Interaction method, device and equipment of intelligent agent and medium

Abstract

This application relates to the field of artificial intelligence technology and discloses an interaction method, device, equipment, and medium for intelligent agents. The interaction method for intelligent agents provided in this application achieves structured storage and fuzzy matching by constructing a memory index matrix that conforms to human cognitive thinking, avoiding the cognitive disconnect problem caused by mechanical classification in traditional technologies. At the same time, it uses a feature code mapping mechanism to reduce the traversal of all data, thereby improving retrieval efficiency and enhancing fuzzy association capabilities.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an interaction method, device, equipment, and medium for intelligent agents. Background Technology

[0002] In related technologies, the memory systems of intelligent agents are disconnected from human cognitive thinking. These memory systems often employ technology-driven, mechanical classification methods, resulting in fragmented and isolated memory information of the same type, high data redundancy, and a lack of hierarchical relationships. This makes it difficult for intelligent agents to effectively establish logical connections between memories during anthropomorphic interactive tasks, leading to responses that lack intuition and contextual relevance. Summary of the Invention

[0003] This application provides an interaction method, apparatus, device, and medium for intelligent agents. It aims to address the problems of low retrieval efficiency in existing intelligent agent memory indexing systems due to the disconnect between classification logic and human cognitive thinking.

[0004] In a first aspect, this application provides an interaction method for an intelligent agent, the method comprising: constructing a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories; receiving interaction information and storing the interaction information based on the cognitive categories and semantic feature categories to obtain a storage memory, wherein the interaction information includes statement information used to train the intelligent agent; receiving a query instruction and matching the query instruction with the statement information in the storage memory to obtain a query result.

[0005] Based on the aforementioned technical means, this application achieves structured management and anthropomorphic mapping of an agent's memory by constructing a memory index matrix based on cognitive categories and semantic feature categories. This improves the agent's memory retrieval efficiency and association accuracy in complex interaction scenarios, enabling it to provide intuitive responses.

[0006] In one optional implementation, receiving interaction information and storing the interaction information based on cognitive category and semantic feature category to obtain a storage memory includes: semantically decomposing the statement information to obtain multiple semantic feature sub-codes, wherein the semantic feature sub-codes include multiple dimensions, which are dimensions corresponding to the row and column dimensions of the memory index matrix; combining the feature sub-codes to obtain feature codes corresponding to the interaction information; and binding the feature codes with storage addresses and storing them to obtain the storage memory.

[0007] Based on the aforementioned technical means, relevant memory content can be quickly and efficiently located and accessed through feature codes, improving the organization efficiency, retrieval efficiency, and accuracy of the memory bank, thereby enhancing the cognitive and interactive capabilities of the intelligent agent.

[0008] In one optional implementation, the feature code is bound to a storage address and then stored to obtain a storage memory, including: obtaining the dimension information of the feature code, wherein the dimension information of the feature code is used to characterize the dimension information of the feature subcodes within the feature code; mapping the dimension information to the row dimension and column dimension in the memory index matrix to obtain the coordinate information of the feature code in the memory index matrix; and binding the coordinate information to a storage address and then storing it.

[0009] Based on the above technical means, the targeting and accuracy of memory retrieval are improved, avoiding the high computational overhead and low efficiency problems caused by traditional full traversal or fuzzy matching. This enables the intelligent agent to retrieve relevant memories more quickly and accurately, thereby achieving a more human-like and smoother interactive response.

[0010] In one optional implementation, the row dimension includes at least one of the following: emotional experience, skill mastery, personality traits, living habits, interpersonal relationships, social rules, social scenarios, career cognition, natural phenomena, environmental interaction, transaction attributes, behavioral logic, value cognition, interest preferences, and special experiences; the column dimension includes at least one of the following: time characteristics, location characteristics, related object characteristics, event content characteristics, emotional tendency characteristics, scenario attribute characteristics, behavioral characteristics, and behavioral outcome characteristics.

[0011] Based on the aforementioned technical means, intelligent agents can transform fragmented interactive information into structured memories with high relevance and rich context. When the intelligent agent needs to retrieve or associate information, it can accurately locate information based on specific cognitive dimensions and semantic features, thereby achieving fuzzy association and cross-domain association capabilities that are closer to human thinking, and improving the intuitiveness and anthropomorphism of the intelligent agent's response.

[0012] In one optional implementation, receiving a query instruction and matching the query instruction with statement information in a storage memory to obtain query results includes: parsing the query instruction to determine the row dimension corresponding to the query instruction; decomposing the query instruction to obtain at least one query feature subcode; performing fuzzy matching on the query feature subcode and semantic feature subcode within the row dimension to obtain the semantic matching degree; when the semantic matching degree is greater than a preset matching degree threshold, determining the statement information corresponding to the semantic feature subcode, and generating query results based on the statement information.

[0013] Based on the aforementioned technical means, when an intelligent agent needs to retrieve or associate information, it can accurately locate information based on specific cognitive dimensions and semantic features, thereby achieving fuzzy association and cross-domain correlation capabilities that are closer to human thinking, and improving the intuitiveness and anthropomorphism of the intelligent agent's response.

[0014] In one optional implementation, generating query results based on statement information includes: restructuring the statement information or adjusting the tone according to the query intent of the query instruction.

[0015] Based on the aforementioned technical means, the naturalness and friendliness of the interaction are improved, enabling the intelligent agent to provide responses that are closer to human thinking and emotions when interacting with humans, thereby enhancing the user experience.

[0016] In an optional implementation, the method further includes: continuously receiving interaction information, and performing the steps of receiving interaction information, storing the interaction information based on cognitive category and semantic feature category to obtain a storage memory; and adjusting a preset matching degree threshold according to the interaction scenario to enhance query accuracy.

[0017] Based on the above technical means, the anthropomorphism and response quality of the interaction are improved, enabling the intelligent agent's memory retrieval system to have adaptive optimization capabilities for the interactive environment.

[0018] Secondly, this application provides an interactive device for an intelligent agent, the device comprising: a construction module for constructing a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories; a storage module for receiving interactive information and storing the interactive information based on the cognitive categories and semantic feature categories to obtain a storage memory; and a query module for receiving query instructions and matching the query instructions with the statement information in the storage memory to obtain query results.

[0019] Thirdly, this application provides an electronic device, including: a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the interaction method of the intelligent agent described in the first aspect or any corresponding embodiment.

