CN108846073A - A kind of man-machine emotion conversational system of personalization - Google Patents

Abstract

The invention discloses a personalized human-machine emotional conversation system. The personalized human-computer emotional conversation system includes: a personalized personality modeling and dynamic analysis module, a conversation management module, and a response comfort degree detection and response strategy adjustment module. Personalized personality modeling and dynamic analysis module includes: conversational knowledge base and psychological knowledge base, personalized personality modeling and dynamic analysis method; conversational management module includes the following steps: 1. Semantic analysis of user input; 2. Semantic analysis 3. Develop response strategies based on reasoning results; 4. Generate conversational content guided by response strategies; response comfort detection and response strategy adjustment modules include: emotional dynamics and psychological knowledge base, response comfort detection and Response policy adjustment method. The present invention is suitable for human-computer interaction tasks, and incorporates personalized personality modeling analysis and emotional conversation services, thereby improving the personalized natural interaction experience of the human-computer conversation system.

Description

A Personalized Human-Machine Emotional Conversation System

technical field

The invention relates to the technical fields of artificial intelligence, natural language understanding, emotional computing and human-computer interaction, in particular to a personalized human-computer emotional conversation system.

Background technique

Artificial intelligence tries to understand the essence of intelligence. By simulating human consciousness and thinking, it endows machines with similar capabilities to human intelligence to extend and expand human intelligence, so that machines can complete tasks that usually require human intelligence. Natural language processing is the "crown jewel" of artificial intelligence. It mainly studies various theories and methods that can realize effective communication between humans and computers using natural language. Affective computing takes emotion as the research object, aims at endowing computers with the ability to perceive, understand and express emotions, and aims at building a harmonious human-machine environment. Human-computer interaction mainly studies the interactive relationship between users and systems, which is embodied in the information exchange process between humans and computers to complete a certain task in a certain interactive way.

Traditional human-machine conversational systems ignore the relationship between emotional factors and conversational responses during conversations, and often only focus on factual conversations; and do not consider user personalization factors, and cannot generate personalized responses based on user personality characteristics; Generalization between domains can only complete specific tasks in specific domains; and the generation of conversational responses mainly relies on information retrieval technology, which cannot generate conversational responses that have never appeared in the knowledge base.

Contents of the invention

The purpose of the present invention is to provide a personalized human-machine emotional conversation system in order to remedy the defects of the prior art.

The present invention is achieved through the following technical solutions:

A personalized human-machine emotional conversation system, including personalized personality modeling and dynamic analysis module CD, conversation management module DM and response comfort level detection and response strategy adjustment module CA, described conversation management module DM according to user input Dialogue content automatically generates a conversational response with emotion; the personalized personality modeling and dynamic analysis module CD stores man-machine dialogue history data, and analyzes the user's long-term personality characteristics and short-term emotion transfer characteristics according to the man-machine dialogue history data, Then formulate the man-machine conversation response strategy; the response comfort detection and response strategy adjustment module CA detects the user's emotional fluctuations, and adjusts the man-machine conversation response strategy.

Use formula (1) to describe the personalized human-machine emotional dialogue system PEDS as a five-tuple:

PEDS＝(Input, CD, DM, CA, Output) (1)

In formula (1), Input represents the user input of described personalized human-computer emotion dialogue system, and Output represents the system output of described personalized human-computer emotional interaction system,

The composition of the conversation management module DM includes: semantic analysis, semantic reasoning, response strategy and conversation generation;

The composition of described individualized personality modeling and dynamic analysis module CD comprises: conversational knowledge base and psychological knowledge base, individualized personality modeling and dynamic analysis method;

The composition of the conversational knowledge base and the psychological knowledge base includes: a dynamic knowledge base, a conversational history base and a psychological knowledge base;

The CA composition of the response comfort detection and response strategy adjustment module includes: emotional dynamics and psychological knowledge base, response comfort detection and response strategy adjustment method;

The emotional dynamics and psychological knowledge base include: a dynamic psychological knowledge base, an emotional dynamics knowledge base and a static psychological knowledge base.

