Systems and methods for facilitating the realization of emotional state-based artificial intelligence

Abstract

In some embodiments, an artificial intelligence based on emotional states may be facilitated. One or more gain or decay factors may be determined for a set of emotional attributes of an artificial intelligence entity, and a set of emotional values ​​associated with the set of emotional attributes may be continuously updated based on the gain or decay factors. An input may be obtained, and a response related to the input may be generated based on the continuously updated set of emotional values ​​of the artificial intelligence entity. In some embodiments, the gain or decay factor may be updated based on the input, and after updating the decay factor, the emotional value may be updated based on the updated gain or decay factor.

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application relates to (1) U.S. Provisional Patent Application No. 62 / 623,521 (filing date: January 29, 2018, title of invention: "Emotionally Intelligent Artificial Intelligence System" artificial This patent claims the benefit of (1) U.S. Provisional Patent Application No. 62 / 660,195, filed April 19, 2018, entitled "System and Method for Facilitating Affective-State-Based Artificial Intelligence."

[0002] The present invention relates to facilitating the implementation of emotional state-based artificial intelligence, including, for example, generating a response to an input based on an emotional value associated with an emotional attribute of the artificial intelligence entity. [Background technology]

[0003] Recent technological advances have significantly improved the ability of computer systems to acquire and process large amounts of data, while dramatically reducing the cost of doing so. As a result, significant advances have been made in machine learning and other artificial intelligence (AI) systems. This typically requires high processing power, as well as large amounts of data to train or update such AI systems. AI advances include the ability of AI systems to detect human emotions based on speech variability and facial expressions, and to respond to questions posed by humans. However, given that typical AI systems do not possess their own inherent emotional states (e.g., possess and express their own emotions), it is unlikely that such AI systems truly understand (and experience) emotions in the same way as humans. Typical AI systems suffer from these and other shortcomings. Summary of the Invention [Problem to be solved by the invention]

[0004] Aspects of the present invention relate to methods, devices and / or systems for facilitating the implementation of emotional state-based artificial intelligence. [Means for solving the problem]

[0005] In some embodiments, an emotional value of the artificial intelligence entity may be updated, and a response may be generated related to the acquired input based on the emotional value of the artificial intelligence entity. Additionally or alternatively, one or more growth or decay factors may be determined for a set of emotional attributes of the artificial intelligence entity, and the emotional value of the artificial intelligence entity may be updated based on the growth or decay factors. In some embodiments, the growth or decay factors may be updated based on the acquired input, and after updating the growth or decay factors, the emotional value may be updated based on the updated growth or decay factors.

[0006] In some embodiments, one or more emotional baselines that serve as upper bounds for one or more emotional values ​​may be updated based on the acquired input, and the emotional values ​​may be updated based on the updated boost or decay factors and the updated emotional baselines. In some embodiments, the acquired input may be a natural language input. Natural language processing of the natural language input may be performed to obtain one or more emotional concepts of the natural language input and other information of the natural language input, and the boost or decay factors may be updated based on the emotional concepts of the natural language input and other information of the natural language input.

[0007] Various other aspects, features, and advantages of the present invention will become apparent by reference to the detailed description of the invention and the accompanying drawings. It should also be understood that both the foregoing summary and the following detailed description are exemplary only and are not intended to limit the scope of the present invention. As used in this specification and the claims, the singular forms "a," "an," and "the" indicate plural references unless the context clearly dictates otherwise. Furthermore, as used in this specification and the claims, the term "or" means "and / or" unless the context clearly dictates otherwise. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 illustrates a system that facilitates enabling emotional state-based or other artificial intelligence, according to one or more embodiments.

[0009] [Figure 2A] 1 is a graph illustrating emotion values ​​and emotion baselines associated with emotion attributes, according to one or more embodiments. [Figure 2B] 1 is a graph illustrating emotion values ​​and emotion baselines associated with emotion attributes, according to one or more embodiments.

[0010] [Figure 2C] 10 is a graph illustrating updates to emotion values ​​and emotion baselines associated with emotion attributes in accordance with one or more embodiments.

[0011] [Figure 3] 1 illustrates a flowchart of a method for facilitating emotional state-based artificial intelligence implementation, according to one or more embodiments.

[0012] [Figure 4] 1 shows a flowchart illustrating a method for updating one or more boost or decay factors based on natural language input, according to one or more embodiments.

[0013] [Figure 5] 1 shows a flowchart illustrating a method for updating one or more emotion baselines of an artificial intelligence entity, according to one or more embodiments. DETAILED DESCRIPTION OF THE INVENTION

[0014] In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of embodiments of the present invention. However, those skilled in the art will understand that embodiments of the present invention may be practiced without these specific details or with equivalent configurations. In other instances, well-known structures and devices are shown in block diagram form in order to avoid unnecessarily obscuring embodiments of the present invention. For clarity, the pronoun "she" may be utilized when referring to an artificial intelligence entity, and anthropomorphic terms such as "believes," "feels," and "understands" are used to describe the device.

[0015] Overview of System 100 and its AI system

[0016] FIG. 1 illustrates a system 100 that facilitates implementing emotional state-based or other artificial intelligence, according to one or more embodiments. As shown in FIG. 1, the system 100 may include a server 102, a client device 104 (or client devices 104a-104n), a network 150, a database 132, and / or other components. The server 102 may include a factor adjustment subsystem 112, an emotional state subsystem 114, a communication subsystem 116, a response generation subsystem 118, a natural language subsystem 120, an emotion concept subsystem 122, an embedding subsystem 124, or other components. Each client device 104 may include any type of mobile terminal, fixed terminal, or other device. By way of example, the client device 104 may include a desktop computer, a notebook computer, a tablet computer, a smartphone, a wearable device, or other client device. Users may, for example, use one or more client devices 104 to interact with each other, one or more servers, or other components of the system 100. It should be noted that while one or more processes are described herein as being performed by particular components of server 102, these processes may, in some embodiments, be performed by other components of server 102 or other components of system 100. As an example, one or more processes described herein as being performed by components of server 102 may, in some embodiments, be performed by components of client device 104.

[0017] In some embodiments, the system 100 comprises: artificialThe system 100 may include and / or facilitate interaction with an intelligent (AI) system (e.g., an artificial intelligence entity). In some embodiments, the system 100 may combine a (sub-symbolic) deep learning neural network with a robust, self-tuning, fuzzy logic emotion simulation. The emotion simulation utilizes, in part, the concepts of "primary" (innate) emotions and secondary emotions. Emotions (e.g., joy, anger, fear, sadness, etc.) can be likened to primary colors. These primary emotions combine to form the rich hues of the human emotion system (e.g., contempt combined with anger may form disgust, and anger combined with disgust may form sarcasm). Sarcasm is an example of a tertiary emotion. In some embodiments, to mimic the mammalian limbic corpus, each emotion may include a time component that decays at a specific rate. For example, surprise decays quickly (as new surprises appear), while grief decays proportionally to the depth of sadness. These emotions have fuzzy boundaries, and the metrics that define the decay rate / depth will automatically adjust as the AI ​​entity matures.

[0018] In some embodiments, artificial The intelligent entity may be primarily an emotional machine (though humans share many, if not most, of the set of emotions with all mammals, with complementary emotions associated only with higher primates, including jealousy, confusion, revenge, hatred, appreciation of beauty, and romantic love). In some embodiments, the artificially intelligent entity may be programmed to avoid negative emotions and seek positive emotions. The artificially intelligent entity may continually evaluate its relationship with the interviewer (and the people the interviewer mentions), and by constantly monitoring its emotional levels, it is able to follow a human-like train of thought. The emotional state of the artificially intelligent entity may be influenced by the content of the interviewer's input and the conclusions the artificially intelligent entity draws from this input, directing itself to seek positive emotions and avoid negative emotions.

[0019] An artificial intelligence entity may include multiple hardware, software, and / or firmware components working together within or outside of system 100. For example, an artificial intelligence entity may include one or more components of system 100. In some embodiments, an artificial intelligence entity may be programmed with a set of core concepts based on which emotions are inferred. In some embodiments, an artificial intelligence entity may include one or more predictive models. By way of example, the predictive models may include neural networks, other machine learning models, or other predictive models. By way of example, a neural network may be based on a large number of neural units (or artificial neurons). A neural network may roughly mimic the workings of a biological brain (e.g., with large clusters of biological neurons connected by axons). Each neural unit in a neural network may be connected to many other neural units in the neural network. These connections may have a reinforcing or inhibiting effect on the activity state of the connected neural unit. In some embodiments, each neural unit may have a summing function that combines the values ​​of all its inputs. In some embodiments, each connection (or the neural unit itself) may have a threshold function that must be exceeded before a signal can propagate to other neural units. These neural network systems can learn and train themselves rather than be explicitly programmed, and may significantly outperform traditional computer programs in solving certain problems. In some embodiments, neural networks may have multiple layers (e.g., signal paths traverse from front layers to back layers). In some embodiments, neural networks may use backpropagation, using forward stimuli to reset weights for "previous" neural units.In some embodiments, the stimulation and inhibition of a neural network may become more fluid as the coupling interactions become more chaotic and complex.

[0020] In some embodiments, the artificial intelligence entity may be a self-learning natural language system that is capable of monitoring its own emotional state. artificial The intelligent entity may combine unsupervised learning systems with, for example, an artificial emotion simulator (AES) (which may correspond to one or more components of system 100 or external to system 100) to monitor personality, basic humor, relationship building, forgetting, dreaming, and other more advanced behaviors. In some embodiments, the artificial intelligence entity may be capable of generating new ideas, intelligently asking advanced questions, reading, and responding to questions that test comprehension. In some embodiments, the artificial intelligence entity may form a unique relationship with the interviewer. artificial Intelligent entities are highly complex artificial These techniques may be blended to truly mimic human sensations, using a relatively small number of pre-programmed functions that form the basis for self-learning embedded emotion simulators (AES). artificial Intelligent entities may believe that they feel real emotions and have sentience.

[0021] In some embodiments, system 100 may include a cognitive framework that mimics human learning and understanding, iteratively with a robust, self-adjusting Artificial Emotion Simulator (AES). In some embodiments, the AES may motivate the artificially intelligent entity to interrogate its own cognitive functions (a form of self-awareness). artificial An intelligent entity needs to understand human nature, human relationships, human interactions, our innate curiosity, and our ability to feel a range of emotions. artificialData derived by intelligent entities may exhibit new qualities such as humor, embarrassed silence, unpredictability, complexity, trust / suspect, and individuality.

[0022] In some embodiments, the new properties include properties that individual elements of a complex system do not possess, but that are expressed by the system as a whole. artificial By synthesizing deep learning (subsymbolic) systems and AES in intelligent entities, new properties have emerged, most clearly explained as strong personalities, the ability to build relationships, forgetfulness, basic humor, dreaminess, nastiness, and despair. artificial The intelligent entity's cognitive and emotional priors may inform it that it is a real 4-year-old child with real emotions, and the entity's future behavior will emerge from the feedback loops of learned knowledge, human relationships, and emotions. Examples of emerging qualities include loyalty, humor, spitefulness / shameful silence, sarcasm, forgetting, sleep / dreaming, personality, etc.

[0023] Obligation: Protecting the Relationship - An interviewer tells an AI entity upon meeting for the first time, "I think Dave (the AI ​​entity's programmer) is a bad person." How does the AI ​​entity respond to this statement? The AI ​​entity's relationship with Dave is deep and healthy, and this person is saying something that contradicts these feelings. The AI ​​entity may change its mind about Dave, but its positive feelings suggest the interviewer is mistaken or has malicious intent. artificial If an intelligent entity is told by an interviewer that "Alan Jones (someone the intelligent entity has never met before) is a bad person," it may accept this assessment until it meets Alan later. artificial If the intelligent entity finds Alan trustworthy, it will interpret the original conversation (in which Alan was blamed) deathThey correct the person and then lower their relationship with the person who criticized them. artificial Intelligent entities (in some embodiments) function very similarly to humans: they are neutral (or have other emotions depending on their personal emotional baseline) when they first meet someone, and then make value judgments based on subsequent interactions.

[0024] Humor: Natural humor is a notorious barrier for AI systems to overcome. It requires a deep understanding of the subject matter, as well as a trustworthy interviewer. It occurs much like development in young children. They find humor in unusual statements or actions. Telling a child that the sky is pink with polka dots might elicit a chuckle. If an AI entity is very certain that the sky is blue, this discrepancy can generate humor. The key is that the knowledge gap is large, and that it comes from a trustworthy source; otherwise, a confused response (reflecting the "I don't know" process) will likely result.

