Large model-based intention recognition method and device, electronic equipment and storage medium

Abstract

The invention provides an intention recognition method and device based on a large model, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. The method comprises the steps of generating a current query statement according to a scene context and current dialogue content through a large model, matching the current query statement with an intention library, and selecting candidate intentions from the intention library according to a matching result; and carrying out current round processing on the candidate intention and the current query statement, if the processing result of the current round does not meet the intention identification requirement, fusing the processing result of the current round into the scene context to update the current query statement, and carrying out next round processing according to the updated current query statement. And the target intention meeting the intention recognition requirement is obtained.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, specifically the field of artificial intelligence technology. Specifically, it relates to an intent recognition method, apparatus, electronic device, and storage medium based on a large model. Background Technology

[0002] Large-scale models, with their powerful language understanding and generation capabilities, have become a significant force driving technological progress. Based on these large-scale models, agent technology is gradually emerging as a key to enhancing the interactive capabilities of AI (Artificial Intelligence). Agents can simulate human behavior and decision-making processes, achieving more natural and efficient interactions and injecting intelligent vitality into various application scenarios.

[0003] Users have placed new demands on the intent understanding capabilities of intelligent conversational assistants, thus posing numerous challenges to intent recognition tasks. Summary of the Invention

[0004] This disclosure provides an intent recognition method, apparatus, electronic device, and storage medium based on a large model.

[0005] According to one aspect of this disclosure, an intent recognition method based on a large model is provided, comprising:

[0006] The large model generates a current query statement based on the scene context and the current dialogue content, matches the current query statement with the intent library, and selects candidate intents from the intent library based on the matching results.

[0007] The candidate intent and the current query statement are processed in the current round. If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement. The next round of processing is then performed based on the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

[0008] According to one aspect of this disclosure, an intent recognition device based on a large model is provided, comprising:

[0009] The candidate intent selection module is used to generate a current query statement based on the scene context and the current dialogue content using a large model, match the current query statement with the intent library, and select candidate intents from the intent library based on the matching results.

[0010] The target intent acquisition module is used to process the candidate intent and the current query statement in the current round. If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement, and the next round of processing is performed according to the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

[0011] According to another aspect of this disclosure, an artificial intelligence-based intelligent agent is provided, comprising:

[0012] The input module is used to receive input information;

[0013] The processing module is configured to execute the method provided in any embodiment of this disclosure based on the input information received by the input module to obtain output information;

[0014] The output module is used to output the output information obtained by the processing module.

[0015] According to another aspect of this disclosure, an electronic device is provided, the electronic device comprising:

[0016] At least one processor; and

[0017] A memory communicatively connected to the at least one processor; wherein,

[0018] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the methods provided in any embodiment of this disclosure.

[0019] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods provided in any embodiment of this disclosure.

[0020] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the methods provided according to any embodiment of this disclosure.

[0021] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become apparent from the following description. Attached Figure Description

[0022] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0023] Figure 1This is a flowchart of an intent recognition method based on a large model provided according to an embodiment of this disclosure;

[0024] Figure 2a This is a flowchart of another intent recognition method based on a large model provided according to embodiments of this disclosure;

[0025] Figure 2b This is a schematic diagram illustrating the principle of intent recognition based on a large model according to an embodiment of this disclosure;

[0026] Figure 3 This is a flowchart of yet another intent recognition method based on a large model provided according to embodiments of this disclosure;

[0027] Figure 4 This is a schematic diagram of the structure of an intent recognition device based on a large model according to an embodiment of the present disclosure;

[0028] Figure 5 This is a block diagram of an electronic device used to implement an intent recognition method based on a large model according to embodiments of the present disclosure. Detailed Implementation

[0029] Figure 1 This is a flowchart illustrating a large-model-based intent recognition method according to embodiments of this disclosure. This method is applicable to situations where intent recognition is performed based on a large model for dialogue content generated during interaction between an intelligent conversational assistant and a user. This method can be executed by a large-model-based intent recognition device, which can be implemented in software and / or hardware and can be configured within the intelligent agent of an electronic device. Reference Figure 1 One example of an intent recognition method based on a large model in this embodiment may include:

[0030] S101, using a large model, generate a current query statement based on the scene context and the current dialogue content, match the current query statement with the intent library, and select candidate intents from the intent library based on the matching results;

[0031] S102, the candidate intent and the current query statement are processed in the current round. If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement, and the next round of processing is performed according to the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

[0032] The intelligent conversational assistant interacts with the user to obtain the current conversation content and uses a large model combined with the context to perform intent recognition. Context refers to the set of related information input into the large model to assist in intent understanding, including the current reading page environment (such as page function, device type, etc.), user characteristics, and conversation process information. The current query statement is a prompt input into the large model, used to trigger the intent recognition process.

