JP2023156489A - Natural language processing system, natural language processing method, and natural language processing program - Google Patents

Abstract

To provide a natural language processing system, a method, and a program that can set positive and negative emotions of a scene, a situation, and a subject model from an input sentence, and infer an optimal emotion according to the situation rather than a simple emotion corresponding to a word, and understand a meaning of the sentence.SOLUTION: A natural language processing system comprises an input device to which sentences are input by characters or voice, and a control device which has a CPU and a main storage memory and resolves the sentences input from the input device into words to analyze a meaning thereof. In the main storage memory, a word object that has a word corresponding to a subject who understands a natural language, has a mental state, and has a positive emotion that is favorable to oneself and a negative emotion that is unfavorable to oneself is generated. The control device generates the word object of the subject from a word of the input sentence, determines the mental state of the subject by determining positive or negative emotions of the subject, and performs semantic analysis by predicting an action of the subject.SELECTED DRAWING: Figure 3

Description

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

This application claims priority based on the international application PCT/JP2019/022680 filed on June 7, 2019, and incorporates all the contents described in the international application.

In recent years, the number of systems that allow interaction using human language (natural language) has been increasing. For example, in Japanese Patent Application Laid-Open No. 2006-178063 (Patent Document 1), an emotion word dictionary is created that categorizes words into three emotions: negative, positive, and neutral, and the user's emotions are determined based on the emotion words included in the user's utterances. An interaction processing device is disclosed that estimates the answer and provides a response based on the guess.

Japanese Patent Application Publication No. 2006-178063

According to the technology disclosed in Patent Document 1, when an emotional word appears, a response corresponding to the emotional word is simply made. Therefore, the dialogue processing device disclosed in Patent Document 1 does not understand and output the meaning of a sentence. As a result, there is a problem in that the responses are inconsistent and a natural conversation cannot continue.

In the first place, emotions do not have a one-to-one correspondence with words; even the same word can produce different emotions depending on the situation and perspective. For example, when you receive a present, you feel "happy" and you also feel "thankful" to the person who gave you the gift. If you create an emotion dictionary that matches the word ``present'' with positive emotions, you can read the positive in the sentence ``I received a present from my father.'' "I have to do it," is not the answer.

The fundamental problem with this is that the field of natural language processing has not yet achieved semantic understanding or even a definition of semantic understanding.

An 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 analyzes input natural language sentences, 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 generates scene data representing a part of the world in which the subject acts. A scene setting section that sets subject data in data, a situation change setting section that analyzes input natural language sentences and sets changes in the situation of the scene, and creates emotions favorable to the subject based on the changes in the situation of the scene. It has a plus/minus emotion setting unit that determines whether a positive emotion or a negative emotion that is an unfavorable emotion for the subject has occurred, and a meaning analysis unit that performs meaning analysis based on the generated positive/minus emotion. The situation change setting unit obtains, as numerical values, the values that act on the subject data in a state before the situation change and the values that act on the subject data in a state after the situation change. The plus/minus emotion setting unit calculates the plus/minus emotion by calculating the difference between the numerical value of the previous state and the numerical value of the subsequent state.

Furthermore, the natural language processing method according to the present invention analyzes input natural language sentences and 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. , a step of setting subject data in the scene data, a step of analyzing the input natural language sentence and setting a situation change of the scene, and a step of setting a situation change that is favorable for the subject based on the situation change of the scene. The present invention includes a step of determining whether a positive emotion, which is a negative emotion, or a negative emotion, which is an unfavorable emotion for the subject, has occurred, and a step of performing semantic analysis based on the generated positive emotion. In the step of setting the situation change, 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 are obtained as numerical values. In the determining step, the difference between the numerical value of the previous state and the numerical value of the subsequent state is calculated to calculate the plus/minus emotion.

Further, the natural language processing program according to the present invention causes a computer to analyze input natural language sentences and generate subject data representing a subject who understands the natural language, and scene data representing a part of the world in which the subject acts. a scene setting section that generates and sets subject data in the scene data; a situation change setting section that analyzes input natural language sentences and sets changes in the situation of the scene; A natural language for functioning as a plus/minus emotion setting unit that determines whether a positive emotion that is an emotion or a negative emotion that is an unfavorable emotion for the subject has occurred, and a semantic analysis unit that performs meaning analysis based on the generated positive or negative emotion. It is a processing program. The situation change setting unit obtains, as numerical values, the values that act on the subject data in a state before the situation change and the values that act on the subject data in a state after the situation change. The plus/minus emotion setting unit calculates the plus/minus emotion by calculating the difference between the numerical value of the previous state and the numerical value of the subsequent state.

This natural language processing system sets the positive and negative emotions of the scene, situation, and subject model from the input sentence, so it estimates the optimal emotion according to the situation, rather than the simple emotion that corresponds to the word. Can understand the meaning of sentences.

