CN112949857A - Conversation strategy generation method for simulating user emotion - Google Patents
Conversation strategy generation method for simulating user emotion Download PDFInfo
- Publication number
- CN112949857A CN112949857A CN202110312778.6A CN202110312778A CN112949857A CN 112949857 A CN112949857 A CN 112949857A CN 202110312778 A CN202110312778 A CN 202110312778A CN 112949857 A CN112949857 A CN 112949857A
- Authority
- CN
- China
- Prior art keywords
- user
- emotion
- conversation
- dialog
- strategy
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
- 230000008451 emotion Effects 0.000 title claims abstract description 100
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000002996 emotional effect Effects 0.000 claims abstract description 47
- 230000008859 change Effects 0.000 claims abstract description 27
- 238000004088 simulation Methods 0.000 claims abstract description 23
- 230000009471 action Effects 0.000 claims abstract description 20
- 230000008569 process Effects 0.000 claims abstract description 10
- 239000011159 matrix material Substances 0.000 claims description 18
- 239000013598 vector Substances 0.000 claims description 18
- 230000004044 response Effects 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 12
- 230000003993 interaction Effects 0.000 claims description 9
- 230000002787 reinforcement Effects 0.000 claims description 7
- 238000012549 training Methods 0.000 claims description 7
- 230000007613 environmental effect Effects 0.000 claims description 6
- 230000036651 mood Effects 0.000 claims description 6
- 230000014509 gene expression Effects 0.000 claims description 4
- 206010029216 Nervousness Diseases 0.000 claims description 3
- 230000008921 facial expression Effects 0.000 claims description 3
- 238000005562 fading Methods 0.000 claims description 3
- 230000002028 premature Effects 0.000 claims description 3
- 230000000638 stimulation Effects 0.000 claims description 3
- 230000003313 weakening effect Effects 0.000 claims description 3
- 206010016322 Feeling abnormal Diseases 0.000 claims description 2
- 230000006399 behavior Effects 0.000 description 8
- 230000000694 effects Effects 0.000 description 5
- 238000005457 optimization Methods 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 230000008447 perception Effects 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000007717 exclusion Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a dialogue strategy generation method for simulating user emotion, which comprises the following steps of firstly, establishing a user character and emotion representation system; then, representing the environment trigger event; then, establishing a functional representation of emotional changes in the conversation; and finally, selecting the dialogue action according to the emotion of the user. By simulating more complex emotional states and change processes for people with different characters, the generalization capability of the conversation strategy model is improved, and the conversation system is allowed to make targeted conversation strategies for users with different personalities; the dialog strategy generation method can simulate the user to terminate the dialog at any time in the dialog, and does not always assume the user as an ideal individual which always fits the dialog, so that the simulation behavior which is closer to the actual user than the user simulator of the existing dialog system can be generated, and the dialog strategy for the robot can be trained and evaluated in a simulation framework which is closer to the real environment.
Description
Technical Field
The invention relates to a conversation strategy generation method for simulating user emotion.
Background
The man-machine conversation system can be divided into a single round and a multi-round according to the number of interaction rounds. A single round of dialog is often a traditional question-answering system, where the questions include the necessary information needed to find the answer, and the system gives the answer based on the information in the question. The form of multi-turn dialogue is closer to the real-person interaction habit, but the realization difficulty is also higher. The dialogue system selects a specific action response according to the dialogue state at each moment, the type of the response is limited and can be enumerated, the dialogue system assumes that the state sequence S of a multi-turn dialogue is { S1, S2, S3, … … Sn }, the system correspondingly needs to take the response sequence A { A1, A2, A3 … … An }, and the dialogue strategy is to establish a mapping relation from S to A, so that the dialogue effect is best.
Task-oriented dialog strategy optimization can be generally regarded as supervised learning or reinforcement learning tasks, in a supervised learning-based method, a strategy model is trained to simulate the behavior of an expert, and the method usually needs to be trained by means of a large amount of data marked by domain experts, so that for a specific task domain, expensive and time-consuming data collection and marking are usually needed; furthermore, supervised learning-based approaches lack the ability to explore unknown dialog state spaces, which limits the ability of the policy model to find the best dialog policy.