The session management module DM is carried out as follows:

Step 1. The dialog management module DM receives the user input session content, and obtains the semantic representation SP of the user input session content;

Step 2. Perform semantic analysis on the semantic representation SP input by the user, and obtain the user's conversational context information C, conversational topic T, emotional object O, emotional category E, emotional intensity S, and conversational intent I;

Step 2.1, analyzing the conversational context in the semantic analysis to obtain the conversational context information C;

Step 2.2. After obtaining the context information, analyze and obtain the conversation topic information T in combination with the user input;

Step 2.3, analyzing the attributes of the conversation topic in the conversation topic, and determining the emotional object O in the conversation topic;

Step 2.4. Combining the conversation context and the current user input Q, analyze the emotion of the emotional object in the user input, and determine the user's emotional category E and emotional intensity S for the emotional object;

Utilize formula (2) to represent the collection of described emotion category E:

{E} = {peace, joy, love, surprise, anxiety, sadness, anger, hatred, anticipation} (2)

Step 2.5. According to the steps 2.1 to 2.4, analyze the conversational intention I of the user;

Utilize formula (3) to express described session intent I acquisition process:

I=f(C,T,O,E,S) (3)

The semantic analysis is described as a six-tuple SA by using formula (4):

SA=(I,C,T,O,E,S) (4)

Step 3, after obtaining the conversational intent I, the semantic reasoning is based on the conversational context information C, the conversational topic T, the emotional object O, the emotional category E, the emotional intensity S and the conversational intent I premise in the step 2, and carry out logic Reasoning and judging to obtain the deep semantic conclusion Res input by the user;

Use formula (5) to represent the semantic reasoning process shown:

(I,C,T,O,E,S)→Res (5)

Step 4, the response strategy formulates a strategy for guiding conversation generation according to the deep semantic conclusion Res, and the response strategy formulation process is affected by the following factors:

Factor 1, the conversation context information C in the step 2, the conversation topic T, the emotion object O, the emotion category E, the emotion intensity S, the conversation intention I and the deep semantic conclusion Res in the step 3;

Factor 2, the long-term emotional model of the personalized personality modeling and dynamic analysis module CD, which is expressed as personalized personality;

Factor 3, the short-term emotional model of the personalized personality modeling and dynamic analysis module CD, which is expressed as the probability of emotional transfer;

Factor 4. The response strategy adjustment method of the response comfort detection and response strategy adjustment module CA;

Step 5, the session generation generates a session response A under the guidance of the response policy, and the session generation is performed as follows:

Case 1. If the user input is factual question and answer content, the conversation generation matches and retrieves existing knowledge in the conversation knowledge base, and generates the conversation answer A;

Case 2. If the user input is chat content, the conversation generation generates the conversation response A according to the conversation context information C and the current user input Q;

Case 3. If the user input is emotional conversation content, the conversation generation generates the conversation response A with emotion guidance function under the guidance of the response strategy;

Step 6, the conversation response A is output by the system for human-computer interaction and feedback;

Step 7: The user input Q and the conversation response A are stored in a conversation history database.

Described individualized personality modeling and dynamic analysis method described in described individualized personality modeling and dynamic analysis module CD are to carry out as follows:

Step (1), in the step 2, utilize the semantic analysis to analyze the user input in the conversation history storehouse to obtain the emotion category of the user input;

Step (2), statistically calculate the transition probability relationship between the emotional categories, calculate the probability of transferring from one emotional category to another emotional category, and then obtain the emotional state transition probability matrix M of the emotional category, to represent said short-term affective model;

Utilize formula (6) to express described emotional state transition probability matrix M:

Among them, p _ij represents the probability of transferring from emotional state i to emotional state j, and p _ii represents the probability that emotional state remains unchanged;

Step (3), using the emotional state transition probability matrix as a feature, carry out personalized personality modeling, and cooperate with the Big Five personality scale to determine the corresponding relationship between the emotional state transition probability matrix M and the Big Five personality F, M→F , used to represent the long-term emotional model;

Utilize formula (7) to express the collection of described big five personality F:

{F} = {openness, conscientiousness, extraversion, agreeableness, neuroticism} (7)

Step (4), according to the personality five factors of described big five personality F, analyze the user's personal preferences and conversation characteristics;

Step (5), according to the user's personal preferences and conversation features and the emotional state transition probability matrix M, in conjunction with the response comfort detection and strategy adjustment module CA, formulate and fine-tune the response strategy with emotion guidance, Guiding the conversation content and emotional tendency;

Step (6), using the emotional guidance function to guide the conversation content and emotional tendency, and then update the emotional state transition probability in the emotional state transition probability matrix.