FIG. 1 is a block diagram of a chatbot system according to the first embodiment of the present invention. FIG. 2 is a block diagram of the semantic understanding program. FIG. 3 is a flowchart showing the processing flow of the semantic understanding program. FIG. 4 is a diagram showing the data structure of the 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 is a diagram showing the data structure of a word object. FIG. 8 is a diagram showing the data structure of events. FIG. 9 is a diagram showing the behavior data structure. FIG. 10 is a diagram showing the contents of event data. FIG. 11 is a diagram showing a standard response dictionary. FIG. 12 is a diagram showing a mental state dictionary. FIG. 13 is a diagram showing the behavior of the action word "aru". FIG. 14 is a diagram showing the front and back states. FIG. 15 is a diagram showing an emotion word response dictionary. FIG. 16 is a block diagram of a robot according to a second embodiment of the present invention. FIG. 17 is a diagram showing the appearance of the robot. FIG. 18 shows the event data structure. FIG. 19 is a diagram showing the self-emotion management section. FIG. 20 is a diagram showing the data structure of the virtual partner's mind. FIG. 21 shows the data structure of the town's Has-a tree. FIG. 22 is a diagram showing two before and after states derived from the action of "running".

The present invention relates to a method for understanding the meaning of natural language, that is, a method for making a computer understand the meaning of sentences in the same way as humans do. In the field of natural language processing, even the definition of semantic understanding has not yet been determined, so the inventor started by defining two aspects of semantic understanding.

The first definition is information reduction. Understanding meaning can be said to involve performing some kind of information processing. Consider numerical calculation as an example of information processing. "1+1=2" becomes the information "2" by processing the originally existing information "1+1". As a result of information processing, the amount of information has been reduced. The information about what was added was lost, but instead it was summarized as the total value. Information processing is the process of deleting unnecessary information and organizing it into necessary information. The inventor thought that the same could be said for understanding the meaning of natural language. To understand the meaning of a sentence, the amount of information must be reduced compared to before understanding the meaning. The idea is that if you can eliminate unnecessary elements in a sentence and summarize it into necessary elements, you have understood the meaning.

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

There must be some kind of psychological driving force behind a person's actions or words. These are people's various mental states, from those rooted in instincts such as ``I'm hungry'' and ``sleepy,'' to emotions such as ``anger,'' ``gratitude,'' and ``frustration.'' If you understand this mental state, you can predict the person's next words and actions. In other words, if a mental situation can be expressed accurately, that can be said to be an understanding of meaning. The core of the present invention is to represent a person's mental state as a data structure, read sentences in natural language, and estimate what kind of mental state the person is in. In other words, it can be said that the same ``mind'' as a human being is realized by a computer program. For example, if that person's mental state is ``grateful,'' it can be predicted that he or she will say ``thank you.'' In other words, it can be said that behavioral prediction, which is the second definition of semantic understanding, has been achieved. This is what it means to use a computer to understand the meaning of natural language.

A 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. A chatbot system 1 shown in FIG. 1 includes an input device 2 such as a keyboard through which a user inputs characters, a display device 3 that displays dialogue content, and a display device 3 that analyzes the meaning of input sentences and generates response sentences. and a control device 4 for displaying the information.

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 semantic understanding program of the present invention is loaded.

FIG. 2 is a block diagram showing the configuration of the semantic understanding program 9 of the present invention. The meaning understanding program 9 includes a scene setting section 10, a situation change setting section 11, a plus/minus emotion setting section 12, and a meaning analysis section 13.

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

When the user finishes inputting the input sentence from the input device 2, the user presses the enter key. The control device 4 waits for the input sentence to be finalized in step S14, and when the input sentence is determined, the user's input sentence is displayed on the display device 3 as the user input sentence 5 in step S15.

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

In step S17, a change in the mental state is determined as a result of the semantic analysis, and if there is a change, the process proceeds to step S18, and an emotional response corresponding to the mental state is performed. For example, if the user's state of mind is "joy," the corresponding response would be, "That's great."

If there is no change in mental state, the process advances to step S19 and a standard response is made. For example, prompt the next conversation with something like, "So what happened?"

Next, what kind of data the semantic analysis program has and how it is processed will be explained with reference to FIGS. 4 to 15. The semantic analysis program manages the contents of the conversation as story data 20. The story data 20 is composed of one or more scene data 21 as shown in FIG. As shown in FIG. 5, the scene data 21 includes time data 26, place data 27, event data 28, appearing object data 29, and appearing character 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.

FIG. 6 shows a concept tree 40 of "pudding" as an example of a concept tree. The concept tree 40 is composed of a series of higher-level concepts and lower-level concepts, and the highest level concept here is a concrete object 41. A subordinate concept of the concrete object 41 is an inanimate object 42, a subordinate concept thereof is food 43, a subordinate concept thereof is cooking 44, a subordinate concept thereof is confectionery 45, and a subordinate concept thereof is pudding 46. In this example, the concept tree is constructed in a single column, but the concept tree may be branched.