[0020] Fourthly, this application provides a computer-readable storage medium storing computer instructions for causing a computer to execute the interaction method of the intelligent agent described in the first aspect or any corresponding embodiment.

[0021] Fifthly, this application provides a computer program product, including computer instructions for causing a computer to execute the interaction method of an intelligent agent according to the first aspect or any corresponding embodiment. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a schematic flowchart of a first type of interaction method for an intelligent agent according to an embodiment of this application; Figure 2 This is a second flowchart illustrating the interaction method of an intelligent agent according to an embodiment of this application; Figure 3 This is a schematic diagram of a memory index matrix according to an embodiment of this application; Figure 4 This is a schematic diagram of the third process of the interaction method of the intelligent agent according to the embodiments of this application; Figure 5 This is a structural block diagram of an interactive device for an intelligent agent according to an embodiment of this application; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0026] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0027] In related technologies, the memory system of intelligent agents is disconnected from human cognitive thinking in terms of indexing and classification logic. This leads to the agent's inability to accurately associate memories when handling cross-domain, anthropomorphic interactive tasks, resulting in a lack of intuitive responses. Simultaneously, retrieval efficiency is low, computational power consumption is high, and traditional retrieval models have low error tolerance and struggle to simulate human fuzzy associations. Memory data is severely fragmented, with weak correlations, high redundancy, and a lack of a unified structured system, making it difficult for intelligent agents to achieve anthropomorphic fuzzy association capabilities, resulting in stiff interactive responses.

[0028] To address this, this application proposes an interaction method for intelligent agents. For example... Figure 1 The methods include: Step S101: Construct a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories.

[0029] The memory index matrix, constructed as a two-dimensional structure, serves to organize and index the various types of information received and stored by the agent. The rows and columns of this matrix correspond to different information classification dimensions, thus providing a foundation for rapid information retrieval and association.

[0030] Cognitive categories are defined as the multiple classifications contained in the row dimensions of the memory index matrix. These categories are designed to simulate the human thought process of understanding and generalizing the world; for example, they can include emotions, skills, habits, relationships, etc., and are used to categorize interactive information at a macro level.

[0031] Semantic feature categories are defined as multiple classifications contained in the column dimensions of the memory index matrix. These categories are used to extract key features from the deep meaning of language or text, such as time, place, event, sentiment, etc., for analyzing and categorizing interactive information at a micro level.

[0032] Specifically, the memory index matrix is ​​designed as a two-dimensional structure, with rows representing multiple cognitive categories and columns representing multiple semantic feature categories. For example, in one implementation, these cognitive and semantic feature categories can be predefined by the system designer and preset into the agent system as a list. Specifically, a series of labels representing different cognitive domains can be manually entered as row dimensions, such as "personal experience," "professional knowledge," and "social rules"; simultaneously, a series of labels representing different semantic attributes can be manually entered as column dimensions, such as "time," "location," "person," and "event." This forms a preset framework for information classification and indexing. The construction of this matrix lays the foundation for the subsequent structured storage of information.

[0033] Step S102: Receive interaction information and store the interaction information based on cognitive category and semantic feature category to obtain a storage memory bank, wherein the interaction information includes statement information used to train the agent.

[0034] Interaction information refers to all data received by the agent during its interaction with the external environment or user. This information forms the basis for the agent's learning and memory, and is usually presented in the form of utterances to train the agent to understand and respond to user intentions.

[0035] A storage memory refers to the complete set of interactive information stored by an intelligent agent system after structured processing. The information in this memory is organized within a memory index matrix to facilitate subsequent retrieval, matching, and utilization.

[0036] Specifically, the interactive information received by the agent is usually presented in the form of statements, such as user-input text dialogues or speech transcriptions, primarily for training the agent. Upon receiving the interactive information, the system stores this information based on predefined cognitive and semantic feature categories, thus creating a storage memory. One implementation method involves extracting keywords from the received statements and then directly matching the extracted keywords with predefined cognitive and semantic feature categories in the memory index matrix. For example, if a statement contains keywords such as "yesterday" and "meeting," it might be categorized into the storage area where the "time" semantic feature category and the "work cognition" cognitive category intersect. Subsequently, the statement is directly stored in the corresponding storage address within that area. In this way, the interactive information is initially structured and stored in the storage memory.

[0037] Step S103: Receive the query command and match the query command with the statement information in the storage memory to obtain the query result.

[0038] A query instruction is a request issued by a user or system to an agent to obtain specific information. This instruction typically includes the user's intent and the query content, which the agent needs to parse and search for matching information in its memory to generate a response.

[0039] Statement information, a specific form of interactive information, is typically expressed as natural language text. This statement information is the core content for intelligent agents to perform semantic understanding, memory storage, and query matching.

[0040] Specifically, after receiving a query command, the agent matches it against the stored information in its memory to obtain the query results. For example, when a user enters a query command such as "about yesterday's meeting," the system can parse the query command and extract keywords such as "yesterday" and "meeting." The system then iterates through the stored information, searching for statements containing the matching keywords. Once a matching statement is found, it is output as the query result. Through this matching mechanism, the agent can respond to user queries and provide the corresponding stored content.

[0041] It is understood that this application achieves structured management and anthropomorphic mapping of agent memory by constructing a memory index matrix based on cognitive categories and semantic feature categories. This solves the problems of existing systems lacking intuitiveness and accuracy in handling cross-domain, anthropomorphic interaction tasks, due to disconnected classification logic, low retrieval efficiency, insufficient structure of memory data, and a lack of anthropomorphic fuzzy association capabilities. It improves the agent's memory retrieval efficiency and association accuracy in complex interaction scenarios, enabling it to provide intuitive responses.

[0042] In some of the solutions mentioned above in this application, a memory index matrix is ​​proposed to be constructed and interactive information is stored based on cognitive category and semantic feature category in order to realize the memory management of the agent. However, in this process, there is still a lack of specific implementation paths for how to transform unstructured sentence information into feature codes that can be accurately mapped to the matrix dimension, and how to realize the structured storage of memory through efficient encoding methods. This makes it difficult to directly transform the original sentence information into searchable index data, which limits the construction efficiency of the memory bank and the accuracy of subsequent retrieval.

[0043] In response, this application further proposes an interaction method for intelligent agents, such as... Figure 2 As shown, the method includes: Step S201: Construct a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories. See details... Figure 1 Step S101 in the embodiment will not be described again here.

