JP6343823B2 - Emotion estimation system by machine dialogue and its program - Google Patents

Description

  The present invention relates to a dialogue system for dialogue between a person and a computer and a dialogue program that operates the dialogue system.

  In recent years, an increasing number of patients are anxious about mental health, which is in a mood state, and accordingly, mental health care services that lead mental health to a good state are attracting attention. Usually, the diagnosis target person visits the therapist, the therapist interviews the diagnosis target person, and the therapist diagnoses and conveys the result. However, this diagnosis result greatly depends on the knowledge and experience value of the therapist, and it takes time to make an accurate inquiry.

  Therefore, a diagnosis using a computer network has been proposed. For example, Patent Document 1 discloses a technique for improving a user's lifestyle using a computer connected to a network. According to this, advice is output from the computer based on information input by the user. Such a system is not intended for treatment, but is intended to be improved by the use of a patient before becoming ill with minor symptoms.

  By the way, regarding the input method to a computer, a multimodal interface which is a technique for inputting and outputting multimedia information such as audio information and image information has been developed (see Patent Document 2).

JP 2001-22837 A Japanese Patent No. 3844874

  However, in Patent Document 1, there is little daily information input from the user, and the lifestyle habit improvement support device cannot grasp much of the user's life content. When there is little information input in this way, it is only possible to roughly analyze the current state of the subject, such as the subject's emotions, and it is difficult to analyze with high accuracy. On the other hand, Patent Document 2 does not show a specific application example using a multimodal interface. Therefore, the input information cannot be used effectively, and the emotion of the subject cannot be analyzed with high accuracy.

  There are many challenges in mental health care services, including those mentioned above. For example, when an inquiry is performed, the diagnosis results vary depending on the knowledge and experience of the person who makes the inquiry. In addition, since there is a large difference in the degree of self-recognition and communication ability of the person being interrogated, it takes time to collect information and the diagnosis is delayed, which causes a decrease in diagnostic accuracy. If the person being diagnosed has a mental disorder such as a mood disorder, self-recognition itself may adversely affect the physical condition.

  The person to be diagnosed easily feels a psychological burden when asked by the other person for personal information, and the burden of time, labor, money, etc. for receiving an inquiry is large. Mood disorders have characteristics that are not clear when they are affected, and it is difficult to identify themselves, especially in the early stages. Therefore, in order to make a correct diagnosis, information that grasps the actual situation of life as well as the examination scene is necessary.

  Of the information about the recovery of the person being diagnosed, the duration of positive transformation of self-cognition is important for judgment, but this information can be obtained by self-reporting. In addition, coping with problems that lower mood requires grasping of actual behavior, most of which is based on self-reporting. In obtaining such information, since there is a large part that depends on the knowledge and experience of the interviewer, improper interviewing has a risk of misunderstanding behavior facts and further lowering the mood of the subject.

  Although the mood state of the person to be diagnosed changes ups and downs in time series, there is no means for constantly measuring it. If the person being diagnosed has few contacts with the person due to depression, etc., a simple method for grasping the mood state will be developed because frequent interviews are not possible. It is useful.

  In cognitive behavioral therapy, which is effective in treating mood disorders, we discuss the thought patterns (automatic thinking) that the person takes without being conscious, but self-reported automatic thinking is not necessarily the most effective information for treatment. Not exclusively. Since the therapist does not know all the information about the person's life, it takes time to identify important issues, and differences due to the technology of the therapist tend to occur. For correction of automatic thinking, it is necessary for the therapist to observe and guide the practice situation in daily life. However, since intervention is possible only intermittently, it takes time for the treated person to become independent.

  Behavioral therapy, which is often used to treat anxiety and developmental disorders, has a method of changing the content of actions step by step toward a major goal. However, since behavior records are mainly self-reported, confidence in accuracy is required. Sex is unstable. Although there are autonomous training methods that control the autonomic nervous functions such as breathing and pulse by themselves, there is no method for recognizing physical changes uniquely or in time series.

  Behavioral analysis is performed when behavioral therapy is performed, but the collection of behavior is mainly based on the self-report of the person, so the accuracy is low, which may lead to a diagnosis error. Diagnosed subjects with undeveloped language skills, such as preschool children, cannot understand changes in psychological state by language alone, making it difficult to find disabilities. When the person being diagnosed is a minor and the supervisor such as a parent cannot observe the behavior, there is no system for objectively recording even if there is a problem in the behavior.

  If the person being diagnosed has the belief that “it is bad to express self-perception of emotions”, it is difficult to obtain language expressions about emotions, so changes in psychological state are objective vital signs. I have to figure it out. When the person to be diagnosed is alone, there is no objective information in the home, so in order to know the objective changes seen by others, it depends on the social nature of the person. Even if the person to be diagnosed has a disorder such as a language disorder that makes it difficult to translate his / her state, it is difficult to obtain information necessary for the diagnosis.

  When the person being diagnosed is receiving treatment for symptoms, assistance from a guardian may be required, but due to time constraints, assistance may not always be possible. If the person to be diagnosed is a worker who is on leave, supervision of the supervisor of the affiliated company is required, but grasping the person's status is personal.

  Although experiences such as domestic violence in childhood may be related to later stress disorders or depression, there is no unified means of measuring experiences that can cause personal trauma, and there is no indication of symptoms. Ease of discovery depends on the social nature of the individual. Bullying in schools and companies is a serious problem that can lead to suicide, but the detection and response of signs is largely dependent on the ability to gather information from teachers and supervisors. In addiction behavior problems such as alcoholism and bulimia, the behavior record is self-reported and the accuracy is low, and deviation from objective information tends to be a problem.

  An object of the present invention is to provide a dialogue system and a dialogue program that do not require knowledge and experience values, can analyze a subject's state with high accuracy, and can generate an accurate response.