As shown in FIG. 7, word objects have attributes. Attributes are managed using attribute names and attribute values, and attribute names include color, shape, name, etc. In the case of pudding, the name is ``pudding,'' the shape is ``round,'' and the taste is ``sweet.''

The concept tree is constructed such that lower-level concepts have (inherit) the attributes of higher-level concepts. For example, if the taste (attribute name) of confectionery is set to "sweet (attribute value)," the taste of pudding that is subordinate to it is also set to "sweet." Note that if there are exceptions, it is also possible to overwrite the attribute value. The time data 26, location data 27, appearance data 29, and appearance character data 30 set in FIG. 5 are configured such that such word objects can be set.

When people talk, they say things like, ``This happened yesterday.'' It contains one story. The story takes place in some kind of world. The world here has a three-dimensional spatial expanse, includes people and things, and is one in which time is flowing. As an example of the world, it can be the current world that includes the person having the conversation, or it can be another world that is different from the real world, such as an old story.

The story data 20 is composed of a plurality of scene data 21 as shown in FIG. The scene data 21 can be said to be a cross-section of the world cut out by time and place. The scene data 21 includes time data 26, location 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 location data 27, a word object that is a subordinate concept of the location concept is set. This scene is generated and set by the scene setting section 10.

For example, suppose there is an input sentence such as "Once upon a time, there was an old man and an old woman." The semantic analysis program determines that a new story has started, and first generates new story data 20, then generates scene data 21, and adds it to the story data 20. Next, "once upon a time" is set in the time data 26 of the scene data 21, "some place" is set in the place data 27, and "grandfather" and "grandmother" are set in the character data 30.

The difference between the character data 29 and the character data 30 is whether they have a heart or not. Character data 30 refers to people with hearts, but this is not limited to people; for example, animals that appear in folk tales, imaginary creatures, and even objects can be character data if they have a heart. Set it to 30. Having a mind means being able to have emotions such as happiness and sadness, and mental states such as effort. In the present invention, a person with a mind existing in the actual world is called a subject, and the data expressed as data 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 role of sentences can be broadly divided into two: setting the situation and changing the situation. Setting the situation means setting the appearing object data 29 and appearing character data 30 in the scene data 21, and setting attributes such as names in the appearing object data 29 and appearing character data 30. In other words, it is to set up the situation that is the premise of the story that is about to begin.

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

Adjectives also play a role in setting the situation. For example, in the case of the sentence "flowers are white," the predicate "white" means setting the attribute value "white" to the color attribute of the subject "flower."

In the case of the sentence ``My grandfather went to the mountains,'' the role of the sentence is determined by the predicate ``go,'' and the predicate ``go'' plays a role in changing the situation. ``Go'' has the function of moving the ``subject'' to the ``destination.'' Changing the location means changing the scene, so when this sentence is read, new scene data 21 is generated, the destination "Mountain" is set in the location data 27, and "Mountain" is set in the character data 30. Set "Grandfather". This is a change of situation sentence.

Furthermore, since a change in situation is also an event, it is also set in the event data 28 of the scene data 21. FIG. 8 is a diagram showing the data structure of the event data 28. The event data 28 includes an event entity 50 in which characters or objects appearing in the event are set, a before/after state 51 for managing the before and after states of the event entity 50, and actions 60 of the event entity 50.

FIG. 9 is a diagram showing the data structure of the action 60. The actions 60 are grouped around action words 61 that indicate the action of the sentence. An agent is usually a predicate of a sentence. In the case of ``My grandfather went to the mountains,'' the predicate is ``go,'' so the operator 61 is ``go.'' The data structure of the action 60 differs depending on the action word, and the data structure of the action word "go" has the contents shown in FIG. The action word "go" has a subject 62, a destination 63, a starting point 64, and a meaning 65.

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

When the meaning understanding program obtains the data structure of the action word "go", it compares it with the sentence "My grandfather went to the mountain" and sets "Grandpa" as the subject 62 and "Mountain" as the destination 63. Since the starting point 64 cannot be read from this sentence, the starting point 64 is left blank, but "home" may be set as the default value. This is how action 60 is completed.

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

In the case of "go", the meaning 65 includes the subject 62 in the event subject 50 of the event data 28, the "place" in the preceding attribute 52 and the posterior attribute 54, the content of the place of departure 64 in the preceding attribute value 53, and the content of the starting place 64 in the preceding attribute value 53. Attribute value 55 describes the content to set the contents of destination 63. The situation change setting unit 11 of the meaning understanding program 9 reads this meaning and sets it according to the content. That is, the event subject 50 of the before and after state 51 is set as "old man", and the after attribute value 55 is set as "mountain". Note that if the starting place 64 is "home", "home" is set as the previous attribute value 53. In this way, event data 28 is completed.