The response comfort detection and strategy adjustment method described in the response comfort detection and strategy adjustment module CA is performed in the following steps:

Semantic analysis described in step 1), step 2, obtains described emotion category E _t , E _t-1 , E _t-2 of user t time, t-1 time, t-2 time...tn time series input content ...E _tn and the emotional intensity S _t , S _t-1 , S _t-2 ...S _tn ;

Step 2), according to the transfer relationship of the emotional category in the time series in the step 1), and the fluctuation of the emotional intensity, calculate the response comfort;

Equation (8) is used to define the emotional energy function to characterize the emotional energy intensity EE in the time series:

EE=f(E,S) (8)

Wherein, if the emotion category E∈{happiness, love, expectation}, the value of the emotion energy function is positive; if the emotion category E∈{anxiety, sadness, anger, hatred, surprise}, the emotion The value of the energy function is negative; if the emotional category E∈{calm}, the value of the emotional energy function is 0;

Utilize formula (9) to define the zero-crossing rate R of described emotion intensity in described time series:

R=N/n (9)

Wherein, N represents the positive and negative conversion times of sentiment function value in described time series, and n represents described time series length;

Use formula (10) to define the rate of change of the emotional energy function D, to represent the changing trend of emotional energy:

Using formula (11) to express the response comfort level as a triplet CL:

CL=(EE,R,D) (11)

Step 3), detecting an abnormal emotional state according to the response comfort level, and adjusting the response strategy based on the user input, the system output and the emotional dynamics and psychological knowledge base;

According to the emotional guidance function, focus on adjusting the response strategy for the following situations:

Case 1), the emotional energy function value in the time series is negative for a long time, which is characterized by the negative emotion of the emotional state for a long time;

Case 2), the value of the emotional energy function in the time series changes from positive to negative, which is characterized by the transfer of the emotional type from positive emotion to negative emotion;

Case 3), the zero-crossing rate R in the time series is higher than a certain set threshold, which is characterized by an unstable emotional state;

Case 4), the rate of change of the emotional energy function in the time series is negative, which is characterized by a tendency for the emotional state to shift to negative emotion or to increase the intensity of negative emotion.

The advantages of the present invention are:

1. The present invention introduces emotional factors into the man-machine conversation system by extracting the emotional semantic information input by the user, which overcomes the defect of ignoring the emotional factors in the prior art, thereby making the human-computer interaction more natural and harmonious, and the application field is wider.

2. In the present invention, through the personalized personality modeling and dynamic analysis module, the conversation system can carry out personalized modeling and analysis of user personality, thereby generating personalized conversation responses, and then can carry out personalized emotional guidance for specific user groups.

3. The present invention can detect the quality of the system response in the last round of conversation and improve the quality of the conversation response by adjusting the response strategy of the next round through the response comfort detection and response strategy adjustment module, which can solve the problem of multiple rounds of emotional conversation.

4. The conversation management module of the present invention uses a generative model to generate conversation responses other than the knowledge base, so that the system is suitable for human-machine emotional conversation in the open field, and realizes the evolution of the emotional conversation system through the accumulation of knowledge.

Description of drawings

Fig. 1 is the overall structure and flow diagram of the present invention.

Detailed ways

As shown in Figure 1, a personalized human-computer emotional conversation system is suitable for human-computer interaction tasks, including a personalized personality modeling and dynamic analysis module CD, a conversation management module DM, and a response comfort detection and response strategy adjustment module CA, the conversation management module DM automatically generates a conversation response with emotion according to the dialogue content input by the user; the personalized personality modeling and dynamic analysis module CD stores the man-machine conversation history data, according to the man-machine conversation The historical data analyzes the user's long-term personality characteristics and short-term emotional transfer characteristics, and then formulates a human-computer conversation response strategy; the response comfort detection and response strategy adjustment module CA detects user emotional fluctuations, and adjusts the human-computer conversation response strategy.