[0044] Step S202: Receive interaction information and store the interaction information based on cognitive category and semantic feature category to obtain a storage memory bank. The interaction information includes statement information used to train the agent. Specifically, this includes: Step S2021: Semantically decompose the statement information to obtain multiple semantic feature subcodes. The semantic feature subcodes include multiple dimensions, which are dimensions corresponding to the row and column dimensions of the memory index matrix. The semantic feature subcodes include hash subcodes.

[0045] Specifically, this step aims to transform unstructured natural language statements into structured, quantifiable semantic units. For example, Natural Language Processing (NLP) techniques, such as dependency parsing, named entity recognition, and sentiment analysis, can be used to extract elements like subject, predicate, object, time, location, and sentiment from the statement. Each extracted element or its attribute can serve as a semantic feature sub-code. Alternatively, a pre-trained language model can be used to encode the statement, and then clustering or dimensionality reduction techniques can be used to map the statement's semantic vectors to a predefined semantic feature space, thereby obtaining semantic feature sub-codes representing different semantic dimensions. These semantic feature sub-codes include multiple dimensions, and these dimensions correspond to the row and column dimensions of the memory index matrix. This ensures that the semantic decomposition results match the structure of the predefined memory index matrix, providing a unified coordinate system for subsequent storage and retrieval. For example, when designing the semantic decomposition algorithm, the decomposed semantic elements can be directly categorized into the row dimensions of the memory index matrix, such as cognitive categories, and the column dimensions, such as semantic feature categories. Furthermore, semantic feature subcodes include hash subcodes, which are used to hash the semantic feature subcodes to achieve data compression, encryption, and fast lookup, while also providing a basis for fuzzy matching. For example, a hash function can be applied to each semantic feature subcode to generate a fixed-length hash value as the hash subcode; or Locality Sensitive Hashing (LSH) technology can be used so that semantically similar feature subcodes still have a high probability of producing the same or similar hash values ​​after hashing.

[0046] Step S2022: Combine the feature subcodes to obtain the feature code corresponding to the interaction information; wherein, the feature code includes a fuzzy hash short code.

[0047] Specifically, this step aims to integrate multiple independent semantic feature subcodes into a unified feature code that comprehensively represents interactive information, facilitating overall storage and retrieval. For example, all semantic feature subcodes, or their hash subcodes, can be concatenated, XORed, or weighted summed according to a preset order or weight to form a longer feature code. This feature code includes fuzzy hash short codes, which can tolerate minor changes in input data to a certain extent, allowing relevant memories to be found through fuzzy matching even if the queried information is not completely consistent with the stored information, simulating the fuzzy association ability of humans. For example, when generating the feature code, fuzzy hashing algorithms such as perceptual hashing can be used, which can generate similar hash values ​​based on content similarity.

[0048] Step S2023: After binding the feature code with the storage address, store it to obtain the storage memory.

[0049] Specifically, this step aims to establish a mapping between signature codes and actual storage locations, enabling agents to quickly locate and access the stored original interaction information using the signature codes. For example, a key-value storage system can be used, storing the generated signature code as the key and the address where the original interaction information is stored as the value. Alternatively, a distributed file system or database can be used to assign a unique storage address to each piece of interaction information, storing that address along with the corresponding signature code in an index table.

[0050] Step S203: Receive the query command and match it with the statement information in the storage memory to obtain the query results. See details... Figure 1 Step S103 in the embodiment will not be described again here.

[0051] In one example, taking the memory event of a user actively rescuing a stray cat during a rainstorm, captured by an intelligent agent, semantic parsing and encoding are performed. First, the event is semantically decomposed, extracting core semantic elements such as time (rainy day), location (outdoors), associated object (stray cat), event content (actively rescuing small animals), emotional tendency (kindness, compassion), and scene attributes (outdoor scene in extreme natural weather). Then, each core element is used to generate an independent hash subcode, which is then concatenated and combined into a fixed-length fuzzy hash short code. Finally, this fuzzy hash short code is bound to the physical storage address of the complete event data to build a fast retrieval link.

[0052] It is understandable that, through the above technical solution, the intelligent agent can semantically decompose sentence information and obtain multiple semantic feature sub-codes corresponding to the row and column dimensions of the memory index matrix, thus realizing the transformation of unstructured natural language information into structured, quantifiable semantic units. This allows interactive information to be accurately mapped to the intelligent agent's pre-defined cognitive and semantic space, laying the foundation for subsequent structured storage. Simultaneously, the introduction of hash sub-codes and fuzzy hash short codes not only effectively compresses the data scale and improves processing efficiency, but more importantly, through the characteristics of fuzzy hashing, the generated feature codes can tolerate a certain degree of semantic difference. This provides a data foundation for the intelligent agent to achieve human-like fuzzy associative retrieval, overcoming the limitations of traditional exact matching retrieval, such as low error tolerance and inability to simulate human memory retrieval characteristics. Finally, the generated feature codes are bound to storage addresses for storage, establishing a direct mapping relationship from high-dimensional semantic features to physical storage locations. This enables intelligent agents to quickly and efficiently locate and access relevant memory content through feature codes when faced with massive amounts of interactive information, improving the organization efficiency, retrieval efficiency, and accuracy of the memory bank, thereby enhancing the cognitive and interactive capabilities of the intelligent agent.

[0053] In some of the solutions mentioned above in this application, a storage memory bank is constructed by binding the feature code with the storage address. However, in this process, simply binding the feature code with the storage address lacks a mechanism that can effectively associate the abstract semantic feature code with the structured space of the memory index matrix. This results in the agent not knowing the specific location of the memory in the multidimensional cognitive space when storing the memory, making it difficult to quickly locate the memory using the structured characteristics of the matrix during subsequent retrieval. This leads to the blindness of memory storage and the bottleneck of retrieval efficiency.

[0054] In response, this application further proposes to bind the feature code with the storage address before storing it to obtain a storage memory, including: Step a1: Obtain the dimensional information of the feature code, wherein the dimensional information of the feature code is used to characterize the dimensional information of the feature subcodes within the feature code.

[0055] Step a2: Map the dimensional information to the row and column dimensions in the memory index matrix to obtain the coordinate information of the feature code in the memory index matrix.

[0056] Step a3: Bind the coordinate information to the storage address and then store it.