The character data 30 includes a mental state section 31 in which a mental state is set. The mental state section 31 has a plus/minus emotion section where one's own plus/minus emotions are set. Positive emotions are emotions that are favorable to oneself, and negative emotions are emotions that are unfavorable to oneself. For example, negative emotions are things that are unpleasant, such as ``I'm hungry'' or ``danger,'' and positive emotions are things that are pleasant, such as ``I'm full,'' or ``I feel safe.''

The basic principle of biological behavior is to avoid discomfort (negative emotions) and seek pleasure (positive emotions). This is a basic principle held by all living things, not just humans. Note that positive and negative emotions are generated by the subject alone, and include physical ones such as ``it hurts'' and ``feels good,'' and mental ones such as ``happy'' and ``sad.''

Now, with reference to the flowchart of FIG. 3 again, the operations from step S16 onwards showing semantic analysis will be explained.

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

FIG. 10 is a diagram showing the contents of the event data 28 of the scene data 21.

First, when we search for ``Nice to meet you'' in the concept tree, we find that it is less than the greeting concept, so we search the operative word dictionary for ``greeting'' and obtain the data structure of ``greeting.'' Then, the subject 71 of the "greeting" data structure is set as "Taro" and the greeting content 72 is set as "nice to meet you, Taro" to complete the greeting action 70. Furthermore, the operative word ``greeting'' does not change the context of the subject of the event, so its meaning is left blank. Since the meaning is blank, the situation change setting unit 11 does not set anything in the previous and subsequent states 51.

Here, the process proceeds to the next step S17, and the plus/minus emotion setting unit 12 determines whether or not there has been a change in the mental state of the character data 30. In this case, since there is no change in the preceding attribute value and subsequent attribute value of the event data 28, the semantic analysis unit 13 determines that there is no change in the mental state and proceeds to step S19, which indicates the next standard response.

In step S19, which indicates a standard response, a standard response dictionary shown in FIG. 11 is used to respond. The standard response dictionary is a dictionary that pairs the other party's utterance and its response. In this case, the utterance of "Taro" is "Nice to meet you," so the corresponding response is "Nice to meet you, Taro-kun." . Also, after greeting someone, follow up by asking, ``Has anything happened recently?'' to keep the conversation going. These are displayed on the display device 3 as a response sentence 6.

In response to this, Taro might reply, ``Today, I went to the convenience store and drew a lottery.'' In step S16 showing 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 as "convenience store", and changes the location to the event data 28. , "Taro goes to the convenience store" is set. Subsequently, event data 28 is further generated, and "Taro draws a lottery" is set. In this way, a plurality of event data 28 may be set in one scene data 21.

In this case, the situation change setting unit 11 determines that the following attribute 54 is "place" and the subsequent attribute value 55 is "convenience store" as the state before and after the event.

Once the event data 28 is set, the plus/minus emotion setting unit 12 determines a change in the mental state in step S18. The mental state is determined using the mental state dictionary shown in FIG. The mental state dictionary has numerical values set according to the character's attributes, and if the numerical value is positive, it is a positive emotion, and if the numerical value is negative, it is a negative emotion, and the magnitude of the emotion is indicated by the magnitude of the value. It will be done.

In the mental state dictionary shown in FIG. 12, the expected positive and negative emotional values are set when the character assumes the attribute. For example, an amusement park such as Disneyland is a fun place, so the characters' positive and negative emotions when they are at that place have a high value, such as +7. In this case, the value of the convenience store at the location is +1. If the value at the original location is 0, the difference is +1.

Here, even if "+3" is set in advance as the magnitude of the reference value for emotion generation, since the magnitude is less than +3, it is determined that no emotion occurs. The next step, ``drawing a lottery,'' does not include the action ``drawing a lottery'' in the mental state dictionary, so this is also a step in which no emotion occurs and there is no change in mental state, indicating a standard response. Proceed to S19.

Since there is no response corresponding to "draw a lottery" in the standard response dictionary, "So what happened?" is displayed on the display device 3 as the response sentence 6 when the response is not applicable anywhere.

In response to this, Taro replies, ``I won the pudding.'' In step S16 showing semantic analysis, first, as shown in FIG. 13, an action 90 of "winning pudding" is created. In action 90, the operative word 61 is "Ataru", the subject 62 is "Taro", and the object 66 is "pudding". In the meaning, it is described that the context 51 is set so that the property of the subject becomes the object. According to this meaning, when the situation change setting unit 11 converts the action 90 into the before and after state 51, as shown in FIG. 14, the property of "Taro" becomes "pudding" in the rear attribute value.