In order to achieve the above object, the dialogue system of the present invention includes an acquisition unit that acquires an input sentence input by a subject,
A response master that defines emotion states and character strings
A storage unit for storing a change in the emotional state over time as an emotional state history;
A controller that calculates the emotional state from the input sentence, updates the emotional state history, and generates a response sentence by reading and connecting a plurality of character strings from the response master based on the emotional state history; Prepare.

  In the above interactive system, the control unit may determine a plurality of sets of character strings to be read from the response master in accordance with the most recent combination of a predetermined number of emotional states.

Further, in the above dialogue system, the emotional state is represented by positive, neutral, or negative,
Emotion word master with emotion values according to emotion words
The control unit calculates a sum of emotion values of the input sentence based on the emotion word master, normalizes the sum of the emotion values,
When θ ≦ | E | ≦ 1 (E: value obtained by normalizing the sum of emotion values, θ: threshold value (positive value)), the emotional state is positive if E is positive, and negative if E is negative It has been determined that
When -θ <E <0, it is determined that the emotional state continues the current state, or has transitioned to neutral or negative,
In the case of 0 ≦ E <θ, it may be determined that the emotional state continues the current state or has transitioned to neutral or positive.

In the above dialogue system, the calculation formula for specifying the emotional state is:
M _k (S ₀ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 0} , M _k-1 (S ₁ ) + L _k ^{1 → 0} , M _k-1 (S _-1 ) + L _k ^{-1 → 0} )
M _k (S ₁ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 1} , M _k-1 (S ₁ ) + L _k ^{1 → 1} , M _k-1 (S _-1 ) + L _k ^{-1 → 1} )
M _k (S _-1 ) = max (M _k-1 (S ₀ ) + L _k ^{0 → -1} , M _k-1 (S ₁ ) + L _k ^{1 → -1} , M _k-1 (S _-1 ) + L _k ^{-1 → -1} )
M _k (S _i ): Path metric value of emotional state at time k
L _k ^{i → j} : a branch metric value based on the state transition probability of transition from state i to state j at time k,
L _k ^{i → j} = 1 / (-ln P _k ^{i → j} )
i: Either -1 (negative), 0 (neutral), or 1 (positive)
L _k ^{-1 → -1} : Branch metric value where emotional state does not change from negative
L _k ^{-1 → 0} : Branch metric value where the emotional state transitions from negative to neutral
L _k ^{-1 → 1} : Branch metric value where emotional state transitions from negative to positive
L _k ^{0 → -1} : Branch metric value where emotional state transitions from neutral to negative
L _k ^{0 → 0} : Branch metric value where emotional state does not change from neutral
L _k ^{0 → 1} : Branch metric value where emotional state transitions from neutral to positive
L _k ^{1 → -1} : Branch metric value where emotional state transitions from positive to negative
L _k ^{1 → 0} : Branch metric value where emotional state transitions from positive to neutral
L _k ^{1 → 1} : Branch metric value where emotional state does not change from positive
P _k ^{-1 → 1} : Probability of emotional state transition from negative to positive
P _k ^{-1 → 0} : Probability of emotional state transition from negative to neutral
P _k ^{-1 → -1} : Probability that emotional state does not change from negative
P _k ^{0 → 1} : Probability of emotional state transitioning from neutral to positive
P _k ^{0 → 0} : Probability that emotional state does not change from neutral
P _k ^{0 → -1} : Probability of emotional state transitioning from neutral to negative
P _k ^{1 → 1} : Probability that emotional state does not change from positive
P _k ^{1 → 0} : Probability of emotional state transitioning from positive to neutral
P _k ^{1 → -1} : It may be the probability that the emotional state changes from positive to negative.

  In the above interactive system, when a plurality of paths are calculated, a survival path that is one connected path by selecting an emotional state having the maximum path metric value at each time when the paths converge. May be determined.

In the above dialogue system, the emotional state is expressed as a ratio of nodes including positive, neutral, and negative.
Emotion word master with emotion values according to emotion words
The control unit calculates a sum of emotion values of the input sentence based on the emotion word master, normalizes the sum of the emotion values,
When θ ≦ | E | ≦ 1 (E: value obtained by normalizing the sum of emotion values, θ: threshold value (positive value)), the emotional state is positive if E is positive, and negative if E is negative It has been determined that
When -θ <E <0, it is determined that the emotional state continues the current state and has transitioned to neutral and negative,
If 0 ≦ E <θ, it is determined that the emotional state continues the current state and transitions to neutral and positive,
A calculation formula for specifying the emotional state is:
R _k ⁱ = M _k (S _i ) / (M _k (S ₀ ) + M _k (S ₁ ) + M _k (S _-1 ))
i: Either -1 (negative), 0 (neutral), or 1 (positive)
R _k ⁱ : Ratio of each node in the emotional state at time k
M _k (S _i ): Path metric value of emotional state at time k
M _k (S ₀ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 0} , M _k-1 (S ₁ ) + L _k ^{1 → 0} , M _k-1 (S _-1 ) + L _k ^{-1 → 0} )
M _k (S ₁ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 1} , M _k-1 (S ₁ ) + L _k ^{1 → 1} , M _k-1 (S _-1 ) + L _k ^{-1 → 1} )
M _k (S _-1 ) = max (M _k-1 (S ₀ ) + L _k ^{0 → -1} , M _k-1 (S ₁ ) + L _k ^{1 → -1} , M _k-1 (S _-1 ) + L _k ^{-1 → -1} )
It is good also as being.

In the above dialogue system, a character input unit and / or a voice input unit are provided,
The input sentence acquired by the acquisition unit may be text data input from the character input unit by the target person or text data obtained by converting voice data input from the voice input unit by the target person.

In the above dialogue system, a personal characteristic master is provided.
The control unit decomposes the input sentence into parts of speech, associates the decomposed word with a predetermined emotion value and stores it in the personal characteristic master, and adds the information of the personal characteristic master to the sum of the emotion values. You may make it calculate.

  In the above interactive system, the control unit may generate the response sentence using a word stored in the personal characteristic master.