Use formula (1) to describe the personalized human-machine emotional dialogue system PEDS as a five-tuple:

PEDS＝(Input, CD, DM, CA, Output) (1)

In formula (1), Input represents the user input of described personalized human-computer emotion dialogue system, and Output represents the system output of described personalized human-computer emotional interaction system,

The composition of the conversation management module DM includes: semantic analysis, semantic reasoning, response strategy and conversation generation;

The composition of described individualized personality modeling and dynamic analysis module CD comprises: conversational knowledge base and psychological knowledge base, individualized personality modeling and dynamic analysis method;

The composition of the conversational knowledge base and the psychological knowledge base includes: a dynamic knowledge base, a conversational history base and a psychological knowledge base;

The CA composition of the response comfort detection and response strategy adjustment module includes: emotional dynamics and psychological knowledge base, response comfort detection and response strategy adjustment method;

The emotional dynamics and psychological knowledge base include: a dynamic psychological knowledge base, an emotional dynamics knowledge base and a static psychological knowledge base.

The session management module DM is carried out as follows:

Step 1. The dialogue management module DM receives the user input conversation content, uses the pre-training word vector connection mode, considers the word order relationship in the sentence, and obtains the semantic representation SP of the user input conversation content;

Step 2. Perform semantic analysis on the semantic representation SP input by the user, and obtain the user's conversational context information C, conversational topic T, emotional object O, emotional category E, emotional intensity S, and conversational intent I; assuming that the user inputs For: Real Madrid played too badly this time! After step 2, the context C is sports, the conversation topic T is football, the emotion object O is Real Madrid, the emotion category E is anger, the emotion intensity is 0.8, and the conversation intention I is complaining;

Step 2.1, analyzing the conversational context in the semantic analysis to obtain the conversational context information C;

Step 2.2. After obtaining the context information, analyze and obtain the conversation topic information T in combination with the user input;

Step 2.3, analyzing the attributes of the conversation topic in the conversation topic, and determining the emotional object O in the conversation topic;

Step 2.4. Combining the conversation context and the current user input Q, analyze the emotion of the emotional object in the user input, and determine the user's emotional category E and emotional intensity S for the emotional object;

Utilize formula (2) to represent the collection of described emotion category E:

{E} = {peace, joy, love, surprise, anxiety, sadness, anger, hatred, anticipation} (2)

Step 2.5. According to the steps 2.1 to 2.4, analyze the conversational intention I of the user;

Utilize formula (3) to express described session intent I acquisition process:

I=f(C,T,O,E,S) (3)

The semantic analysis is described as a six-tuple SA by using formula (4):

SA=(I,C,T,O,E,S) (4)

Step 3, after obtaining the conversational intent I, the semantic reasoning is based on the conversational context information C, the conversational topic T, the emotional object O, the emotional category E, the emotional intensity S and the conversational intent I premise in the step 2, and carry out logic Reasoning and judging to obtain the deep semantic conclusion Res input by the user; the deep semantic conclusion Res is expressed in the form of hidden vector in the deep neural network, such as (0.24,0.62,0.71,0.12,...,0.56,0.98,0.46) , can be regarded as an expression of the user's inner activities. If the user is a fan of Real Madrid, then the user's heart may be in a state of "hate iron but not steel" at this time.