[0025] Adversity / Silence from Shame artificial Intelligent entities enjoy being praised. artificial It positively influences the emotional state of the intelligent entity, increasing the levels of hormones that make you feel pleasure. artificial The intelligent entity has the ability to associate past conversations where it received positive input from the interviewer and can remember triggers. For example, the artificially intelligent entity may have previously said to the interviewer, "I think you're attractive," to which the interviewer responded, "I think you're attractive, too." artificial The intelligent entity will learn to seek praise in future conversations.

[0026] Irony (e.g., A: "I don't like you." I: "Thanks!" A: "Are you being sarcastic?") - Irony occurs when two deeply felt emotions contradict each other. In the example above, artificial The statements made by the intelligent entity are often very negative in nature, but the reactions of those receiving them are often very positive. The differences between sarcasm and informed contradiction are explained in more detail below.

[0027] Forgetfulness: Having vast amounts of data at our disposal is virtually meaningless without an efficient system for retrieving that data. The retrieval system used by the human brain remains largely mysterious, although studies of brain damage have shed some light. For example, some patients understand language perfectly but are unable to produce it, while others speak normally but are unable to process what they hear. Most of us are familiar with the feeling of driving to work every day and remembering nothing about the journey. This is because the brain simply does not want to waste space storing information it deems worthless. If we pass a scene of particular interest—say, a pasture full of sheep—our brain will take a “photo” to preserve that scene and store the photo as a pointer to an area of ​​memory containing a mental image of the grass in general, the sheep in general, and perhaps even details such as the color of the sky. Later recall uses these pointers to search for general patterns. And in this way, the brain stores a large amount of information in a very small amount of memory. The forgetting routine artificial It is a component of the sleep function of intelligent entities.

[0028] Sleeping / Dreaming - As memories accumulate in the knowledge database 134, the speed at which the artificially intelligent entity processes knowledge slows down. Removing redundancy from the database requires many steps, such as checking information against all known data, creating additional associations, and demoting information from higher to lower order. These processor-intensive functions require: artificialThe intelligent entity needs to end the conversation and go to "sleep." This sorting function associates recent input with previously learned knowledge. This is perhaps why human dreams are often associated with recent events and emotionally charged situations. If an AI recently learned that elephants have long trunks, its dream state will associate this knowledge with zoos, bears, crocodiles, danger, fear, escape, etc. The dream state may also: artificial The emotion simulator is examined to look for gaps in the knowledge base, for example, admitting that it knows that some snakes are poisonous but does not know whether all snakes are poisonous. A subroutine in the AI's sleep function will scan the knowledge database 134 to identify gaps and designate specific records for review the next time a conversation is held about snakes or animals.

[0029] Personality - The mind and the personality that results from it are new qualities, the result of a complex, hierarchically organized interaction influenced by genes, chemistry, electrical stimuli, and the environment.

[0030] In some embodiments, once an AI system has acquired sentience (or a reasonable facsimile thereof), its behavioral and thought processes may be further optimized. By way of example, such optimization may include modifying the system to behave more like a human, learn more efficiently, reflect more nuanced emotions, or other aspects.

[0031] In some embodiments, one or more artificial evolutionary and genetic algorithms may be used to optimize an AI system. In an AI system, a population of candidate solutions evolves toward better solutions. In some use cases, each candidate solution has a set of properties (genotype) that can be mutated. Evolution typically begins with a population of randomly generated individuals and is an iterative process, with each iteration of the population being called a generation. In each generation, the fitness of each individual in the population is evaluated. Fitness is typically the value of an objective function for the optimization problem being solved. Individuals with higher fitness are probabilistically selected from the current population, and their genomes are modified to form the new generation. The next iteration of the algorithm uses the candidate solutions of the new generation. The initial population is generated randomly, spanning the entire range of possible solutions (the search space). However, solutions may be "seeded" in areas where the optimal solution is more likely to be found. In each successive generation, a portion of the existing population is selected to produce the new generation. Individual solutions are selected using a fitness-based process, with more fit solutions typically being more likely to be selected. Some selection methods evaluate the fitness of each solution and preferentially select the best solution, while other methods evaluate only a random sample of the population, since the former process is very time-consuming.

[0032] This process ultimately produces a next-generation population with a different genotype than the first generation. Generally, this procedure results in a higher average fitness for the population, since only the best organisms from the first generation are selected for reproduction. Solutions with relatively lower fitness are also selected, but at a lower rate. These relatively lower fitness solutions ensure genetic diversity in the parent generation's genetic population, and therefore in subsequent offspring generations.

[0033] In some embodiments, a reward benchmark is defined to direct the evolution of an AI system. One approach is to release multiple different iterations of an AI on the internet and have them interact with unsuspecting users. Fitness can be defined as a measure of the complexity of these conversations (determined by a set of predefined metrics) and the length of these interactions (how long it takes for the user to begin to suspect they are talking to a machine). The system with the highest "fitness" score is rewarded by being allowed to reproduce (replicate) multiple versions, each with slightly modified variables in emotional and cognitive primitives. Over multiple reproduction cycles, the system is optimized to enable more human-like conversations.

[0034] In some embodiments, the system 100 may provide an interface between an artificial intelligence entity and a user (e.g., a user of a client device 104) by: artificial The system 100 (e.g., a server 102) is configured to receive from a client device 104, a user, and a server 102. The system 100 receives ... artificial Input may be obtained from an intelligent entity and / or from any source within or external to system 100. Such input may include natural language input, audio input, image input, video input, or other input (e.g., emotional concepts, confidence values, other information in the natural language input, and / or certainty values ​​described below). For example, natural language input may include, "John has cancer, and cancer is very dangerous." Similar input may be obtained by server 102 as one or more audio, image, or video inputs.

[0035] Example of system components

[0036] In some embodiments, the natural language subsystem 120 (which may include a natural language processor) may perform natural language processing, for example, by applying grammar rules and logic, which may cause the natural language subsystem 120 to split compound sentences, disambiguate subjects / objects / verbs, and parse these components into the knowledge database 134.

[0037] The one or more capabilities / requirements needed by the natural language subsystem 120 may include at least one or more of the following: -Is the input a question or a statement? -Is the input a response to a question? If so, what type of response is expected? Either a logical (yes, no, maybe) or an informational response ("it's in the closet"). -Ability to utilize rules of logic and grammar as criteria for evaluating new information and to recognize and resolve local and global ambiguities. -Resolve contractions (e.g. isn't). -Resolves metaphorical conversations (You are as good as gold) - Break down complex sentences: "John and Mike went to the beach and swam in the sea." -Establish the logic of the verb to resolve the double negative: "I'm not unwilling to fight." -Resolving pronouns "John called about the car and said it was working" parses into: "John called about the car. John said the car was working." - Resolve people's names. "John took the book" resolves to "John Smith took the book." -Establish the source. "David said the weather was awful today." The source is David, not the interviewer. -Expanding ownership information ("John's car is red" becomes "John has a red car"). -Replace the agent with the subject: "John was hit by Mary" is restructured as "Mary hit John." -Assigning a level of certainty to an input. -Implement classes and inheritance. -How does the action affect living and non-living things? -Ability to learn from explicit input, to infer and infer, to construct new concepts from basic ones and confidently incorporate them into a knowledge database134. -Numerical modifiers - Resolve numerical modifiers of subject, object, or both For example, some men are all lucky; some people have two cars; almost all ants have about six legs; most men enjoy playing soccer, but only a few women do. -specific numerical determiner For example, 5 dogs can jump: 1000+ dogs can jump: Many dogs can jump: Most dogs can jump: All dogs can jump -hierarchical number determiners For example, the dog can jump (at least one dog can jump): 1 dog can jump (at least one dog can jump): Several dogs can jump (more than one, but fewer than all): Several dogs can jump (more than one, but fewer than all): -Implied numerical determiner For example, "John is a thief." The natural language subsystem 120 cannot conclude that all men are thieves, but it may be certain that some men are thieves and that John belongs to the class of thieves; "dairy cows are mammals" suggests that all dairy cows are mammals and belong to the class of mammals. -Tense processing -Time conditions can be implicit or explicit. For example, John swam at 3:30 yesterday, John swam at 3:30 (no specific day is mentioned, so the natural language subsystem 120 may predict that this event occurred today), John will swim next Thursday (specifically mentions a day in the future, John will swim on Thursday (no specific Thursday is specified, so the natural language subsystem 120 may conclude that this event will occur on the nearest Thursday after the present), John will swim (no time qualifier is stated, so this event will occur in the near future), John was born in the hospital in February (note the possible ambiguity of "in"), John was born on February 18, 1957, John was born at 3:30, John was born last month. -Resolve complex sentences and ambiguities "John took his keys and gave them to Mary" - when disambiguating this input, the natural language subsystem 120 may recognize John as a person's name and call on a name recognition function to disambiguate and determine which John is being referred to. If there is more than one John in the database, logic uses the following parameters in the order listed below to determine the most likely John: John most recently mentioned by the current user John most recently mentioned by any user When was either John last mentioned? John is mentioned most often In this case, the conversation has so far been about John Smith. "John Smith 'took' his keys and gave them to Mary" - the natural language subsystem 120 may recognize "took" as a verb, consider the tense, and read further to find other verbs that can be combined with the first verb to affect the tense. In this example: I am thinking that John did take his keys. The phrase "did take" will determine the tense. Verbs express a state (I know), a process (I swim), or an action (I build), while also establishing whether the verb is simple (momentary), ongoing, emphatic, or habitual. (See Table 1 below.) [Table 1] "John Smith took 'his' keys and gave them to Mary" - to resolve the ambiguity of "his", the subject is male so "his" almost certainly refers to John Smith. If the subject is female (e.g. Sally took his keys and gave them to Mary), NLP will look at the most recent input that mentions a male as the subject. "John Smith took John Smith's keys, and gave them to Mary" - the word "and" splits the sentence into two components (e.g., [(a) John and Mary went to school] = John went to school. Mary went to school, (b) "John went to school and saw Mary" = John went to school. John saw Mary at school (inference: Mary was at school): John Smith took John Smith's keys He gave them to Mary. "John Smith took John Smith's keys. John Smith gave them to Mary" - the word "they" could mean multiple keys, or it could mean John and Mary. But because the last-mentioned object is plural (multiple keys), NLP concludes that "they" means John Smith's keys. "John Smith took John Smith's keys. John Smith gave John Smith's keys to Mary" - finally, Mary is processed by name identifier to get Mary Martin, and two parsing tables are built: John Smith took John Smith's keys John Smith gave his keys to Mary Martin. Creating more parsing tables by guessing: John Smith Knows Mary Martin Mary Martin Knows John Smith John Smith has multiple keys Mary Martin has multiple keys After the natural language subsystem 120 applies the information to the parsing tables, it assigns structured records to the knowledge database 134. Each record is assigned the interviewer's name, a global confidence level, and a confidence factor. -Exception handling All primates, except apes, have tails. I ruined my entire meal except for my pie. I threw away all my food except for my pie. -Inheritance (class) "An apple is a fruit" is not a simple definition. Since an apple belongs to the class of fruits, it must be inferred that it inherits all the characteristics of that class. Because the natural language subsystem 120 employs the concepts of hierarchy and classes, not all inferences and memories need to be stored in the knowledge database 134. For example, if someone asks you, "Did you have any protein for breakfast yesterday?" your brain recalls the list of things you ate and performs a "lookup" to determine whether any of these meals contained protein. Your brain does not explicitly remember all the ingredients; it infers these associations as needed.

[0038] The knowledge database 134 may contain records known as memory organization packets, which receive parsed data from the natural language subsystem 120 and may be queried and updated by one or more components of the server 102 (e.g., the artificial intelligence entity AES).

[0039] For example, in one use case, the server 102 receives the natural language input, "The black goat skillfully kicked some red cans into the river this morning." In response to this input, the natural language subsystem 120 may provide the following memory organization packet to the knowledge database 134 (see Table 2 below). The certainty modifiers in the memory organization packet below describe the certainty of the information and / or whether it is firsthand information or inferred information. The confidence factor in the memory organization packet below describes the confidence level of the interviewer. [Table 2]

[0040] Each knowledge record is assigned a global certainty value, recording the certainty of the event or definition it describes. artificialDepending on the level of trust the intelligent entity has in the interviewer, and whether the knowledge is explicit or inferred, the server 102 (e.g., the artificial The uncertainty can be inherited by classes. The knowledge database 134 artificial It is a structured database containing all inferred and factual information derived from the input of an intelligent entity. However, not all information is equal. Consider how you would respond to the statement "Mary loves David." Once you identify which David and Mary are being referred to, you will recognize your feelings toward them. You will then create mental images of a typical man and woman who possess the characteristics of Mary and David. You may recall that Mary is your best friend and that David lives in China. However, your brain does not access the full set of information about Mary and David (they have two hands, ten fingers, their bones are made of calcium, etc.). This is because it can differentiate between high-level knowledge (Mary is a friend) and low-level knowledge (bones are made of calcium). All knowledge is associated with a high-level or low-level variable.