[0057] Specifically, when acquiring the dimensional information of a feature code, this dimensional information refers to the data extracted from the feature sub-codes obtained after semantic decomposition of the interaction information. This data indicates the attribution or classification attributes of the semantic content represented by the feature sub-code within a predefined cognitive framework, i.e., the row and column dimensions of the memory index matrix. This dimensional information is metadata carried by the feature sub-codes within the feature code, corresponding to the structure of the memory index matrix. It serves as a crucial bridge for establishing the connection between semantic features and structured storage space. One implementation approach is to pre-design the structure of the feature code during the generation of the feature code, such as a fuzzy hash short code, so that specific segments or sub-codes directly encode their respective cognitive and semantic feature categories. When acquiring the dimensional information, these pre-defined segments can be directly parsed. Another implementation approach is to use a predefined mapping table or mapping function. After the feature code is generated, the system can query a pre-established mapping table based on the overall structure of the feature code or specific feature sub-codes within it. This mapping table records the dimensional information corresponding to different combinations of feature codes or feature sub-codes.

[0058] Mapping dimensional information to the row and column dimensions of the memory index matrix yields the coordinates of the feature code within the matrix. This step aims to transform the abstract dimensional information extracted from the feature code into concrete, addressable two-dimensional coordinates, such as row and column numbers, within the memory index matrix. Through this mapping, the feature code corresponding to each interaction can obtain a definite "location" in the multidimensional cognitive space, thus achieving structured and spatialized storage of memory. One implementation is that if the dimensional information itself is a row / column index value of the memory index matrix—for example, if the cognitive category is encoded as an integer from 0 to N, and the semantic feature category is encoded as an integer from 0 to M—then these integers can be directly used as coordinate information. For example, if the dimensional information is cognitive category ID and semantic feature category ID, it can be directly mapped to matrix coordinates (row ID, column ID). Another implementation is that when the dimensional information is not a direct index value, it can be mapped through semantic matching. For example, the dimensional information might be descriptive labels; the system semantically compares these labels with the row and column dimensions of the memory index matrix to find the most matching row and column, thereby determining the coordinate information.

[0059] Finally, the coordinate information is bound to the storage address before storage. This step establishes a connection between the logical "memory location," i.e., the coordinate information, and the physical "data storage location," i.e., the storage address. Through this binding, when an agent needs to retrieve a specific memory, it can first deduce the coordinate information based on its semantic features, and then directly locate the physical address where the memory is stored using the coordinate information, thereby achieving efficient and accurate memory retrieval. One implementation method is to maintain a hash table or dictionary, where the key is the coordinate information, for example, a tuple or string consisting of row and column numbers, and the value is the corresponding storage address. When storage is needed, the coordinate information is used as the key, and the storage address is used as the value in the hash table. Another implementation method is, in a database system, to use the coordinate information as the primary key or create an index, and the storage address as a field of the corresponding record. When storing data, the coordinate information and the storage address are written to the database together.

[0060] In one example, the semantic attribution dimension of the memory event "a user actively rescues a stray cat during a rainstorm" is analyzed. This event simultaneously belongs to three cognitive categories: emotional experience, natural phenomenon, and environmental interaction. The hash code and coordinate information corresponding to the event are synchronously mapped to the intersection of the three cognitive rows and corresponding multi-semantic feature columns in the memory index matrix, completing the multi-dimensional associated storage of a single memory, which is in line with the human memory characteristic that a single event can be associated with multiple cognitions.

[0061] It is understood that, through the above technical solution, this application, based on binding feature codes to storage addresses to construct a storage memory bank, further introduces a dimension mapping mechanism. Specifically, by obtaining the dimensional information of the feature codes, the semantic attributes of the feature sub-codes within the feature codes can be accurately characterized, and these are mapped to the row and column dimensions of the memory index matrix, thereby generating unique coordinate information for each feature code in a multidimensional cognitive space. Subsequently, this coordinate information is bound to the actual storage address and stored. This solves the problem of lack of spatial positioning in memory storage. It establishes a direct association between the originally abstract semantic feature codes and the structured memory index matrix, enabling each piece of interaction information to find its precise logical location on the agent's "cognitive map" during storage. This structured storage method not only avoids the blindness of memory storage, but more importantly, it provides an efficient path for subsequent memory retrieval. When an agent receives a query command, it can first parse the query command and deduce its corresponding coordinate information. Then, it can directly use the coordinate information to quickly locate the storage address, thereby improving the targeting and accuracy of memory retrieval. This avoids the high computational overhead and low efficiency problems caused by traditional full traversal or fuzzy matching, enabling the agent to retrieve relevant memories more quickly and accurately, thus achieving a more human-like and smoother interactive response.

[0062] In some of the solutions mentioned above in this application, a memory index matrix is ​​proposed to be constructed for the structured storage of interactive information. However, in practical applications, if there is a lack of clear definition of the connotation of the matrix dimension, the intelligent agent will lack a unified cognitive standard when classifying memories. This will result in the stored memory data still having semantic ambiguity or inaccurate classification when it is associated across domains, making it difficult to achieve anthropomorphic deep cognition.

[0063] In this regard, this application further clarifies the dimensional definition of the memory index matrix, such as... Figure 3 As shown, the row dimension includes at least one of the following: emotional experience, skill mastery, personality traits, living habits, interpersonal relationships, social rules, social scenarios, career cognition, natural phenomena, environmental interaction, transaction attributes, behavioral logic, value cognition, interest preferences, and special experiences; the column dimension includes at least one of the following: time characteristics, location characteristics, related object characteristics, event content characteristics, emotional tendency characteristics, scenario attribute characteristics, behavioral characteristics, and behavioral outcome characteristics.

[0064] Specifically, row dimensions represent the macroscopic cognitive categories by which an agent classifies and categorizes information. They aim to provide an agent with a high-level, anthropomorphic cognitive framework, enabling information to be classified into specific psychological, social, or knowledge domains. For example, core concepts in human cognitive psychology, such as emotional experiences and personality traits, can be predefined as row dimensions through expert systems or manual annotation. Alternatively, unsupervised learning methods, such as clustering algorithms, can be used to analyze large amounts of interaction data to automatically discover and abstract representative cognitive categories as row dimensions.