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

The emotional word response is performed based on the emotional word response dictionary shown in FIG. In the emotional word response dictionary, if the response word is positive, the response word is "It was good," and if it is negative, the response word is "It was disappointing." This time, since a positive emotion has occurred, a response sentence 6 saying "That's good" is displayed on the display device 3.

The purpose of people having a conversation is because they have something to say. What I want to say is that something like this happened yesterday. In a nutshell, it is the feeling of having experienced something fun or something sad. When emotions are broken down, the most basic emotions are positive and negative emotions. In other words, it is something that is good or bad for that person. Furthermore, the generation of emotions can be said to be the smallest unit of what one wants to say, and can also be said to be the smallest unit of the "story" of the present invention. Therefore, if you cannot understand the emotion from the other person's utterances, you can judge that the "story" is not complete yet, and you can respond with "So what happened?" to encourage them to continue. . In this way, natural conversation becomes possible.

To put it simply, the meaning understanding method of the present invention is to process the content of a sentence so as to reflect the positive and negative emotions of the subject. For example, what can be gleaned from a sentence such as ``Taro drew a lottery at a convenience store and won a pudding'' is derived from data such as the concept tree and the attributes of word objects, such as ``pudding is sweet and round.'' The convenience store clerk would have taken the pudding out of the refrigerator.'' There are an infinite number of possible meanings for this sentence.

This is processed in a way that reflects 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 this end, I defined the meaning of the verb ``ataru'' to mean a change in the characters' possessions. This is because people experience positive and negative emotions: they become happy when they gain something of value, and they become sad when they lose something of value.

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

Next, behavior prediction, which is the second definition of understanding the meaning of natural language, will be explained using the second embodiment.

FIG. 16 is a block diagram showing a second embodiment in which the present invention is incorporated into a robot 100, and FIG. 17 is an external appearance thereof. The robot 100 is a humanoid robot that is approximately the same size as a human and has approximately the same functions as a human. That is, the camera 101 provided in the eye 107 photographs the outside world, the microphone 102 listens to sound, and the speaker 103 provided in the mouth 108 emits sound. Further, the limbs 109 are driven by a motor 104, and are also provided with a contact sensor 105, etc., which are controlled by a control device 106.

In the control device 106, a semantic understanding program similar to that of the first embodiment is executed. However, in this embodiment, the input is audio input from the microphone 102, and the output is audio output from the speaker 103. Therefore, the input sentences are converted into text, and the response sentences are also converted into speech and output, allowing the user to converse in a manner closer to that of humans. Furthermore, the camera 101 and the sensor 105 can also acquire external conditions other than audio. What is further different is the content of the semantic 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. A psychological program is a program that can understand various human psychology, and it is also possible to understand social emotions. Social emotions are emotions that occur when you and the other person are involved, and include ``gratitude,'' ``anger,'' and ``shame.''

As an example, we will explain how to understand the emotion of "gratitude." Here, it is assumed that the name of the robot 100 is "Robota" and that Taro has given a present to Robota. For example, let's say that Taro hands over a present, saying, ``Robota, I'll give you a present.''

FIG. 18 shows the data structure of the event 110 of the semantic understanding program of this embodiment. The event 110 has a self 111 and an opponent 112 as characters. The self is the person on whom the psychological program is executed, and the other is the person with whom one is having a conversation.

When the input sentence ``Robota, I'll give you a present'' is obtained, the predicate ``give'' is searched in the concept tree, determined to be a subordinate concept of ``yield'', and the action of event 110 is generated. Since the action word 120 is "Yuzuru", the data structure of "Yuzuru" is obtained from the action word dictionary. For the obtained data structure, the operator 120 is set to "Yuzuru", the subject 121 is set to "Taro", the "Yuzuru party" is set to "Robota", and the "Yuzuru item" is set to "Present". The action is now complete.

Then, according to the meaning 124 of the action, the state before and after the event 110 is set. The own before and after state is a data structure that shows one's own state before and after an action, and in this case, the front attribute name 130 and the rear attribute name 132 are "possessions", and the front attribute value 131 is empty, The rear attribute value 133 is "present".

If you search for "present" in the mental state dictionary, the value will be +8, for example, and Robota's plus/minus emotion will be +8, which is greater than the standard value, so Robota will have a positive emotion, such as "happy". will occur.

What occurred here were ``my'' feelings. Next, when checking the action, the subject 121 is "Taro." In other words, the person who took the action was the person with whom he was currently conversing. If you feel positive emotions as a result of the other person's actions, it means that the other person is the cause of your positive emotions.

Here, we will create what is called a social emotion dictionary. This is a registration of patterns in which social emotions, such as "gratitude" and "anger," occur depending on the person and the other person. In the case of "gratitude," if a positive emotion occurs in oneself, and the cause of the positive emotion is the other person, the emotion of "gratitude" will occur toward the other person. In this case, since Taro is the one who gave him the present, Robota will feel a feeling of "gratitude" towards Taro. Also, by registering ``Thank you'' as the word to say to the other person when the emotion of ``gratitude'' occurs, Robota can receive a present and say ``Thank you'' to Taro.