Further, in the above dialogue system, the acquisition unit is configured to obtain biological information including at least one or more information among audio information, facial expression information, facial temperature information, eye movement information, pulse information, heart rate information, respiratory information, and blood pressure information. Acquired,
The control unit may convert the biometric information into an emotion value, and calculate the sum of the emotion values in consideration of the converted emotion value.

  In the above interactive system, an output unit that is an output destination of the response sentence may be provided.

  The interactive program of the present invention operates any of the interactive systems described above.

  According to the present invention, it is possible to provide a dialogue system and a dialogue program that do not require knowledge and experience values, can analyze the state of a subject with high accuracy, and generate an accurate response.

It is a block diagram which shows the structure of the dialogue system of one Embodiment of this invention. It is a flowchart which shows operation | movement of the dialogue system of one Embodiment of this invention. It is an example of the personal characteristic master of one Embodiment of this invention. It is an example of the emotion word master of one Embodiment of this invention. It is an example of the inversion word master of one Embodiment of this invention. It is an example of the trellis diagram which shows the change of the emotional state of one Embodiment of this invention. It is an example of the trellis diagram which shows the change of the emotional state of one Embodiment of this invention. It is an example of the response master of one Embodiment of this invention.

  FIG. 1 is a block diagram showing a configuration of an interactive system according to an embodiment of the present invention. The dialogue system 10 includes a central device 12 capable of storing and calculating data, such as a server device, a terminal device 13, 14 such as a personal computer, and a mobile terminal device 15, 16 such as a mobile phone, in a communication network 11 such as the Internet. The character string converting device 26 is connected. Printing devices 17 to 19 such as printers are connected to the central device 12 and the terminal devices 13 and 14, respectively. Further, cameras 20 and 23, microphones 21 and 24, and speakers 22 and 25 are connected to the terminal devices 13 and 14.

  Each of the terminal devices 13 and 14 and the mobile terminal devices 15 and 16 includes input units 13a, 14a, 15a, and 16a including keys and a touch panel, and output units 13b and 14b including a liquid crystal display device that displays characters and images. 15b and 16b. The mobile terminal devices 15 and 16 also include cameras 15c and 16c, microphones 15d and 16d, and speakers 15e and 16e. Here, the input unit serves as a character input unit for inputting characters. The microphone also serves as a voice input unit that detects the voice of the subject. In addition, the camera and the microphone serve as sensors that detect biological information. The biological information is, for example, at least one or more information among speech information, facial expression information, facial temperature information, eye movement information, pulse information, heart rate information, respiratory information, and blood pressure information based on words uttered by the subject.

  The character string converting device 26 is a device that converts voice information into a character string. In addition, the technique developed conventionally can be employ | adopted suitably for the technique which makes speech information into a character string.

  In the configuration of FIG. 1, a dialogue program is stored in the central device 12, and the terminal devices 13 and 14 and the mobile terminal devices 15 and 16 transmit and receive information to and from the central device 12 and output from the output unit or speakers. Or information input from the input unit, camera, or microphone is transmitted to the central device 12 or the like.

  Note that the printing devices 17 to 19 are not necessarily required. The terminal devices 13 and 14 and the mobile terminal devices 15 and 16 may be any one or more, and the central device 12 includes, for example, an arithmetic device (control unit 12b), a database (storage unit 12a), and an acquisition unit 12c. A plurality of devices may be linked such as being separate. Further, the system of the present invention does not necessarily need to be configured by a plurality of devices connected by the communication network 11 as shown in FIG. 1, and may have a configuration in which the devices are directly connected without the communication network 11. All the configurations may be incorporated in one apparatus.

  Next, automatic response sentence generation in the dialogue system 10 will be described. According to the dialogue system 10, a response sentence suitable for the current situation is created for each individual (subject).

  FIG. 2 is a flowchart showing the operation of the dialogue system 10. The broken lines in the figure indicate the data flow. Here, the case where Mr. N (subject) who is rehabilitating uses the dialogue system 10 will be described as an example.

  First, for example, when Mr. N's activities start, such as when Mr. N wakes up, the camera 20 installed near the bed takes a picture of Mr. N's movement and passes through the terminal device 13 to the central device 12. It is detected that the user has woken up by transmitting data. In addition, for example, when Mr. N who gets up operates the input unit 15a of the mobile terminal device 15 and checks in the system, the data is transmitted to the central device 12 and is detected to be up. The position of the subject detected by GPS (Global Positioning System) can also be used for check-in. The data of Mr. N's activity is stored in the storage unit 12a of the central device 12.

  When waking up is detected or check-in is performed as described above, the processing of the flowchart of FIG. 2 is started. And it progresses to step S10, and the control part 12b of the central apparatus 12 produces | generates a question sentence, and makes it display on the output part 13b of the terminal device 15. FIG. The question sentence can be appropriately selected from a plurality of example sentences previously stored in the storage unit 12a. Note that step S10 is not always necessary, and may be started from the following step S11.

  From step S10, the process proceeds to step S11. When the user operates the input unit 15a of the mobile terminal device 15 or the like or inputs text or voice through the microphone, step S12 or step S13 is performed according to the input data. Proceed to If the input data is text data, the process proceeds to step S12, the text data is transmitted from the mobile terminal device 15 or the like operated by Mr. N to the central apparatus 12, and the process proceeds to step S14 where the acquisition unit 12c of the central apparatus 12 The text data is acquired as an input sentence.

  On the other hand, if the input data is voice data, the process proceeds to step S13, where the voice data is transmitted from the microphone 21 or the like that picks up the voice uttered by Mr. N to the character string forming device 26, and the process proceeds to step S15. The voice data is converted into text data by the converting device 26, and the process proceeds to step S14 where the acquisition unit 12c of the central device 12 acquires the converted text data as an input sentence.