Based on the reinforcement learning approach, the robot can improve the dialog strategy according to reward signals from the environment without any expert generated examples. Unfortunately, reinforcement learning based methods require a large number of human-computer interaction samples for model optimization. This is too expensive and impractical, especially when cold starting a strategy model, to overcome this problem many researchers use user simulators to train reinforcement learning based conversational robots. The goal of user simulation is to generate natural and reasonable dialogs so that the robot can explore and learn from trajectories that may not exist in the collected observation data, overcoming the major limitations of supervised learning based approaches.
However, the user simulator cannot generate any obvious negative reward (penalty) for the apparently unreasonable behavior of the conversation robot (such as repeatedly inquiring known information, answering questions, etc.), and the human user may show dissatisfaction in the conversation, and a common performance is to terminate the intolerable conversation as early as possible. Many task-oriented dialog strategy models based on neural networks have been studied in the industry to help users complete tasks, but there is still a lack of relevant research concerning user emotion. The main reason is that deep learning based methods typically require large amounts of data for model learning, but existing publicly labeled task-oriented dialog data typically does not contain affective information; meanwhile, the existing user simulator also ignores the simulation of the user emotion, so that it is difficult to effectively study a task-oriented dialog strategy generation method relating to the user emotion.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides a conversation strategy generation method for simulating the emotion of a user.
The purpose of the invention is realized by the following technical scheme:
a conversation strategy generation method for simulating user emotion is characterized in that: the method comprises the following steps:
s101) establishing a user character and emotion representation system;
the character personality and emotion are digitally represented, the personality is divided into five dimensions of openness, responsibility, extroversion, hommization and nervousness according to a five-dimensional personality theory, six emotions of joy, anger, sadness, surprise, fear and disgust are selected as emotional dimensions according to the types of common facial expressions, and the personality P of the user is represented as a one-dimensional vector:
P＝[popening device，pBlame for，pOuter cover，pIt is suitable for,pSpirit of the invention] (1)
Wherein each p in the brackets represents the weight corresponding to each character dimension, and the numeric values are real numbers in the interval of [0,1 ];
the emotional state Et of the user at time t is represented as a one-dimensional vector:
Et＝[Ethappiness，EtAnger，EtSade with，EtSurprise that,EtTerrorism，EtAnaerobic type] (2)
Wherein, each Et in the brackets represents the intensity corresponding to each emotion category, the numeric values of the Et are real numbers in the interval of [0,1], and the sum of the Et is 1;
s102) representing an environment trigger event;
identifying typical events in the conversation process and generating corresponding emotion changes according to the occurrence conditions of the events; the general event factors are:
7) and (3) performing super-many rounds: the goal of task-oriented dialog is to help the user reach the goal with as few turns as possible; lengthy conversations cause user dissatisfaction;
8) and (4) relevant responses: the robot has reasonable response and strong pertinence, so that the emotion of the user tends to be positive and positive;
9) irrelevant responses: if the robot responses are irrelevant to the user words, the user satisfaction is reduced, and the emotion is biased to be negative;
10) repeating the inquiry: if the robot asks again for the information which has been asked and answered, the user satisfaction is reduced, and the emotion is biased to be negative;
11) active interaction: a well-designed robot actively provides information related to tasks, and facilitates the completion of the tasks; such events improve the user experience, with a positive impact on mood;
12) reasonably proposing: when the available resources do not meet the requirements or constraint conditions of the user, the robot provides other choices, and the emotion of the user is changed by the events according to different rationality and feasibility of the choices;