[0033] For example, the context and current dialogue content are input into a large model. The model can then use an intent recognition prompt module to rewrite the current dialogue content and generate the current query. This process transforms the user's colloquial and fragmented input into structured prompts, lowering the input threshold and improving the convenience of the intelligent conversational assistant. Furthermore, based on the large model, the current query is semantically matched with various intents in the intent library. Through a recall and ranking mechanism, the top k candidate intents are selected from the intent library, where k is a preset positive integer. The recall rate must be higher than a corresponding threshold to ensure that no potentially relevant intents are missed.

[0034] The Reasoning Acting (ReAct) mechanism is an agent-based decision-making framework that deeply integrates reasoning and action. It achieves this deep integration through a dynamic cycle of "thinking-action-observation." The agent acquires multiple candidate intents selected from an intent database by a large model and dynamically analyzes these intents using the Reasoning Acting (ReAct) mechanism. During the current processing round, the agent processes the candidate intents and the current query statement to generate the current round's processing result. It then determines whether the current round's processing result meets the intent recognition requirements, for example, by calculating the semantic similarity between the candidate intent and the current query statement as a confidence score, and checking if the confidence score exceeds a preset confidence threshold. If the current round's processing result contains a candidate intent with a confidence score exceeding the confidence threshold, then that candidate intent meets the intent recognition requirements and is directly output as the target intent. The target intent can be a creation intent such as a generation or editing intent, or a search intent or an operational intent, etc. This approach automatically transforms users' conversational and fragmented input into actionable creative tasks. It not only supports traffic distribution for core functions such as intelligent PPT generation, full-network document retrieval, AI writing assistance, and academic Q&A, but also effectively addresses key challenges such as understanding professional terminology, adapting to format specifications, and optimizing multi-round interactions.

[0035] If the current round of processing determines that none of the candidate intents meet the intent recognition requirements, multiple rounds of iterative processing are performed: the current round's processing result is integrated into the scene context to update the current query statement, and the next round of processing is performed based on the updated current query statement; this process is repeated until the target intent that meets the intent recognition requirements is output. Alternatively, based on the current round's result, the system can invoke the authoring tools in the interactive interface or initiate content clarification, feeding back the corresponding processing result to the next round, thus progressively exploring complex user needs and improving the accuracy and efficiency of intent recognition. The reasoning-action loop of the reflection mechanism enables the system to proactively identify issues such as missing content and logical flaws, and provides professional suggestions by calling resources such as knowledge bases and template libraries.

[0036] The technical solution provided in this disclosure generates a current query statement by combining a large model with the scene context and the current dialogue content; semantically matches the current query statement with each intent in the intent library to filter multiple candidate intents; the intelligent agent uses a reflective mechanism to perform multiple rounds of iterative processing until it outputs the target intent that meets the intent recognition requirements. This mechanism can reduce the user input threshold, improve the convenience of intelligent conversation assistants, and at the same time improve the accuracy and efficiency of intent recognition.

[0037] Figure 2a This is a flowchart of another intent recognition method based on a large model provided according to embodiments of this disclosure, used to further limit the matching between the current query statement and the intent database. Combined with... Figure 2a and Figure 2b One example of an intent recognition method based on a large model in this embodiment may include:

[0038] S201, Input the current dialogue content and the scene context into the large model for conversation recognition to obtain the conversation type;

[0039] S202, based on the session type and the scene context, rewrite the current dialogue content to obtain the current query statement;

[0040] S203, Based on the large model, the current query statement is matched with the intent library, and the candidate intent is selected from the intent library according to the matching result;

[0041] S204, perform current round processing on the candidate intent and the current query statement. If the processing result of the current round does not meet the intent recognition requirements, then integrate the processing result of the current round into the scene context to update the current query statement, and perform the next round processing based on the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

[0042] In this embodiment, candidate intents are filtered from the intent library through the following process: the current dialogue content and scene context are input into a large model for conversation recognition. It can be determined whether the current dialogue content and historical dialogue content belong to the same topic; if the current dialogue content is semantically relatively independent and has no topic continuity with historical dialogue content, it is determined to be a single-turn conversation; if the current dialogue content has topic continuity with historical dialogue content, it is determined to be a multi-turn conversation. Combining the conversation type and scene context, the current dialogue content is intelligently rewritten in a structured manner to generate the current query statement.

