JP7514019B2 - Natural language processing system, natural language processing method and natural language processing program - Google Patents

Description

This invention relates to a natural language processing system, a natural language processing method, and a natural language processing program.

This application claims priority to international application PCT/JP2019/022680, filed on June 7, 2019, and incorporates by reference all of the contents of said international application.

In recent years, there has been an increase in systems that can converse in human language (natural language). For example, Japanese Patent Application Laid-Open No. 2006-178063 (Patent Document 1) discloses a dialogue processing device that creates an emotion word dictionary that classifies words into three emotions: negative, positive, and neutral, infers the user's emotions from the emotion words contained in the user's speech, and responds based on that.

JP 2006-178063 A

According to the technology of Patent Document 1, when an emotion word appears, a response corresponding to it is simply made. Therefore, the dialogue processing device disclosed in Patent Document 1 does not understand the meaning of the sentence and output it. This causes the problem that the response is incoherent and the conversation does not continue naturally.

First of all, emotions do not correspond one-to-one to words, and the same word can give rise to different emotions depending on the situation and perspective at the time. For example, when you receive a present, you feel "happy" and also feel "gratitude" toward the person who gave you the present. If you create an emotion dictionary that matches the word "present" with the positive emotion, you can read the positive in the sentence "I received a present from my dad" and say "That's great," but you won't reply, "I should thank my dad."

The fundamental problem with this is that in the field of natural language processing, semantic understanding has not yet been achieved, and semantic understanding has not even been defined.

The object of the present invention is to provide a method for semantic understanding in natural language processing.

The natural language processing system according to the present invention has a scene setting unit that analyzes an input natural language sentence, generates subject data representing a subject who understands the natural language and scene data representing a part of the world in which the subject acts, and sets the subject data to the scene data, a situation change setting unit that analyzes the input natural language sentence and sets a situation change in the scene, a plus/minus emotion setting unit that determines whether a positive emotion, which is a favorable emotion for the subject, or a negative emotion, which is an unfavorable emotion for the subject, has occurred based on the situation change in the scene, and a semantic analysis unit that performs semantic analysis based on the generated plus/minus emotions. The situation change setting unit obtains, as numerical values, the value acting on the subject data in the state before the situation change and the value acting on the subject data in the state after the situation change. The plus/minus emotion setting unit calculates the difference between the numerical value in the before state and the numerical value in the after state to calculate the plus/minus emotion.

The natural language processing method according to the present invention includes the steps of: analyzing an input natural language sentence, generating subject data representing a subject who understands the natural language, and scene data representing a part of the world in which the subject acts, and setting the subject data in the scene data; analyzing the input natural language sentence and setting a situation change in the scene; determining whether a positive emotion, which is a favorable emotion for the subject, or a negative emotion, which is an unfavorable emotion for the subject, has occurred based on the situation change in the scene; and performing semantic analysis based on the positive and negative emotions that have occurred. The step of setting the situation change obtains, as numerical values, a value acting on the subject data in the state before the situation change and a value acting on the subject data in the state after the situation change. The step of determining calculates the difference between the numerical value of the before state and the numerical value of the after state to calculate the positive and negative emotions.

The natural language processing program of the present invention is a natural language processing program for causing a computer to function as a scene setting unit that analyzes an input natural language sentence, generates subject data representing a subject who understands natural language and scene data representing a part of the world in which the subject acts, and sets the subject data to the scene data, a situation change setting unit that analyzes the input natural language sentence and sets a situation change in the scene, a plus/minus emotion setting unit that determines whether a positive emotion, which is a favorable emotion for the subject, or a negative emotion, which is an unfavorable emotion for the subject, has occurred based on the situation change in the scene, and a semantic analysis unit that performs semantic analysis based on the generated plus/minus emotions. The situation change setting unit obtains, as numerical values, the value acting on the subject data in the state before the situation change and the value acting on the subject data in the state after the situation change. The plus/minus emotion setting unit calculates the difference between the numerical value in the before state and the numerical value in the after state to calculate the plus/minus emotion.

This natural language processing system determines the scene, situation, and positive and negative emotions of the subject model from the input sentence, so it can understand the meaning of the sentence by inferring the most appropriate emotion according to the situation, rather than a simple emotion corresponding to the word.

FIG. 1 is a block diagram of a chatbot system according to a first embodiment of the present invention. FIG. 2 is a block diagram of the semantic understanding program. FIG. 3 is a flowchart showing the process flow of the meaning understanding program. FIG. 4 is a diagram showing the data structure of a story. FIG. 5 is a diagram showing the data structure of a scene. FIG. 6 is a diagram showing the data structure of a concept tree. FIG. 7 shows the data structure of a word object. FIG. 8 is a diagram showing the data structure of an event. FIG. 9 is a diagram showing the data structure of an action. FIG. 10 is a diagram showing the contents of event data. FIG. 11 is a diagram showing the standard response dictionary. FIG. 12 is a diagram showing a mental state dictionary. FIG. 13 is a diagram showing the action of the action word "hit". FIG. 14 is a diagram showing the front-rear state. FIG. 15 is a diagram showing an emotion word response dictionary. FIG. 16 is a block diagram of a robot according to the second embodiment of the present invention. FIG. 17 is a diagram showing the external appearance of the robot. FIG. 18 shows the data structure of an event. FIG. 19 is a diagram showing the self-emotion management section. FIG. 20 shows a virtually constructed data structure of the other person's mind. FIG. 21 shows the data structure of a town Has-a tree. FIG. 22 is a diagram showing two forward and backward states derived from the action of "running."