in conjunction with whether the above event occurs, the trigger event ATt at time t is represented as a one-dimensional vector:
ATt＝[Ttultra-much，TtCorrelation，TtIs irrelevant，EtRepetition of,TtActive，TtAdvising] (3)
Wherein each Tt in the brackets indicates whether each corresponding event occurs at the time t, the value is 0 or 1, 0 indicates that the event does not occur, and 1 indicates that the event occurs;
s103) establishing a function representation of emotion change in the conversation;
if only the influence of the user's personality P and the environmental trigger event ATt is considered, the change in the mood of the user at time t, Vt, is theoretically represented as a function of the arguments P and ATt:
Vt＝((P·M)*ATt)·W (4)
wherein M, W is two weight matrices, the dimension of M is the number of trigger event categories x the number of emotion categories, and the dimension of W is the number of user character dimensions x the number of trigger event categories; in the formula, the dot product operation in the matrix operation is represented, and the operation of multiplying corresponding elements of two vectors is represented;
the emotional state of the user gradually changes along with the time, and the emotion of the current moment is always influenced by the emotional state of the previous moment, so that the historical emotional information of the user needs to be considered for updating the emotional state; the emotional change is divided into an external part and an internal part, wherein the external part is caused by the stimulation of the environmental trigger event, the internal part is caused by the gradual weakening and fading of the original emotional state along with the time, and therefore, the emotional state at the moment t is calculated by the following formula:
Et＝Et-1+F(E1:Et-1，P，Vt)+G(Et-1，P) (5)
f is an updating function for calculating the actual change amount of the emotion of the user according to the historical emotional state E1 Et-1, the character P of the user and the current emotional change Vt, and G is a decay function for simulating the influence of internal factors according to the emotional state Et-1 and the character P of the user at the previous moment;
s104) selecting a dialogue action in combination with the user emotion.
Further, in the above dialog strategy generation method for simulating the emotion of the user, in step S103, formula (4) is shown in (4), M is a matrix of 6 × 6, and W is a matrix of 5 × 6.
Further, the dialog strategy generation method for simulating the emotion of the user is described above, wherein in step S103), one linear implementation form of formula (5), F (E1: Et-1, P, Vt), and G (Et-1, P) is:
F(E1:Et-1，P，Vt)＝P·K*Vt (6)
G(Et-1，P)＝[-C1,-C2,……，-Clen(Et)]*Et-1 (7)
wherein K is a weight matrix, and the dimension is the number of character categories multiplied by the number of emotion categories; len (Et) represents the length of the Et vector, i.e., the number of emotions; c1To Clen(Et)The attenuation of each emotion is shown, and the value ranges are [0, 1%]Real numbers within the interval; in the formula, the expression represents dot product operation in matrix operation, and the expression represents multiplication operation of corresponding elements of two vectors.
Further, a dialog strategy generation method for simulating the emotion of the user is described above, wherein K is a 5 × 6 matrix.
Further, in the above method for generating a dialog strategy for simulating a user emotion, step S104), selecting a dialog action in combination with emotion requires a task completion type dialog strategy as a basis, and the main influence of emotion is to take a special action of stopping a dialog in advance when the user feels bad; the basic task completion form is a dialogue strategy implementation form based on a simple rule dialogue strategy, a finite state automaton-based dialogue strategy, a form filling-based dialogue strategy or a reinforcement learning training model; the policies all support the selection of certain system actions by dialog state during the dialog without taking into account emotional impact.
Further, a dialog strategy generation method for simulating a user's emotion is described above, wherein a situation in which a strong negative emotion causes a premature termination of a dialog, i.e. four negative emotion components Et at time t, is simulated in a simple mannerAnger、EtSade with、EtTerrorism、EtAnaerobic typeWhen the sum of the user simulation result exceeds a certain threshold value, the user simulator generates a termination action to finish the conversation in advance; otherwise, the user simulation isAnd adopting system actions of a basic task completion conversation strategy for interaction.