[0043] In related intent recognition technologies, large models are used for end-to-end semantic understanding of conversation content. While this demonstrates excellent generalization ability and contextual understanding advantages, it struggles to effectively track and maintain the conversation state in multi-turn dialogue scenarios, leading to the loss of historical information and biases in intent understanding. This disclosure provides a three-stage architecture of "conversation recognition," "current dialogue rewriting," and "reflection mechanism," overcoming the limitations of end-to-end large models or fixed multi-turn templates. While maintaining high accuracy, it significantly reduces system latency and computational resource consumption.

[0044] In one optional implementation, rewriting the current dialogue content according to the session type and the scene context to obtain the current query statement includes: when the session type is a single-turn session, expanding the statement in the current dialogue content according to at least one of the task scenario, source path, and user static attributes in the scene context to obtain the current query statement.

[0045] For single-turn conversation scenarios, intelligent expansion can be performed on statements with brief or implicit intents in the current dialogue content based on at least one piece of information from the task scenario, source path, or user static attributes within the scenario context. Specifically, semantic completion is performed on statements lacking key entities (such as time, location, or object of operation) or logical relationships (such as cause and effect or condition) in the current dialogue content, generating complete and explicit query instructions as the current query statement, thereby improving the accuracy of candidate intents.

[0046] In one optional implementation, rewriting the current dialogue content according to the session type and the scene context to obtain the current query statement includes: when the session type is a multi-turn session, extracting historical dialogue content with topic continuity with the current dialogue content from the historical dialogue records included in the scene context; clarifying the referential relationship in the current dialogue content according to at least one of the historical dialogue content, the task scenario in the scene context, the source path, and the user's static attributes, to obtain the current query statement.

[0047] For multi-turn conversation scenarios, historical dialogue content with topical continuity with the current dialogue content is extracted; that is, the parts belonging to the same topic as the current dialogue content are extracted from historical dialogue records. Combining historical dialogue content on the same topic with at least one of the following: task scenario, source path, or user static attributes, the referential relationships in the current dialogue content are clarified, realizing the resolution of contextual dependencies and providing rich dialogue background for subsequent processing. For example, based on historical dialogue content on the same topic, and combined with task scenario, source path, or user static attributes, contextual dependency features such as pronoun references and ellipsis are identified in the current dialogue content. By analyzing semantic logic, contextual clues, and user characteristics, ambiguous referential expressions are transformed into specific and clear query content, enabling the large model to accurately understand user intent and provide answers that conform to the coherence and consistency of multi-turn dialogues.

[0048] In one optional implementation, the method further includes: obtaining the historical dialogue intent corresponding to the historical dialogue content; and adding an initial intent to the current query statement based on the historical dialogue intent. For example, to further enrich the basis for subsequent intent recognition, the historical dialogue intent obtained by pre-recognizing the historical dialogue content can be obtained and used as the initial intent of the current query statement. It should be noted that the current query statement and the historical dialogue intent may be the same or different.

[0049] Single-turn and multi-turn recognition analyzes the user's continuous creative instructions to intelligently determine the continuity and completeness of the creative task, providing the system with a coherent understanding of the creative process. Query rewriting, based on semantic understanding of the creative scenario, standardizes and semantically completes the fragmented creative needs of users, significantly improving the clarity and executability of creative instructions. This mechanism endows the creative process with memory and state tracking capabilities through single-turn and multi-turn recognition, and strengthens the accuracy of creative instruction expression through query rewriting; together, they provide a high-quality input foundation for candidate intent screening.

[0050] In one optional implementation, the task scenario indicates the scenario of the current reading page where the current session is located, and the task scenario includes device type and page function scenario; the source path is used to describe the path trajectory to the current reading page; the user static attributes include user preference information or historical behavior attributes.