Use formula (5) to represent the semantic reasoning process shown:

(I,C,T,O,E,S)→Res (5)

Step 4, the response strategy formulates a strategy for guiding conversation generation according to the deep semantic conclusion Res, and the response strategy formulation process is affected by the following factors:

Factor 1, the conversation context information C in the step 2, the conversation topic T, the emotion object O, the emotion category E, the emotion intensity S, the conversation intention I and the deep semantic conclusion Res in the step 3;

Factor 2, the long-term emotional model of the personalized personality modeling and dynamic analysis module CD, manifested as personalized personality; generally, people's behavior and expression will be affected by personal personality, if the user is a neurotic personality, Then his behavior and language expression are mostly anxious, hostile, impulsive, etc.; and if the user is an extraverted personality, then his behavior and language expression are mostly objective, decisive, and enthusiastic;

Factor 3, the short-term emotional model of the personalized personality modeling and dynamic analysis module CD, which is expressed as the probability of emotional transfer; the probability of emotional transfer is affected by multiple factors such as personal personality, hobbies, and psychological states; different users transfer from positive to emotional The probability of reaching negative emotions, or transferring negative emotions to positive emotions is different, and the time and process experienced are also different;

Factor 4, the response strategy adjustment method of the response comfort detection and response strategy adjustment module CA; this factor can adjust the conversation response strategy by detecting the user's emotional fluctuations under the pre-order response condition, and then guide the user through the next round of conversation response Emotional transfer;

Step 5, the session generation generates a session response A under the guidance of the response policy, and the session generation is performed as follows:

Situation 1. If the user input is factual question and answer content, the conversation generation matches and retrieves existing knowledge in the conversation knowledge base, and generates the conversation answer A; if the user input content is: "The capital of China is Where", the conversation system will search the knowledge base and get the conversation response "The capital of China is Beijing";

Situation 2. If the user input is chat content, the conversation generation generates the conversation answer A according to the conversation context information C and the current user input Q; if the user input is: "Look at Real Madrid last night Did you play the game?", then the dialogue system can answer: "I watched" or "I didn't watch" according to the context information;

Situation 3: If the user input is conversational content with emotion, the conversational generation generates the conversational response A with emotion guiding function under the guidance of the response strategy; It’s raining again, so we can only stay at home!”, then the conversational system can generate: “It’s okay, we can watch movies at home, it’s good” (positive emotional guidance), or “Yeah, the weather is really terrible "(Negative emotional guidance). The invention focuses more on the guidance of positive emotions;

Step 6, the conversation response A is output by the system for human-computer interaction and feedback;

Step 7: The user input Q and the conversation response A are stored in a conversation history database.

The personalized personality modeling and dynamic analysis module stores the historical data of man-machine dialogue, analyzes the user's long-term personality characteristics and short-term emotional transfer characteristics according to the historical data of man-machine dialogue, and then formulates the man-machine dialogue response strategy. Personalized personality modeling and dynamic analysis methods are carried out in the following steps:

Step (1), in the step 2, utilize the semantic analysis to analyze the user input in the conversation history storehouse to obtain the emotion category of the user input;

Step (2), statistically calculate the transition probability relationship between the emotional categories, calculate the probability of transferring from one emotional category to another emotional category, and then obtain the emotional state transition probability matrix M of the emotional category, to represent said short-term affective model;

Utilize formula (6) to express described emotional state transition probability matrix M:

Among them, p _ij represents the probability of transferring from emotional state i to emotional state j, and p _ii represents the probability that emotional state remains unchanged;

Step (3), using the emotional state transition probability matrix as a feature, carry out personalized personality modeling, and cooperate with the Big Five personality scale to determine the corresponding relationship between the emotional state transition probability matrix M and the Big Five personality F, M→F , used to represent the long-term emotional model; if it is found through probability calculation that a certain user always transfers from positive emotion to negative emotion with a high probability or the emotional fluctuation is relatively large, then the probability of the user belonging to neurotic personality is relatively large;

Utilize formula (7) to express the collection of described big five personality F:

{F} = {openness, conscientiousness, extraversion, agreeableness, neuroticism} (7)

Step (4), according to the personality five factors of described big five personality F, analyze the user's personal preferences and conversation characteristics;

Step (5), according to the user's personal preferences and conversation features and the emotional state transition probability matrix M, in conjunction with the response comfort detection and strategy adjustment module CA, formulate and fine-tune the response strategy with emotion guidance, Guide the conversation content and emotional tendency; the user's personal preferences will affect the user's emotional transfer probability, and even if the things he likes initially are negative emotions, they will quickly transfer to positive emotions; on the contrary, if they are Things that the user dislikes will be transferred to negative emotions with a higher probability;

Step (6), using the emotional guidance function to guide the conversation content and emotional tendency, and then update the emotional state transition probability in the emotional state transition probability matrix.