This invention is about a method for semantic understanding of natural language, that is, a method for making a computer understand the meaning of a sentence in the same way as a human being does. In the field of natural language processing, even a definition of semantic understanding has not yet been established, so the inventors began by defining semantic understanding in two ways.

The first definition is the reduction of the amount of information. Understanding meaning can be said to be some kind of information processing. As an example of information processing, consider numerical calculations. In "1 + 1 = 2," the original information "1 + 1" is processed to become the information "2." As a result of information processing, the amount of information is reduced. The information about what was added together is lost, and instead is summarized into the sum of the sums. Information processing can be said to be the deletion of unnecessary information and the compilation of necessary information. The inventor considered the same to be true for understanding the meaning of natural language. Understanding the meaning of a sentence should mean that the amount of information has been reduced compared to before the meaning was understood. The idea is that if unnecessary elements in a sentence can be eliminated and the necessary elements can be compiled, the meaning can be understood.

The second definition is predicting human behavior. Being able to understand the meaning of words in the same way as a human being means that when you converse with a person, you can respond in the same way as a human being. In other words, if you can predict how a person would speak, you can say that you understand the meaning.

There must be some psychological driving force behind a person's actions or speech. These include various mental states, ranging from instinctual feelings such as "I'm hungry" or "I'm sleepy" to emotions such as "anger," "gratitude" and "regret." If we can understand this mental state, we can predict what that person will say or do next. In other words, if we can accurately express a person's mental state, this is what we call semantic understanding. The core of this invention is to express a person's mental state as a data structure, read sentences in natural language, and infer what mental state they are in. In other words, this can be said to realize the same "mind" as a person in a computer program. And, for example, if a person's mental state is "gratitude," we can predict that they will say "thank you." In other words, we can say that we have achieved behavior prediction, which is the second definition of semantic understanding. This is what it means to have a computer understand the meaning of natural language.

The first embodiment of the present invention will be described below. FIG. 1 is a block diagram of a chatbot system 1 according to the first embodiment of the present invention. The chatbot system 1 shown in FIG. 1 has an input device 2 such as a keyboard through which a user inputs characters, a display device 3 that displays the contents of the dialogue, and a control device 4 that performs a semantic analysis of the input sentence, generates a response sentence, and displays it on the display device 3.

The display device 3 is realized by an LCD (Liquid Crystal Display) or the like, and displays an input sentence 5 from the user and a response sentence 6 from the system. The control device 4 has a CPU (Central Processing Unit) 7 and a main memory 8 into which the meaning understanding program of the present invention is loaded.

Figure 2 is a block diagram showing the configuration of the meaning understanding program 9 of the present invention. The meaning understanding program 9 has a scene setting unit 10, a situation change setting unit 11, a positive/negative emotion setting unit 12, and a meaning analysis unit 13.

Figure 3 is a flowchart showing the processing flow of the meaning understanding program. The overall processing flow of the chatbot system 1 will be explained according to Figure 3.

When the user has finished inputting the input sentence from the input device 2, he or she presses the confirmation key. The control device 4 waits for the input sentence to be confirmed in step S14, and when it is confirmed, in step S15, the control device 4 displays the user's input sentence on the display device 3 as the user input sentence 5.

In step S16, semantic analysis of the input sentence is performed. In the semantic analysis, first, word division processing of the input sentence is performed. For example, a sentence such as "Taro went to school" is divided into "Taro went to school." If the input sentence is in Japanese, this is performed by morphological analysis. The subsequent semantic analysis processing will be explained in detail later.

In step S17, a change in the user's mental state is determined based on the results of the semantic analysis, and if there is a change, the process proceeds to step S18, where an emotional response is made according to the user's mental state. For example, if the user's mental state is "joy," the corresponding response would be "That's good."

If there is no change in the patient's mental state, proceed to step S19 and give a standard response. For example, prompt the patient to continue the conversation by asking, "So, what happened?"

Next, with reference to Figures 4 to 15, we will explain what kind of data the semantic analysis program has and how it processes it. The semantic analysis program manages the content of a conversation as story data 20. As shown in Figure 4, the story data 20 is made up of one or more scene data 21. As shown in Figure 5, the scene data 21 has time data 26, location data 27, event data 28, character data 29, and cast member data 30. These data 26 to 30 set in the scene data 21 are set as word objects. A word object is data that defines a word, and is managed by a concept tree.

Figure 6 shows a concept tree 40 for "pudding" as an example of a concept tree. The concept tree 40 is composed of higher-level concepts and lower-level concepts linked together, with the highest-level concept here being a concrete object 41. The lower-level concept of the concrete object 41 is an inanimate object 42, which has a lower-level concept of food and drink 43, which has a lower-level concept of cooking 44, which has a lower-level concept of sweets 45, which has a lower-level concept of pudding 46. In this example, the concept tree is composed in a single line, but the concept tree may also be branched.