In this way, the meaning understanding program of this embodiment can understand the emotion of "gratitude." You can also say "thank you" when you receive a gift. This is the same behavior that humans take. In other words, it can be said that human behavior can be predicted, and the second definition of semantic understanding defined at the beginning has been achieved.

Now let's consider a case where the present is not what Robota wants. For example, let's say that Taro thought that Robota would probably like dry cell batteries, so he gave him a dry cell battery. Robota doesn't run on batteries, so he won't be happy if he receives one. Even in such a situation, if you were a human, you would understand the other person's goodwill and say "thank you."

Therefore, the next step is to help students understand such complex emotions. To this end, as shown in FIG. 19, the user has a self-emotion management section 150 that manages his or her own positive and negative emotions. The own emotion management unit 150 has emotions 151 that actually occurred in oneself and emotions 152 that the other party intended. In this case, my actual emotions are the positive and negative emotions of Robota who received the battery, so they are negative. The other party's intention is the intention of Taro (the other party) who thought that Robota (himself) would be happy (positive emotion) if he received the battery, so the emotion 152 that the other party (Taro) intended is positive.

The meaning understanding program sets the self-emotion management section 150 after setting the self-previous and post-self states. Then, it is registered in the social emotion dictionary that if the emotion 152 intended by the other person is positive, the emotion of "gratitude" will occur. In this way, you will be able to understand more complex emotions, such as understanding the other person's feelings and prioritizing their feelings over your own when saying "thank you."

Note that whether to give priority to one's own feelings or to the other party's intentions may be determined by setting parameters in a psychological program. Depending on this parameter, if you tend to prioritize others, you will be considered "kind," and if you tend to prioritize yourself, you will be considered "selfish." In this way, the robot's "personality" can also be set.

This kind of function can be said to be the function of guessing the other person's mind or imagining the events that the other person is thinking about. When you imagine an event that the other person thinks about, it can be said that you are simulating the event that the other person thinks about. In order to realize this function, the other person's mind is constructed virtually. This is shown in FIG. In the virtual opponent's mind 160, an event 161 that the other party thinks is calculated. An event 161 considered by the other party has a context 162 of the other party. The person that the other person thinks of is the person that the other person sees, so it becomes oneself. Therefore, the attribute value set in the opponent's front and back state 162 is the attribute value that the opponent considers to be his or her own.

To explain using the previous example, you are Robota, your opponent is Taro, and Robota (yourself) imagines the virtual mind 160 of Taro (your opponent). Events 161 that Taro thinks about are set in the virtual mind 160 of Taro. When Taro presents Robota with a battery, Robota's posterior attribute value increases by the value of the battery that Robota received. The value of the battery that Robota received is not the value that Robota actually feels, but the value of Robota's battery that Taro imagines. Taro imagines the value of batteries that Robota feels from Taro's experience and knowledge. For example, ``Since he is a robot, he would be happy if he received a battery.'' This can be achieved by creating a mental state dictionary for each subject based on Taro's knowledge, and recording things like ``The value of a dry cell battery to the robot (subject) is +5.'' Then, the emotion that the other person (Taro) assumes for himself (Robota) will be +5 when he receives the battery. This means that as a result of the other person's (Taro) actions, one's (Robota)'s positive and negative emotions become positive, and this can be said to be the other person's intention. If the emotions intended by the other person are positive, the other person acted with the intention of making the other person happy, and the emotion of ``gratitude'' toward the other person occurs. On the other hand, if the feelings intended by the other person are negative, the other person acted with the intention of making you uncomfortable, and the emotion of anger will occur.

Guessing the other person's mind is not limited to when the other person acts, but also includes guessing one's own evaluation and value from the other person's perspective. In other words, you can guess what the other person thinks about you. Once we can do this, we can generate things like the gaze of others, and we can also generate the feeling of ``shame.'' In any group or society, there are minimum standards that must be met if you belong to that society. For example, ``Junior high school students should know this level of kanji,'' or ``If you're on the track and field team, you should be able to run 50 meters in less than 6 seconds.'' Suppose that a certain society sets a minimum standard that must be met, such as, ``If so-and-so should at least be able to do this.'' This standard is derived from experience. Then, when you, as a member of that society, do not meet the standards, and you simulate how others would feel if you did not meet the standards, the feeling of ``shame'' will occur. If you do this, you will be able to estimate your worth in the eyes of others, such as ``I can't do this,'' and you will feel embarrassed.

People form a cohesive society by having this emotion of shame. This is because, in order to belong to that society, there is a psychological need to at least meet the standards, and this becomes the driving force for behavior to meet those standards. By possessing this emotion of "shame," we can create artificial intelligence with a mind that can integrate into human society.