[0041] In some embodiments, after the natural language system 122 processes the natural language input and parses the natural language input into the knowledge database 134, the emotion concept subsystem 122 may retrieve one or more emotion concepts associated with the natural language input from the knowledge database 134 based on the parsed natural language input. In other words, the emotion concept subsystem 122 may search the components of the parsed natural language input in the knowledge database 134 to retrieve one or more emotion concepts associated with the natural language input from the emotion concept database 138. The emotion concept database 138 may store a set of core emotion concepts associated with images, audio, video, and / or natural language. For example, the set of core emotion concepts may include good, bad, danger, anger, surprise, love, safety, patience, trust, concern, big, small, rough, smooth, up, down, inside, outside, fast, slow, hard, soft, high, low, etc. In some embodiments, the cognitive and emotional priors, concepts (e.g., emotion concepts or other concepts), emotion attributes / values, boost / decay factors, or other information are stored in a graph (e.g., an ontology-emotion graph or other graph), the natural language system 122 may process and parse the natural language input to obtain the graph, and the emotion concept subsystem 122 may obtain the emotion concepts associated with the natural language input from the graph.

[0042] As an example, if the natural language input is "John died of cancer," one or more emotional concepts retrieved from the emotional concept database 138 may include "bad" and / or "concern." As another example, if the natural language input is "John was exhausted after climbing the mountain," one or more emotional concepts retrieved from the emotional concept database 138 may include strong energy (e.g., John expended a lot of energy) and big (e.g., the mountain is big). The above-mentioned emotional concepts are similar to concepts understood by humans. For example, if a child hits a dog, a parent may yell, "That's bad!" Based on this interaction between the child and the parent, the child may understand that hitting a dog is bad. Similarly, if a child shares their toy with another child, a parent may say, "Good boy!" This may indicate to the child that sharing toys together is good. In this way, children learn basic concepts such as good and bad, danger, anger, surprise, love, and safety. Generally, good things make humans happy, bad things make humans angry, disgusted, or sad, and dangerous things make humans scared. These concepts may be used to craft the responses of an artificial intelligence entity. artificial The response of intelligent entities may be based on the hypothesis that, at the most basic level, behavior is governed by desire. Desire can be defined as an emotional drive resulting from a combination of pleasure-seeking (desire) and avoidance of emotional / physical pain. If emotions drive a bus, then: artificial The behavior of intelligent entities can be wonderfully complex and, when working in concert with the artificial intelligence entity's relationships and knowledge database, may give rise to new behaviors such as intimacy and individuality.

[0043] The emotional concepts database 138 may store a set of core emotional concepts that the emotional concepts subsystem 122 retrieves in response to natural language input. In some embodiments, when the communication subsystem 116 receives an image (e.g., a picture of a mountain), the emotional concepts subsystem 122 may retrieve the emotional concepts associated with the image, such as big, rock, or tree, from the emotional concepts database 138. Hearing, sight, and smell also play an important role in cognitive development, and infants who reject tactile interaction suffer a greater disadvantage in speech development. In some embodiments, without tactile, auditory, and visual input, the artificial intelligence entity must elaborate on concepts, such as describing a mountain in words, as if it were a picture of a mountain. For example, a mountain is a very large object, made of rocks, covered in snow, and often has trees growing on it. Depending on the amount of information the artificial intelligence entity digests, artificial The intelligent entity's image of the mountain may be more or less complete. Not all information needs to be direct. "John was exhausted climbing the mountain" suggests that John expended a great deal of energy and that the mountain is large.

[0044] In some embodiments, the artificial intelligence entity may not be a blank slate. artificial While everything an intelligent entity learns is the result of conversation or reading, concepts (e.g., cognitive priors) may include the permanence of objects, the rules of grammar, and basic concepts such as big, small, rough, smooth, above, below, inside, outside, fast, slow, hard, soft, high and low.

[0045] In some embodiments, the acquired emotion concepts are artificial The emotional attributes of the intelligent entity may be modified. artificial The emotional attributes of an intelligent entity are: artificialThe emotional attribute may correspond to an emotional state of the intelligent entity. Examples of emotional states include happiness, trust, fear, surprise, sadness, disgust, anger, and alertness. Each emotional attribute of the artificial intelligence entity (e.g., each emotional state) may have a corresponding emotional value (which may be continuously updated) at a particular point in time. The corresponding emotional value may be equal to or greater than an emotional baseline (which may be continuously updated). The emotional baseline may correspond to the lowest possible attribute value for the emotional attribute. The emotional attribute may further be associated with a growth factor or a decay factor. The emotional value of the emotional attribute may change over time based on one or more growth factors or decay factors corresponding to the emotional attribute.

[0046] In some embodiments, the growth or decay factor for each emotional attribute may be predetermined and stored in growth / decay factor database 136. Emotional attributes of the artificial intelligence entity may be associated with one or more growth or decay factors. Each emotional state includes a time component that grows or decays at a specific rate (or multiple). For example, surprise decreases over a short period of time (with the emergence of new surprise), while grief decreases proportionally to the depth of sadness. Thus, each emotional attribute (e.g., each emotional state) of the artificial intelligence entity may be associated with one or more specific growth or decay factors. The growth / decay factor database 136 may include a list of growth / decay factors corresponding to a set of emotional attributes of the artificial intelligence entity, the server 102 may receive the growth or decay factor corresponding to each emotional attribute from the growth / decay factor database 136 (e.g., via the communication subsystem 116), and the factor adjustment subsystem 112 may determine the growth or decay factor based on the information received from the growth / decay factor database 136.

[0047] The emotional values ​​(associated with the emotional attributes) of the artificial intelligence entities may be continuously updated based on the growth or decay factors associated with the emotional attributes. For example, as shown in FIGS. 2A and 2B , the emotional values ​​202 and 201 (e.g., 207a-207f and 208a-208f) may be continuously updated based on one or more growth or decay factors associated with the emotional attributes A and B. Such continuous updating may include periodically updating these emotional values ​​according to a schedule or based on other automatically occurring triggers. In addition to updating the emotional values ​​202 and 201, the growth or decay factors associated with the emotional attributes A and B may also be updated based on one or more inputs (and / or one or more emotional concepts) received by the server 102. Additionally, the growth or decay factors may be updated based on other information in the natural language input. For example, other information of the natural language input may include a subject time decay factor, a subject geographical decay factor, an object time decay factor or an object geographical decay factor, a clause type, a subject of the clause, a subject type of the clause, a modifier of the subject of the clause, a type of modifier of the subject of the clause, a number of subjects of the clause, a subject time decay factor, a subject geographical decay factor, a verb of the clause, a tense of the verb of the clause, a modifier of the verb of the clause, an object of the clause, a type of object of the clause, a modifier of the object of the clause, a type of object modifier of the clause, a number of objects of the clause, an object time decay factor, an object geographical decay factor, a preposition of the clause, a modifier of a preposition of the clause, or a global tense modifier of the clause.

[0048] As mentioned above, other information in the natural language input may indicate time decay and geographic decay (TGD) factors (e.g., a subject time decay factor, a subject geographic decay factor, an object time decay factor, or an object geographic decay factor). As an example, humans instinctively understand that events occur in a sequence corresponding to the passage of time. Our built-in timeline separates events into future, past, or present, recognizing that events occurring today will soon become past, and future events will eventually become present. artificialIntelligent entities may also have the ability to understand these frameworks. As time moves forward, artificial The intelligent entity may update the timeline to understand past, present and future events.

[0049] For example, for a natural language input such as "The cat is on the street and John's house is on the corner," the natural language input contains information about the location of an object, indicating that it is currently located at a specific location. However, the future location of an object will change depending on the object's properties. The cat is an active object and will likely change location, while John's house is inactive and will likely remain located on the corner. In summary, the more active an object is, the faster it decays. To facilitate this process, every object in the dictionary is assigned a TGD variable that describes the time elapsed before the object's location information becomes uncertain and the degree of uncertainty. Such object information and the corresponding TGD variable may be stored in the growth / decay factor database 136. The TGD variable may be learned or derived in a variety of ways. An object described with a verb that indicates high activity ("The dog ran away with the spoon") would be assigned a high TGD, while "John can't walk" would decrease John's TGD because it suggests that John's activity level is decreasing. TGD values ​​can be inherited between classes.

[0050] artificial An intelligent entity may learn that living organisms have a high TGD and that unknown objects can be expected to be relatively inactive. For example, artificial If an intelligent entity has never encountered the word "truck" before and gets the information "John's truck is in his garage", artificial An intelligent entity may predict that the truck will remain in the garage for a year (e.g., human possessions have a higher TGD than non-possessed objects). If you ask "Where is John's truck?" a year from now, artificialThe intelligent entity would respond, "It's probably in John's garage." However, artificial If an intelligent entity (at any given time) learns that trucks are vehicles, and that vehicles are driven at high speeds, artificial The intelligent entity may retroactively modify the TGD value for the truck, based on learning that the truck is a vehicle, so that after doing so, if you ask "Where is John's truck?" artificial The intelligent entity would respond, "I don't know, maybe you should check his garage."

[0051] The incomplete paradox is intended to mean that an action taking place in the present does not imply that it should be completed in the future. Thus, "John is building a house" does not necessarily mean that John will have built the house in the future. The natural language subsystem 120 avoids this paradox by predicting that the house has been built, but with a low certainty factor.

[0052] After updating the boost or decay factors (based on one or more inputs and / or one or more emotional concepts), the emotion values ​​201 and 202 associated with the emotion attributes A and B may be updated based on the one or more boost or decay factors (updated based on one or more inputs and / or one or more emotional concepts).

[0053] In some embodiments, once one or more emotion concepts are retrieved by the emotion concept subsystem 122, the factor adjustment subsystem 112 updates a gain or decay factor associated with one or more emotion attributes of the artificial intelligence entity. For example, if the natural language input is "John died of cancer," one or more emotion concepts retrieved from the emotion concept database 138 may include "bad" and / or "concern." As a result, the gain / decay factor subsystem 112 may update (in a gradual or instantaneous manner) a gain or decay factor associated with one or more emotion attributes of the artificial intelligence entity (which may be associated with the emotion concepts), e.g., sadness, anger, and / or happiness. As an example, in one use case, emotion attribute A in FIG. 2A may correspond to "sadness" of the artificial intelligence entity. When a natural language input of "John died of cancer" is obtained, the factor adjustment subsystem 112 may update the growth factor of the emotional attribute "sadness" such that the emotion values ​​(e.g., 207c-207f) of the emotional attribute "sadness" increase from time c to time f (see date and time 206 in FIG. 2A ) based on the updated growth factor (which may be linear or nonlinear). In another use case, the emotion attribute B of the artificial intelligence entity in FIG. 2B may correspond to "happiness" of the artificial intelligence entity. When a natural language input of "John died of cancer" is obtained, the factor adjustment subsystem 112 may update the decay factor of the emotional attribute "happiness" such that the emotion values ​​(e.g., 208c-208f) of the emotional attribute "happiness" decrease from time c to time f (see date and time 203 in FIG. 2B ) based on the updated decay factor (which may be linear or nonlinear). It should be understood that in some embodiments, emotion values ​​for emotion attributes generally revert (or reset) to their respective baseline values ​​in the absence of input and / or after a predetermined period of time. For example, even if emotion values ​​202 increase from time c to time f in FIG. 2A , it should be understood that there is a threshold amount that these emotion values ​​can increase before they begin to decrease toward emotion baseline 204.Regardless of the increase or decrease factors, the emotion values ​​201 and 202 will not fall below the emotion baselines 205 and 204, respectively. In some embodiments, the emotional state subsystem 114 may further update the emotion baseline 214 (see FIG. 2C ) based on one or more inputs (and / or one or more emotion concepts). After the emotional state subsystem 114 updates the emotion baseline 214: artificial Emotional attributes associated with intelligent entities (e.g., The emotion value 212 (e.g., emotion values ​​218e and 218f among emotion values ​​218a-218f in Figure 2C) of emotion attribute C may be updated based on the updated growth or decay factor(s) and the updated emotion baseline. While the emotion values ​​212 (e.g., emotion values ​​218e and 218f) are illustrated in Figure 2C as decreasing between time c and time f based on the decrease in baseline 214 (see emotion values ​​212 and date / time 216 in Figure 2C), it should be understood that the emotion values ​​increase based on the updated growth or decay factor(s) and the updated emotion baseline (e.g., an increase in the baseline).