Compared with the prior art, the invention has obvious advantages and beneficial effects, and is embodied in the following aspects:
the invention remarkably improves the generalization ability of a conversation strategy model by simulating more complex emotional states and change processes for people with different characters, and allows a conversation system to make targeted conversation strategies for users with different personalities; the dialog strategy generation method can simulate the user to terminate the dialog at any time in the dialog, and does not always assume the user as an ideal individual which is always matched with the dialog, so that the simulation behavior which is closer to the actual user than the user simulator of the existing dialog system can be generated, and the dialog strategy of the robot can be trained and evaluated in a simulation frame which is closer to the real environment;
the emotion change simulation method is event-driven independent of knowledge in a specific field, has good field adaptability, and can be transferred to other business fields with little change;
the invention also provides a visual angle for evaluating the conversation strategy of the robot, the method of the invention can observe and analyze the optimization process of the conversation strategy of the robot from the perspective of conversation events and emotions, and the training effect is evaluated by the reduction of the negative experience conversation quantity; therefore, the emotion simulation method is creatively introduced into the user simulator of the conversation system, the generalization performance of the robot conversation strategy is improved, high-value conversation samples can be generated in a mode closer to real user behaviors, and meanwhile the method has good field adaptability, is convenient to popularize into the specific vertical industry field and has wide application value in various human-computer conversation systems.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by the practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description.
Drawings
FIG. 1: the invention is a flow diagram.
Detailed Description
In order to more clearly understand the technical features, objects, and effects of the present invention, specific embodiments will now be described in detail.
As shown in fig. 1, a dialog strategy generation method for simulating user emotion specifically includes the following steps:
s101) establishing a user character and emotion representation system;
the users faced by the conversation system are diversified, have different character traits, and can generate different emotions for the response of the robot in the conversation so as to make different responses; simulating the emotion of a user in a conversation, wherein character characters and emotion need to be digitally represented;
typically, the character is divided into five dimensions of openness, responsibility heart, extroversion, humanity and nervousness according to the five-dimensional personality theory; in addition, six emotions, namely joy, anger, sadness, surprise, fear and disgust, are selected as emotional dimensions according to the category of common facial expressions;
the character has stability, and usually does not change in a long period of time; the emotion changes with time in the conversation, and thus the character characters and the emotion are respectively expressed as follows:
the user profile P is represented as a one-dimensional vector:
P＝[popening device，pBlame for，pOuter cover，pIt is suitable for,pSpirit of the invention] (1)
Wherein each p in the brackets represents the weight corresponding to each character dimension, and the numeric values are real numbers in the interval of [0,1 ];
the emotional state Et of the user at time t is represented as a one-dimensional vector:
Et＝[Ethappiness，EtAnger，EtSade with，EtSurprise that,EtTerrorism，EtAnaerobic type] (2)
Wherein, each Et in the brackets represents the intensity corresponding to each emotion category, the numeric values of the Et are real numbers in the interval of [0,1], and the sum of the Et is 1;
s102) representing an environment trigger event;
some events that occur during a conversation have a significant impact on the emotional state of the user, such as: answering questions; in order to simulate the emotion perception of a real user on an external event, a typical event in a conversation process needs to be identified, and corresponding emotion change is generated according to the occurrence condition of the event;
the emotional triggering events in a task-oriented dialog depend on the specific task domain; except for some mutual exclusion cases, it may be assumed that all emotional triggers may occur simultaneously; the general event factors are:
1) and (3) performing super-many rounds: the goal of task-oriented dialog is to help the user reach the goal with as few turns as possible; lengthy conversations can cause user dissatisfaction; the acceptable maximum round limit can vary with the complexity of the task in a specific field, and should be combined with actual selection, such as 15 rounds;
2) and (4) relevant responses: the robot has reasonable response and strong pertinence, and the emotion of the user tends to be positive and positive;
3) irrelevant responses: if the robot responses are irrelevant to the user words, the user satisfaction is reduced, and the emotion is biased to be negative;
4) repeating the inquiry: if the robot asks again for the information which has been asked and answered, the user satisfaction is reduced, and the emotion is biased to be negative;
5) active interaction: the well-designed robot can actively provide information related to tasks, and promote the completion of the tasks; such events improve the user experience, with a positive impact on mood;