Next, let's talk about "anger." For example, let's say someone says to Robota, "Ponkotsu." Robota uses speech recognition to convert the sounds it hears into character strings and then analyzes the meaning. The meaning of the word ``crank'' is that the performance of the target is poor. Since the target is Robo-ta, Robo-ta judges that he has been spoken ill of and feels negative emotions. When Robota took a look, image recognition confirmed that Jiro had spoken.

If this situation is set in the personal emotion management section 150, the personal emotion becomes negative. Further, since the other party said it, it is determined that the other party said the statement with intention, and the own emotion 152 that the other party intended becomes negative. In the social emotion dictionary, if the emotion 152 intended by the other person is negative, it is registered that the emotion of "anger" will occur toward the other person who has performed the action. As a result, in this case, the emotion of ``anger'' is generated, and a response corresponding to the emotion of anger is generated, such as ``What?!''.

Now, suppose that the person who has spoken ill of you says, "I'm sorry." When we search for "I'm sorry" in the concept tree, we find that it is below the "apology" concept, so we create an "apology" action. When the action word is set as "apology" and the event is set based on the meaning of "apology", the feelings 152 of one's self intended by the other party become ±0. Furthermore, the person who apologizes is in a lower position relative to the person to whom he or she apologizes. In other words, ``apology'' has the function of showing that there is no malicious intent by lowering one's own position. Here, when the meaning understanding program reanalyzes, the emotion 152 intended by the other person becomes 0, so anger does not occur, that is, the anger subsides. In this way, you will be able to understand the psychology of ``apology.''

Next, we will explain behavior prediction based on social emotions. People try to maintain a psychological balance with their partner. For example, in the case of the emotion ``anger'', depending on the other person, one's own negative emotion may be felt. Achieving balance means that if you have negative emotions due to the other person, by making the other person have negative emotions as well, you will both be negative and a balance will be achieved. Therefore, when the emotion of "anger" occurs, the psychological program searches for a statement that will cause the other party to generate negative emotions as the next response.

This can be achieved by performing the reverse process. In other words, in event 110, assume that the other person's before and after situation is not your own, look for a subsequent attribute value that will cause the other party to feel negative, and then search for an action that will set the attribute value. be. In other words, this can be said to be about thinking from the other person's perspective.

As an example, let's say someone calls you an idiot, and you feel negative emotions. It is conceivable to search for behaviors that will cause the other person to experience the same negative feelings as you, and for example, to respond by calling the other person "stupid." This can be said to be an understanding of the meaning of ``retaliation'' and ``retaliation.'' This is also an example of semantic understanding as action prediction.

Even when feelings of gratitude arise, we try to maintain a balance and act in a way that makes the other person feel positive emotions. For example, when you receive a birthday present, you feel positive, so you take action to make the other person feel positive, such as giving them a present on their birthday. This can be said to be an act that understands the meaning of "giving something back." In this way, by taking actions to balance the positive and negative emotions received from the other person, the psychological program becomes closer to the human heart.

Next, we will explain the most important social emotions, ``good'' and ``bad.'' ``Goodness'' can be defined as ``one should act in such a way that the other person's positive and negative emotions become positive.'' In other words, doing something that makes others happy is ``good.'' When multiple actions are possible, the meaning of "good" can be understood by configuring a psychological program to select the action that will give the other person positive feelings. This is because being able to take the same actions as humans (behavior prediction) can be said to understand meaning. Similarly, "evil" can be defined as choosing an action that causes negative feelings in the other person, when multiple actions are possible.

Let's now consider the judgment of right and wrong when the other person does not exist. For example, this is the case with the act of throwing garbage on the street. This is clearly "bad" behavior, but there is no clear opponent. In such cases, the relevant third party is set as the other party. The relevant third party is the entity affected by the act, such as the person using the road in the case of ``throwing garbage on the road''.

Here, a method for extracting related third parties will be explained. Words are managed not only in the concept tree but also in the Has-a tree. The Has-a tree is a data structure that manages words in terms of whole and part relationships.

FIG. 21 shows the data structure of the Has-a tree 141 of the town 140. Assume that a town 140 is found when searching the Has-a tree that includes the road 142. It is assumed that the town 140 has residents 143, roads 142, schools 144, hospitals 145, etc. When a subject included in the town 140 is searched, a resident 143 is obtained, and this resident 143 is set as a third party related to the road. Next, perform the action of ``throwing away garbage''. Then, with the town 140 as the object appearing in the event, the post-attribute value of the post-attribute "environment" becomes "defacement" in the post-state of the town 140. Next, the forward and backward states of the residents 143 of the town 140 in this case are set. Since the environment of the town 140 is "contaminated" and in a negative state, it is calculated that the emotions of the residents 143 included in the town 140 are also negative emotions. From the above, the act of ``throwing garbage on the road'' can be said to be a ``bad'' act because the residents 143 of the town 140 related to the road experience negative feelings.