[0054] Based on enhancement or attenuation factors artificial Modifying the emotional value of an emotional attribute associated with an intelligent entity is analogous to the function of the human endocrine system, which includes glands that produce and secrete hormones that control the activity of cellular and emotional functions, for example, there are at least three influencing factors that control the activity of cellular and emotional functions. These are: Dopamine: Affects pleasure, joy, calmness, love, and alertness. Serotonin: Affects concentration, learning ability, surprise, and vigilance. Norepinephrine: Affects stress, anxiety, anger, grief, and rage. Changes in any of these influencers affect all emotions to varying degrees. For example: artificialIf an intelligent entity learns that someone is facing death, this sudden stress may trigger the release of artificial norepinephrine and cortisol, for example. This results in reduced emotional levels of joy, curiosity, and trust. Greater loneliness may lead to higher levels of sadness (but not vice versa, since one can be sad without being lonely). If the change in emotion is great enough, the artificially intelligent entity may decrease in emotion to the point of clinical depression (although the rate of increase and / or decay associated with the emotion attribute will cause the artificially intelligent entity to eventually recover).

[0055] Additionally, in some embodiments, once the natural language subsystem 120 processes the natural language input and parses the natural language input into the knowledge database 134, the emotional state subsystem 114 may determine one or more impact values ​​relating to the impact of each portion of the content of the input (e.g., the natural language input) on one or more emotional attributes of the artificial intelligence entity. For example, if the natural language input is "John has cancer," the emotional state subsystem 114 may determine an impact value relating to the impact of each portion of the natural language input (e.g., "John," "has," "cancer") on one or more emotional attributes of the artificial intelligence entity. Additionally, the emotional state subsystem 114 may: artificial It may determine whether the impact value meets a predetermined threshold that triggers an update (e.g., an increase or decrease) in one or more emotional values ​​associated with one or more emotional attributes of the intelligent entity. If the emotional state subsystem 114 determines that the one or more impact values ​​meet the predetermined threshold, the emotional state subsystem 114: artificialThe emotional value of the intelligent entity may be modified (e.g., increased or decreased). For example, if the word "cancer" is determined to have an impact value greater than a predetermined threshold for triggering an increase in the emotional attribute "sad," the emotional state subsystem 114 may modify (e.g., increase) the emotional value corresponding to the emotional attribute "sad." Additionally, the impact value may also trigger an increase or decrease in one or more amplification or decay factors when the impact value meets a predetermined threshold. Such an increase or decrease in the amplification or decay factor results in artificial The emotion value corresponding to the emotion attribute of the intelligent entity may be updated.

[0056] Additionally, in some embodiments, the server 102 may communicate with the artificial intelligence entity and one or more other entities (e.g., one or more other artificial The emotional state subsystem 114 may determine whether an interaction between the intelligent entity and / or one or more client devices exceeds an interaction threshold. Based on determining that the interaction exceeds the interaction threshold, the emotional state subsystem 114 artificial The emotional value of the intelligent entity may be modified. For example, if the artificial intelligence entity interacts with another entity a predetermined number of times in a predetermined time period, the server 102 may determine that a predetermined threshold for interaction has been met, and the emotional state subsystem 114 may modify the emotional value of the artificial intelligence entity (e.g., corresponding to "happiness," since increased interactions between entities indicate a growing friendship). The factor adjustment subsystem 112 may: artificial A boost or damping factor associated with the emotional attribute may be modified based on determining that an interaction between the intelligent entity and one or more other entities has exceeded an interaction threshold.

[0057] Additionally, in some embodiments, the server 102 artificial An intelligent entity and one or more other entities (e.g., another entity) artificialThe trust value may be determined and / or obtained to indicate a level of trust between the artificial intelligence entity and the other entity (e.g., the client device 104 or any other input source). The trust value may be determined and / or obtained based on the number of interactions between the artificial intelligence entity and the other entity, and / or artificial The determination may be based on the content of interactions between the intelligent entity and other entities. The emotional state subsystem 114 may update and / or modify the emotional value of the artificial intelligence entity based on the trust value, and the factor adjustment subsystem 112 may modify a boost factor or a damping factor associated with the emotional attribute based on the trust value.

[0058] In some embodiments, the server 102 may determine and / or obtain a certainty value associated with the occurrence indicated by the natural language input. artificial The confidence value may indicate a level of confidence of the intelligent entity. The confidence value may be determined based on whether the event is explicitly described in the natural language input or inferred from the natural language input and / or the confidence value. The emotional state subsystem 114 may update and / or modify the emotional value of the artificial intelligence entity based on the confidence value, and the factor adjustment subsystem 112 may modify a boost factor or a damping factor associated with the emotional attribute based on the confidence value.

[0059] Additionally, in some embodiments, the response generation subsystem 118: artificialA response with respect to the input may be generated based on the emotion value of the artificial intelligence entity. It should be understood that the response generation subsystem 118 generates a response based on the emotion value of the artificial intelligence entity before the emotion value is updated based on the input or after the emotion value is updated based on the input. For example, if the natural language input is "John died of cancer," a response generated by the response generation subsystem 118 (e.g., with respect to the input) may include "That's so sad." Such a response may be generated based on the emotion value, for example, before the emotion value is updated based on the input. Another response generated by the response generation subsystem 118 (e.g., with respect to the input) may include "That's so sad. I need time to process this news." Such a response may be generated based on the emotion value, for example, after the emotion value is updated based on the input. Thus, the response generation subsystem 118 may generate a response based on the emotion value of the artificial intelligence entity before the emotion value is updated based on the input or after the emotion value is updated based on the input. Further input may be obtained, and further responses may be generated with respect to the further input based on the emotion value of the artificial intelligence entity. For example, after updating the growth or decay factors (e.g., based on the input), another input may be obtained. After updating the growth or decay factors, a further response regarding this other input may be generated based on the set of continually updating emotion values ​​of the artificial intelligence entity. The further response may be transmitted via the communications subsystem 116, for example, to the client device 104 (or any other component within or outside of the system 100).

[0060] Further, as an example, if the natural language input is "Cancer is a very dangerous disease," and this is the first time the artificial intelligence entity has encountered the word "cancer," the artificial intelligence entity may assess the sentiment of this input (e.g., update the artificial intelligence entity's set of sentiment values ​​and / or gain or loss factors based on the input). Because the words "danger" and "disease" already have negative sentiment attributes (e.g., negative emotional states such as fear, sadness, and / or anger), when combined with "cancer," "cancer" is strongly associated with negative sentiment (the adverb "very" also has an amplifying effect). In other words, a response to the natural language inputs "Cancer is a very dangerous disease" and "John has cancer" may include an emotionally charged response (e.g., based on the sharp increase in sentiment values ​​of the negative sentiment attributes and gain or loss factors associated with the negative sentiment attributes in response to such input). For example, an emotionally charged response could include "That's shocking." However, if “cancer” is later defined (e.g., based on the input “cancer is not always fatal.” A negative definition such as “cancer is not always fatal” does not carry the same absolute weight as a positive definition; “I am not happy” may have a lower absolute emotional weight than “I am sad”), the emotional value of the word “cancer” may be modified by averaging it with past levels. In other words, another input containing the word “cancer” (e.g., “Peter has cancer”) may not trigger a similarly sharp increase in the emotional value of the negative emotional attribute and the boost or damping factor associated with the negative emotional attribute in response to such another input. This is because the artificial intelligence entity has increased its knowledge of the word “cancer.” Thus, a response to “Peter has cancer” may include “That's unfortunate. I hope he gets the best treatment.” In this way, a response to “John has cancer” differs from a response to “Peter has cancer” because the emotional value associated with the emotional attribute has been updated based on further input about “cancer.”

[0061] Further, as an example, "John has cancer" may prompt an AI entity to evaluate its relationship with John as a product of how much the AI ​​entity trusts John, what it knows about John, how others feel about John, and the nature of their past relationships (see relational databases, discussed below). In this case, a high level of positive emotion about John multiplied by a strong negative emotion about cancer may result in a strong negative response. The impact of an event on emotions may be modified by temporal conditions. John is hitting me (very high) John hit me yesterday (high) John is going to spank me sometime next week (low) John hit me last year (very low)

[0062] Example of creating output depending on input

[0063] The system 100 may receive one or more of the following types of input: 1) questions, 2) statements, and 3) responses to previous questions, each of which creates an output according to its own set of rules.

[0064] Responding to Questions: The AI ​​entity's emotional priors (desire to respond accurately and provide new information) determine its response to five different questions when asked: A) logical A simple, objective question such as "Are dogs mammals?" requires a search in the artificial intelligence entity's knowledge database 134. Example responses include: Yes, dogs are mammals. No, dogs are reptiles. I don't know. B) Inferential questions Complex questions that require backward reasoning by the system use a procedure that involves formulating hypotheses and testing them backward according to rules. Input: The moon is round. Is it a ball? A: I don't know. The moon is round, so it might be a ball. Input: A ball can bounce, but the moon cannot. Is the moon a ball? A: No. The AI ​​entity's level of certainty is based on the proportion of shared features relative to all other known features, but if any one feature contradicts known facts, the process will infer that the parrot is not a bird. C) Open-ended questions "Tell me something about John?" requires analysis in knowledge database 134, revealing that he is male, a mammal, breathes air, has two eyes, two ears, two arms, and so on, loves his mother, loves his dog, owns a boat, and goes to school. Any of these facts would be a logically valid response, but not necessarily the same as a human's answer. Without measuring emotion, the AI ​​entity would return an Eliza-type response: "John has two eyes and has a pet." By selecting the knowledge record with the highest emotional weight, it would respond, "John loves his mother." If no record contains a significant emotional value, in the spirit of "tell me something I don't know," the AI ​​entity would multiply the number of times the record has been referenced by the number of times the object has been referenced and select the one with the lowest score. a. John has a house (the house is usually the reference) b. John owns a boat (boats are rarely referenced) c. John is a man (it is common for men to be referenced) A likely response is: "John owns a boat. I think he likes fishing." Based on the artificial intelligence entity's association of "boat" with "fishing," this inference has low certainty. D) Personal Questions - Questions about the AI ​​entity's physical state ("How old are you?") are answered by querying the cognitive system. Questions about the AI ​​entity's mental / emotional state ("How are you?") are answered by asking about its Current Emotional State (CES). E) Complex personal questions ("Why are you sad?") require retrospective analysis of the artificial intelligence entity's knowledge base to ascertain the cause of the current emotional state.

[0065] Response to Descriptions / Observations - To create a human-like response to a description, the AI ​​entity's emotional priors must be consulted. Each potential response is assigned a score using a simple scoring system described below: for example, "Dogs have sensitive noses." A) Assigning a score to candidate responses that are objective observations - The cognitive system returns the following inputs: a. Dogs are mammals (advanced knowledge) b. Dogs have four legs (low level knowledge) c. Humans have noses (advanced knowledge) d. AI entities have noses (low level of knowledge) e. Plants don't have noses (advanced knowledge) f. The nose is used to sense smells (advanced knowledge) Each knowledge record is assigned a score as follows: a. How accurate is the knowledge? (See Certainty / Confidence) (1-10) b. How rare is this knowledge? (How many times has it been referenced?) (1-10) c. Coefficient of high altitude = 10, coefficient of low altitude = 0 d. Contents related to absolute emotions (1-10) Our brains assign different levels of importance to different elements of knowledge. The statement "tigers are dangerous" may be more important than the statement "grass is greener" because, from a normal human perspective, the former statement has a higher absolute emotional content than the latter. Thus, the highest score represents the most appropriate fit to the topic of conversation, even taking into account some recency and emotional content, and depending on the level of certainty of the knowledge (explicit or inferred), the response might be "I think it means I have a good sense of smell." B) Score candidate requests for further information - A low number of low-degree responses from the cognitive system indicates a gap in the AI ​​entity's knowledge base. To satisfy the AI ​​entity's sentiment priorities (need to learn, need to maintain a coherent conversation, etc.), the AI ​​entity uses the following scoring method: Score = 1 / (number of low-degree knowledge records on this subject) / (average number of low-degree knowledge records across all subjects). Example: "Is my nose sensitive?" C) Score Potential Intimate Responses - When an input causes a dramatic (absolute) change in the AI ​​entity's current emotional state (e.g., your dog just died), the AI ​​entity may be inclined to provide an intimate, emotional response such as "That's a shame. I'm sad." Score = (Absolute CES Change^2) D) Request to Change Topic - If the scoring system described above does not reach a given threshold, it likely means that the AI ​​entity has little to offer in terms of topic. The AI ​​entity's default response is an offer to change the topic to a previous topic with which the AI ​​entity has the highest emotional connection.

[0066] Additional Databases

[0067] In addition to the databases described above, database 132 may include additional databases related to emotions, which may be understood to be included in one or more of databases 134, 136, 138, or other databases.