  From step S14, the process proceeds to step S16 and step S17. In step S16, the control unit 12b decomposes the input sentence into parts of speech, extracts necessary words / phrases from the decomposed sentence, and associates predetermined emotion values with the words / phrases and stores them in the personal characteristic master stored in the storage unit 12a. To do. This personal characteristic master is used later when calculating the sum of emotion values. FIG. 3 shows an example of the personal characteristic master. Corresponding to the ID number of the subject, the standard type of the word, the part of speech information, the detailed part of speech information, the emotion value, and the stored date and time are written. Although the emotion value is uniformly set to 0.5, for example, a different value may be assigned according to a word or a part of speech. The words to be stored are preferably nouns or verbs.

  On the other hand, in step S17, the control unit 12b calculates the sum of emotion values from the input sentence. The storage unit 12a stores an emotion word master in which an emotion value according to an emotion word is recorded, and an inversion word master in which an emotion value according to an inversion word is recorded. The sum of emotion values of the input sentence is calculated based on the personal characteristic master. Specifically, the input sentence is decomposed into parts of speech, each phrase is searched with each master, and if they match, the emotion value is read and added to calculate the sum. In addition, you may include all in an emotion word master, without distinguishing an emotion word master, an inversion word master, and a personal characteristic master.

  FIG. 4 shows an example of the emotion word master. Corresponding to the ID number of the subject, emotion words, part of speech information, detailed part of speech information, and emotion values are written. Although the emotion value is negative in FIG. 4, naturally the emotion value is also included.

  FIG. 5 shows an example of the inverted word master. Corresponding to the ID number of the subject, inversion words, part of speech information, detailed part of speech information, and emotion values are written. An inverted word is a word that reverses the meaning of a sentence.If another word in a sentence containing an inverted word is positive, it changes the sentence to a negative meaning, while another word in a sentence containing an inverted word If the word is negative, the sentence is changed to a positive meaning.

  For example, if the input sentence is “Today, the golf schedule is crushed and boring.” In step S17, the control unit 12b determines in step S17 the emotion value “0” corresponding to “golf” from the personal characteristic master of FIG. .5 ”, the emotion value“ −0.5 ”corresponding to“ Crush ”from the inverted word master of FIG. 5, and“ −0.3 ”corresponding to“ boring ”from the emotion word master of FIG. Is calculated as -0.3. When an inverted word is included, instead of using an inverted word master in which an emotion value is written as shown in FIG. 5, processing such as reversing the sign of the emotion value may be performed.

  Biometric information may be taken into account when calculating the sum of emotion values. In this case, the biometric information detected by the camera or the microphone is converted into an emotion value using a master in which the biometric information and the emotion value are associated, and the total of the emotion values is calculated by adding (for example, adding) the converted emotion values. calculate.

  Next, the process proceeds from step S17 to step S18, where the calculated sum of emotion values is stored in the history of the sum of emotion values in the storage unit 12a, and the process proceeds to step S19. In step S19, the control unit 12b makes a soft decision by estimating the current emotional state of the subject from the total of the stored emotion values. In this embodiment, the emotional state is handled as changing with time, and the emotional state at a certain time (for example, when the subject inputs an input sentence) is expressed as positive, neutral, or negative.

  To softly determine the emotional state, first, the summation of emotional values is normalized. Thus, the value E obtained by normalizing the sum of emotion values is −1 ≦ E ≦ 1. Then, it applies to three patterns according to the value E which normalized the sum total of the emotion value. The first is the case of θ ≦ | E | ≦ 1 (θ: threshold value (positive value)), and the emotional state is determined to be positive if E is positive and negative if E is negative. The second is a case where -θ <E <0, and it is determined that the emotional state continues the current state or has transitioned to neutral or negative. The third is a case of 0 ≦ E <θ, and it is determined that the emotional state continues the current state or has transitioned to neutral or positive.

Then, in order to specify the emotional state, the path metric value M _k (S _i ) of each state (S ₀ , S ₁ , S ₋₁ ) at time k is used, and the calculation formula is
M _k (S ₀ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 0} , M _k-1 (S ₁ ) + L _k ^{1 → 0} , M _k-1 (S _-1 ) + L _k ^{-1 → 0} )
M _k (S ₁ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 1} , M _k-1 (S ₁ ) + L _k ^{1 → 1} , M _k-1 (S _-1 ) + L _k ^{-1 → 1} )
M _k (S _-1 ) = max (M _k-1 (S ₀ ) + L _k ^{0 → -1} , M _k-1 (S ₁ ) + L _k ^{1 → -1} , M _k-1 (S _-1 ) + L _k ^{-1 → -1} )
And

L _k ^{i → j} is a branch metric value based on the state transition probability of transition from state i to state j at time k, and its calculation formula is
L _k ^{i → j} = 1 / (-ln P _k ^{i → j} )
i: Either -1 (negative), 0 (neutral), or 1 (positive)
L _k ^{-1 → -1} : Branch metric value where emotional state does not change from negative
L _k ^{-1 → 0} : Branch metric value where the emotional state transitions from negative to neutral
L _k ^{-1 → 1} : Branch metric value where emotional state transitions from negative to positive
L _k ^{0 → -1} : Branch metric value where emotional state transitions from neutral to negative
L _k ^{0 → 0} : Branch metric value where emotional state does not change from neutral
L _k ^{0 → 1} : Branch metric value where emotional state transitions from neutral to positive
L _k ^{1 → -1} : Branch metric value where emotional state transitions from positive to negative
L _k ^{1 → 0} : Branch metric value where emotional state transitions from positive to neutral
L _k ^{1 → 1} : Branch metric value where emotional state does not change from positive
P _k ^{-1 → 1} : Probability of emotional state transition from negative to positive
P _k ^{-1 → 0} : Probability of emotional state transition from negative to neutral
P _k ^{-1 → -1} : Probability that emotional state does not change from negative
P _k ^{0 → 1} : Probability of emotional state transitioning from neutral to positive
P _k ^{0 → 0} : Probability that emotional state does not change from neutral
P _k ^{0 → -1} : Probability of emotional state transitioning from neutral to negative
P _k ^{1 → 1} : Probability that emotional state does not change from positive
P _k ^{1 → 0} : Probability of emotional state transitioning from positive to neutral
P _k ^{1 → -1} : The probability that the emotional state changes from positive to negative.