[0065] Meanwhile, column dimensions represent semantic feature categories that provide fine-grained descriptions and contextual constraints for information. Their role is to provide specific contextual information and detailed descriptions for remembered data, thereby enhancing the structure and retrieval accuracy of the remembered data. For example, Natural Language Processing (NLP) techniques can be used to extract entity information such as time features, location features, and associated object features from received interactive information and use these as column dimensions. Alternatively, sentiment analysis models can be used to identify sentiment characteristics in interactive information, or event extraction techniques can be used to identify event content features and behavioral outcome features, and these can be used as column dimensions.

[0066] It is understood that, through the aforementioned technical solution, this application constructs a classification system that aligns with human cognitive thinking by refining the row and column dimensions of the memory index matrix. The pre-defined row dimensions, such as emotional experiences and personality traits, provide the intelligent agent with a unified and multi-dimensional cognitive standard, enabling it to summarize information from different perspectives, just like humans, when storing and classifying interactive information, effectively avoiding arbitrariness and inconsistency in classification. The introduction of column dimensions, such as time features and emotional tendency features, ensures that each piece of memory information not only has a macroscopic cognitive classification but also a detailed contextual semantic description. This cross-definition of row and column dimensions enhances the structure and semantic clarity of memory data, reduces ambiguity during information storage, and ensures the accuracy of classification. Ultimately, this allows the intelligent agent to transform fragmented interactive information into structured memories with high relevance and rich context. When an intelligent agent needs to retrieve or associate information, it can accurately locate information based on specific cognitive dimensions and semantic features, thereby achieving fuzzy association and cross-domain correlation capabilities that are closer to human thinking, and improving the intuitiveness and anthropomorphism of the intelligent agent's response.

[0067] In some of the embodiments described above in this application, it is proposed to construct a memory index matrix and store interactive information based on cognitive category and semantic feature category to realize intelligent agent memory management. However, when querying massive interactive information, if only full matching or simple exact matching is performed, it will not only lead to a large consumption of computing resources, but also make it difficult to simulate the characteristic of human memory to make fuzzy associations based on specific contexts, resulting in insufficient accuracy and anthropomorphism of the query results.

[0068] In response, this application further proposes an interaction method for intelligent agents, such as... Figure 4 As shown, the method includes: Step S401: Construct a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories. See details... Figure 1 Step S101 in the embodiment will not be described again here.

[0069] Step S402: Receive interaction information and store the interaction information based on cognitive category and semantic feature category to obtain a storage memory bank. The interaction information includes statement information used to train the agent. Specifically, this includes: Step S4021 involves semantically decomposing the statement information to obtain multiple semantic feature sub-codes. These semantic feature sub-codes include multiple dimensions, which correspond to the row and column dimensions of the memory index matrix. The semantic feature sub-codes also include hash sub-codes. (See details...) Figure 2 Step S2021 in the embodiment will not be repeated here.

[0070] Step S4022: Combine the feature subcodes to obtain the feature code corresponding to the interaction information; wherein, the feature code includes a fuzzy hash short code. See details. Figure 2 Step S2022 in the embodiment will not be repeated here.

[0071] Step S4023 involves binding the feature code to the storage address and then storing it to obtain the storage memory. See details in the reference section. Figure 2 Step S2023 in the embodiment will not be repeated here.

[0072] Step S403: Receive a query command and match the query command with the statement information stored in the memory to obtain the query results. Specifically, this includes: Step S4031: Parse the query command and determine the row dimension corresponding to the query command.

[0073] Step S4032: Deconstruct the query instruction to obtain at least one query feature subcode; the query feature subcode includes a hash subcode.

[0074] Step S4033: Within the row dimension, perform fuzzy matching on the query feature subcode and the semantic feature subcode to obtain the semantic matching degree.

[0075] Step S4034: When the semantic matching degree is greater than the preset matching degree threshold, determine the statement information corresponding to the semantic feature subcode, and generate query results based on the statement information.

[0076] Specifically, parsing query commands and determining their corresponding row dimensions aims to map the user's query intent to a predefined cognitive category space within the memory index matrix, thereby effectively narrowing down the scope of subsequent searches. For example, Natural Language Processing (NLP) techniques can be used to perform semantic analysis on query commands, identifying keywords, phrases, or sentence structures, and matching them with predefined cognitive categories, i.e., row dimensions. Alternatively, a pre-trained text classification model can be used to categorize the overall semantics of the query command into the most relevant row dimension. Furthermore, combining historical user interaction data and the user's past query category preferences can help determine the row dimension to which the current query command belongs.

[0077] Deconstructing a query instruction to obtain at least one query feature sub-code, including a hash sub-code, is a method to transform the query instruction in natural language form into a structured feature representation that can be efficiently processed by a computer. For example, the query instruction can be segmented into words, and then a hash function can be applied to each word or phrase to generate a corresponding hash sub-code. Alternatively, an N-gram model can be used to extract N-gram features from the query instruction and then hash-encode these features. Another approach is to use a deep learning model to convert the query instruction into a semantic vector, and then hash a specific dimension or combination thereof of this semantic vector to obtain a hash sub-code.

[0078] Within the row dimension, fuzzy matching is performed on query feature subcodes and semantic feature subcodes to obtain the semantic matching degree, aiming to simulate the associative rather than absolute consistency-based thinking pattern in human memory retrieval. For example, the Hamming distance algorithm can be used to calculate the similarity between the query feature subcode and the semantic feature subcode stored in the memory bank, and this can be used as the semantic matching degree. Alternatively, Locality Sensitive Hash (LSH) technology can be used to map similar hash subcodes to the same bucket, achieving efficient fuzzy matching through in-bucket comparison. In addition, a semantic vector space model can be combined to calculate the cosine similarity between the semantic vector corresponding to the query feature subcode and the semantic vector corresponding to the stored semantic feature subcode, in order to more accurately measure the semantic matching degree.

[0079] When the semantic matching degree exceeds a preset matching degree threshold, determining the sentence information corresponding to the semantic feature subcode and generating query results based on this information ensures the quality and relevance of the retrieval results while filtering out low-relevance noise. For example, when the matching degree reaches the preset threshold, the system can directly return the original sentence information bound to the semantic feature subcode as the query result. Alternatively, when multiple sentences are matched, they can be sorted according to their matching degree, and the sentences with the highest matching degree can be selected for display. Furthermore, based on the query intent of the query command, the matched sentence information can be reorganized, summarized, or its tone adjusted to generate more natural and context-appropriate query results.