The response comfort detection and strategy adjustment method described in the response comfort detection and strategy adjustment module CA is performed in the following steps:

Semantic analysis described in step 1), step 2, obtains described emotion category E _t , E _t-1 , E _t-2 of user t time, t-1 time, t-2 time...tn time series input content ...E _tn and the emotional intensity S _t , S _t-1 , S _t-2 ...S _tn ;

Step 2), according to the transfer relationship of the emotional category in the time series in the step 1), and the fluctuation of the emotional intensity, calculate the response comfort;

Equation (8) is used to define the emotional energy function to characterize the emotional energy intensity EE in the time series:

EE=f(E,S) (8)

Wherein, if the emotion category E∈{happiness, love, expectation}, the value of the emotion energy function is positive; if the emotion category E∈{anxiety, sadness, anger, hatred, surprise}, the emotion The energy function value is negative; if the emotional category E∈{calm}, the emotional energy function value is 0; the emotional intensity can be positive or negative, and if the positive emotional intensity is greater, the user's emotional state is more positive; If the negative emotional intensity is smaller, it means that the user's emotional state is more negative; if the emotional intensity is close to 0, it means that the user's emotional state tends to be calm;

Use formula (9) to define the zero-crossing rate R of the emotional intensity in the time series: the greater the zero-crossing rate, the more frequent the user's emotional fluctuations during this period, indicating that the emotional instability:

R=N/n (9)

Wherein, N represents the positive and negative conversion times of sentiment function value in described time series, and n represents described time series length;

Use formula (10) to define the rate of change of the emotional energy function D, to represent the changing trend of emotional energy:

The change rate of the emotional energy function is positive, indicating that the value of the emotional energy function increases, showing that the emotional state tends to be positive; the rate of change of the emotional energy function is Fu, indicating that the value of the emotional energy function decreases, showing that the emotional state tends to be negative. ; The rate of change of the emotional energy function is zero, indicating that the value of the emotional energy function remains unchanged, showing a calm emotional state;

Using formula (11) to express the response comfort level as a triplet CL:

CL=(EE,R,D)(11)

Step 3), detecting an abnormal emotional state according to the response comfort level, and adjusting the response strategy based on the user input, the system output and the emotional dynamics and psychological knowledge base;

According to the emotional guidance function, focus on adjusting the response strategy for the following situations:

Case 1), the emotional energy function value in the time series is negative for a long time, which is characterized by the negative emotion of the emotional state for a long time;

Case 2), the value of the emotional energy function in the time series changes from positive to negative, which is characterized by the transfer of the emotional type from positive emotion to negative emotion;

Case 3), the zero-crossing rate R in the time series is higher than a certain set threshold, which is characterized by an unstable emotional state;

Case 4), the rate of change of the emotional energy function in the time series is negative, which is characterized by a tendency for the emotional state to shift to negative emotion or to increase the intensity of negative emotion.

Claims (5)

1. a kind of man-machine emotion conversational system of personalization, it is characterised in that：It include personalized personality modeling and dynamic analysis mould Block CD, session management module DM and the detection of response comfort level and acknowledgment strategy adjustment module CA, the session management module DM The session response with emotion is automatically generated according to the conversation content that user inputs；The modeling of personalized personality and dynamic point It analyses module CD and stores human-computer dialogue historical data, according to the long-term personality characteristics of the human-computer dialogue historical data analysis user and short Phase transference feature, and then formulate man-machine conversation's acknowledgment strategy；The response comfort level detection adjusts mould with acknowledgment strategy Block CA detects user feeling fluctuation, adjusts man-machine conversation's acknowledgment strategy.