As shown in Figure 7, word objects have attributes. Attributes are managed by attribute name and attribute value, and attribute names include color, shape, name, etc. In the case of pudding, the name would be "pudding", the shape would be "round", and the taste would be "sweet".

The concept tree is structured so that subordinate concepts have (inherit) the attributes of superior concepts. For example, if the taste of sweets (attribute name) is set to "sweet (attribute value)", then the taste of its subordinate pudding will also be "sweet". Note that, in exceptions, it is possible to overwrite attribute values. The time data 26, place data 27, character data 29, and cast data 30 set in Figure 5 are structured so that such word objects can be set.

When people talk, for example, they might say, "This happened yesterday." This story contains a story. A story unfolds in some kind of world. The world we are talking about here is one that has a three-dimensional spatial extent, contains people and objects, and has a flow of time. An example of a world could be the present world, right now, which contains the person who is having the conversation, or it could be a different world different to the real world, such as a fairy tale.

The story data 20 is composed of multiple scene data 21 as shown in FIG. 4. The scene data 21 can be thought of as a cross-section of the world cut into time and place. The scene data 21 includes time data 26, place data 27, and the like. In the time data 26, a word object that is a subordinate concept of the time concept is set, and in the place data 27, a word object that is a subordinate concept of the place concept is set. This scene is generated and set by the scene setting unit 10.

For example, suppose there is an input sentence that reads, "Once upon a time, there lived an old man and an old woman." The semantic analysis program determines that a new story has begun, and first generates new story data 20, then generates scene data 21 and adds it to the story data 20. Next, it sets the time data 26 of the scene data 21 to "Once upon a time," the place data 27 to "A certain place," and sets "Grandfather" and "Grandmother" to the character data 30.

The difference between character data 29 and character data 30 is whether they have a mind or not. Anyone with a mind is called character data 30, but this does not necessarily mean a person; for example, animals that appear in fairy tales, imaginary creatures, and even objects can be called character data 30 if they have a mind. Having a mind means that they can have emotions such as happiness or sadness, or mental states such as making an effort. In this invention, a person with a mind who exists in the real world is called a subject, and the data representing that in a program is called subject data.

When the meaning understanding program reads a sentence, it analyzes the sentence, generates story data 20 and scene data 21, and sets the situation of the scene data 21. In other words, the meaning understanding process can be said to be a process of converting a sentence expressed in natural language into story data 20.

A sentence consists of a subject, predicate, object, etc., and the role of the sentence is determined by the predicate. The roles of a sentence can be broadly divided into two: setting the situation and changing the situation. Setting the situation means setting character data 29 and/or character data 30 in scene data 21, and setting attributes such as names in character data 29 and/or character data 30. In other words, it means setting the situation that is the premise for the story that is about to begin.

For example, in the sentence "There was a grandfather," the role of the sentence is determined by the predicate "is." The predicate "is" has the role of setting the situation, and more specifically, it has the function of making the subject exist in the scene. In this case, the subject is "grandfather," and since "grandfather" is a subordinate concept of a person, a character called "grandfather" is generated as subject data and set as character data 30 in scene data 21.

Adjectives also play a role in setting a situation. For example, in the sentence "The flower is white," the predicate "white" sets the color attribute of the subject "flower" to the attribute value "white."

In the sentence "The grandfather went to the mountain," the role of the sentence is determined by the predicate "go," which acts as a change in situation. "Go" has the function of moving the "subject" to the "destination." A change of location means a change of scene, so when this sentence is read, new scene data 21 is generated, the destination "mountain" is set in the location data 27, and "grandfather" is set in the character data 30. This is a sentence of change of situation.

Since a change in situation is also an event, it is also set in the event data 28 of the scene data 21. Figure 8 is a diagram showing the data structure of the event data 28. The event data 28 has an event subject 50 in which a character or object is set, a before and after state 51 that manages the state before and after the event subject 50, and an action 60 of the event subject 50.

Figure 9 shows the data structure of an action 60. Action 60 is organized around an action word 61 that indicates the action of a sentence. An action word is usually the predicate of a sentence. In the case of "The grandfather went to the mountain," the predicate is "go," so the action word 61 is "go." The data structure of action 60 differs for each action word, and the data structure of the action word "go" is as shown in Figure 9. The action word "go" has a subject 62, a destination 63, a starting point 64, and a meaning 65.

The data structures of all action words are stored in the action word dictionary, and the data structure of "go" can be obtained by searching the action word dictionary with an action word such as "go".

When the semantic understanding program obtains the data structure of the action word "go", it matches it with the sentence "The grandfather went to the mountain" and sets "grandfather" to subject 62 and "mountain" to destination 63. Since the starting point 64 cannot be read from this sentence, it is left blank, but it is also possible to set "home" as a default value. In this way, action 60 is completed.

The contents of the meaning 65 of the action 60 are a program that sets the contents of the action 60 to the event subject 50 and the preceding and following states 51 of the event data 28, and are written in, for example, a scripting language. The meaning understanding program can interpret and execute this scripting language. For the meaning 65, such a scripting language may be written in text format, or an execution program may be set.