Then, the branch metric value L _k ^{i → j} is added for each state transition (for example, every time the input sentence is input). At this time, if the paths merge, the maximum value is selected from those values. If a plurality of paths are calculated, a surviving path that is one connected path is determined by selecting an emotional state having the maximum path metric value at each time when the paths converge. If the path does not converge, it is sufficient to truncate appropriately and select the path with the maximum path metric at that time. For example, it may be truncated at 2 to 8 state transitions, preferably 3 to 6 times, more preferably about 5 times. Alternatively, the path metrics in each state may always be relatively compared, and may be forcibly determined by truncating when they are several percent larger than others. Thereby, the determination of the emotional state at the past time is finalized at the stage where the path converges or is truncated. Then, when the surviving path is determined or truncated, the phase is temporarily ended, and a value obtained by adding the branch metric value may be set to 0.

  Here, the determination of the emotional state will be described along a specific example. The logic of path transition for state transition can be represented by a trellis diagram as shown in FIGS. Here, the threshold θ is set to 0.25, for example. In addition, the probability that the emotional state does not change state is 0.5, the probability that state transition is from neutral to positive or negative, and conversely, the probability that state transition is from positive or negative to neutral is 0.4, and the probability that state transition is from positive to negative On the contrary, the probability of state transition from negative to positive is set to 0.1.

By determining the probability of each state transition as described above, the path value (L _k ^{i → j} ) is determined as follows.
L _k ^{0 → 0} = L _k ^{1 → 1} = L _k ^{-1 → -1} = 1 / (-ln 0.5) = 1.43
L _k ^{0 → 1} = L _k ^{1 → 0} = L _k ^{0 → -1} = L _k ^{-1 → 0} = 1 / (-ln 0.4) = 1.11
L _k ^{1 → -1 =} L _k ^{-1 → 1} = 1 / (-ln 0.1) = 0.43

Hereinafter, FIG. 6 will be described as an example. First, the initial state is time 0, the emotional state is neutral Nt, and the pass initial value is 0. When an input sentence is input at time 1 and the value E (normalized emotion value) obtained by normalizing the sum of emotion values is +0.5, 0.25 ≦ | E | ≦ 1 and E is positive. It is determined that the emotional state has changed to positive. In other words, the state transition is from neutral to positive. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 0 + 1.11 = 1.11.

Subsequently, when an input sentence is input at time 2 and the normalized emotion value E is −0.5, since 0.25 ≦ | E | ≦ 1 and E is negative, the emotional state transitions to negative. judge. In other words, it has transitioned from positive to negative. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 1.11 + 0.43 = 1.55.

Subsequently, when an input sentence is input at time 3 and the normalized emotion value E is −0.3, since 0.25 ≦ | E | ≦ 1 and E is negative, the emotional state transitions to negative. judge. In other words, there is no transition from negative. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 1.55 + 1.43 = 2.97.

Subsequently, when an input sentence is input at time 4 and the normalized emotion value E is −0.1, since −0.25 <E <0, the emotional state continues the current state, It is determined that the transition is negative. That is, the transition from negative to neutral or no transition from negative. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 2.97 + 1.11 = 4.08 if the transition is to neutral, and 2.97 + 1.43 if the transition is not from negative. = 4.4. Therefore, the emotional state is uncertain at time 4.

Subsequently, when an input sentence is input at time 5 and the normalized emotion value E is +0.2, since 0 ≦ E <0.25, the emotional state continues the current state, or is neutral or positive. It is determined that a transition has occurred. In other words, there is no transition from negative, transition from negative to neutral, transition from negative to positive, transition from neutral, transition from neutral to positive, and a total of five paths can be taken. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 4.4 + 1.43 = 5.83 if no transition is made from negative. In addition, when transitioning to neutral, there are two paths. If the transition is from negative to neutral, 4.4 + 1.11 = 5.51, and if not transitioning from neutral, 4.08 + 1.43 = 5.5. 51. Then, the maximum value is taken out of the two values, both of which are the same 5.51 here. In the case of transition from positive to negative, there are two paths. If the transition from negative to positive is 4.4 + 0.43 = 4.83, and if the transition is from neutral to positive, 4.08 + 1.11 = 5.19. The maximum value of the two values is taken as 5.19. Therefore, the emotional state is not yet determined at time 5.

Subsequently, when an input sentence is input at time 6 and the normalized emotion value E is +0.5, since 0.25 ≦ | E | ≦ 1 and E is positive, it is determined that the emotion state has changed to positive. To do. In other words, the transition from negative to positive, transition from neutral to positive, or no transition from positive, and the path has converged to positive. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 5.83 + 0.43 = 6.26 if the transition is from negative to positive, and 5 if the transition is from neutral to positive. .51 + 1.11 = 6.62, and 5.19 + 1.43 = 6.62 if there is no transition from positive. The maximum value among the three is taken as 6.62. Since the path has converged at time 6, a surviving path is determined. That is, since the maximum value (5.51 or 5.19 or 5.83) _obtained by adding the branch metric value L _k at time 5 is 5.83, the negative having the value is determined at time 5. Judged as emotional state. Further, since the path from time 4 to time 5 is the only path from time 4 to negative, it can be determined that the emotional state at time 4 is negative. Therefore, since the surviving path is determined at time 6, the emotional state at times 4 to 6 is determined. Since the surviving path is determined, the phase is temporarily terminated at time 8 and the value _obtained by adding the branch metric value L _k is set to 0.

Subsequently, when an input sentence is input at time 7 and the normalized emotion value E is −0.9, since 0.25 ≦ | E | ≦ 1 and E is negative, the emotional state transitions to negative. judge. That is, the state transition is from positive to negative. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 0 + 0.43 = 0.43.