6) reasonably proposing: when the available resources do not meet the requirements or constraint conditions of the user, the robot can provide other choices, and the emotion of the user is changed by the events according to different rationality and feasibility of the choices;
in conjunction with whether the above event occurs, the trigger event ATt at time t is represented as a one-dimensional vector:
ATt＝[Ttultra-much，TtCorrelation，TtIs irrelevant，EtRepetition of,TtActive，TtAdvising] (3)
Wherein each Tt in the brackets indicates whether each corresponding event occurs at the time t, and the value is 0 (not occurring) or 1 (occurring);
s103) establishing a function representation of emotion change in the conversation;
if only the influence of the user's personality P and the environmental trigger event ATt is considered, the change in the mood of the user at time t, Vt, is theoretically represented as a function of the arguments P and ATt:
Vt＝((P·M)*ATt)·W (4)
wherein M, W is two weight matrices, the dimension of M is the number of trigger event categories x the number of emotion categories, and the dimension of W is the number of user character dimensions x the number of trigger event categories; m is a 6 x 6 matrix, W is a 5 x 6 matrix, the initial value of the elements is determined according to experience perception, and the subsequent performance of the combination system is finely adjusted and optimized; in the formula, the dot product operation in the matrix operation is represented, and the operation of multiplying corresponding elements of two vectors is represented;
the emotional state of the user generally changes gradually along with the time, and the emotion of the current moment is always influenced by the emotional state of the previous moment; updating the emotional state needs to consider historical emotional information of the user; the emotional change is divided into an external part and an internal part, wherein the external part is caused by the environmental trigger event stimulation, the internal part is caused by the gradual weakening and fading of the original emotional state along with the time, and therefore, more specifically, the emotional state at the moment t is calculated by the following formula:
Et＝Et-1+F(E1:Et-1，P，Vt)+G(Et-1，P) (5)
wherein F is an updating function for calculating the actual change amount of the emotion of the user according to the historical emotional state E1 Et-1, the character P of the user and the current emotional change Vt, and G is a decay function for simulating the influence of internal factors according to the emotional state Et-1 and the character P of the user at the previous moment;
specifically, one linear implementation form of F (E1: Et-1, P, Vt), G (Et-1, P) is:
F(E1:Et-1，P，Vt)＝P·K*Vt (6)
G(Et-1，P)＝[-C1,-C2,……，-Clen(Et)]*Et-1 (7)
wherein K is a weight matrix, and the dimension is the number of character categories multiplied by the number of emotion categories; k is a 5 x 6 matrix, the initial value of the element is determined according to experience perception, and the subsequent performance of the combination system is finely adjusted and optimized; len (Et) represents the length of the Et vector, i.e., the number of emotions; c1To Clen(Et)The attenuation of each emotion is shown, and the value ranges are [0, 1%]Real numbers within the interval; in the formula, the dot product operation in the matrix operation is represented, and the operation of multiplying corresponding elements of two vectors is represented;
s104) selecting the dialogue action according to the emotion of the user
Selecting a conversation action in combination with emotion requires a task completion type conversation strategy as a basis, and the main influence of emotion is to adopt a special action of stopping conversation in advance when a user feels very bad; the basic task completion type dialogue strategy has various implementation forms, and can be a simple rule-based dialogue strategy, a finite state automaton-based dialogue strategy, a form filling-based dialogue strategy or a reinforcement learning training model-based dialogue strategy and the like; the policies all support the selection of certain system actions during a conversation by conversation state, without regard to emotional impact.
In particular, the simulation of the situation in which a strong negative emotion leads to premature termination of the conversation, i.e. four negative emotion components Et at time t, in a simple mannerAnger、EtSade with、EtTerrorism、EtAnaerobic typeWhen the sum exceeds a certain threshold (0.5), the user simulator generates a termination action to end the conversation in advance; otherwise, the user simulation carries out interaction on the system action adopting the basic task completion conversation strategy.
In conclusion, the invention remarkably improves the generalization capability of the conversation strategy model by simulating more complex emotional states and change processes for people with different characters, and allows the conversation system to make specific conversation strategies for users with different personalities; the dialog strategy generation method of the invention can imitate that the user terminates the dialog at any time in the dialog, rather than always assuming the user as an ideal individual which always fits the dialog, so that the simulation behavior which is closer to the actual user than the user simulator of the existing dialog system can be generated, thereby allowing the training and evaluation of the dialog strategy for the robot in the simulation framework which is closer to the real environment.