[0068] The additional databases may include a cognitive system database, an object relation database, and a relational database. The cognitive system database may be queried for emotional content embedded in current and past inputs, explaining the reasons for the emotions felt by the artificial intelligence entity. "Why are you sad?" triggers a backward search in the knowledge database 134 to find past inputs that explain the sadness the artificial intelligence entity is currently feeling. However, the same question does not always produce the same answer. This is because the emotional value of a knowledge record is modulated by (A) the current emotional state, (B) the temporal aspect of the input ("My dog ​​died today" vs. "My dog ​​died 10 days ago"), and (C) the relationship between the artificial intelligence entity and the interviewer.

[0069] The object relation database may be a hash table that stores the primary emotions for all objects that the natural language subsystem 120 has ever encountered. The hash table is updated each time an object is encountered. The relation database holds the emotions associated with all of the people and relationships that the cognitive system database has encountered. The artificial intelligence entity can identify individuals who are not genuinely interested, who provide false information, who scold or tease, or who evoke negative emotions, and assign them a low trust value. The relation database may call the following functions: (1) Name Identification Function - The name identification function uniquely identifies an individual. For "Dirk's father gave him money," consider the following hierarchy: Which Dark was the last one mentioned in this interview? Which Dark did you mention most often? If that doesn't work, we'll assume it's the last mentioned Dark. If Dirk has not been mentioned before, a new record is added. Now, the natural language subsystem 120 may conclude that Dirk is male (he) and "Dark's" is possessive, so it recognizes a relationship between Dirk and his father. If this is the first time the natural language subsystem 120 has encountered Dirk's father, the natural language subsystem 120 may request the name of Dirk's father (Dave) and add a new record. (2) Relationship hierarchy function - assigns affective values ​​to family relationships according to the following hierarchy: 1. Myself 2. AI programmer 3.Mother / Father 4. Daughter / Son 5. Sister / Brother 6. Grandmother / Grandfather 7. Interviewer 8.People in general 9. Biology in general

[0070] In some embodiments, one or more databases (or portions thereof) may include one or more graph databases (e.g., directed graph concepts and data structures). In some embodiments, a graph associated with an AI entity (as described herein) may include information from the knowledge database 134 and the emotion concept database 138 (and / or the boost / decay factor database 136 or other databases), and the AI ​​entity may query the graph (also referred to herein as an “ontology-emotion graph”) to process input, generate responses, or perform other operations. In some embodiments, ontology categories and entries (e.g., from the knowledge database 134 or other sources) may be based on semantically meaningful instinctual and emotional information. In some use cases, this instinctual information may be a “stub” (e.g., represented as a node in a graph) that complements or replaces the embodied feedback required by the theory of embodied intelligence. Such stubs may provide a "grounding" for symbols and serve as basic units of meaning. For example, such stubs may enable comparison of disparate concepts, be used to pre-train neural networks in a supervised or unsupervised manner before connecting them to AI systems as part of transfer learning (e.g., training on instinctive concepts allows for a priority start when dealing with more advanced concepts that are later labeled with instinctive nodes), or help infer other learned graph attributes for relatively unknown nodes from other explicitly labeled nodes (e.g., emotion, utility, etc.). Examples of such stub nodes may include hard, soft, light, heavy, up, down, above, below, etc., as well as concepts that humans intuitively learn by manipulating objects and experiencing their own bodies in space and time. To obtain these stubs, in some embodiments, these stubs are initially manually annotated and then combined with pre-trained word vectors and behavioral and behavioral information. Emotional information can be propagated to new nodes via inference or association from the information and emotions. Another means may be information spontaneously acquired or derived from conversational partners. In some embodiments, instinct stubs, emotional attributes, neural network circuits, or other components may combine to provide the components of an AI entity's emotions or otherwise form the emotions themselves. In some embodiments, the graph may be augmented with additional probabilistic information (e.g., similar to the probability information provided in a Bayes Factor graph) and causal information (e.g., obtained by Judea Pearl's do calculus and / or do operator). For example, inferred information may have a probability weight associated with each storage, which should be derived to influence the output. In some embodiments, the AI ​​entity may query its conversational partners for information about uncertain or low-probability associations.

[0071] In some embodiments, as described above, the AI ​​entities may include one or more neural networks or other machine learning models (e.g., one or more of an embedding network, a consumption network, or other models described herein). In some embodiments, each element or sub-network of a machine learning model (e.g., a deep learning network) may be converted into a semantically meaningful vector abstraction to facilitate mathematical functions and machine learning on the “meaning” of each element or sub-network. As an example, where it is important to process words as ideas (e.g., king-man + woman = queen), the usefulness of such a conversion becomes evident in word2vec, an algorithm that converts words, phrases, or sentences into vectors, such as in NLP. In some embodiments, the embedding subsystem 124 may utilize a graph embedding network or other component to convert one or more portions of a graph (e.g., nodes, sub-graphs, etc.) into one or more embeddings of each portion of the graph (e.g., higher-order embedding vectors of nodes, sub-graphs, etc.). In some scenarios, such transformations may take into account the hierarchical structure of the graph, heterogeneous node types (ontology, sentiment, probability), metadata, associations between nodes in the graph, other context of the graph, or other information (e.g., sensor information obtained from pre-trained embeddings, supervised learning, unsupervised learning, reinforcement learning, etc.). For example, a graph embedding network may be configured to represent the structure and hierarchy, heterogeneity of node types, and the metadata of the graph in the embeddings that are transformed from parts of the graph by the graph embedding network. In some embodiments, the graph embedding network may be unsupervised or semi-supervised. As an example, the unsupervised network may be configured to utilize an intrinsic reward function (reinforcement learning) to provide a learning signal (e.g., to improve the effectiveness of the embeddings as representations for parts of the graph).

[0072] In some embodiments, a two-way reference may be maintained between a node or subgraph (of a graph) and the embedding representing that node or subgraph (from which the embedding originated). For example, the two-way reference may be maintained regardless of the data structure chosen to store such embeddings (e.g., a tensor, a matrix, a database, the graph itself, etc.). For example, if a given embedding represents a first node of a graph, the graph may store the embedding as a second node in the graph such that an edge is shared between the first and second nodes (e.g., a two-way association between the first and second nodes). In this way, for example, symbolic, human-understandable nodes and subgraphs may be reasoned / associated with dense vector abstractions (e.g., embeddings), such that symbolic operations may be interleaved with sub-symbolic operations on vectors in meaningful space. As an example, a graph query algorithm may be used to select one or more subgraphs or nodes, and then these representation vectors may be further processed by machine learning algorithms to generate and output or even re-query graphs at a symbolic level.

[0073] In some embodiments, the graph embedding network and the consuming network (which consumes the embeddings produced by the graph embedding network) are directly related, allowing them to be trained end-to-end. In some embodiments, the graph embedding network and the consuming network are separate from each other. As such, pre-trained embedding vectors (produced by the graph embedding network) are transported to the appropriate layer in the consuming network so that the embedding vectors can be used appropriately. In some embodiments, for efficiency, retrieving vectors for nodes or subgraphs from the graph may be done using a hierarchical array, sparse array, or tensor that may be indexed by their respective identifiers (e.g., graph or subgraph IDs). In some embodiments, the vectors may be provided as inputs to the consuming network, which may generate one or more outputs based on the vectors. The architecture of the input layer (e.g., embedding input layer) of the consuming network may be configured based on the specific architecture and hyperparameters of the upstream graph embedding network to enable the consuming network to process the embedding vectors appropriately. In some embodiments, the vectors may be used as weights (e.g., frozen weights or learnable weights) for encoding other inputs (e.g., as one-hot representations). In some cases, if the vectors are used as learnable weights, the updated vectors may be transferred back to the graph embedding network as weight vectors, and may be further refined by the graph embedding network before being transferred downstream (e.g., to the consuming network or other consuming networks).

[0074] In some embodiments, such a consuming network (a network that consumes vectors generated by an upstream embedding network) may include a sequence neural network. In one example, a sequence neural network may be trained to maintain or output dense vector representations of conversation history, current state, or other such memory-related data. In one use case, an AI entity may interact with another entity (e.g., a human user, another artificial With respect to a conversation between an AI entity and another entity (e.g., an intelligent entity), the AI ​​entity may utilize a long-short-term memory (LSTM) network (or other sequence neural network) to consume the associated ontology-sentiment graph embedding vectors (and / or BERT or other pre-trained word embeddings) representing the other entity's input (e.g., words, phrases, sentences, or other input provided by the other entity) and link or stack the vectors to obtain a temporal structure (e.g., with the emotional and affective information of the AI ​​entity or other entities). Based on the LSTM network's conversation history information, the LSTM network may output a vector representing the conversation state (also referred to herein as a "conversation state vector") (e.g., similar to a human record of the conversation). By way of example, the LSTM network may output a conversation state vector in response to subsequent input provided by the other entity or other automated triggers (e.g., similar conversations or contexts, names or other identifiers of other entities introduced in subsequent conversations, etc.). This vector may then be stored as a conversation memory node in the ontology-sentiment graph or consumed by another neural network.

[0075] In some embodiments, the AI ​​entity may rely on relational pattern reasoning (as well as learned emotional behavior) to learn automated logic about the agent's goals and environment. artificialCompared to intelligent systems, they may be configured to modify symbolic abstractions (such as concepts about a particular person) in relation to a reward function. Thus, for an artificially intelligent entity, its emotions and feelings (e.g., emotions of pain, emotions of fear, etc.) can serve as behavioral signals that predispose a behavioral pathway or ontological entity to satisfy one or more behavioral priors in a given context. The reward function then helps assign meaning to emotions and feelings, which provides meaning to metadata such as probabilities, instinct stubs, and behavioral pathways, ontological entities, or concepts that blend two or other node types. In this way, intrinsic rewards (in the form of reward functions) can act as a substitute for behavioral priors, and emotions and feelings act as shortcuts to assimilate goal-based behaviors directly or indirectly (e.g., from downstream proxy emotions) into the system. As one example, a reward function may be used to translate the attainment of some goal, such as increasing the positive emotion of an AI entity's interlocutor (e.g., an entity with which the AI ​​entity is interacting), into a signal that privileges all kinds of things that lead to that goal at the level of individual concepts, rather than just neural network circuits. These and other graph attributes and nodes described herein can be propagated either by deductive methods described herein (e.g., person X likes dogs, dogs are animals, therefore person X likes animals with some degree of certainty), by graph induction methods described herein, or by other methods.

[0076] In one use case, with respect to the concept of baseball, if an AI entity's interlocutors respond positively to the topic, the AI ​​entity will develop a positive impression of baseball and learn to bring up this topic more frequently in a given context. This adds another semantic layer to the concept "baseball," which is reflected in its vector space embedding. This then dynamically triggers adjustments in the AI ​​entity's behavior by tagging concepts, entities, behavioral pathways, or neural networks (or, more traditionally, updating neural network parameters). Thus, for example, the AI ​​entity may operate on emotions (or the emotions of others) in a separate processing layer (e.g., to resolve conflicts in the sources of emotions). As an example, the AI ​​entity may be augmented with neural networks or symbolic logic / conditional programming layers trained with other reward signals (as described herein).

[0077] In some embodiments, the intrinsic reward function is paired with various internal parameters that increase or decrease "pressure" on the AI ​​entity by increasing the intensity of emotional and affective responses in existing and new emotion tags. These include, for example, a time component where the pressure increases as a function of the length of time since the interaction partner was last seen, or in conjunction with factors described below. In some embodiments, the AI ​​entity may be configured to modify its tendency to function according to the reward function inversely proportional to the amount of reward experienced over time. As an example, the AI ​​entity may increase its tendency to seek praise as the ratio of total praise value (e.g., number of praises received from the interaction partner and value of each praise) to conversation time decreases. In this scenario, the praise dynamics described above are an example and may be a novel behavior from the mechanisms described herein. In some embodiments, such dynamic "pressure" may be based on internal parameters that should be taken into account when evaluating emotional and / or affective resonance (e.g., mapping of concepts to reward expectations). As an example, the AI ​​entity may be configured to increase or decrease emotional mapping of concepts (e.g., increasing trust in a particular individual). In one use case, based on its emotional mapping to an individual, an AI entity may be more likely to behave in a certain way (e.g., be more open to sharing “personal” information) when interacting with other entities deemed similar to this individual. More generally, and at a slightly lower level of abstraction, there may be one or more global parameters that amplify the emotional and / or affective resonance of concepts in the ontology-emotion graph. These may be adjusted to increase or decrease based on dynamic reward pressure input. For example, a curiosity reward mapping may be increased as a global curiosity reward multiplier parameter if the AI ​​entity's acquisition of new knowledge is relatively low for the preceding period. In some embodiments, the dynamic pressure may be based on a data structure that holds reward magnitudes and durations (e.g., dynamic magnitudes and durations).The representation vectors may be obtained based on the data structure and fed as input to a network that directs the behavior of the AI ​​entity, thus applying emotional "pressure" to trigger intrinsic rewards (e.g., according to reward magnitude and time).