Subsequently, when an input sentence is input at time 8 and the normalized emotion value E is +0.7, since 0.25 ≦ | E | ≦ 1 and E is positive, it is determined that the emotion state has changed to positive. To do. That is, the state transition is from negative to positive. At this time, the value _obtained by adding the branch metric value L _k for each state transition is 0.43 + 0.43 = 0.86. In this way, the emotional state at each time when the input sentence is input is determined.

  In FIG. 7, since the emotional state has been undecided for five times from time 4 to time 8, the path is truncated at time 8 and the path with the maximum path metric at that time is selected. Thereby, it is determined that the emotional state at times 4 to 8 is negative.

  Returning to the description of FIG. 2, the process proceeds from step S19 to step S20 to determine whether or not the emotional state has been determined. If it has not been determined, the process proceeds to step S21, and if it has been determined, the process proceeds to step S22. That is, as described in FIG. 6, it is determined that the emotional state is not yet determined at the time points 4 and 5, and the process proceeds to step S <b> 21, and the emotional state is determined at the time points 1 to 3 and 6 to 8. It discriminate | determines and progresses to step S22.

  When the process proceeds from step S20 to step S21, the control unit 12b reads the question text stored in the storage unit 12a, and returns to step S10 to display the question text on the output unit 13b of the terminal device 15. . Note that the question sentence may be generated using a dialogue algorithm described later.

  On the other hand, when the process proceeds from step S20 to step S22, the temporal change of the emotional state is stored in the storage unit 12a as the emotional state history. That is, the determined emotional state is stored as an emotional state history corresponding to the time. And it progresses to step S23 and processes an emotional state log | history with a dialogue algorithm. That is, the control unit 12b generates a response sentence by reading and connecting a plurality of sets of character strings from the response master based on the emotional state history. For example, the control unit 12b reads a predetermined number of recent emotion states from the emotion state history, and determines a plurality of sets of character strings to be read from the response master according to the combination. As the determination algorithm, for example, the order of combination, the number of emotion states, the degree of undulation of emotion states, and the like can be considered.

  The response master describes a plurality of phrases (character strings) for generating a response sentence, and is stored in the storage unit 12a. Each phrase may be divided into submasters for each type. FIG. 8 shows an example of the response master. Submasters include a conflict master, promotion master, materialization master, empathy master, interpretation master, and the like, and a plurality of phrases are written in each submaster. For example, in the master master, “Oh!” Is written as the phrase 1, “Hoo.” Is written as the phrase 2, “Oh.” Is written as the phrase 3, and the like.

  For example, when generating a response sentence according to the latest five emotional states, analyze the order of the combination, the number of each emotional state, the degree of undulation of the emotional state, etc., and which submaster of which phrase in what order Decide whether to create a response sentence by connecting. As a result, for example, if it is determined that phrase 3 of the master master + phrase 3 of the sympathy master + phrase 2 of the materialization master are connected in this order, these phrases are read out and “Oh. A response sentence is generated.

  If there is a phrase that returns a parrot using a word in the input sentence in the response sentence, it is appropriately read from the personal characteristic master and inserted into the response sentence. For example, when generating a response sentence “Oh! Golf. That's good!” For the input sentence “Go on golf on weekends”, read “Golf” from the personal characteristics master. Insert.

  From step S23, it progresses to step S10 and a response sentence is displayed on the output part 15b of the portable terminal device 15 which Mr. N is operating. Note that the output destination of the response sentence may be the speaker 15e. That is, the response sentence may be output by voice from the speaker 15e. Further, both display and sound may be output. Conversion of the character string to be output to speech can be performed using the character string converting device 26.

  And after displaying a response sentence, you may make it complete | finish this flowchart, when there is no next input from a subject in step S11. Further, this flowchart may be ended when the subject inputs the end or when it is detected that the subject has gone to bed. In addition, in the case of a processing result that a response sentence is not generated by the processing in step S23, this flowchart may be ended.

  Thus, the dialogue system that automatically outputs a response does not require knowledge and experience values, and can analyze the state of the subject with high accuracy and generate and output an accurate response. Therefore, the subject feels as if he / she is interacting with someone through the terminal device.

  The dialogue contents (input sentence, response sentence, question sentence) may be stored in the central device 12 and / or the terminal device operated by the subject. As a result, the history of dialogue can be confirmed later.

  Next, another method for softly determining the emotional state will be described. The processing in step S19 in FIG. 2 can be replaced with the following processing.

  The controller 12b makes a soft decision by estimating the current emotional state of the subject from the sum of the stored emotion values. Here, it is assumed that the emotional state changes over time, and the emotional state at a certain time (for example, when the subject inputs an input sentence) is represented by the ratio of nodes consisting of positive, neutral, and negative.

  To softly determine the emotional state, first, the summation of emotional values is normalized. Thus, the value E obtained by normalizing the sum of emotion values is −1 ≦ E ≦ 1. Then, it applies to three patterns according to the value E which normalized the sum total of the emotion value. The first is the case of θ ≦ | E | ≦ 1 (θ: threshold value (positive value)), and the emotional state is determined to be positive if E is positive and negative if E is negative. The second is a case where -θ <E <0, and it is determined that the emotional state continues the current state and has transitioned to neutral and negative. The third is a case of 0 ≦ E <θ, and it is determined that the emotional state continues the current state and has transitioned to neutral and positive.

And in order to identify the emotional state, using the ratio of each node occupying the emotional state at time k, the calculation formula is
R _k ⁱ = M _k (S _i ) / (M _k (S ₀ ) + M _k (S ₁ ) + M _k (S _-1 ))
i: Either -1 (negative), 0 (neutral), or 1 (positive)
R _k ⁱ : Ratio of each node in the emotional state at time k
M _k (S _i ): The path metric value of the emotional state at time k.