The emotion change simulation method is event-driven independent of specific field knowledge, has good field adaptability, and can be transferred to other business fields with little change.
The invention also provides a visual angle for evaluating the conversation strategy of the robot, and the method can observe and analyze the optimization process of the conversation strategy of the robot from the perspective of conversation events and emotions and evaluate the training effect by negatively experiencing the reduction condition of the conversation quantity. Therefore, the emotion simulation method is creatively introduced into the user simulator of the conversation system, the generalization performance of the robot conversation strategy is improved, high-value conversation samples can be generated in a mode closer to real user behaviors, and meanwhile the method has good field adaptability, is convenient to popularize into the specific vertical industry field and has wide application value in various human-computer conversation systems.
The invention simulates more complex emotional states and change processes for people with different characters, remarkably improves the generalization capability of a conversation strategy model, and allows a conversation system to make targeted conversation strategies for users with different personalities; the dialog strategy generation method can simulate the user to terminate the dialog at any time in the dialog, and not always presume the user as an ideal individual which is always matched with the dialog, and can generate the simulation behavior which is closer to the actual user than the user simulator of the existing dialog system, thereby allowing the dialog strategy of the robot to be trained and evaluated in a simulation framework which is closer to the real environment. The emotion change simulation method is event-driven independent of specific field knowledge, has good field adaptability, and can be transferred to other business fields with little change. The visual angle of the robot conversation strategy is evaluated, the process of optimizing the robot conversation strategy can be observed and analyzed from the perspective of conversation events and emotions through the method, and the training effect is evaluated through the reduction condition of negative experience conversation quantity. The emotion simulation method is creatively introduced into the user simulator of the conversation system, so that the generalization performance of the robot conversation strategy is improved, a high-value conversation sample can be generated in a mode closer to the real user behavior, and the method has good field adaptability and is convenient to popularize into the specific vertical industry field.
It should be noted that: the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention; while the foregoing is directed to embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.
Claims (6)
1. A conversation strategy generation method for simulating user emotion is characterized in that: the method comprises the following steps:
s101) establishing a user character and emotion representation system;
the character personality and emotion are digitally represented, the personality is divided into five dimensions of openness, responsibility, extroversion, hommization and nervousness according to a five-dimensional personality theory, six emotions of joy, anger, sadness, surprise, fear and disgust are selected as emotional dimensions according to the types of common facial expressions, and the personality P of the user is represented as a one-dimensional vector:
P＝[popening device，pBlame for，pOuter cover，pIt is suitable for,pSpirit of the invention] (1)
Wherein each p in the brackets represents the weight corresponding to each character dimension, and the numeric values are real numbers in the interval of [0,1 ];
the emotional state Et of the user at time t is represented as a one-dimensional vector:
Et＝[Ethappiness，EtAnger，EtSade with，EtSurprise that,EtTerrorism，EtAnaerobic type] (2)
Wherein, each Et in the brackets represents the intensity corresponding to each emotion category, the numeric values of the Et are real numbers in the interval of [0,1], and the sum of the Et is 1;
s102) representing an environment trigger event;
identifying typical events in the conversation process and generating corresponding emotion changes according to the occurrence conditions of the events; the general event factors are:
1) and (3) performing super-many rounds: the goal of task-oriented dialog is to help the user reach the goal with as few turns as possible; lengthy conversations cause user dissatisfaction;
2) and (4) relevant responses: the robot has reasonable response and strong pertinence, so that the emotion of the user tends to be positive and positive;