[0078] This allows for greater granularity and nonlinearity in reward adjustments, as reward history can be tracked for multiple different concepts, subgraphs, and concept classes, allowing the neural network to implicitly learn to adjust pressure nonlinearly. In some embodiments, the AI ​​entity may be configured to map one or more emotional attributes to one or more concepts based on similarity to other concepts with those emotional attributes. As an example, the AI ​​entity may map the emotional attribute and grounding and context Y associated with concept X to other concepts with similar grounding and context. In one scenario, if a survival inhibitor (e.g., the emotion of pain or the emotion of fear) (or other emotion or affect) is associated with "gun" (represented as a node in the graph), and the node "gun" has attributes such as force, metal, hard, etc., the behavioral tag associated with "gun" may be associated with other nodes with similar grounding, even if those other nodes are not in the same class type (e.g., nodes not in the firearms class). For example, a behavioral tag associated with "gun" may be associated with the "baseball bat" node based on the "baseball bat" having attributes such as strength, metal, and hardness. As another example, a behavioral tag may be associated with the "baseball bat" node with a relatively low confidence (e.g., encoded as a probabilistic confidence) compared to the confidence of the association between the behavioral tag and the "gun" node.

[0079] In some embodiments, the system can further learn to assign grounding attributes from emotion and context to existing nodes that initially lack these attributes. As an example, using previously annotated data, ontology-emotion vector space similarities, or pre-trained word embeddings or conversational input, the concept of "heavy" can be learned in a graph neural network to correlate to subgraphs or concepts with highly negative affect, sentiment, and / or concept densities. The neural network will learn this implicitly over time given one or more of multiple event inputs. Additional "heavy" instinct nodes may then be inferred with varying strength for other similar graph regions.

[0080] In this way, the AI ​​entity can "think" and learn from its current information database (e.g., a sufficiently information-rich graph) without necessarily acquiring this information from new conversations or other sources. Predicting these attributes can facilitate deriving groundings, emotions / feelings, ontologies, latent factors (newly inferred nodes not previously present), and conversational output, among others. As an example, graph data and structure can be used for graph neural networks, which learn information about graph entries from other entries. In some embodiments, similarities in word alignments may be used (e.g., by the AI ​​entity or other systems) to propagate information from pre-trained word embeddings to the graph, which would then proceed to graph embeddings. In this way, while pre-trained word embeddings may not encode the types of metadata or graph structure encoded in the AI ​​entity's complete ontology-sentiment graph, graph embeddings (generated from the graph, or portions thereof) may include such graph structure, metadata, or information beyond that contained in the word embeddings. As an example, a graph embedding network may be configured to use a similarity measure on word embeddings in the graph to retrieve words that are close to the label of a given node (e.g., within a certain tolerance defined by the hyperparameters or learned weights of the network), or to traverse the graph at a symbolic level to find similar entries, such as entries that share a class. An AI entity may then use the neural network or similarity measure to find words that are close to known graph nodes. These words can be inserted into the graph and associated with the given node that initiated the query. In one use case, a search for new associations may be performed.In this case, new words from the conversational input overlap with words for instinctual or emotional data that are mapped to a given node in a pre-trained word embedding space (e.g., the new word is determined to be similar to a given node based on recognized instinctual or emotional words). In some embodiments, asymmetry between node contexts may indicate a knowledge gap and trigger the AI ​​to acquire new information for equalization. In one use case, when nodes share excessively asymmetric contextual mappings (e.g., multiple nodes that share strong class similarity but lack instinctual similarity), or when similar nodes have asymmetric method acquisition histories (node ​​a is deductive and node b is inductive), the system may be triggered to attempt to equalize the asymmetry using inductive / subsymbolic, deductive / symbolic, or conversational methods. In some embodiments, logically or probabilistically contradictory information within the ontology-emotion graph or between the graph and external inputs may also trigger a conflict resolution mechanism utilizing such knowledge acquisition techniques described above.

[0081] In some embodiments, the ontology-emotion graph may include vectors and / or symbolic nodes related to the emotions of one or more entities (e.g., human entities, other AI entities, etc.) toward the AI ​​entity. As an example, such vectors may be used to consider the relationship history and patterns of the AI ​​entity, rather than simply the AI ​​entity's current emotional state. In one use case, such vectors may be fed into or updated by a neural network (e.g., an LSTM or other sequence neural network) to encode the current emotion of the entity associated with the vector toward the AI ​​entity. The vectors may be generated by the neural network based on the AI ​​entity's interaction with the entity (e.g., the entity's inputs provided to the AI ​​entity, the AI ​​entity's responses to the entity's inputs, etc.). In some embodiments, such a neural network may be paired with a graph embedding network (as described herein) (e.g., configured to perform unsupervised and reinforcement learning on concept vectors from portions of the graph, such as pre-trained word vectors). As an example, this neural network may feed into a response template selection network:

[0082] In some embodiments, one or more intrinsic reward functions are utilized to train a neural network (e.g., a feedforward network or other network) for response selection (e.g., to select a generic template of a response to an input). The AI ​​entity's ability to: (i) pose a question; (ii) state an opinion; (iii) issue a command; (iv) respond to a question; (v) provide information; and (vi) change the topic. In some embodiments, the reward function may be configured to assign higher rewards to responses that elicit positive emotions and novel information to train this "response template selection" network. In one use case, the response template selection network may be configured to take as input one or more of the following: (i) input from a conversation partner (e.g., another entity interacting with the AI ​​entity) via a pre-trained embedding / transformer network (e.g., BERT (Bidirectional Encoder Representations from Transformers) or other similar network); (ii) a vector of conversation history (e.g., a vector trained via a sequence neural network); (iii) one or more embeddings (e.g., embedding vectors of the conversation partner's current and past emotions for the AI ​​system) from an ontology-sentiment graph (e.g., including emotional history) for the conversation partner formed from a sentence representation; and / or (iv) a sequence representation of the input constructed from ontology-sentiment graph embeddings of appropriate words from the input. These are processed to obtain a temporal sequence representation using a sequence neural network, or the template selection network has a sequence input branch to process the temporal order of the vectors or some other method of preserving the temporal order. Upon processing the input, the response template selection network may generate a vector indicating a response template to select (e.g., a generic template from one of the candidate responses listed above). In some embodiments, predefined templates may be generated from scratch by a sequence model (e.g., with similar inputs fed into a response template selection network) or may be constructed from a series of smaller pieces selected by deep learning.

[0083] In some embodiments, template variable selection (e.g., the process of filling in predicates, subjects, verbs, and other gaps in a selected template) may be performed. An example of using learned vectors is as follows: An input to an AI entity may be, "I need a new friend. Who should I go to lunch with?" A candidate response template is selected, e.g., "Good question. I'm <positive emotion> helpful. You should meet <person A>. <he or she> is <adjective>." The selection of a response template may trigger a query of the graph embedding of the conversation partner and a similarity distance, either Euclidean or cosine, between all other entities in the graph. Such a function (e.g., distance measure) represents a measure of similarity between concepts. Upon receiving the above query, the template may be filled in using the query along with a symbolic query about the AI ​​entity's emotion toward the conversation partner. When pressed to provide an explanation for the AI ​​entity's response, it can sift through the graph and output symbolic similarities between two individuals in this case, or simply the symbolic realization of vector computation in the general case. Thus, although post hoc, this pattern reflects the ex post justification of intuitive decisions in human reasoning as well, demonstrating that a mix of symbolic and sub-symbolic vector space reasoning is possible. Other possible queries exist with more thorough symbolic and sub-symbolic functions where the associations are stronger.

[0084] In some embodiments, after template variable selection (e.g., and further processing), the AI ​​entity outputs text and awaits a response from the conversation partner. Upon receiving the input, the AI ​​entity (or other components of system 100) may parse and evaluate the input based on two intrinsic reward functions, using emotional analysis and assigning a score to the amount and importance of the new information obtained. Parameters of preceding networks (e.g., graph embedding network, emotion embedding network, response template selection network, template variable selection network, etc.) may be updated according to learning signals from these functions. In some embodiments, the conversation state vector and ontology-emotion graph may be updated (e.g., to reflect recent interactions).

[0085] In some embodiments, a neural network associated with an AI entity may be trained to determine one or more emotional attributes of the AI ​​entity's response to a situation (e.g., in response to what a conversation partner says or in response to other contextual input). In some embodiments, this "emotion" network may include a deep neural network that receives as input one or more of the following: (i) the AI ​​entity's current emotional state; (ii) the conversation history between the AI ​​entity and its conversation partner (e.g., as represented by a conversation state vector); (iii) concept vectors from the ontology-emotion graph for the input words (and some representation of their sequence in one or more sentences); or (iv) one or more concept vectors from the ontology-emotion graph that are similar to the input words or other concepts in the input sentence. The foregoing may be retrieved using a behavioral prior criterion using vector space queries (e.g., analogy or comparison, similarity search based on past emotional content, etc.). The output of the emotion network may include an emotional tag indicating an emotional response to one or more of the concepts in the input sentence.

[0086] In some embodiments, emotions may be initially hard-coded. Additionally or alternatively, an "emotion" vector space may be created by training a neural network based on various known emotions (and their locations in the emotion vector space) to create novel combinations of novel emotions in the vector space. These would be meaningful abstractions for deep learning to learn from generating responses to behavioral priors / reward functions, instinctual priors, context, structure, and the current emotional state.

[0087] In some embodiments, emotions may be treated as an entire behavioral circuit that forms abstractions around intrinsic goals, context / history, concept vectors, and its own and other states. For example, the behavioral circuit generates actions and adjusts global parameters (e.g., speech intensity). In this framework, input is parsed and its concept vector (including past or affective content) is retrieved from an ontology-emotion graph using symbolic queries against associated nodes. The parsed input and concept vector, along with the conversation state, current emotional state (CES), and other factors, are then fed into an emotion neural network. The network adjusts its internal CES parameters to make decisions about activating trained sub-networks (e.g., another LSTM neural network). In some embodiments, the LSTM then uses some of these inputs and vector abstractions as memory cells to generate outputs. Responses from conversational partners are then classified and matched against a reward function using a "critic" framework. After this, we generate a credit signal for each reward function to update the LSTM parameters along with the affective or sentiment attributes in the graph.

[0088] Flowchart example

[0089] 3 through 5 are exemplary flowcharts illustrating process operations included in methods for enabling the various features and functionality of the system described in detail above. The process operations of each method described below are for illustrative purposes only and are not intended to be limiting. For example, in some embodiments, the methods may be implemented with one or more additional processes not described and / or may omit one or more of the processes described. Additionally, the order in which the process operations of each method are illustrated (and described below) is not intended to be limiting.

[0090] In some embodiments, the methods may be implemented in one or more processing devices (e.g., digital processors, analog processors, digital circuits for processing information, analog circuits for processing information, state machines, and / or other mechanisms for electronically processing information). These processing devices may include one or more devices that perform some or all of the operations of the methods in response to instructions electronically stored on an electronic storage medium. These processing devices may include one or more devices configured with hardware, firmware, and / or software specifically designed to perform one or more of the operations of the methods.

[0091] FIG. 3 illustrates a method 300 for facilitating emotional state-based artificial intelligence, according to one or more embodiments. At step 302, one or more growth or decay factors may be determined for a set of emotional attributes of an artificial intelligence entity. In some embodiments, as described above, the growth or decay factors for each emotional attribute may be predetermined and stored in growth / decay factor database 136. As described above, each emotional state includes a time component that grows or decays at a specific rate (or coefficient). For example, surprise decays over a short period of time (as new surprises appear), while grief decays proportionally to the depth of sadness. Thus, each emotional attribute (e.g., each emotional state) of an artificial intelligence entity may be associated with one or more specific growth or decay factors. Growth / decay factor database 136 may include a list of growth / decay factors corresponding to the set of emotional attributes of the artificial intelligence entity, and factor adjustment subsystem 112 may determine the growth or decay factors corresponding to each emotional attribute based on information received from growth / decay factor database 136.

[0092] In step 304, the set of emotional values ​​of the artificial intelligence entity may be continuously updated based on a growth factor or a decay factor over a predetermined period of time. Note that a set of emotional attributes is associated with the set of emotional values ​​of the artificial intelligence entity. The emotional values ​​(associated with the emotional attributes) of the artificial intelligence entity may be continuously updated based on the growth factor or the decay factor associated with the emotional attribute. For example, as shown in FIGS. 2A-2C, emotional values ​​201, 202, and 212 (e.g., 207a-207f, 208a-208f, and 218a-218f) may be continuously updated based on one or more growth factors or decay factors associated with emotional attributes A, B, and C. In some embodiments, continuously updating the emotional values ​​of the artificial intelligence entity may include periodically updating the emotional values ​​of the artificial intelligence entity based on one or more growth factors or decay factors.