Similarly, the act of picking up trash on the road can be judged as a "good" act because it makes the town 140 beautiful and the residents 143 of the town 140 feel positive emotions. In this way, it becomes possible to judge whether an action is right or wrong by calculating the positive and negative feelings of a third party regarding the action when the action is performed. In this way, it can be said that the robot 100 with the psychological program can coexist with humans in human society because the psychological program can judge the good and bad of human society.

Next, I will explain the understanding of the meaning of the word ``effort'', which is a psychological situation. For example, suppose Taro is practicing running. FIG. 22 shows two before and after states derived from the action of "running". Anteroposterior state A indicates that running causes fatigue in the physical state. Front-back state B shows that running increases the speed of the legs. The negative emotion of fatigue arises from the front/back state A, and the positive emotion of becoming faster at the front/back state B. The psychological state of ``effort'' can be defined as a psychological state in which a person intentionally chooses to take action even if negative emotions occur in order to achieve positive emotions.

Furthermore, by repeating this action, it can be calculated that your feet will become faster. Repeating an action means taking time. In other words, this is where the concept of "time" comes into being. Also, assume that the condition is set to run 50 meters in 7 seconds or less. If we set a behavioral pattern that repeats the action of running to satisfy this condition, this means that we have defined the psychological state of "making efforts toward a goal," and the conditions set in this behavioral pattern are This is the meaning of the word "goal".

In this way, we can understand psychological states such as ``effort'' and ``goals.'' In addition, "support" can be defined as words and actions that support a person's behavior. An example of "support" is the words and actions of "do your best." For example, if you recognize someone who is working towards a goal, you can tell them to do their best.

Next, let's define the word "regret." For example, let's say you set a goal of passing a university entrance exam and put effort into studying toward that goal. However, as a result, he did not get into university. In this case, it can be assumed that the student did not study enough. In other words, it can be calculated that the number (amount) of studying was insufficient during the period leading up to university entrance exams. If you increase the amount of studying, you can generate a hypothetical event in which you passed the exam and compare it with the event that actually occurred. When comparing the two, the difference can be calculated as the number of studying behaviors. In other words, by changing past actions, negative emotions that actually occurred can 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 feeling of ``regret,'' such as the feeling that if we had studied harder, we would have been able to get into university.

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

In each embodiment of the present invention, semantic understanding is executed by a control device integrated with the chatbot system 1 or the robot 100, but the present invention is not limited to this. The configuration may be such that the semantic understanding is performed on a server or cloud, and the results are returned to the chatbot system 1 or the robot 100.

Furthermore, 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 be applied to automatic document summarization or 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 not by the above description but by the claims, and is intended to include meanings equivalent to the claims and all changes within the scope.

1 Chatbot system, 9 Meaning understanding program, 10 Scene setting section, 11 Situation change setting section, 12 Plus/minus emotion setting section 13 Semantic analysis section, 100 Robot

Claims (3)

Natural language processing comprising: an input device into which written or spoken sentences are input; and a control device having a CPU and main memory, and which breaks down the input sentences from the input device into words and analyzes their meanings. A system,
In the main memory, there is a word object that has a word corresponding to a subject who understands natural language, has a mental state, and has positive emotions that are favorable emotions for oneself, and negative emotions that are unfavorable emotions for oneself. is generated,
The control device generates a word object of the subject from the words of the input sentence, determines the subject's positive emotion or negative emotion, thereby determining the subject's mental state and predicting the subject's behavior. A natural language processing system that performs semantic analysis by Natural language processing comprising: an input device into which written or spoken sentences are input; and a control device having a CPU and main memory, and which breaks down the input sentences from the input device into words and analyzes their meanings. A natural language processing method used in a system,
In the main memory, there is a word object that has a word corresponding to a subject who understands natural language, has a mental state, and has positive emotions that are favorable emotions for oneself, and negative emotions that are unfavorable emotions for oneself. is generated,
The control device generates a word object of the subject from the words of the input sentence, determines the subject's positive emotion or negative emotion, thereby determining the subject's mental state and predicting the subject's behavior. A natural language processing method characterized by comprising a step of performing semantic analysis by performing semantic analysis. Natural language processing comprising: an input device into which written or spoken sentences are input; and a control device having a CPU and main memory, and which breaks down the input sentences from the input device into words and analyzes their meanings. A natural language processing program used in the system,
In the main memory, there is a word object that has a word corresponding to a subject who understands natural language, has a mental state, and has positive emotions that are favorable emotions for oneself, and negative emotions that are unfavorable emotions for oneself. is generated,
computer,
The control device generates a word object of the subject from the words of the input sentence, determines the subject's positive emotion or negative emotion, thereby determining the subject's mental state and predicting the subject's behavior. A natural language processing program that functions to perform semantic analysis by

Images