[0051] The device types can include PCs (Personal Computers), apps (Applications), or WAP (Wireless Application Protocol) interfaces on mobile browsers. Page functional scenarios can include the homepage of the interactive interface, document browsing pages, editing pages, etc. The source path can include the path from the search interface to the reading page; or the path from the previous reading page within the interactive interface to the current reading page, etc. The interactive interface can be a knowledge service product, including online creation, editing, retrieval, reading, downloading, and sharing functions. User static attributes refer to relatively stable user characteristic information that does not easily change with the user's current session behavior. User preference information can include document type preferences, topic preferences, or reading habit preferences; historical behavior attributes can include historical search keywords, frequency of use of frequently used functions, or historical document reading time, etc.

[0052] By combining the current dialogue content with at least one of the following contextual elements—task scenario, source path, user static attributes, or historical dialogue content—session recognition is performed on the current dialogue content. For single-turn conversations, the system intelligently expands the user's short or implicit intent statements by combining at least one of the following elements—task scenario, source path, or user static attributes—transforming them into complete and explicit instructions. For multi-turn conversations, deep contextual understanding technology is applied based on historical dialogue content and at least one of the following elements—task scenario, source path, or user static attributes—to dynamically parse the referential relationships in the user's expression, reconstructing fragmented natural language into semantically complete query requests. By standardizing and semantically completing the user's fragmented creative needs to obtain the current query content, the clarity and executability of creative instructions are significantly improved. Simultaneously, the query rewriting based on deep semantic understanding ensures the accurate transmission of creative instructions, preventing content deviations caused by unclear expression from the outset.

[0053] The technical solution provided in this disclosure performs conversation recognition by inputting the current dialogue content and scene context into a large model. The context-preserving capability of multi-turn recognition ensures stylistic consistency and logical coherence throughout the creation process, significantly improving the professionalism and completeness of the output content. Simultaneously, by rewriting the current dialogue content based on the conversation type and scene context to obtain the current query statement, the accurate transmission of creation instructions is ensured, avoiding content deviations caused by unclear expression from the outset, and providing coherent contextual understanding support for subsequent intent recognition. Furthermore, this processing significantly improves the clarity and executability of creation instructions by explicitly displaying implicit intents and clarifying referential relationships, enabling the system to possess a coherent understanding capability of the creation process, thereby improving the accuracy and efficiency of intent understanding.

[0054] In one optional implementation, before matching the current query statement with the intent library, the method further includes adding an initial intent to the current query statement by at least one of the following: if the intelligent conversation assistant is triggered by a target button in the interactive interface, then based on a preset association between buttons and intents, the intent associated with the target button is taken as the initial intent; if the input box corresponding to the current dialogue content includes a prompt word for any intent, then that intent is taken as the initial intent; wherein, the current dialogue content is the content obtained after the user performs an input operation on the input box.

[0055] For example, initial intents can also be obtained in the following ways: One is intent labeling based on cached associations, which pre-defines the association between buttons and intents. When the AI ​​assistant session is triggered by a target button in the interactive interface, the intent associated with the target button is labeled as the associated initial intent. Another is intent labeling based on forced hit rules, which pre-defines the association between intents and prompt words. When a preset prompt word is included in the input box of the current dialogue content, such as "Help me generate a PPT, the topic is *", the intent associated with that prompt word is used as the initial intent. Historical intents pre-obtained from historical dialogue content can also serve as a supplementary source of initial intents. By introducing initial intents as auxiliary information into the current query statement, the contextual features of subsequent intent matching can be enriched, further improving the accuracy and efficiency of candidate intent retrieval.

[0056] Figure 3 This is a flowchart of another intent recognition method based on a large model provided according to embodiments of this disclosure, used to further define the reflective mechanism of the agent. See also Figure 3 One example of an intent recognition method based on a large model in this embodiment may include:

[0057] S301, using a large model, generate a current query statement based on the scene context and the current dialogue content, match the current query statement with the intent library, and select candidate intents from the intent library based on the matching results;

[0058] Execute S302-S306 in the current round to obtain the processing result of the current round;

[0059] S302, perform semantic similarity calculation between the candidate intent and the current query statement to obtain the confidence score;

[0060] S303, perform conflict detection based on the candidate intent, the current query statement and the historical dialogue content, and obtain the context consistency verification result;

[0061] S304, perform ambiguity identification on the candidate intent, the current query statement and the historical dialogue content, and obtain the ambiguity identification result;

[0062] S305, determine whether the candidate intent meets the dynamic supplementation condition or the dynamic elimination condition, and obtain the determination result;

[0063] S306, Update the candidate intent based on the determination result;

[0064] S307, if the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement, and the next round of processing is performed according to the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

[0065] In an optional implementation, the method further includes: if there is at least one candidate intent with a confidence level higher than the confidence level threshold in the processing results of the current round, then the at least one candidate intent is output as the target intent, and the corresponding confidence level is output.