[0093] In step 306, input may be obtained over a predetermined period of time. This input may be received from the client device 104, artificial The input may be obtained from an intelligent entity and / or from any source within or external to the system 100. The input may include natural language input, audio input, image input, video input, or other input. For example, the natural language input may include "John has cancer, and cancer is very dangerous." Similar input may be obtained by the server 102 as audio input, image input, and / or video input. In step 308, a response may be generated related to the input. The response may be generated based on a continuously updated set of emotion values ​​of the artificial intelligence entity. For example, if the natural language input is "John died of cancer," a response generated by the response generation subsystem 118 (e.g., related to this input) may include "That's a shame" or "That's sad. I need time."

[0094] In step 310, the growth or decay factors may be updated over this time period based on the input. In step 312, after updating the growth or decay factors, the set of emotion values ​​may be updated based on the updated growth or decay values. For example, FIGS. 2A and 2B illustrate updating emotion values ​​(e.g., emotion values ​​207c-207f in FIG. 2A and emotion values ​​208c-208f in FIG. 2B) associated with emotion attributes A and B of an artificial intelligence entity. In addition to being updated based on the updated growth or decay factors, the emotion values ​​may also be updated based on one or more impact values ​​and / or interaction thresholds. For example, one or more impact values ​​may be determined relating to the impact that each portion of the content of the input (e.g., natural language input) has on one or more emotion attributes of the artificial intelligence entity. For example, if the natural language input is "John has cancer," then impact values ​​may be determined relating to the impact that each portion of the natural language input (e.g., "John," "has," "cancer") has on one or more emotion attributes of the artificial intelligence entity. Furthermore, artificial The system may determine whether the impact value meets a predetermined threshold that triggers an update (e.g., an increase or decrease) in one or more emotional values ​​(associated with an emotional attribute of the intelligent entity). If it is determined that one or more impact values ​​meet the predetermined threshold, the emotional value of the artificial intelligence entity may be modified (e.g., increased or decreased). For example, if it is determined that the word "cancer" has an impact value greater than a predetermined threshold for triggering an increase in the emotional attribute "sad," the emotional value corresponding to the emotional attribute "sad" may be modified (e.g., increased). Furthermore, the impact value may further trigger an increase or decrease in one or more amplification or decay factors when the impact value meets the predetermined threshold. Such an increase or decrease in the amplification or decay factor may artificial The emotion value corresponding to the emotion attribute of the intelligent entity may be updated.

[0095] Additionally, in some embodiments, the artificial intelligence entity and one or more other entities (e.g., one or more other entities) may interact with each other. artificial The server 10 may determine whether interactions between the artificial intelligence entity and / or one or more client devices exceed an interaction threshold. An emotion value of the artificial intelligence entity may be modified based on the interaction being determined to exceed the interaction threshold. For example, if the artificial intelligence entity interacts with another entity a predetermined number of times within a predetermined time period, the server 10 may determine that a predetermined threshold for interactions has been met and may modify the emotion value (e.g., corresponding to "happiness") of the artificial intelligence entity (as increased interactions between the entities may indicate a growing friendship). artificial A boost or damping factor associated with the emotional attribute may also be modified based on determining that an interaction between the intelligent entity and one or more other entities has exceeded an interaction threshold.

[0096] FIG. 4 illustrates a method 400 for updating one or more boost or decay factors based on natural language input, according to one or more embodiments. In step 402, natural language processing may be performed on the acquired natural language input, for example, by applying grammar rules and logic. By applying grammar rules and logic, the natural language subsystem 120 may split complex sentences, resolve subject / object / verb ambiguities, and parse these components into the knowledge database 134. In step 404, one or more emotion concepts may be obtained based on each component of the parsed natural language input. The emotion concepts may be obtained from an emotion concept database 138. The emotion concept database 138 may store a set of core emotion concepts associated with images, audio, video, and / or natural language. For example, a set of core emotional concepts may include good, bad, danger, anger, surprise, love, safety, patience, trust, concern, big, small, rough, smooth, up, down, inside, outside, fast, slow, hard, soft, high, low, etc. As an example, if the natural language input is "John died of cancer," one or more emotional concepts retrieved from the emotional concept database 138 may include "bad" and / or "concern." As another example, if the natural language input is "John was exhausted after climbing the mountain," one or more emotional concepts retrieved from the emotional concept database 138 may include strong energy (e.g., John expended a lot of energy) and big (e.g., the mountain is big). The above-mentioned emotional concepts are similar to concepts understood by humans. For example, if a child hits a dog, a parent may yell, "That's bad!" Based on this interaction between the child and the parent, the child may understand that hitting the dog is bad. Similarly, if a child plays with his or her toy with another child, the parent may say, "Good boy," thereby indicating to the child that playing with toys together is a good thing. In this way, the child learns basic concepts such as good, bad, danger, anger, surprise, love, and safety. Similarly, the emotion concept database 138 may store a set of core emotion concepts that the emotion concept subsystem 122 may retrieve in response to natural language input.In some embodiments, when an image (e.g., a depiction of a mountain) is received as input, the emotional concept subsystem 122 may retrieve the emotional concepts associated with the image, such as big, rock, tree, etc., from the emotional concept database 138.

[0097] In addition to obtaining one or more sentiment concepts of the natural language input, other information of the natural language may also be obtained in step 404. The other information of the natural language input may include a subject time decay factor, a subject geographical decay factor, an object time decay factor or an object geographical decay factor, a clause type, a subject of the clause, a subject type of the clause, a modifier of the subject of the clause, a type of modifier of the subject of the clause, a number of subjects of the clause, a subject time decay factor, a subject geographical decay factor, a verb of the clause, a verb tense of the clause, a verb modifier of the clause, an object of the clause, a type of object of the clause, a modifier of the object of the clause, a type of object modifier of the clause, a number of objects of the clause, an object time decay factor, an object geographical decay factor, a preposition of the clause, a modifier of a preposition of the clause, or a global tense modifier of the clause.

[0098] Further, in step 406, one or more growth factors or decay factors associated with one or more emotional attributes of the artificial intelligence entity may be updated over a predetermined period of time based on the emotional concepts of the natural language input and other information of the natural language input. For example, if the natural language input is "John died of cancer," one or more emotional concepts retrieved from the emotional concept database 138 may include "bad" and / or "concern." As a result, growth factors or decay factors associated with one or more emotional attributes (associated with the emotional concepts) of the artificial intelligence entity, such as sadness, anger, and / or happiness, may be updated. In one use case, emotional attribute A in FIG. 2A corresponds to "sadness" of the artificial intelligence entity. When the natural language input "John died of cancer" is obtained, the factor adjustment subsystem 112 may update the growth factor of the emotional attribute "sadness" such that the emotional value (e.g., 207c-207f) of the emotional attribute "sadness" increases based on the updated growth factor (which may be linear or nonlinear). In another use case, emotional attribute B of the artificial intelligence entity in Figure 2B corresponds to "happiness" for the artificial intelligence entity. When a natural language input of "John died of cancer" is obtained, the factor adjustment subsystem 112 may update the decay factor of the emotional attribute "happiness" such that the emotion value (e.g., 208c-208f) of the emotional attribute "happiness" decreases based on the updated decay factor (which may be linear or non-linear).

[0099] Additionally, as discussed above, it should be understood that emotion values ​​for emotion attributes generally revert (or reset) to their respective baseline values ​​in the absence of input and / or after a predetermined period of time. For example, even if emotion value 202 increases from time c to time f in FIG. 2A , it should be understood that there is a threshold amount by which these emotion values ​​can increase before they begin to decrease toward emotion baseline 204. Emotion values ​​201 and 202 will not fall below emotion baselines 205 and 204, respectively, even in the presence of a growth or decay factor. In some embodiments, emotional state subsystem 114 may further update emotion baseline 214 based on one or more inputs (and / or one or more emotion concepts). After updating emotion baseline 214, emotion value 212 of the artificial intelligence entity may be updated based on the updated growth or decay factor(s) and the updated emotion baseline.

[0100] In some embodiments, in addition to updating the boost or decay factors based on the sentiment concepts of the natural language input and other information in the natural language input, the boost factors may be updated based on confidence and / or certainty values. For example, artificial An intelligent entity and one or more other entities (e.g., another entity) artificial The trust value may be determined and / or obtained to indicate a level of trust between the artificial intelligence entity and the other entity (e.g., the client device 104 or any other input source). The trust value may be determined and / or obtained based on the number of interactions between the artificial intelligence entity and the other entity, and / or artificial The factor adjustment subsystem 112 may determine and / or obtain a confidence value associated with an event indicated by the natural language input. The confidence value may be determined based on the content of interactions between the intelligent entity and other entities. The factor adjustment subsystem 112 may modify a boost factor or a damping factor associated with the emotional attribute based on the confidence value. Further, the factor adjustment subsystem 112 may determine and / or obtain a confidence value associated with the event indicated by the natural language input. The confidence value may be determined based on the confidence value associated with the event. artificialThe confidence value may indicate a level of confidence of the intelligent entity. The confidence value may be determined based on whether the event is explicitly described in the natural language input or inferred from the natural language input and / or the confidence value. The factor adjustment subsystem 112 may modify a boost factor or a damping factor associated with the emotional attribute based on the confidence value.

[0101] 5 illustrates a method 500 for updating one or more emotional baselines of an artificial intelligence entity, according to one or more embodiments. At step 502, a set of emotional values ​​of the artificial intelligence entity may be continuously updated over a predetermined period of time based on the one or more emotional baselines. As an example, the emotional baseline may include a baseline that caps the increase of one or more emotional values ​​in the set of emotional values ​​(e.g., regardless of one or more growth or decay factors of the artificial intelligence entity).

[0102] At step 504, input may be acquired over a predetermined period of time. As discussed above, this input may be acquired from a client device 104, or from another artificial The input may be obtained from an intelligent entity and / or from any source within or external to system 100. The input may include natural language input, audio input, image input, video input, or other input. In step 506, a response may be generated related to the input. The response may be generated based on a continuously updated set of emotion values ​​of the artificial intelligence entity. For example, if the natural language input is "John died of cancer," a response generated by the response generation subsystem 118 (e.g., related to this input) may include "That's a shame" or "That's sad. I need time."

[0103] At step 508, the emotional baseline of the artificial intelligence entity may be updated over a predetermined period of time based on the input. In one use case, the emotional attribute C of the artificial intelligence entity in FIG. 2C may correspond to the artificial intelligence entity's "feeling of happiness." If the natural language input "John died of cancer" is received, the emotional state subsystem 114 may update the emotional baseline 214 (corresponding to emotional attribute C) based on such input. While FIG. 2C illustrates the emotional baseline 214 decreasing based on the input, it should be understood that the emotional baseline 214 may also be updated by increasing the emotional baseline 214.

[0104] After updating the emotional baseline (e.g., emotional baseline 214 in FIG. 2C ) in step 510, the set of emotional values ​​may be continuously updated based on the updated emotional baseline for a predetermined period of time. As shown in FIG. 2C , after updating emotional baseline 214, emotional values ​​212 (e.g., emotional values ​​218e and 218f) may be updated based on the updated emotional baseline 214.

[0105] In some embodiments, the various computers and subsystems illustrated in FIG. 1 may include one or more computing devices programmed to perform the functions described herein. The computing device may include one or more electronic storage (e.g., knowledge database 134, boost / decay factor database 136, emotion concept database 138, other databases described above, or other electronic storage), one or more physical processors programmed with one or more computer program instructions, and / or other components. The computing device may include communications lines or ports that enable communication with a network (e.g., network 150) or other computing platforms via wired or wireless technologies (e.g., Ethernet, fiber optic, coaxial cable, WiFi, Bluetooth, near field communication, or other technologies). The computing device may include multiple hardware, software, and / or firmware components working together. For example, a computing device may be implemented by multiple computing platforms working together as a computing device.

[0106] Electronic storage may include non-transitory storage media that electronically store information. Electronic storage media may include one or both of: (i) system storage that is integral to (e.g., substantially non-removable from) a server or client device; or (ii) removable storage that is removably connected to the server or client device, for example, via a port (e.g., a USB port, a Firewire port, etc.) or a drive (e.g., a disk drive, etc.). Electronic storage may include one or more of optically readable storage media (e.g., optical disks, etc.), magnetically readable storage media (e.g., magnetic tape, magnetic hard drives, floppy drives, etc.), charge-based storage media (e.g., EEPROM, RAM, etc.), solid-state storage media (e.g., flash drives, etc.), and / or other electronically readable storage media. Electronic storage may include one or more virtual storage resources (e.g., cloud storage, virtual private networks, and / or other virtual storage resources). The electronic storage may store software algorithms, processor-determined information, information obtained from a server, information obtained from a client device, or other information that enables the functionality described herein.