[0080] In one example, when a user asks, "Do you think I'm a compassionate person?", the agent first analyzes the query intent, determining that it belongs to two cognitive categories: emotional experience and value perception. It then narrows down the search by locking onto two rows of data in the memory index matrix. Next, it extracts the feature hash codes corresponding to the query, including compassion, kindness, and positive behavior, and performs a fuzzy match with the hash codes of all memories in that row. Previously, the memory code for "rescuing stray cats during a rainstorm" highly overlapped with the query code and was determined to be a valid associated memory. Finally, the agent combines this positive emotional experience to generate a response that aligns with human cognition, accurately addressing the user's question intent.

[0081] In another example, when a user queries: "What's a good thing to do on a rainy day?", the agent first analyzes the query intent, determining that it belongs to one of two cognitive categories: natural phenomena and environmental interaction, and then locks the corresponding category row in the memory index matrix. It then extracts query features such as rainstorms, outdoor activities, and scene-related behaviors, and retrieves relevant memories of "rescuing stray cats on a rainy day" through fuzzy matching. Building upon general rainy day protection suggestions, the agent further expands the content by incorporating associated memories, providing personalized, anthropomorphic responses such as "You're likely to encounter stranded stray animals outdoors during a rainstorm; it's appropriate to offer assistance." This achieves a cross-dimensional associative effect that is difficult to achieve with traditional memory architectures.

[0082] It is understood that, through the above technical solutions, this application can improve the retrieval efficiency and anthropomorphic interaction level of intelligent agents in complex memory banks. First, by parsing the query instruction and determining the corresponding row dimension, the query intent is mapped to a specific cognitive category space, effectively narrowing the retrieval scope and avoiding an indiscriminate full traversal of the entire memory bank, thereby significantly reducing computational load and improving retrieval efficiency. Second, by decomposing the query instruction to obtain query feature subcodes containing hash subcodes, the query instruction in natural language form is transformed into a feature representation that can be efficiently processed by a computer, laying the foundation for subsequent rapid comparison. Furthermore, within the determined row dimension, fuzzy matching is performed between the query feature subcode and the stored semantic feature subcode. This process simulates the associative rather than absolute consistency-based thinking mode in human memory retrieval, enabling the intelligent agent to extract memory fragments from the memory bank that are highly relevant to the current context but not entirely consistent in expression, enhancing the flexibility and anthropomorphism of the interaction. Finally, by setting a preset matching threshold, the reliability of the search results is ensured, low-relevance noise information is effectively filtered out, and the quality of the output content is guaranteed. This enables the intelligent agent to respond in a way that is more in line with human cognitive logic based on the matched memory content, thereby solving the problems of high computational resource consumption, difficulty in simulating human fuzzy association, and insufficient accuracy and anthropomorphism of the search results when querying massive interactive information.

[0083] In some of the solutions mentioned above in this application, query results are obtained by matching query commands with the storage memory. However, in this process, the intelligent agent can only perform mechanical memory retrieval and cannot dynamically adjust the extracted sentence information according to the user's real-time interactive intent. This results in a monotonous language style and a stiff tone, which is difficult to meet the human needs for anthropomorphic interaction in different social scenarios.

[0084] In response, this application further proposes generating query results based on statement information, including: Step b1: Based on the query intent of the query command, reorganize the statement information or adjust the tone.

[0085] The "query intent" of a query refers to the underlying purpose or expectation expressed by the user through the query. Various technical means can be employed to accurately identify this intent. For example, Natural Language Processing (NLP) techniques can be used to perform semantic analysis on the query, including but not limited to intent classification, entity recognition, and sentiment analysis, to infer whether the user wants to obtain information, express emotions, seek advice, or perform other actions. Furthermore, contextual information, such as previous conversation history, user profiles, and the current interaction scenario, can be combined to make a more accurate judgment of the query intent. For instance, if a user makes a query after expressing a certain emotion, their intent may lean towards seeking comfort or a solution.

[0086] "Sentence restructuring" refers to adjusting the structure of a query after obtaining information that matches the query, rather than directly outputting the original sentence, to better suit the query intent and the current context. This can be achieved in several ways. One method is syntactic transformation, such as changing the active or passive voice of the original sentence, or adjusting the word order to better suit spoken language or specific emphasis needs. Another method is word substitution or paraphrasing, which involves replacing some words or phrases while maintaining the original meaning to change the style, tone, or clarity of the sentence, such as from formal to informal, or vice versa. Information can also be summarized or elaborated based on the query intent; for example, it can be simplified when the user needs a quick overview, or expanded when the user needs in-depth discussion.

[0087] "Pitch adjustment" is mainly aimed at scenarios where intelligent agents output query results in the form of speech. It aims to convey specific emotions, attitudes or emphasis by changing the prosodic features of speech such as pitch, speed, volume and timbre.

[0088] It is understandable that, through the above technical solution, this application achieves a leap from simple memory retrieval to intelligent interactive expression. After the intelligent agent obtains the matching sentence information, the system no longer directly outputs the raw data, but instead, through in-depth analysis of the query intent contained in the query command, it performs targeted reorganization or tone adjustment of the sentence information. This deep integration of static memory content with dynamic interactive context allows the intelligent agent to flexibly change its expression based on the user's current psychological state, social needs, or conversational atmosphere. By reorganizing the sentences, the logical structure of the information can be optimized to match the user's understanding habits, ensuring the accuracy and fluency of information transmission; by adjusting the tone, the intelligent agent can be endowed with more emotionally expressive abilities, making its output language style more diverse and its tone more natural. This effectively solves the problem of the intelligent agent's stiff response and lack of anthropomorphic features, improving the naturalness and affinity of the interaction, thus enabling the intelligent agent to provide responses closer to human thought and emotion when interacting with humans, enhancing the user experience.

[0089] In some of the solutions mentioned above in this application, a memory index matrix is ​​proposed to store and match interactive information to realize the memory retrieval of the agent. However, in this process, as the interactive information continues to accumulate, the fixed matching degree threshold is difficult to adapt to the dynamically changing interactive scenarios, resulting in large fluctuations in query accuracy in different contexts, and it is impossible to always maintain the optimal retrieval accuracy.

[0090] In response, this application further proposes an interaction method for intelligent agents, which further includes: Step c1 involves continuously receiving interactive information, performing the receiving of interactive information, and storing the interactive information based on cognitive category and semantic feature category to obtain the storage memory bank.