In the case of "go", the meaning 65 describes that the subject 62 should be set in the event subject 50 of the event data 28, the previous attribute 52 and the next attribute 54 should be set to "location", the previous attribute value 53 should be set to the contents of the departure point 64, and the next attribute value 55 should be set to the contents of the destination 63. The situation change setting unit 11 of the meaning understanding program 9 reads this meaning and sets it according to the contents. That is, the event subject 50 of the previous and following states 51 is set to "grandfather" and the next attribute value 55 is set to "mountain". If the departure point 64 is "house", then "house" is set in the previous attribute value 53. In this way, the event data 28 is completed.

The character data 30 has a mental state section 31 in which the mental state is set. The mental state section 31 has a positive/negative emotion section in which one's positive and negative emotions are set. A positive emotion is an emotion that is favorable to oneself, and a negative emotion is an emotion that is unfavorable to oneself. For example, unpleasant emotions such as "I'm hungry" or "danger" are negative emotions, while pleasant emotions such as "I'm full" or "I feel relieved" are positive emotions.

The basic principle of behavior for living things is to avoid discomfort (negative emotions) and seek pleasure (positive emotions). This is a basic principle shared not only by humans, but by all living things. Furthermore, positive and negative emotions arise independently from the subject, and can be physical, such as "pain" or "good," or mental, such as "happy" or "sad."

Now, referring again to the flowchart in Figure 3, we will explain the operations from step S16 onwards, which shows semantic analysis.

Suppose the sentence "Nice to meet you, I'm Taro" is first input. Then, as an initial setting, the scene setting unit 10 generates story data 20 and scene data 21, and adds the scene data 21 to the story data 20. It then analyzes the input sentence and extracts the subject "Taro". When "Taro" is searched in the concept tree, it is found to be below the "person's name" concept, so it is determined that "Taro" is a human name, character data 30 is generated, and "Taro" is set as the name attribute. Furthermore, since the predicate is "desu", it can be determined that "Taro" exists, and the generated "Taro" is set as character data 30 of scene data 21.

Figure 10 shows the contents of the event data 28 for this scene data 21.

First, by searching the concept tree for "Nice to meet you," it is found to be below the greeting concept, so the action word dictionary is searched for "greeting" to obtain the data structure of "greeting." Then, the subject 71 of the data structure of "greeting" is set to "Taro," and the greeting content 72 is set to "Nice to meet you, I'm Taro," completing the greeting action 70. Also, since the action word "greeting" does not affect the context of the event subject, its meaning is left blank. Because the meaning is blank, the situation change setting unit 11 does not set anything in the context 51.

Then, proceed to the next step S17, where the plus/minus emotion setting unit 12 determines whether there has been a change in the mental state of the character data 30. In this case, since there is no change in the previous attribute value or the subsequent attribute value of the event data 28, the semantic analysis unit 13 determines that there has been no change in the mental state, and proceeds to step S19, which indicates the next standard response.

Step S19, which indicates a standard response, uses the standard response dictionary shown in FIG. 11 to respond. The standard response dictionary is a dictionary in which the other person's utterance and its response are paired. In this case, since "Taro" uttered "Nice to meet you," the corresponding "Nice to meet you, Taro" is returned as the response. After the greeting, the system continues by asking "Has anything happened recently?" to keep the conversation going. These are displayed on the display device 3 as response sentences 6.

In response to this, suppose Taro replies, for example, "Today I went to the convenience store and drew a lottery." In step S16, which indicates semantic analysis, it is determined that the location has changed from "going to the convenience store," and the scene setting unit 10 generates new scene data 21, sets the location to "convenience store," and sets the event data 28 to "Taro goes to the convenience store." Next, further event data 28 is generated and set to "Taro draws a lottery." In this way, multiple event data 28 may be set for one scene data 21.

In this case, the situation change setting unit 11 finally determines the state before and after the event, with the subsequent attribute 54 being "location" and the subsequent attribute value 55 being "convenience store."

Once the event data 28 has been set, in step S18 the positive/negative emotion setting unit 12 determines the change in the mental state. The mental state is determined by the mental state dictionary shown in FIG. 12. In the mental state dictionary, values corresponding to the attributes of the characters are set as numerical values, and if the numerical value is positive, it is a positive emotion, and if it is negative, it is a negative emotion, and the magnitude of the emotion is indicated by the magnitude of the value.

In the mental state dictionary in Figure 12, the values of positive and negative emotions that are expected when a character takes on that attribute are set. For example, an amusement park such as Disneyland is a fun place, so the positive and negative emotions of a character when in that place will be a high value such as +7. In this case, the value of the location convenience store is +1. If the original location value is 0, the difference is +1.

Here, if the magnitude of the threshold value for emotion generation was set in advance to "+3", it is determined that no emotion is generated because the magnitude does not reach +3. As for the next action "draw lots", since "draw lots" does not exist in the mental state dictionary, this also does not generate an emotion and it is determined that there is no change in the mental state, and the process proceeds to step S19 which indicates a standard response.

Since there is no response corresponding to "draw lots" in the standard response dictionary, if there is no corresponding response, "So what happened?" is displayed on the display device 3 as response sentence 6.