3) irrelevant responses: if the robot responses are irrelevant to the user words, the user satisfaction is reduced, and the emotion is biased to be negative;
4) repeating the inquiry: if the robot asks again for the information which has been asked and answered, the user satisfaction is reduced, and the emotion is biased to be negative;
5) active interaction: a well-designed robot actively provides information related to tasks, and facilitates the completion of the tasks; such events improve the user experience, with a positive impact on mood;
6) reasonably proposing: when the available resources do not meet the requirements or constraint conditions of the user, the robot provides other choices, and the emotion of the user is changed by the events according to different rationality and feasibility of the choices;
in conjunction with whether the above event occurs, the trigger event ATt at time t is represented as a one-dimensional vector:
ATt＝[Ttultra-much，TtCorrelation，TtIs irrelevant，EtRepetition of,TtActive，TtAdvising] (3)
Wherein each Tt in the brackets indicates whether each corresponding event occurs at the time t, the value is 0 or 1, 0 indicates that the event does not occur, and 1 indicates that the event occurs;
s103) establishing a function representation of emotion change in the conversation;
if only the influence of the user's personality P and the environmental trigger event ATt is considered, the change in the mood of the user at time t, Vt, is theoretically represented as a function of the arguments P and ATt:
Vt＝((P·M)*ATt)·W (4)
wherein M, W is two weight matrices, the dimension of M is the number of trigger event categories x the number of emotion categories, and the dimension of W is the number of user character dimensions x the number of trigger event categories; in the formula, the dot product operation in the matrix operation is represented, and the operation of multiplying corresponding elements of two vectors is represented;
the emotional state of the user gradually changes along with the time, and the emotion of the current moment is always influenced by the emotional state of the previous moment, so that the historical emotional information of the user needs to be considered for updating the emotional state; the emotional change is divided into an external part and an internal part, wherein the external part is caused by the stimulation of the environmental trigger event, the internal part is caused by the gradual weakening and fading of the original emotional state along with the time, and therefore, the emotional state at the moment t is calculated by the following formula:
Et＝Et-1+F(E1:Et-1，P，Vt)+G(Et-1，P) (5)
f is an updating function for calculating the actual change amount of the emotion of the user according to the historical emotional state E1 Et-1, the character P of the user and the current emotional change Vt, and G is a decay function for simulating the influence of internal factors according to the emotional state Et-1 and the character P of the user at the previous moment;
s104) selecting a dialogue action in combination with the user emotion.
2. A conversation strategy generating method simulating emotion of a user according to claim 1, wherein: in step S103), formula (4) in (4), M is a 6 × 6 matrix, and W is a 5 × 6 matrix.
3. A conversation strategy generating method simulating emotion of a user according to claim 1, wherein: step S103), one linear implementation form of formula (5), F (E1: Et-1, P, Vt), G (Et-1, P) is:
F(E1:Et-1，P，Vt)＝P·K* Vt (6)
G(Et-1，P)＝[-C1,-C2,……，-Clen(Et)]*Et-1 (7)
wherein K is a weight matrix, and the dimension is the number of character categories multiplied by the number of emotion categories; len (Et) represents the length of the Et vector, i.e., the number of emotions; c1To Clen(Et)The attenuation of each emotion is shown, and the value ranges are [0, 1%]Real numbers within the interval; in the formula, the expression represents dot product operation in matrix operation, and the expression represents multiplication operation of corresponding elements of two vectors.
4. A conversation strategy generating method simulating emotion of a user according to claim 3, wherein: k is a 5 x 6 matrix.
5. A conversation strategy generating method simulating emotion of a user according to claim 1, wherein: step S104), selecting a conversation action by combining emotion, wherein a task completion type conversation strategy is needed as a basis, and the emotion has the main influence of adopting a special action of stopping conversation in advance when a user feels bad; the basic task completion form is a dialogue strategy implementation form based on a simple rule dialogue strategy, a finite state automaton-based dialogue strategy, a form filling-based dialogue strategy or a reinforcement learning training model; the policies all support the selection of certain system actions by dialog state during the dialog without taking into account emotional impact.