[0066] The agent employs a reflective mechanism to analyze candidate intents in multiple rounds, using a closed-loop iterative design. The processing result of each round is fed back to the next round until a target intent that meets the intent recognition requirements is output. In each round, the agent performs the following tasks to obtain the processing result of the current round:

[0067] A. Confidence Calculation: Calculate the semantic similarity between multiple candidate intents and the current query statement to obtain the confidence score; if the confidence score of any candidate intent is higher than the preset confidence score threshold, then the candidate intent is output as the target intent, and the corresponding confidence score is output at the same time.

[0068] B. Contextual consistency verification: Perform conflict detection on candidate intents, current query statements, and historical dialogue content, such as semantic conflicts, temporal conflicts, or scene conflicts. For example, if a PC intent appears on a WAP end, there is a scene conflict. Determine whether there is a logical conflict between the candidate intent and the dialogue, and use this as the result of the contextual consistency verification.

[0069] C. Ambiguity Identification and Processing: Based on candidate intent, current query statement, and historical dialogue content, ambiguity identification is performed for lexical ambiguity, structural ambiguity, or contextual ambiguity, and the ambiguity identification results are obtained.

[0070] D. Dynamic adjustment of candidate intents: Determine whether the candidate intents meet the conditions for dynamic addition or dynamic elimination, and add new candidate intents or eliminate some candidate intents based on the determination results.

[0071] Based on the processing results of the tasks in the current round, the agent determines whether there are any issues with contextual consistency, ambiguity identification and handling, and dynamic adjustment of candidate intents. If issues are found, an interactive clarification process can be triggered, or candidate intents can be supplemented or eliminated to update the current query or candidate intents. The agent then analyzes the candidate intents and the current query in the next round until at least one target intent with a confidence level higher than the confidence threshold and its corresponding confidence level are output. To ensure system processing efficiency, alternative decision schemes are pre-configured to ensure reasonable results are still output when iteration limits are reached; simultaneously, a maximum number of iterations is set to avoid excessive resource consumption.

[0072] By employing a reflective decision-making mechanism, the creative process is structured as a "reasoning-action" loop. Through dynamic invocation of creation tools or content clarification, it progressively explores and satisfies complex creative needs. Compared to the limitations of traditional document creation tools in their single-command response mode, the current technology system has significantly overcome the bottlenecks in understanding continuous creation and implicit needs. Reflective decision-making represents a revolution in the intelligent creation paradigm—no longer limited to single-command responses, but proactively exploring the user's deep creative needs through interactive reasoning, enabling the system to possess creative intelligence that allows for continuous optimization and dynamic adjustment.

[0073] In one optional implementation, the step of integrating the processing result of the current round into the scene context to update the current query statement if the processing result of the current round does not meet the intent recognition requirements includes: if all the confidence scores are lower than the confidence score threshold and the current query statement has conflicts or ambiguities, then obtaining the input new dialogue content; integrating the processing result of the current round and the new dialogue content into the scene context to update the current query statement.

[0074] For example, if the confidence scores of all candidate intents are below the confidence threshold, and there is a conflict in the current query statement in the corresponding context consistency check result, or there is ambiguity in the current query statement in the ambiguity identification and processing result, then an interactive clarification process is triggered. Based on the corresponding conflict or ambiguity, the system interacts with the user to obtain new dialogue content to clarify the conflict or ambiguity. Furthermore, the processing result of the current round and the new dialogue content are integrated into the scene context to update the current query statement, thereby improving the quality of the query statement in the next round and thus improving the accuracy of intent recognition.

[0075] In one optional implementation, determining whether the candidate intent meets the dynamic supplementation condition includes: if the confidence level between the current query statement and the candidate intent is lower than the confidence level threshold in a consecutive preset numerical round, then it is determined that the dynamic supplementation condition is met; updating the candidate intent according to the determination result includes: obtaining new candidate intents from the intent library.

