memory comprising one or more of dialogue context, emotional information, and intention information.

3. The method of claim 1, wherein, Processing, using the decision large model, the perception information and the memory information to obtain the decision information for the dialogue comprises:

4. The method of any one of claims 1 to 3, wherein, executing, using the decision large model, a thinking process based on the dialogue state information to obtain a thinking result; executing, based on the thinking result and one or more of the dialogue context, the emotional information, the intention information, the object characteristics, and the interaction history, an action process to obtain an action direction; executing, based on the action direction, an observation process to obtain an observation result; and obtaining the decision information for the dialogue based on the action direction and / or the observation result. Processing, using the dialogue large model, the decision information for the dialogue and the role characteristic information to generate the reply information for the dialogue comprises:

5. The method of claim 1, wherein, inputting the decision information for the dialogue output by the decision large model and target role characteristic information obtained from the memory system into the dialogue large model to generate reply information conforming to the characteristic information of the target role and the dialogue state information. The training sample of the decision large model comprises one or more of:

6. The method of claim 1, wherein, perception sample information; memory sample information; role thinking logic to be processed and / or decision information to be processed obtained by reverse reasoning based on the perception sample information and / or the memory sample information using a large language model; annotated role thinking logic obtained by annotating the role thinking logic to be processed; annotated decision information obtained by annotating the decision information to be processed. The training sample of the dialogue large model comprises:

7. The method of claim 1, wherein, obtaining role characteristic sample information and role dialogue sample information based on a role dialogue data set, the role characteristic sample information comprising one or more of role knowledge, role background, role personality, role character, and role language habit. ​ 8. The method of claim 1, wherein, The training sample of the perception large model comprises one or more of the following: input sample information; perception information to be processed obtained by using a large language model based on the perception sample information; labeled perception information obtained by labeling the perception information to be processed.

9. A chat robot, comprising: a decision system configured to process perception information and memory information using a decision large model of the decision system to obtain decision information for a conversation, wherein the perception information is obtained by processing input information for the conversation using a perception large model of a perception system, the memory information is obtained by searching information of a conversation initiator in a memory system, and the perception large model is obtained by fine-tuning an initial perception large model; a conversation system configured to process the decision information for the conversation and role feature information in the memory information using a conversation large model of the conversation system to generate reply information for the conversation, wherein the conversation large model is obtained by fine-tuning an initial conversation large model using a question in role feature sample information and role conversation sample information as input features of the model and using a reply in the role conversation sample information as output features of the model.

10. The chat bot of claim 9, wherein, The chat robot further comprises: the perception system configured to obtain the perception information by processing the input information for the conversation, wherein the perception information comprises one or more of emotion information, intent information, and conversation state information of the conversation initiator.

11. The chat bot of claim 9, wherein, The chat robot further comprises: the memory system configured to extract and store memory information, wherein the memory information comprises long-term memory and short-term memory, the long-term memory comprises one or more of object features, interaction history, and role feature information, and the short-term memory comprises one or more of conversation context, emotion information, and intent information.

12. The chat bot of any one of claims 9 to 11, wherein, The decision system comprises: a thinking module configured to use the decision large model to perform a thinking process based on the conversation state information to obtain a thinking result; an action module configured to perform an action process based on the thinking result and one or more of the conversation context, the emotion information, the intent information, the object features, and the interaction history to obtain an action direction; an observation module configured to perform an observation process based on the action direction to obtain an observation result; a decision module configured to obtain the decision information for the conversation based on the action direction and / or the observation result.

13. The chat bot of claim 9, wherein, The conversation system is further configured to input the decision information for the conversation output by the decision large model and target role feature information obtained from the memory system into the conversation large model to generate reply information conforming to the feature information of the target role and the conversation state information.

14. The chat bot of claim 9, wherein, The training sample of the decision large model comprises one or more of the following: perception sample information; memory sample information; role thinking logic to be processed and / or decision information to be processed obtained by using a large language model based on the perception sample information and / or the memory sample information; labeled role thinking logic obtained by labeling the role thinking logic to be processed; labeled decision information obtained by labeling the decision information to be processed.

15. The chat bot of claim 9, wherein, The training sample of the dialogue large model comprises: obtaining role feature sample information and role dialogue sample information based on a role dialogue data set, and the role feature sample information comprises one or more of role knowledge, role background, role personality, role character and role language habit.

16. The chat bot of claim 9, wherein, The training sample of the perception large model comprises one or more of: input sample information; to-be-processed perception information obtained by the large language model based on the perception sample information; annotated perception information obtained by annotating the to-be-processed perception information.

17. An electronic device comprising: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

18. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method according to any one of claims 1-8.

19. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1-8.

Images