In response, Taro responds, "I won a pudding." In step S16, which shows semantic analysis, first, an action 90 "I won a pudding" is created, as shown in FIG. 13. In action 90, the operative word 61 is "to win," the subject 62 is "Taro," and the object 66 is "pudding." The meaning describes that the before and after states 51 are set so that the possession of the subject becomes the object. When the situation change setting unit 11 converts action 90 into before and after states 51 according to this meaning, the possession of "Taro" becomes "pudding" as a posterior attribute value, as shown in FIG. 14.

As shown in FIG. 6, pudding is a subordinate concept of sweets. Looking at the mental state dictionary in FIG. 12, the value of the sweets in possession is +5. This means that Taro's positive/negative emotions have increased by 5, from 0 in the previous state to +5 in the later state. Since this is a value greater than the reference value of +3, the positive/negative emotion setting unit 12 generates and sets a positive emotion. Then, the process proceeds to step S17, where the semantic analysis unit 13 performs an emotional word response.

The emotional response is based on the emotional response dictionary shown in FIG. 15. In the emotional response dictionary, the response word is "That's good" for a positive emotion and "That's unfortunate" for a negative emotion. In this case, a positive emotion has occurred, so the response sentence 6 "That's good" is displayed on the display device 3.

People have a conversation because they have something to say. What they want to say is something that happened yesterday. Simply put, it is emotions, such as something fun or something sad. Breaking down emotions, the most basic emotions are positive and negative emotions. In other words, it is something that is desirable or undesirable for that person. The emergence of emotions can be said to be the smallest unit of what someone wants to say, and also the smallest unit of the "story" of this invention. Therefore, if you cannot read emotions from the other person's speech, you can determine that the "story" is not yet complete, and you can respond by asking, "So what happened?" to encourage them to continue. In this way, natural conversation is possible.

In a nutshell, the meaning understanding method of the present invention processes the content of a sentence so as to break it down into the positive and negative emotions of the subject. For example, from a sentence such as "Taro drew a lottery at the convenience store and won a pudding," it is possible to read from data such as the concept tree and attributes of word objects that the possible meanings of this sentence are infinite, such as "Pudding is sweet and round" or "The convenience store clerk must have taken the pudding out of the refrigerator."

This is then processed in a way that breaks it down into Taro's positive and negative emotions. In other words, the information is organized in a way that leads to whether it was good or bad for Taro. To achieve this, the meaning of the verb "hit" is defined as a change in the possessions of the characters. This is because people feel happy when they obtain something valuable, and sad when they lose something valuable; positive and negative emotions arise.

The meaning of the action word can be defined in various ways, but in this invention, it can be said that the meaning that contributes to the change in the positive and negative emotions of the characters is selected. In this way, by organizing the meaning of the action word into those related to positive and negative emotions, it is possible to narrow down the meaning of the sentence. As a result, it can be said that the amount of information is reduced, which is the first definition of semantic understanding of natural language.

Next, we will explain behavior prediction, which is the second definition of natural language semantic understanding, using the second embodiment.

Figure 16 is a block diagram showing a second embodiment in which the present invention is incorporated into a robot 100, and Figure 17 shows its external appearance. The robot 100 is a humanoid robot that is roughly the same size as a human and has roughly the same functions as a human. That is, it captures images of the outside world with a camera 101 attached to its eye 107, listens to sounds through a microphone 102, and emits sounds from a speaker 103 attached to its mouth 108. In addition, its limbs 109 are driven by motors 104, and it also has contact sensors 105 and the like, all of which are controlled by a control device 106.

In the control device 106, a meaning understanding program similar to that of the first embodiment is executed. However, in this embodiment, the input is voice input from the microphone 102, and the output is voice output from the speaker 103. Therefore, the input sentence is converted from voice to text, and the response sentence is also converted into voice and output, enabling a conversation in a form closer to that of a human. In addition to voice, the camera 101 and the sensor 105 can also acquire information about the external environment. Another difference is the content of the meaning understanding program.

The meaning understanding program of this embodiment is incorporated as part of a psychological program that functions as the "mind" of the robot 100. The psychological program is a program that can understand various human psychological states, and is also capable of understanding social emotions. Social emotions are emotions that exist between oneself and another person, and include things like "gratitude," "anger," and "shame."

As an example, a method for understanding the emotion of "gratitude" will be described. Here, let us assume that the name of the robot 100 is "Robota" and that Taro has given Robota a present. For example, let us assume that Taro gives the present to Robota, saying, "Robota, I'll give you a present."

Figure 18 shows the data structure of an event 110 in the meaning understanding program of this embodiment. The event 110 has two characters, a self 111 and another person 112. The self refers to the person on whose behalf the psychological program is being executed, and the other person refers to the person with whom the self is having a conversation.

When the input sentence "Robota, I'll give you a present" is received, the predicate "give" is searched for in the concept tree, it is determined to be a subordinate concept of "yield", and an action for event 110 is generated. Since the action word 120 is "yield", the data structure for "yield" is obtained from the action word dictionary. For the obtained data structure, "yield" is set as the action word 120, "Taro" as the subject 121, "Robota" as the person to give to, and "present" as the thing to give to. This completes the action.