[0076] If, within a consecutive preset number of rounds (e.g., 3 consecutive rounds), the confidence score between the current query and each candidate intent is below a confidence threshold (e.g., 0.8), then the dynamic supplementation condition is met. In this case, new candidate intents can be retrieved from the intent library, for example, through user interaction, by having the user select a new candidate intent from the intent library. Additionally, if, through user interaction, it is determined that there is an intent whose frequency of use exceeds a preset frequency (e.g., occurrence frequency > 15 times / day) and does not belong to the intent library, then the dynamic supplementation condition is met. In this case, the intent can be added to the intent library, and new candidate intents can be added from the intent library. This dynamic supplementation mechanism for candidate intents effectively adapts to diverse user needs, improves the coverage and accuracy of intent recognition, and thus enhances the system's interactive flexibility and user experience.

[0077] In one optional implementation, determining whether the candidate intent meets the dynamic elimination criteria includes: determining whether the candidate intent meets the dynamic elimination criteria based on usage frequency and confidence level; updating the candidate intent based on the determination result includes: eliminating candidate intents that meet the dynamic elimination criteria.

[0078] If the usage frequency of any candidate intent is lower than the frequency threshold and its confidence level is lower than the lower confidence level, then the candidate intent meets the dynamic elimination criteria and can be eliminated. The lower confidence level is lower than the confidence threshold; for example, the lower confidence level is 0.4 and the confidence threshold is 0.8.

[0079] In one optional implementation, determining whether the candidate intent satisfies the dynamic elimination condition includes: if there is a mutual exclusion conflict between at least two candidate intents, then determining that the candidate intent satisfies the dynamic elimination condition; updating the candidate intent based on the determination result includes: eliminating candidate intents with low confidence.

[0080] If at least two candidate intents have overlapping conflicts, such as the first and second candidate intents having overlapping functional scope or semantic expression, then the confidence levels between the first and second candidate intents are compared, and the candidate intent with the lower confidence level is eliminated. This dynamic elimination mechanism effectively optimizes the candidate intent set, reduces redundancy and conflicts, and improves intent recognition efficiency and system resource utilization.

[0081] The technical solution provided in this disclosure uses a reflection mechanism to construct the creation process as a "reasoning-action" loop, which can dynamically call creation tools or initiate content clarification, progressively explore and meet complex creation needs, and enable the system to have creative intelligence that can be continuously optimized and dynamically adjusted.

[0082] Figure 4This is a schematic diagram of a large-model-based intent recognition device according to an embodiment of this disclosure. This embodiment is applicable to situations where intent recognition is performed based on a large model for dialogue content generated during interaction between an intelligent conversational assistant and a user. This device can be implemented in software and / or hardware and can be configured on mobile devices. Figure 4 As shown, an intent recognition device 400 based on a large model in this embodiment may include:

[0083] The candidate intent selection module 410 is used to generate a current query statement based on the scene context and the current dialogue content through a large model, match the current query statement with the intent library, and select candidate intents from the intent library based on the matching results.

[0084] The target intent acquisition module 420 is used to process the candidate intent and the current query statement in the current round. If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement, and the next round of processing is performed according to the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

[0085] In one optional implementation, the candidate intent selection module 410 includes:

[0086] The conversation recognition unit is used to input the current dialogue content and the scene context into the large model to perform conversation recognition and obtain the conversation type.

[0087] The dialogue rewriting unit is used to rewrite the current dialogue content according to the dialogue type and the scene context to obtain the current query statement;

[0088] The candidate intent selection unit is used to match the current query statement with the intent library based on the large model, and select the candidate intent from the intent library according to the matching result.

[0089] In one optional implementation, the dialogue rewriting unit includes a multi-turn rewriting subunit, which is specifically used for:

[0090] In the case where the conversation type is a multi-turn conversation, extract historical conversation content that has topic continuity with the current conversation content from the historical conversation records included in the scenario context;

[0091] Based on at least one of the historical dialogue content, the task scenario in the scene context, the source path, and the user's static attributes, the referential relationship in the current dialogue content is clarified to obtain the current query statement.

[0092] In one optional implementation, the multi-round rewriting subunit is further configured to:

[0093] Obtain the historical dialogue intent corresponding to the historical dialogue content;

[0094] Add an initial intent to the current query statement based on the historical dialogue intent.

[0095] In one optional implementation, the dialogue rewriting unit includes a single-round rewriting subunit, which is specifically used for:

[0096] When the session type is a single-turn session, the statement in the current dialogue content is expanded based on at least one of the task scenario, source path, and user static attributes in the scenario context to obtain the current query statement.