The calculation formula of M _k (S _i ) is the following (i), (ii), and (iii).
(i) M _k (S ₀ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 0} , M _k-1 (S ₁ ) + L _k ^{1 → 0} , M _k-1 (S _-1 ) + L _k ^{-1 → 0} )
(ii) M _k (S ₁ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 1} , M _k-1 (S ₁ ) + L _k ^{1 → 1} , M _k-1 (S _-1 ) + L _k ^{-1 → 1} )
(iii) M _k (S _-1 ) = max (M _k-1 (S ₀ ) + L _k ^{0 → -1} , M _k-1 (S ₁ ) + L _k ^{1 → -1} , M _k-1 ( S _-1 ) + L _k ^{-1 → -1} )

Then, for each state transition (for example, every time the input sentence is input), the emotional state is represented by the ratio R _k ⁱ of each node. For example, if the emotional state is expressed by (probability of being neutral, probability of being positive, probability of being negative), the probability of being neutral is 40%, the probability of being positive is 50%, and negative. If the probability is 10%, it can be expressed as (40%, 50%, 10%).

The determination of the emotional state will be described using, for example, time 5 in FIG. At time 5, the neutral path metric value M ₅ (S ₀ ) is 5.51, the positive path metric value M ₅ (S ₁ ) is 5.19, and the negative path metric value M ₅ (S ₋₁ ) is 5. .83. Therefore, ΣM ₅ = 5.51 + 5.19 + 5.83 = 16.53, the probability of being neutral is 5.51 / 16.53≈0.33, and the probability of being positive is 5.19 / 16.53. ≈0.31, the probability of being negative is 5.83 / 16.53≈. 0.35. Therefore, the emotional state is represented by (33%, 31%, 35%).

  The emotional state can also be determined by such a method, and according to this method, the emotional state is surely determined when the input sentence is input. Note that the emotional state is always determined means that the emotional state can be always represented by the ratio of each node at each time point, and it is not necessary to truncate.

  In addition, the value of said various parameters is an example, and is not specifically limited. In addition, responses, information on various masters, emotional history, and the like can be appropriately printed by a printing apparatus.

  The mental health care service mentioned above is a service that is performed during the period of studying for the introduction of treatment, such as when the patient is not feeling well, or when rehabilitation work is possible.

  The use of the above dialogue system is as follows. 1. When a general user who has some kind of anxiety knows his emotional state quantitatively through dialogue with a machine without anyone knowing. 2. When a general user who has some kind of anxiety self-recognizes his / her emotional state through machine dialogue without anyone knowing; There may be a case where a general user who has some kind of anxiety tries to improve his / her emotional state through machine dialogue without anybody knowing it.

  The present invention is not limited to mental health care rehabilitation but can also be used for physical rehabilitation. In addition, counseling training for healthcare professionals, orthopedic rehabilitation, gym training menu, diet menu, metabolic elimination menu, health management system for daily measurement and recording of individual health status, school cram school assignment menu, workplace management It can also be applied to support, personnel evaluation system, video conference system, car navigation system, etc.

DESCRIPTION OF SYMBOLS 10 Dialog system 12a Memory | storage part 12b Control part 12c Acquisition part 13a-16a Input part 13b-16b Output part 13.14 Terminal device 15, 16 Portable terminal device 26 Character string conversion apparatus

Claims (12)