[0091] Step c2 involves adjusting the preset matching threshold based on the interaction scenario to enhance query accuracy.

[0092] The steps involved in continuously receiving and storing interactive information, based on cognitive and semantic feature categories, to create a storage memory are designed to ensure the agent's memory is constantly updated and expanded over time, providing the agent with a continuously growing knowledge reserve. Specifically, this step can be implemented in several ways. For example, the agent can use real-time streaming processing, semantically decomposing and generating feature codes for new statements each time it interacts with the user, and storing them in the memory to ensure timely updates. This is suitable for high-frequency, real-time interaction scenarios. Alternatively, the agent can use periodic batch updates, processing accumulated interactive information over a period of time in batches during off-peak hours or at preset time intervals, and storing it in the memory to optimize system resource utilization. This is suitable for scenarios with relatively low real-time requirements but large data volumes.

[0093] The core technical solution of this application is to improve query accuracy by adjusting the preset matching degree threshold based on the interaction scenario. The interaction scenario refers to the comprehensive factors such as the specific environment, contextual information, user intent, dialogue history, and current task type when the agent interacts with the user. The preset matching degree threshold is a value used to determine whether the similarity between the query feature subcode and the semantic feature subcode meets the retrieval standard. Only when the semantic matching degree is greater than or equal to this threshold will the corresponding statement information be considered a valid query result. By dynamically adjusting this threshold, the agent can balance recall and precision in scenarios of different complexities. Specifically, this adjustment process can include the following implementation methods: One method is to use a rule engine for adjustment, that is, to predefine a series of rules and associate specific interaction scenarios, such as users asking factual questions, users expressing emotions, and users seeking advice, with the corresponding matching degree threshold range. When the system identifies that the current interaction scenario conforms to a certain rule, it automatically adjusts the matching degree threshold. For example, in a factual question-and-answer scenario, the threshold may be set to 0.85; in an emotional exchange scenario, the threshold may be set to 0.65. Another approach is to use a machine learning model for dynamic prediction. This involves training a machine learning model that takes various features of the interaction scenario, such as dialogue turns, keywords, user sentiment, and task type, as input and outputs a dynamic matching threshold. The model learns from successful retrieval cases in historical interaction data to predict the most appropriate threshold for the current scenario, thereby maximizing query accuracy.

[0094] It is understandable that, through the aforementioned technical solutions, the intelligent agent can continuously update and expand its memory bank, ensuring the real-time nature and richness of its knowledge reserves. By introducing a dynamic threshold adjustment mechanism, the intelligent agent can intelligently adjust the strictness of memory retrieval based on the current interaction scenario. For example, in scenarios requiring precise answers, the system can increase the matching threshold to ensure high accuracy of the returned results; while in scenarios requiring fuzzy association or creative divergence, the system can appropriately lower the matching threshold to allow for broader associations and richer insights. This method of adjusting thresholds in real time based on the interaction scenario effectively addresses the retrieval bias caused by a single fixed threshold in varied interactions, ensuring that the intelligent agent maintains a high level of query accuracy when handling different types of queries. This enhances the anthropomorphism and response quality of the interaction, enabling the intelligent agent's memory retrieval system to adaptively optimize for the interaction environment.

[0095] In one example, a more specific case will be used to illustrate the above technical solution in greater detail: In an intelligent agent system, the agent aims to provide a more natural and efficient interactive experience by simulating human cognitive processes in storing and retrieving information.

[0096] First, the intelligent agent system constructs a memory index matrix. The rows of this matrix are designed to represent multiple cognitive categories, such as emotional experiences, skill mastery, personality traits, lifestyle habits, interpersonal relationships, and interests. These categories directly correspond to the categorization of everyday human cognitive processes. Simultaneously, the columns are designed to represent multiple semantic feature categories, such as time features, location features, associated object features, event content features, emotional tendency features, behavioral features, and behavioral outcome features. This structured memory index matrix, unlike traditional technology-driven single-label or full-data indexes, establishes a classification logic that aligns with human cognitive thinking from the outset. This lays the foundation for subsequent memory storage and retrieval, effectively solving the problem of the disconnect between index classification logic and human cognitive thinking in existing technologies.

[0097] When user A interacts with the agent, for example, user A says, "I usually get up at seven in the morning and then go for a half-hour jog in the park," the agent receives this interaction, processes it, and stores it. Specifically, the agent first semantically decomposes the statement, obtaining multiple semantic feature sub-codes. For example, "seven in the morning" is decomposed into a time feature sub-code, "get up" into a behavior feature sub-code, and "jog in the park" into a location feature sub-code and a behavior feature sub-code. These semantic feature sub-codes are hash sub-codes, and their dimensions correspond to the row dimensions of the memory index matrix, such as "lifestyle habits," and the column dimensions, such as "time feature," "behavioral feature," and "location feature." Next, these feature sub-codes are combined to form the feature code corresponding to the interaction information; this feature code is a fuzzy hash short code.

[0098] Subsequently, the agent obtains the dimensional information of the feature code, which represents the dimensional attributes of each feature sub-code within the feature code. The agent maps this dimensional information to the row dimensions in the memory index matrix (e.g., associating behaviors like "getting up" and "jogging" with the cognitive category of "lifestyle habits") and the column dimensions (e.g., associating "7 AM" with "time features"), thus obtaining the coordinate information of the feature code in the memory index matrix. Finally, the agent binds this coordinate information to a storage address and stores it, forming a storage memory bank. This storage method ensures that the memory data is no longer a fragmented accumulation, but rather highly structured and closely associated with cognitive categories, reducing the problem of redundant and messy memory data. Compared with existing technologies that rely on full data traversal or precise matching of single labels, this solution achieves refined and structured storage of memory through semantic decomposition, feature code generation, and coordinate mapping, enabling efficient subsequent retrieval.

[0099] When user A sends a query to the agent, such as "What do I usually do in the morning?", the agent receives the query and parses it. First, it determines the main row dimension corresponding to the query, such as "lifestyle habits" or "behavioral logic". Then, the agent breaks down the query to obtain at least one query feature subcode, such as the query time feature subcode corresponding to "morning" and the query behavior feature subcode corresponding to "what do I do", which are also hash subcodes.