[0097] In one optional implementation, the task scenario indicates the scenario of the current reading page where the current session is located, and the task scenario includes device type and page function scenario; the source path is used to describe the path trajectory to the current reading page; the user static attributes include user preference information or historical behavior attributes.

[0098] In an optional implementation, the candidate intent selection module 410 further includes an initial intent addition unit, specifically configured to add an initial intent to the current query statement by at least one of the following:

[0099] If the intelligent conversation assistant is triggered by a target button in the interactive interface, the initial intent will be the intent associated with the target button, based on the preset association between the button and the intent.

[0100] If the input box corresponding to the current dialogue content includes any intention prompt word, then that intention is taken as the initial intention; wherein, the current dialogue content is the content obtained after the user performs an input operation on the input box.

[0101] In one optional implementation, the target intent acquisition module 420 includes:

[0102] The confidence unit is used to calculate the semantic similarity between the candidate intent and the current query statement to obtain the confidence score.

[0103] The conflict detection unit is used to perform conflict detection based on the candidate intent, the current query statement, and the historical dialogue content, and obtain the context consistency verification result.

[0104] An ambiguity identification unit is used to identify ambiguities in the candidate intent, the current query statement, and the historical dialogue content, and obtain an ambiguity identification result.

[0105] The intent condition unit is used to determine whether the candidate intent meets the dynamic addition condition or the dynamic elimination condition.

[0106] A candidate intent update unit is used to update the candidate intent based on the determination result.

[0107] In one optional implementation, the target intent acquisition module 420 includes a query statement update unit, specifically used for:

[0108] If all the confidence scores are below the confidence threshold, and the current query statement contains a conflict or ambiguity, then the new dialogue content is obtained.

[0109] The processing result of the current round and the new dialogue content are integrated into the scene context to update the current query statement.

[0110] In one alternative implementation, the intent condition unit is specifically used for:

[0111] If, in a series of consecutive preset numerical rounds, the confidence level between the current query statement and the candidate intent is lower than the confidence level threshold, then the dynamic supplementation condition is determined to be met.

[0112] The candidate intent update unit is specifically used to: obtain new candidate intents from the intent library.

[0113] In one alternative implementation, the intent condition unit is specifically used for:

[0114] Based on usage frequency and confidence level, determine whether the candidate intent meets the dynamic elimination criteria;

[0115] The candidate intent update unit is specifically used to: remove the candidate intent.

[0116] In one alternative implementation, the intent condition unit is specifically used for:

[0117] If there is a mutual exclusion conflict between at least two candidate intentions, then the candidate intention is determined to satisfy the dynamic elimination condition;

[0118] The candidate intent update unit is specifically used to: remove candidate intents with low confidence.

[0119] In one optional implementation, the target intent acquisition module 420 further includes:

[0120] The target intent acquisition unit is used to output the at least one candidate intent as the target intent and output the corresponding confidence level if there is at least one candidate intent with a confidence level higher than the confidence level threshold in the processing result of the current round.

[0121] The large-model-based intent recognition device provided in this application embodiment can execute the large-model-based intent recognition method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects for executing the large-model-based intent recognition method.

[0122] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0123] According to embodiments of this disclosure, this disclosure also provides an artificial intelligence-based intelligent agent, including: an input module for receiving input information; a processing module for executing an intent recognition method based on a large model provided in any embodiment of this disclosure based on the input information received by the input module to obtain output information; and an output module for outputting the output information obtained by the processing module.

[0124] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0125] Figure 5 This is a block diagram of an electronic device used to implement an intent recognition method based on a large model according to embodiments of the present disclosure.

[0126] Figure 5 A schematic block diagram of an example electronic device 500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0127] like Figure 5 As shown, the electronic device 500 includes a computing unit 501, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. The RAM 503 may also store various programs and data required for the operation of the electronic device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0128] Multiple components in electronic device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows electronic device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0129] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as the large model-based intent recognition method. For example, in some embodiments, the large model-based intent recognition method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the large model-based intent recognition method described above may be performed. Alternatively, in other embodiments, computing unit 501 may be configured to perform an intent recognition method based on a large model by any other suitable means (e.g., by means of firmware).