Then, the state before and after the event 110 is set according to the meaning of the action 124. The state before and after the event 110 is a data structure that indicates the state of the user before and after the action. In this case, the previous attribute name 130 and the next attribute name 132 are "possession", the previous attribute value 131 is empty, and the next attribute value 133 is "present".

When searching for "present" in the mental state dictionary, the result is a value such as +8, and Robota's positive and negative emotions will be +8, which is greater than the base value, so Robota will experience a positive emotion, such as "happy."

What has arisen here is "my" emotion. Next, if we look at the action, the subject 121 is "Taro." In other words, it is the person with whom we are currently speaking who has acted. If we feel positive emotion as a result of the other person's action, then the cause of our positive emotion is the other person.

Here, we create something called a social emotion dictionary. This registers patterns of social emotions that arise when there is a relationship between oneself and another person, such as "gratitude" or "anger." In the case of "gratitude," if a positive emotion arises in oneself and the cause is the other person, the emotion "gratitude" will arise toward that other person. In this case, since it was Taro who gave the present, Robota will feel "gratitude" toward Taro. Also, by registering "thank you" as the words to say to the other person when the emotion "gratitude" arises, Robota will be able to say "thank you" to Taro when he receives a present.

In this way, the meaning understanding program of this embodiment is able to understand the emotion of "gratitude." It is also able to say "thank you" when receiving a gift. This is the same behavior that humans exhibit. In other words, it can be said that it is able to predict human behavior, and that it has achieved the second definition of meaning understanding defined at the beginning.

Now let's consider the case where the present was not something Robota wanted. For example, Taro thinks that Robota will like batteries, so he gives him some as a present. Robota is not powered by batteries, so he will not be happy to receive them. Even in this case, a human would pick up on the other person's kindness and say "thank you."

The next step is to make it possible to understand such complex emotions. To achieve this, as shown in Figure 19, the system has a self-emotion management section 150 that manages one's own positive and negative emotions. The self-emotion management section 150 has emotions 151 that actually occurred to oneself, and one's own emotions 152 that the other person intended. In this case, one's actual emotions are the positive and negative emotions of Robota who received the dry cell battery, and are therefore negative. The other person's intention is Taro (the other person), who thought that Robota (oneself) would be happy (positive emotion) when he received the dry cell battery, and therefore the emotion 152 that the other person (Taro) intended is positive.

The meaning understanding program sets the self-preceding and subsequent states, and then sets this self-emotion management section 150. It then registers in the social emotion dictionary that if the self-emotion 152 intended by the other person is positive, then the emotion of "gratitude" will be generated. In this way, it becomes possible to understand more complex emotions, such as picking up on the other person's feelings and prioritizing the other person's feelings over one's own, such as saying "thank you."

In addition, whether to prioritize one's own feelings or the intentions of others can be determined by setting parameters in the psychological program. If the robot tends to prioritize others, it will be "kind," and if it tends to prioritize itself, it will be "selfish." In this way, the robot's "personality" can also be set.

This function can be said to be the function of surmising the other person's mind, or imagining events that the other person is thinking about. Imagining events that the other person is thinking about can also be said to be simulating events that the other person is thinking about. In order to realize this function, the other person's mind is virtually constructed. This is shown in Figure 20. In the virtual other person's mind 160, the other person's thinking about events 161 are calculated. The other person's thinking about events 161 have the other person's before and after state 162. The other person that the other person thinks of is the other person from the other person's perspective, so it is oneself. Therefore, the attribute values set in the other person's before and after state 162 are the attribute values that the other person thinks of oneself.

To explain using the previous example, the self is Robota and the other person is Taro, and Robota (self) imagines Taro's (other person) imaginary mind 160. An event 161 that Taro is thinking about is set in the imaginary mind 160 of Taro. When Taro gives Robota a battery as a gift, Robota's post-attribute value increases by the value of the battery Robota received. The value of the battery Robota received is not the value Robota actually feels, but the value of Robota's battery that Taro imagines. Taro imagines the value of the battery that Robota feels based on Taro's experience and knowledge. For example, "Because he is a robot, he will be happy to receive a battery." This can be realized by creating a mental state dictionary for each subject as Taro's knowledge, and recording it as "The value of a battery to a robot (subject) is +5." In this way, the other person's (Taro) imagines that the self (Robota)'s emotion is +5 when he receives a battery. In this case, as a result of the other person's (Taro's) actions, your (Robota's) positive or negative emotions become positive, which can be said to be the other person's intention. If the other person's intended emotions are positive, then the other person acted with the intention of making you happy, and so you will feel "gratitude" towards the other person. Conversely, if the other person's intended emotions are negative, then the other person acted with the intention of making you uncomfortable, and so you will feel "anger."

Guessing someone else's feelings does not only mean when they act, but also includes estimating their evaluation and value of you. In other words, you can guess what they think of you. If you can do this, you can generate things like other people's gazes, and you can also generate the emotion of "shame." In any group or society, there are minimum standards that you should meet if you belong to that society. For example, "A middle school student should know this much kanji," or "A member of the track and field club should be able to run 50 meters in under six seconds." Suppose that a minimum standard is set in a particular society, such as "If you are something, you should be able to at least do this." This standard is derived from experience, etc. Then, when you, who belong to that society, do not meet the standard and simulate what others think when they see you, you will generate the emotion of "shame." Then, you can guess your value from others' perspective, such as "Even though you are something, you can't do this," and generate the emotion of "embarrassment."