[0100] Subsequently, within a defined row dimension, such as "lifestyle habits," the agent performs fuzzy matching between these query feature subcodes and the corresponding semantic feature subcodes in the stored memory, and calculates the semantic matching degree. For example, the query "morning" is fuzzily matched with the time feature subcode "7:00 AM" in the stored memory, and the query "what to do" is fuzzily matched with the behavioral feature subcodes "get up" and "jog" in the stored memory. This fuzzy matching mechanism can simulate the fuzzy association ability of humans. Even if the query command and the stored statement information are not completely exact matches, it can still identify highly relevant memories, effectively solving the problems of low retrieval efficiency, high computational consumption, and lack of anthropomorphic fuzzy association ability in existing technologies.

[0101] When the semantic matching degree is greater than a preset matching degree threshold, the agent determines the sentence information corresponding to the semantic feature subcode, namely, "I usually get up at seven o'clock in the morning and then go for a half-hour jog in the park." Based on this sentence information, the agent generates query results. For example, based on the query intent "What do I usually do in the morning?", the agent restructures the retrieved sentence information to generate the response: "You usually get up at seven o'clock in the morning and then go for a half-hour jog in the park." This sentence restructuring based on query intent makes the agent's response more natural and context-aware, further enhancing the human-like interactive experience.

[0102] Furthermore, the intelligent agent system continuously receives new interaction information and executes the aforementioned storage steps, constantly enriching and updating its storage memory. Simultaneously, the system can adjust the preset matching threshold based on the current interaction scenario; for example, it can raise the threshold when high-precision matching is needed and lower the threshold when broader association is required, thereby dynamically enhancing query accuracy and enabling the intelligent agent to provide appropriate responses in different situations.

[0103] This embodiment also provides an interaction device for an intelligent agent, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0104] This embodiment provides an interactive device for an intelligent agent, such as... Figure 5 As shown, it includes: Module 501 is used to construct a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories. The storage module 502 is used to receive interactive information and store the interactive information based on cognitive category and semantic feature category to obtain a storage memory bank; The query module 503 is used to receive query commands and match the query commands with the statement information in the storage memory to obtain query results.

[0105] In some alternative implementations, the storage module 502 includes: The first unit is used to semantically decompose the statement information to obtain multiple semantic feature subcodes. The semantic feature subcodes include multiple dimensions, which are the dimensions corresponding to the row and column dimensions of the memory index matrix.

[0106] The second unit is used to combine the feature subcodes to obtain the feature code corresponding to the interaction information.

[0107] The third unit is used to bind the feature code to the storage address and then store it to obtain the storage memory.

[0108] The interactive device for the intelligent agent provided in this application can execute the interactive method for the intelligent agent provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0109] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0110] The following is a detailed reference. Figure 6 This diagram illustrates a suitable structural schematic for implementing the electronic device described in the embodiments of this application. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0111] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0112] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 609, or installed from memory 608, or installed from ROM 602. When the computer program is executed by processor 601, it performs the functions defined in the interaction method of the intelligent agent according to embodiments of this application.

[0113] Figure 6 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0114] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the interaction method of the intelligent agent shown in the above embodiments is implemented.

[0115] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0116] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. An interaction method for intelligent agents, characterized in that, The method includes: Construct a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories; Receive interaction information and store the interaction information based on the cognitive category and the semantic feature category to obtain a storage memory bank, wherein the interaction information includes statement information for training the agent; Receive a query command and match the query command with the statement information in the storage memory to obtain the query result.

2. The method according to claim 1, characterized in that, The process of receiving interaction information and storing the interaction information based on the cognitive category and the semantic feature category to obtain a storage memory bank includes: The statement information is semantically decomposed to obtain multiple semantic feature subcodes, wherein the semantic feature subcodes include multiple dimensions, which are dimensions corresponding to the row dimensions and column dimensions of the memory index matrix; The feature subcodes are combined to obtain the feature code corresponding to the interaction information; The feature code is bound to the storage address and then stored to obtain the storage memory.

3. The method according to claim 2, characterized in that, The process of binding the feature code with a storage address and storing it to obtain the storage memory includes: Obtain the dimensional information of the feature code, wherein the dimensional information of the feature code is used to characterize the dimensional information of the feature subcode within the feature code; The dimensional information is mapped to the row dimension and column dimension in the memory index matrix to obtain the coordinate information of the feature code in the memory index matrix; The coordinate information is bound to the storage address and then stored.

4. The method according to claim 2, characterized in that, The row dimension includes at least one of the following: emotional experience, skill mastery, personality traits, living habits, interpersonal relationships, social rules, social scenarios, career cognition, natural phenomena, environmental interaction, transaction attributes, behavioral logic, value cognition, interests and preferences, and special experiences. The column dimensions include at least one of the following: time features, location features, associated object features, event content features, sentiment features, scene attribute features, behavioral features, and behavioral result features.

5. The method according to claim 1, characterized in that, The step of receiving a query instruction and matching the query instruction with the statement information in the storage memory to obtain the query result includes: Parse the query command to determine the row dimension corresponding to the query command; The query instruction is decomposed to obtain at least one query feature subcode; Within the row dimension, fuzzy matching is performed on the query feature subcode and semantic feature subcode to obtain the semantic matching degree; When the semantic matching degree is greater than the preset matching degree threshold, the statement information corresponding to the semantic feature subcode is determined, and a query result is generated based on the statement information.

6. The method according to claim 5, characterized in that, The process of generating query results based on the statement information includes: Based on the query intent of the query instruction, the statement information is restructured or the tone is adjusted.

7. The method according to claim 5, characterized in that, The method further includes: The steps are: continuously receiving the interaction information, receiving the interaction information, and storing the interaction information based on the cognitive category and the semantic feature category to obtain a storage memory bank; Based on the interaction scenario, the preset matching threshold is adjusted accordingly to enhance query accuracy.

8. An interactive device for an intelligent agent, characterized in that, The device includes: A construction module is used to construct a memory index matrix, wherein the row dimension of the memory index matrix includes multiple cognitive categories, and the column dimension of the memory index matrix includes multiple semantic feature categories; A storage module is used to receive interaction information and store the interaction information based on the cognitive category and the semantic feature category to obtain a storage memory bank; The query module is used to receive query instructions, match the query instructions with the statement information in the storage memory, and obtain query results.

9. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the interaction method of the intelligent agent as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the interaction method of the intelligent agent as described in any one of claims 1 to 7.

Images