An acquisition unit for acquiring an input sentence input by the target person,
A response master that defines emotion states and character strings
A storage unit for storing a change in the emotional state over time as an emotional state history;
A controller that calculates the emotional state from the input sentence, updates the emotional state history, and generates a response sentence by reading and connecting a plurality of character strings from the response master based on the emotional state history; Prepared,
The emotional state is represented by one of positive, neutral and negative,
Emotion word master with emotion values according to emotion words
The control unit calculates a sum of emotion values of the input sentence based on the emotion word master, normalizes the sum of the emotion values,
When θ ≦ | E | ≦ 1 (E: value obtained by normalizing the sum of emotion values, θ: threshold value (positive value)), the emotional state is positive if E is positive, and negative if E is negative It has been determined that
When -θ <E <0, it is determined that the emotional state continues the current state, or has transitioned to neutral or negative,
When 0 ≦ E <θ, the dialogue system determines that the emotional state continues the current state or has transitioned to neutral or positive. A calculation formula for specifying the emotional state is:
M _k (S ₀ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 0,} M _k-1 (S ₁ ) + L _k ^{1 → 0} , M _k-1 (S _-1 ) + L _k ^{-1 → 0} )
M _k (S ₁ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 1} , M _k-1 (S ₁ ) + L _k ^{1 → 1} , M _k-1 (S _-1 ) + L _k ^{-1 → 1} )
M _k (S _-1 ) = max (M _k-1 (S ₀ ) + L _k ^{0 → -1} , M _k-1 (S ₁ ) + L _k ^{1 → -1} , M _k-1 (S _-1 ) + L _k ^{-1 → -1} )
M _k (S _i ): Path metric value of emotional state at time k
L _k ^{i → j} : a branch metric value based on the state transition probability of transition from state i to state j at time k,
L _k ^{i → j} = 1 / (-ln P _k ^{i → j} )
i: Either -1 (negative), 0 (neutral), or 1 (positive)
L _k ^{-1 → -1} : Branch metric value where emotional state does not change from negative
L _k ^{-1 → 0} : Branch metric value where the emotional state transitions from negative to neutral
L _k ^{-1 → 1} : Branch metric value where emotional state transitions from negative to positive
L _k ^{0 → -1} : Branch metric value where emotional state transitions from neutral to negative
L _k ^{0 → 0} : Branch metric value where emotional state does not change from neutral
L _k ^{0 → 1} : Branch metric value where emotional state transitions from neutral to positive
L _k ^{1 → -1} : Branch metric value where emotional state transitions from positive to negative
L _k ^{1 → 0} : Branch metric value where emotional state transitions from positive to neutral
L _k ^{1 → 1} : Branch metric value where emotional state does not change from positive
P _k ^{-1 → 1} : Probability of emotional state transition from negative to positive
P _k ^{-1 → 0} : Probability of emotional state transition from negative to neutral
P _k ^{-1 → -1} : Probability that emotional state does not change from negative
P _k ^{0 → 1} : Probability of emotional state transitioning from neutral to positive
P _k ^{0 → 0} : Probability that emotional state does not change from neutral
P _k ^{0 → -1} : Probability of emotional state transitioning from neutral to negative
P _k ^{1 → 1} : Probability that emotional state does not change from positive
P _k ^{1 → 0} : Probability of emotional state transitioning from positive to neutral
The interactive system according to claim 1, wherein P _k ^{1 → −1} is a probability that the emotional state changes from positive to negative.   When a plurality of paths are calculated, a survival path that is one connected path is determined by selecting an emotional state having a maximum path metric value at each time when the paths converge. The dialogue system according to claim 2. An acquisition unit for acquiring an input sentence input by the target person,
A response master that defines emotion states and character strings
A storage unit for storing a change in the emotional state over time as an emotional state history;
A controller that calculates the emotional state from the input sentence, updates the emotional state history, and generates a response sentence by reading and connecting a plurality of character strings from the response master based on the emotional state history; Prepared,
The emotional state is represented by the ratio of nodes consisting of positive, neutral and negative,
Emotion word master with emotion values according to emotion words
The control unit calculates a sum of emotion values of the input sentence based on the emotion word master, normalizes the sum of the emotion values,
When θ ≦ | E | ≦ 1 (E: value obtained by normalizing the sum of emotion values, θ: threshold value (positive value)), the emotional state is positive if E is positive, and negative if E is negative It has been determined that
When -θ <E <0, it is determined that the emotional state continues the current state and has transitioned to neutral and negative,
When 0 ≦ E <θ, the dialogue system determines that the emotional state continues the current state and has transitioned to neutral and positive. A calculation formula for specifying the emotional state is:
R _k ⁱ = M _k (S _i ) / (M _k (S ₀ ) + M _k (S ₁ ) + M _k (S _-1 ))
i: Either -1 (negative), 0 (neutral), or 1 (positive)
R _k ⁱ : Ratio of each node in the emotional state at time k
M _k (S _i ): Path metric value of emotional state at time k
M _k (S ₀ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 0} , M _k-1 (S ₁ ) + L _k ^{1 → 0} , M _k-1 (S _-1 ) + L _k ^{-1 → 0} )
M _k (S ₁ ) = max (M _k-1 (S ₀ ) + L _k ^{0 → 1} , M _k-1 (S ₁ ) + L _k ^{1 → 1} , M _k-1 (S _-1 ) + L _k ^{-1 → 1} )
M _k (S _-1 ) = max (M _k-1 (S ₀ ) + L _k ^{0 → -1} , M _k-1 (S ₁ ) + L _k ^{1 → -1} , M _k-1 (S _-1 ) + L _k ^{-1 → -1} )
And
L _k ^{i → j} : a branch metric value based on the state transition probability of transition from state i to state j at time k,
L _k ^{i → j} = 1 / (-ln P _k ^{i → j} )
L _k ^{-1 → -1} : Branch metric value where emotional state does not change from negative
L _k ^{-1 → 0} : Branch metric value where the emotional state transitions from negative to neutral
L _k ^{-1 → 1} : Branch metric value where emotional state transitions from negative to positive
L _k ^{0 → -1} : Branch metric value where emotional state transitions from neutral to negative
L _k ^{0 → 0} : Branch metric value where emotional state does not change from neutral
L _k ^{0 → 1} : Branch metric value where emotional state transitions from neutral to positive
L _k ^{1 → -1} : Branch metric value where emotional state transitions from positive to negative
L _k ^{1 → 0} : Branch metric value where emotional state transitions from positive to neutral
L _k ^{1 → 1} : Branch metric value where emotional state does not change from positive
P _k ^{-1 → 1} : Probability of emotional state transition from negative to positive
P _k ^{-1 → 0} : Probability of emotional state transition from negative to neutral
P _k ^{-1 → -1} : Probability that emotional state does not change from negative
P _k ^{0 → 1} : Probability of emotional state transitioning from neutral to positive
P _k ^{0 → 0} : Probability that emotional state does not change from neutral
P _k ^{0 → -1} : Probability of emotional state transitioning from neutral to negative
P _k ^{1 → 1} : Probability that emotional state does not change from positive
P _k ^{1 → 0} : Probability of emotional state transitioning from positive to neutral
5. The dialogue system according to claim 4, wherein P _k ^{1 → −1} is a probability that the emotional state changes from positive to negative.   The interactive system according to claim 1, wherein the control unit determines a plurality of sets of character strings to be read from the response master according to a combination of a predetermined predetermined number of emotional states. . A character input unit and / or a voice input unit,
The input sentence acquired by the acquisition unit is text data input from the character input unit by the target person, or text data obtained by converting voice data input from the voice input unit by the target person. The dialogue system according to any one of claims 1 to 6. With personal characteristic master,
The control unit decomposes the input sentence into parts of speech, associates the decomposed word with a predetermined emotion value and stores it in the personal characteristic master, and adds the information of the personal characteristic master to the sum of the emotion values. The dialogue system according to claim 1, wherein the dialogue system is calculated.   The dialogue system according to claim 8, wherein the control unit generates the response sentence using a word stored in the personal characteristic master. The acquisition unit acquires biological information including at least one information of voice information, facial expression information, facial temperature information, eye movement information, pulse information, heart rate information, respiratory information, blood pressure information,
Wherein the control unit, the interactive system according to any one of claims 1-9, wherein the biometric information is converted to the emotion value, and calculates the sum of the emotion value by adding the converted emotion value .   The dialogue system according to claim 1, further comprising an output unit that is an output destination of the response sentence.   The dialogue program which operates the dialogue system in any one of Claims 1-11.

Images