People are able to form a cohesive society by being able to feel this emotion of "shame." This is because, in order to belong to a society, a psychological need arises to meet a minimum standard, and this becomes the driving force behind acting in such a way as to meet that standard. By having this emotion of "shame," we can create artificial intelligence with a mind that can blend into human society.

Next, let's explain "anger." For example, suppose someone says "piece of junk" to Robota. Robota uses speech recognition to convert the speech he hears into a string of characters and performs semantic analysis. The word "piece of junk" indicates that the target's performance is poor. As the target is Robota, Robota determines that the person has spoken ill of him, and negative emotions arise. When Robota looks, he can confirm from image recognition that it was Jiro who made the comment.

When this situation is set in the self emotion management unit 150, the self emotion becomes negative. Also, since the other person said it, it is determined that the other person made the statement with intention, and the other person's intended self emotion 152 also becomes negative. In the social emotion dictionary, it is registered that when the other person's intended self emotion 152 is negative, the emotion of "anger" occurs toward the other person who acted in that way. As a result, in this case, the emotion of "anger" occurs, and the person responds in accordance with the emotion of anger, for example, saying "What did you say?!".

Now, let's say that the person you have spoken ill of says "I'm sorry." When you search for "I'm sorry" in the concept tree, you find that it is below the concept of "apology," so you create an action for "apology." When you set the operative word as "apology" and set an event based on the meaning of "apology," the other person's intended emotion 152 becomes ±0. Also, the person who apologizes is in a lower position relative to the person to whom they are apologizing. In other words, an "apology" has the function of showing that there is no malice by lowering one's own position. When the meaning understanding program reanalyzes the situation, the other person's intended emotion 152 becomes 0, so anger does not arise, in other words, the anger subsides. In this way, it becomes possible to understand psychology such as "apology."

Next, we will explain behavioral prediction based on social emotions. People try to maintain psychological balance with others. For example, in the case of the emotion of "anger," the other person causes negative emotions in the person. To achieve balance means that if the other person causes negative emotions in the person, then the other person should also have negative emotions, so that both parties become negative and balance is achieved. Therefore, when the emotion of "anger" arises, the psychological program searches for a statement that will cause the other person to feel negative emotions as the next response.

This can be achieved by reversing the process we have done so far. In other words, in event 110, we imagine the other person's situation before and after the event, not our own, and look for a posterior attribute value that will cause the other person to feel negative emotions, and then look for an action that will set that attribute value. In other words, we can think from the other person's perspective.

As an example, suppose someone calls you an idiot and you feel negative emotions. You might search for actions that would cause the other person to feel the same negative emotions as you, and say back, for example, "You idiot." This can be said to be an understanding of the meaning of "revenge" or "retaliation." This is also an example of meaning understanding as a form of behavioral prediction.

Even when feelings of gratitude arise, they try to balance things out by acting in a way that causes the other person to feel positive emotions. For example, if you receive a birthday present, you will feel positive emotions, so you will act by giving the other person a present on their birthday so that they will also feel positive emotions. This can be said to be acting with an understanding of the meaning of "returning the favor." In this way, by acting to balance the positive and negative emotions received from others, it becomes a psychological program that is closer to the human mind.

Next, we will explain the two most important social emotions, "good" and "evil." "Good" can be defined as "acting in a way that will result in positive positive or negative emotions for the other person." In other words, doing what makes the other person happy is "good." When multiple actions are possible, the meaning of "good" can be understood by configuring a psychological program to select an action that will result in positive emotions for the other person. This is because being able to act in the same way as humans (behavioral prediction) can be said to indicate understanding of the meaning. Similarly, "evil" can be defined as when multiple actions are possible, selecting an action that will result in negative emotions for the other person.

Now let us consider judging right and wrong when there is no other party involved. For example, consider the act of throwing trash on the street. This is clearly a "bad" action, but there is no clear other party involved. In such cases, a related third party is set as the other party. A related third party is an entity that is affected by the action; in the case of "throwing trash on the street," this would be the people who use that street.

Here, we will explain how to extract related third parties. Words are managed not only in the concept tree, but also in a Has-a tree. A Has-a tree is a data structure that manages words in terms of whole-part relationships.

Figure 21 shows the data structure of the Has-a tree 141 of town 140. Let us assume that when searching for a Has-a tree that includes road 142, town 140 is found. Let us assume that town 140 has residents 143, road 142, school 144, hospital 145, etc. By searching for the subjects contained in town 140, resident 143 is obtained, and this resident 143 is set as a third party related to the road. Next, the action of "throwing out garbage" is executed. Then, the subject of the event is set as town 140, and in the post state of town 140, the post attribute value of the post attribute "environment" becomes "pollution". Next, the before and after states of resident 143 of town 140 in this case are set. Since the environment of town 140 is "pollution" and in a negative state, it is calculated that the emotion of resident 143 contained in town 140 will also be negative. From the above, the act of "throwing garbage on the road" can be said to be a "bad" act because it causes negative feelings to the residents 143 of the town 140 related to that road.