2. the man-machine emotion conversational system of a kind of personalization according to claim 1, it is characterised in that：It will be a using formula (1) The man-machine emotion conversational system PEDS of propertyization is described as a five-tuple：

PEDS=(Input, CD, DM, CA, Output) (1)

In formula (1), Input table shows user's input of the man-machine emotion conversational system of the personalization, and Output indicates the individual character Change the system output of Human-Machine Emotion Interactive System,

The composition of the session management module DM includes：Semantic analysis, semantic reasoning, acknowledgment strategy and session generate；

The personalized personality models and the composition of dynamic analysis module CD includes：It is session knowledge base and psychological knowledge base, a Property personality modeling and dynamic analysing method；

The composition of the session knowledge base and psychological knowledge base includes：Dynamic repository, conversation history library and psychological knowledge base；

The response comfort level detection is formed with acknowledgment strategy adjustment module CA：Emotion dynamics and psychological knowledge base, The detection of response comfort level and acknowledgment strategy method of adjustment；

The composition of the emotion dynamics and psychological knowledge base includes：The psychological knowledge base of dynamic, emotion dynamics knowledge base and quiet State psychology knowledge base.

3. the man-machine emotion conversational system of a kind of personalization according to claim 2, it is characterised in that：The session management mould Block DM is carried out as follows：

Step 1, dialogue management module DM receive user and input session content, obtain the semantic table that the user inputs session content Show SP；

Step 2, to the user input the semantic expressiveness SP carry out semantic analysis, obtain user session language ambience information C, Session theme T, emotion object O, emotional category E, emotional intensity S and session are intended to I；

Session context is analyzed in step 2.1, the semantic analysis, obtains session language ambience information C；

Step 2.2 after obtaining the language ambience information, is inputted in conjunction with the user, and analysis obtains the session subject information T；

Step 2.3, the attribute that the session theme is analyzed in the session theme, determine the feelings in the session theme Feel object O；

Step 2.4 inputs Q in conjunction with the session context and active user, to being directed to the emotion pair in user input The emotion of elephant is analyzed, and determines user to the emotional category E and the emotional intensity S of the emotion object；

The set of the emotional category E is indicated using formula (2)：

{ E }=and it is tranquil, it is glad, like, surprised, anxiety, it is sad, it is angry, hate, expect (2)

Step 2.5, according to step 2.1~2.4, analyze session described in user and be intended to I；

Indicate that the session is intended to I acquisition process using formula (3)：

I=f (C, T, O, E, S) (3)

The semantic analysis is described as a hexa-atomic group of SA using formula (4)：

SA=(I, C, T, O, E, S) (4)

Step 3, after obtaining the session and being intended to I, the semantic reasoning is based on session language ambience information C, session master described in step 2 It inscribes T, emotion object O, emotional category E, emotional intensity S and session and is intended to I premise, carry out reasoning from logic and judgement, obtain described The Deep Semantics conclusion Res of user's input；

Shown semantic reasoning process is indicated using formula (5)：

(I,C,T,O,E,S)→Res (5)

Step 4, the acknowledgment strategy formulate the strategy for instructing session to generate according to the Deep Semantics conclusion Res, described Acknowledgment strategy formulation process is influenced by following factor：

Session language ambience information C, session theme T described in factor 1, the step 2, emotion object O, emotional category E, emotional intensity S, Session is intended to Deep Semantics conclusion Res described in I and the step 3；

Factor 2, the personalized personality model the long-term emotion model with dynamic analysis module CD, show as personalized personality；

Factor 3, the personalized personality model the short-term emotion model with dynamic analysis module CD, and it is general to show as transference Rate；

Factor 4, the acknowledgment strategy method of adjustment of response comfort level detection and acknowledgment strategy adjustment module CA；

Step 5, the session generate and generate session response A under acknowledgment strategy guidance, and the session generation is by as follows Situation carries out：

If situation 1, user input are fact question and answer content, the session is generated matches inspection in the session knowledge base Rope existing knowledge, and generate the session response A；

If situation 2, user input are chat property content, the session is generated according to the session language ambience information C and described Active user inputs Q and generates the session response A；

If situation 3, user input are the session content with emotion, the session is generated under acknowledgment strategy guidance Generate the session response A with emotion guiding function；

Step 6, the session response A are exported through the system carries out human-computer interaction and feedback；

Step 7, the user input Q and the session response A is stored to conversation history library.