6. A conversation strategy generating method simulating emotion of a user according to claim 5, wherein: simulation of the situation in which a strong negative emotion leads to premature termination of a conversation, in a simple manner, namely the four negative emotion components Et at time tAnger、EtSade with、EtTerrorism、EtAnaerobic typeWhen the sum of the user simulation result exceeds a certain threshold value, the user simulator generates a termination action to finish the conversation in advance; otherwise, the user simulation carries out interaction on the system action adopting the basic task completion conversation strategy.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110312778.6A CN112949857A (en) | 2021-03-24 | 2021-03-24 | Conversation strategy generation method for simulating user emotion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110312778.6A CN112949857A (en) | 2021-03-24 | 2021-03-24 | Conversation strategy generation method for simulating user emotion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112949857A true CN112949857A (en) | 2021-06-11 |
Family
ID=76228478
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110312778.6A Withdrawn CN112949857A (en) | 2021-03-24 | 2021-03-24 | Conversation strategy generation method for simulating user emotion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112949857A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115457645A (en) * | 2022-11-11 | 2022-12-09 | 青岛网信信息科技有限公司 | User emotion analysis method, medium and system based on interactive verification |
| CN115934909A (en) * | 2022-12-02 | 2023-04-07 | 苏州复变医疗科技有限公司 | Common situation reply generation method, device, terminal and storage medium |
-
2021
- 2021-03-24 CN CN202110312778.6A patent/CN112949857A/en not_active Withdrawn
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115457645A (en) * | 2022-11-11 | 2022-12-09 | 青岛网信信息科技有限公司 | User emotion analysis method, medium and system based on interactive verification |
| CN115934909A (en) * | 2022-12-02 | 2023-04-07 | 苏州复变医疗科技有限公司 | Common situation reply generation method, device, terminal and storage medium |
| CN115934909B (en) * | 2022-12-02 | 2023-11-17 | 苏州复变医疗科技有限公司 | Co-emotion reply generation method and device, terminal and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ritter et al. | Techniques for modeling human performance in synthetic environments: A supplementary review | |
| Ritter et al. | Embodied models as simulated users: Introduction to this special issue on using cognitive models to improve interface design | |
| CN112949857A (en) | Conversation strategy generation method for simulating user emotion | |
| Davidovitch et al. | Simulation-based learning: The learning–forgetting–relearning process and impact of learning history | |
| Woolf | Student modeling | |
| Schaller | The empirical benefits of conceptual rigor: Systematic articulation of conceptual hypotheses can reduce the risk of non-replicable results (and facilitate novel discoveries too) | |
| Hyatt et al. | Computational organizational network modeling: Strategies and an example | |
| Kazemifard et al. | Design and implementation of GEmA: A generic emotional agent | |
| Carmona et al. | Introducing prerequisite relations in a multi-layered Bayesian student model | |
| Dorsey et al. | Mathematical modeling of decision making: a soft and fuzzy approach to capturing hard decisions | |
| Li et al. | Social interaction for efficient agent learning from human reward | |
| Waeselynck et al. | System dynamics applied to the modelling of software projects | |
| Mislevy et al. | Modeling conditional probabilities in complex educational assessments | |
| Yang et al. | Modeling emotional action for social characters | |
| CN115205072A (en) | Cognitive diagnosis method for long-period evaluation | |
| Chao et al. | Impact of cognitive abilities of experts on the effectiveness of elicited knowledge | |
| de Freitas et al. | Emotion in artificial intelligence and artificial life research: Facing problems | |
| Kim | Procedural skills: From learning to forgetting | |
| Naganathan et al. | Intelligent Tutoring System: Predicting Students Results Using Neural Networks. | |
| Blosch et al. | Experimental design and computer‐based simulation: a case study with the Royal Navy | |
| Stearns et al. | Toward Unifying Cognitive Architecture and Neural Task Set Theories. | |
| Gatare et al. | Co-designing nudges for self directed learning within GOAL system | |
| Rank et al. | Comparability is key to assess affective architectures | |
| Wood et al. | 12 Functional Approaches to Representing the Interplay of Situations, Persons, and Behavior | |
| Strachan | Pragmatic User Modelling for Complex Knowledge-Based Systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20210611 |