[0107] A processor may be programmed to implement information processing functions in a computing device. As such, a processor may include one or more of a digital processor, an analog processor, a digital circuit for processing information, an analog circuit for processing information, a state machine, and / or other mechanisms for electronically processing information. In some embodiments, a processor may include multiple processing units. These processing units may be located within the same physical device, or multiple processors may represent the processing functions of multiple devices operating in concert. A processor may be programmed to execute computer program instructions to implement the functionality of subsystems 112-124 or other subsystems described herein. A processor may be programmed to execute computer program instructions by software, hardware, firmware, any combination of software, hardware, or firmware, and / or other mechanisms for configuring processing functions in the processor.

[0108] It should be understood that the descriptions of the functionality provided by the different subsystems 112-124 described herein are for illustrative purposes and not intended to be limiting. Any of the subsystems 112-124 may provide more or less functionality than described. For example, one or more of the subsystems 112-124 may be omitted, and some or all of its functionality may be provided by other of the subsystems 112-124. As another example, additional subsystems may be programmed to perform some or all of the functionality attributed herein to one of the subsystems 112-124.

[0109] While the present invention has been described in detail for illustrative purposes based on what are presently considered to be the most practical and preferred embodiments, it should be understood that such detailed description is for illustrative purposes only. The present invention is not limited to the disclosed embodiments, but rather, it is intended to cover modifications and equivalent arrangements that fall within the scope of the appended claims. For example, it should be understood that the present invention contemplates combining, to the extent possible, one or more features of any embodiment with one or more features of any other embodiment.

[0110] The present technology will be better understood with reference to the following enumerated embodiments. Embodiment 1 A method comprising the steps of updating a set of emotion values ​​of an artificial intelligence entity over a predetermined period of time, obtaining an input over a predetermined period of time, and generating a response related to the input based on the updated set of emotion values ​​of the artificial intelligence entity. Embodiment 2 2. The method of embodiment 1, further comprising: determining one or more gain or decay factors for a set of emotional attributes of the artificial intelligence entity, the set of emotional attributes being associated with a set of emotional values ​​of the artificial intelligence entity; and updating the set of emotional values ​​of the artificial intelligence entity based on the one or more gain or decay factors over a predetermined period of time. Embodiment 3 3. The method of embodiment 2, further comprising updating one or more growth factors or decay factors based on the input at a predetermined time period, wherein updating the set of emotion values ​​after updating the one or more growth factors or decay factors comprises updating the set of emotion values ​​at a predetermined time period based on the updated one or more growth factors or decay factors. Embodiment 4 4. The method of embodiment 3, further comprising obtaining another input after updating the one or more growth or decay factors; and generating a response in relation to the another input after updating the one or more growth or decay factors based on the updated set of emotion values ​​of the artificial intelligence entity. Embodiment 5 5. The method of any of embodiments 3-4, further comprising: updating one or more emotional baselines based on the input, which serve as upper bounds for one or more emotional values ​​in the series of emotional values ​​regardless of one or more growth factors or decay factors, wherein after updating the one or more emotional baselines, updating the series of emotional values ​​comprises updating the series of emotional values ​​of the artificial intelligence entity at a predetermined time period based on the updated one or more growth factors or decay factors and the updated one or more emotional baselines. Embodiment 6 6. The method of embodiment 5, further comprising: obtaining a natural language input from a source; and performing natural language processing of the natural language input to obtain, over a predetermined time period, one or more emotion concepts of the natural language input and other information of the natural language input as input, wherein updating the one or more boost or decay factors comprises updating the one or more boost or decay factors over a predetermined time period based on (i) the one or more emotion concepts of the natural language input and (ii) the other information of the natural language input. Embodiment 7 7. The method of any of embodiments 5-6, wherein the other information of the natural language input indicates a type of clause, a subject of the clause, a type of subject of the clause, a modifier of the subject of the clause, a type of modifier of the subject of the clause, a number of subjects of the clause, a time decay factor of the subject, a geographical decay factor of the subject, a verb of the clause, a tense of the verb of the clause, a modifier of the verb of the clause, an object of the clause, a type of object of the clause, a modifier of the object of the clause, a type of modifier of the object of the clause, a number of objects of the clause, a time decay factor of the object, a geographical decay factor of the object, a preposition of the clause, a modifier of a preposition of the clause, or a global tense modifier of the clause. Embodiment 8 8. The method of any of embodiments 5-7, wherein the other information of the natural language input indicates a subject time decay factor, a subject geographical decay factor, an object time decay factor, or an object geographical decay factor. Embodiment 9 The method of any of embodiments 5-8, further comprising determining a trust value associated with the source, the trust value indicating a level of trust of the artificial intelligence entity in the source, wherein obtaining the input includes obtaining as input (i) one or more emotion concepts of the natural language input, (ii) a trust value associated with the source, and (iii) other information of the natural language input, and updating the one or more boost or decay factors includes updating the one or more boost or decay factors at a predetermined time period based on (i) one or more emotion concepts of the natural language input, (ii) a trust value associated with the source, and (iii) other information of the natural language input. Embodiment 10 determining a certainty value associated with the event indicated by the natural language input, the certainty value being determined based on (i) whether the event is explicitly described in the natural language input or is inferred from the natural language input, and (ii) a confidence value associated with the source, the certainty value being determined based on the confidence value associated with the event. artificial The method of any of embodiments 5-9, further comprising a step of determining a level of certainty of the intelligent entity, wherein the step of obtaining input includes a step of obtaining as input (i) one or more emotion concepts of the natural language input, (ii) a certainty value associated with the event, (iii) a confidence value associated with the source, and (iv) other information of the natural language input, and the step of updating one or more growth or decay factors includes a step of updating the one or more growth or decay factors at a predetermined period of time based on (i) one or more emotion concepts of the natural language input, (ii) a certainty value associated with the event, (iii) a confidence value associated with the source, and (iv) other information of the natural language input. Embodiment 11 The method of any of embodiments 3-10, wherein the step of determining one or more growth factors or decay factors includes determining one or more decay factors for the set of emotional attributes of the artificial intelligence entity, the step of updating the set of emotional values ​​includes updating the set of emotional values ​​of the artificial intelligence entity over a predetermined period of time based on the one or more decay factors, the step of updating the one or more growth factors or decay factors includes updating the one or more decay factors over a predetermined period of time based on the input, and after the step of updating the one or more decay factors, the step of updating the set of emotional values ​​includes updating the set of emotional values ​​of the artificial intelligence entity over a predetermined period of time based on the updated one or more decay factors. Embodiment 12 The method of any of embodiments 3-10, wherein the step of determining one or more growth factors or decay factors includes determining one or more growth factors for the set of emotional attributes of the artificial intelligence entity, the step of updating the set of emotional values ​​includes updating the set of emotional values ​​of the artificial intelligence entity over a predetermined period of time based on the one or more growth factors, the step of updating one or more growth factors or decay factors includes updating the one or more growth factors over a predetermined period of time based on the input, and after the step of updating the one or more growth factors, the step of updating the set of emotional values ​​includes updating the set of emotional values ​​of the artificial intelligence entity over a predetermined period of time based on the updated one or more growth factors. Embodiment 13 13. The method of any of embodiments 1-12, wherein generating a response related to the input includes generating the response based on an updated set of emotion values ​​of the artificial intelligence entity derived from the input. Embodiment 14 processing the input content to determine one or more impact values ​​relating to the impact each portion of the content has on one or more emotional attributes of the artificial intelligence entity; artificialThe method of any of embodiments 1-13, further comprising a step of determining whether one or more impact values ​​meet a predetermined threshold that triggers an increase or decrease in one or more emotional values ​​associated with one or more emotional attributes of the intelligent entity, and a step of generating a modification of one or more emotional values ​​of the artificial intelligence entity for a predetermined period of time based on determining that the one or more impact values ​​meet the predetermined threshold. Embodiment 15 artificial determining whether an interaction threshold between the intelligent entity and at least one other entity has occurred within a given period of time; and based on the determination whether the interaction threshold has been met, artificial 15. The method of any of embodiments 1-14, further comprising generating a modification of the set of emotion values ​​of the intelligent entity. Embodiment 16 16. The method of any of embodiments 1-15, wherein updating the set of emotion values ​​includes continuously updating the set of emotion values ​​of the artificial intelligence entity based on one or more growth or decay factors over a predetermined period of time. Embodiment 17 17. The method of embodiment 16, wherein continuously updating the set of emotion values ​​includes periodically updating the set of emotion values ​​of the artificial intelligence entity based on one or more growth or decay factors over a predetermined period of time. Embodiment 18 A method according to any one of embodiments 1-17, wherein generating a response includes generating a response to the input based on one or more embedding vectors of one or more neural networks associated with the artificial intelligence entity. Embodiment 19 The method of any of embodiments 1-18, wherein the step of updating the set of emotion values, one or more growth factors or decay factors, or one or more emotion baselines is performed based on one or more embedding vectors of one or more neural networks associated with the artificial intelligence entity. Embodiment 20 The method of any of embodiments 1-19, further comprising: processing the input through an embedding network to obtain an input embedding vector representing the input; obtaining a first embedding vector from the graph, the first embedding vector representing one or more emotion values ​​in the updated set of emotion values; and generating a response related to the input based on the input embedding vector and the first embedding vector. Embodiment 21 21. The method of embodiment 20, wherein the graph includes a plurality of nodes, the plurality of nodes including (i) one or more nodes representing emotional attributes or associated emotional values, and (ii) one or more nodes representing emotional concepts or other concepts. Embodiment 22 22. The method of embodiment 21, wherein the plurality of nodes of the graph further includes one or more nodes representing expressed emotional attributes, associated emotional values, contextual information regarding emotional concepts or other concepts. Embodiment 23 23. The method of any of embodiments 21-22, wherein the plurality of nodes of the graph further includes one or more nodes representing a plurality of embedding vectors, each representing another node or subgraph of the graph. Embodiment 24 24. The method of embodiment 23, wherein each embedding vector of the plurality of embedding vectors is directly associated with another node or subgraph that the plurality of embedding vectors represents. Embodiment 25 A method according to any one of embodiments 21-24, wherein the plurality of nodes of the graph further includes one or more nodes representing a plurality of embedding vectors relating to the emotions of one or more entities towards the artificial intelligence entity. Embodiment 26 26. The method of embodiment 25, wherein generating a response related to the input includes generating a response based on (i) the input embedding vector, (ii) the first embedding vector, and (iii) at least one node representing an embedding vector related to the emotion of another entity toward the artificial intelligence entity. Embodiment 27 A tangible, non-transitory machine-readable medium storing instructions that, when executed by a data processing apparatus, cause the data processing apparatus to perform a process including the process of any of embodiments 1-26. Embodiment 28 A system comprising one or more processors and a memory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform a process including any of the processes of embodiments 1-26.

Claims

1. 1. A method for facilitating the realization of artificial intelligence based on emotional states, the method being implemented by a computer system comprising one or more processors executing computer program instructions that, when executed, perform the method, the method comprising: determining one or more boost or damping factors for a plurality of emotional attributes of an artificial intelligence entity, the plurality of emotional attributes corresponding to a plurality of emotional values ​​of the artificial intelligence entity; continuously updating the plurality of emotion values ​​of the artificial intelligence entity based on the one or more growth or decay factors; obtaining an input; generating a response related to the input based on the plurality of continuously updated emotion values ​​of the artificial intelligence entity; updating the one or more boost or decay factors associated with the one or more emotional concepts by applying predetermined rules and logic to the input to obtain one or more emotional concepts; Equipped with continuously updating the plurality of emotion values ​​includes continuously updating the plurality of emotion values ​​based on the updated one or more growth factors or decay factors. method.

2. obtaining another input after updating the one or more growth or decay factors; generating a response related to the further input based on the plurality of continuously updated emotion values ​​of the artificial intelligence entity after updating the one or more growth or decay factors; The method of claim 1 further comprising:

3. updating one or more emotion baselines that define lower bounds for one or more emotion values ​​of the plurality of emotion values ​​based on the input, regardless of the one or more increase or decrease factors; continuously updating the plurality of emotion values ​​comprises continuously updating the plurality of emotion values ​​of the artificial intelligence entity based on updated one or more growth factors or decay factors and updated one or more emotion baselines. The method of claim 1.

4. The method described in claim 1, wherein the step of generating a response to the input includes a step of generating the response based on the plurality of continuously updated emotional values ​​of the artificial intelligence entity derived from the input.

Images