Similarly, picking up trash on the street can be judged as a "good" deed because it makes the town 140 beautiful and causes the residents 143 of the town 140 to feel positive emotions. In this way, the goodness or badness of an action can be judged by calculating the positive or negative emotions of a third party regarding the action when it is performed. In this way, the psychological program can judge what is good and bad in human society, and so it can be said that the robot 100 equipped with the psychological program can coexist with humans in human society.

Next, we will explain the understanding of the meaning of the word "effort", which is a psychological state. For example, let's say Taro is practicing running. Figure 22 shows two before and after states derived from the action of "running". Before and after state A shows that running makes the body tired. Before and after state B shows that running makes the legs faster. Before and after state A gives rise to the negative emotion of fatigue, while before and after state B gives rise to the positive emotion of faster legs. The psychological state of "effort" can be defined as the psychological state of choosing an action in order to achieve a positive emotion, even if it gives rise to a negative emotion.

Furthermore, it can be calculated that by repeating this action, one's running will become faster. Repeating an action means taking time. In other words, this is where the concept of "time" comes into play. Let's also say that a condition is set to run 50 meters in under 7 seconds. If a behavioral pattern is set to repeat the action of running in order to meet this condition, this defines the psychological state of "making an effort toward a goal," and the condition set in this behavioral pattern becomes the meaning of the word "goal."

In this way, we can understand psychological states such as "effort" and "goals." Furthermore, we can define words and actions that encourage someone who is engaging in such behavior as "support." One example of "support" is the word "hang in there." For example, when we recognize someone who is working hard toward a goal, we can say "hang in there!" to them.

Next, let's define the word "regret." For example, suppose you set a goal, such as getting into college, and make an effort to study towards it. However, suppose you don't get into college as a result. In this case, we can infer that you didn't study enough. In other words, we can calculate that you didn't study enough in the time leading up to the college entrance exam. If you were to increase the amount of studying, you can generate a hypothetical event in which you would have passed and compare it with the actual event. By comparing the two, you can calculate that the difference is the number of studying actions. In other words, by changing your past actions, the negative emotions that actually occurred become positive emotions. If we call this pattern "regret," we can say that we understand the meaning of "regret." In other words, we can understand the emotion of "regret," such as wishing that if only you had studied more, you would have been able to get into college.

In this way, the meaning understanding program of the present invention can understand various psychological situations.

In each embodiment of the present invention, semantic understanding is performed by a control device integrated with the chatbot system 1 or the robot 100, but the present invention is not limited to this, and for example, semantic understanding may be performed on a server or cloud connected via the Internet, and the results may be returned to the chatbot system 1 or the robot 100.

In addition, although each embodiment of the present invention has been described as a conversational system, the present invention is not limited to conversation or dialogue, and the meaning understanding program of the present invention may also be applied to automatic document summarization and machine translation.

It should be understood that the embodiments disclosed herein are illustrative in all respects and are not restrictive in any respect. The scope of the present invention is defined by the claims, not the above description, and is intended to include all modifications within the meaning and scope of the claims.

1 Chatbot system, 9 Meaning understanding program, 10 Scene setting section, 11 Situation change setting section, 12 Positive and negative emotion setting section, 13 Meaning analysis section, 100 Robot

Claims (3)

A natural language processing system comprising: an input device into which a sentence is input by characters or voice; and a control device having a CPU and a main memory, and configured to break down the sentence input from the input device into words,
The input sentence input to the input device includes a character having a mental state;
a character object corresponding to the character is generated by the control device in the main memory, the character object having attributes including positive emotions that are favorable to the character and negative emotions that are unfavorable to the character ;
The control device generates the character object in the main memory based on the input sentence , sets an attribute of the positive emotion or the negative emotion to the character object, and predicts the behavior of the character based on the positive emotion and the negative emotion, thereby understanding the meaning of the input sentence. 1. A natural language processing method for use in a natural language processing system having an input device into which a sentence by characters or voice is input, and a control device having a CPU and a main storage memory, and for breaking down the sentence input from the input device into words, comprising:
the input sentence input to the input device includes a character having a mental state;
a character object corresponding to the character is generated by the control device in the main memory, the character object having attributes including positive emotions that are favorable to the character and negative emotions that are unfavorable to the character ;
The natural language processing method includes:
A natural language processing method comprising the steps of: the control device generating the character object in the main storage memory based on the input sentence; setting an attribute of the positive emotion or the negative emotion to the character object ; and understanding the meaning of the input sentence by predicting the behavior of the character based on the positive emotion and the negative emotion . A natural language processing program for use in a natural language processing system having an input device into which a sentence by characters or voice is input, and a control device having a CPU and a main memory, and for breaking down the sentence input from the input device into words,
The input sentence input to the input device includes a character having a mental state;
a character object corresponding to the character is generated by the control device in the main memory, the character object having attributes including positive emotions that are favorable to the character and negative emotions that are unfavorable to the character ;
Computer,
A natural language processing program for causing the control device to function to generate a character object in the main storage memory based on the input sentence, set an attribute of the positive emotion or the negative emotion to the character object, and understand the meaning of the input sentence by predicting the behavior of the character based on the positive emotion and the negative emotion .

Images