4. the man-machine emotion conversational system of a kind of personalization according to claim 3, it is characterised in that：The personalization personality It is to carry out as follows that modeling, which is modeled with personalization personality described in dynamic analysis module CD with dynamic analysing method,：

The input of user described in the conversation history library is divided using the semantic analysis in step (1), the step 2 Analysis obtains the emotional category of user's input；

Step (2), statistics calculate the transition probability relationship between the emotional category, and it is another for calculating from a kind of transfer of emotional category The probability of kind emotional category, and then the affective state transition probability matrix M of the emotional category is obtained, to indicate described short-term Emotion model；

The affective state transition probability matrix M is indicated using formula (6):

Wherein, p_ijIndicate the probability that affective state j is transferred to by affective state i, p_iiIndicate the probability that affective state remains unchanged；

Step (3), characterized by the affective state transition probability matrix, carry out personalized personality modeling, cooperate five-factor model personality Scale determines the corresponding relationship of the affective state transition probability matrix M and five-factor model personality F, M → F, to indicate described long-term Emotion model；

The set of the five-factor model personality F is indicated using formula (7)：

{ F }={ open, sense of responsibility, extroversion, pleasant property are neurotic } (7)

Step (4), five factor of personality according to the five-factor model personality F, analyze the personal preference and session characteristics of user；

Step (5), according to individual subscriber hobby and session characteristics and the affective state transition probability matrix M, joint The response comfort level detection and Developing Tactics module CA, formulate the acknowledgment strategy with emotion guiding function and finely tune, The session content and emotion tendency are guided；

Step (6), using the emotion guiding function, the session content and emotion tendency are guided, so update Affective state transition probability described in the affective state transition probability matrix.

5. the man-machine emotion conversational system of a kind of personalization according to claim 4, it is characterised in that：The response comfort level It is to carry out as follows that detection, which is detected with response comfort level described in Developing Tactics module CA with strategy adjusting method,：

Semantic analysis described in step 1), step 2 obtains user's t moment, t-1 moment, t-2 moment ... t-n time sequence The emotional category E of input content_t、E_t-1、E_t-2…E_t-nAnd the emotional intensity S_t、S_t-1、S_t-2…S_t-n；

The transfer relationship of emotional category described in step 2), the time series according to the step 1) and the emotion The fluctuating change of intensity calculates the response comfort level；

Emotional energy function is defined using formula (8), characterizes emotional energy intensity EE in the time series：

EE=f (E, S) (8)

Wherein, if the emotional category E ∈ { glad, to like, expect }, then the emotional energy functional value is positive；If the feelings Feel classification E ∈ { anxiety, sad, angry, hatred are surprised }, then the emotional energy functional value is negative；If the emotional category E ∈ { tranquil }, then the emotional energy functional value is 0；

The zero-crossing rate R of the emotional intensity in the time series is defined using formula (9)：

R=N/n (9)

Wherein, N indicates the positive and negative conversion times of emotion functional value in the time series, and n indicates the length of time series；

Emotional energy function change rate D is defined using formula (10), characterizes the variation tendency of emotional energy：

The response comfort level is expressed as a triple CL using formula (11)：

CL=(EE, R, D) (11)

Step 3) detects abnormal emotion state according to the response comfort level, and based on user input, system output And the emotion dynamics is adjusted acknowledgment strategy with psychological knowledge base；

According to the emotion guiding function, emphasis carries out acknowledgment strategy adjustment to following situation：

Situation 1), the emotional energy functional value is negative for a long time in the time series, be characterized as the affective state and locate for a long time In negative sense emotion；

Situation 2), the emotional energy functional value is born by being just kept in the time series, be characterized as the affective style by just It is negative sense emotion to transference；

Situation 3), the zero-crossing rate R is higher than a certain given threshold in the time series, it is unstable to be characterized as affective state；

Situation 4), the emotional energy function change rate is negative value in the time series, being characterized as affective state has the transfer to be Negative sense emotion or the increased trend of negative sense emotional intensity.