[0130] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0131] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0132] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0133] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0134] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0135] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0136] Artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies mainly include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0137] Cloud computing refers to a technology system that enables access to a shared pool of physical or virtual resources via a network. These resources can include servers, operating systems, networks, software, applications, and storage devices, and can be deployed and managed on demand and in a self-service manner. Cloud computing technology can provide efficient and powerful data processing capabilities for applications such as artificial intelligence and blockchain, as well as for model training.

[0138] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0139] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. Intent recognition methods based on large models, including: The large model generates a current query statement based on the scene context and the current dialogue content, matches the current query statement with the intent library, and selects candidate intents from the intent library based on the matching results. The candidate intent and the current query statement are processed in the current round. If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement. The next round of processing is then performed based on the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

2. The method according to claim 1, wherein, The process of generating a current query statement based on the scene context and current dialogue content using a large model, matching the current query statement with an intent library, and selecting candidate intents from the intent library based on the matching results includes: The current dialogue content and the scene context are input into a large model for session recognition to obtain the session type; Based on the session type and the scenario context, the current dialogue content is rewritten to obtain the current query statement; Based on the large model, the current query statement is matched with the intent library, and the candidate intent is selected from the intent library according to the matching result.

3. The method according to claim 2, wherein, The step of rewriting the current dialogue content according to the session type and the scene context to obtain the current query statement includes: In the case where the conversation type is a multi-turn conversation, extract historical conversation content that has topic continuity with the current conversation content from the historical conversation records included in the scenario context; Based on at least one of the historical dialogue content, the task scenario in the scene context, the source path, and the user's static attributes, the referential relationship in the current dialogue content is clarified to obtain the current query statement.

4. The method according to claim 3, further comprising: Obtain the historical dialogue intent corresponding to the historical dialogue content; Add an initial intent to the current query statement based on the historical dialogue intent.

5. The method according to claim 2, wherein, The step of rewriting the current dialogue content according to the session type and the scene context to obtain the current query statement includes: When the session type is a single-turn session, the statement in the current dialogue content is expanded based on at least one of the task scenario, source path, and user static attributes in the scenario context to obtain the current query statement.

6. The method according to claim 3 or 5, wherein, The task scenario indicates the scenario of the current reading page in which the current session is located. The task scenario includes the device type and page function scenario. The source path is used to describe the path trajectory to the current reading page. The user static attributes include user preference information or historical behavior attributes.

7. The method according to claim 2, wherein before matching the current query statement with the intent library, the method further includes adding an initial intent to the current query statement by at least one of the following: If the intelligent conversation assistant is triggered by a target button in the interactive interface, the initial intent will be the intent associated with the target button, based on the preset association between the button and the intent. If the input box corresponding to the current dialogue content includes any intention prompt, then that intention is taken as the initial intention; wherein, The current dialogue content is the content obtained after the user performs an input operation on the input box.

8. The method according to claim 1, wherein, The current round processing of the candidate intent and the current query statement includes: The confidence score is obtained by calculating the semantic similarity between the candidate intent and the current query statement; Conflict detection is performed based on the candidate intent, the current query statement, and the historical dialogue content to obtain the context consistency verification result; The candidate intent, the current query statement, and the historical dialogue content are subjected to ambiguity identification to obtain the ambiguity identification result. Determine whether the candidate intents meet the dynamic addition or dynamic elimination conditions to obtain the determination result; The candidate intent is updated based on the determination result.

9. The method according to claim 8, wherein, If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement, including: If all the confidence scores are below the confidence threshold, and the current query statement contains a conflict or ambiguity, then the new dialogue content is obtained. The processing result of the current round and the new dialogue content are integrated into the scene context to update the current query statement.

10. The method according to claim 8, wherein, Determining whether the candidate intent meets the dynamic supplementation conditions includes: If, in a series of consecutive preset numerical rounds, the confidence level between the current query statement and the candidate intent is lower than the confidence level threshold, then the dynamic supplementation condition is determined to be met. Updating the candidate intent based on the determination result includes: New candidate intents are obtained from the intent library.

11. The method according to claim 8, wherein, Determining whether the candidate intent meets the dynamic elimination criteria includes: Based on usage frequency and confidence level, determine whether the candidate intent meets the dynamic elimination criteria; Updating the candidate intent based on the determination result includes: removing candidate intents that meet the dynamic removal conditions.

12. The method according to claim 8, wherein, Determining whether the candidate intent meets the dynamic elimination criteria includes: If there is a mutual exclusion conflict between at least two candidate intentions, then the candidate intention is determined to satisfy the dynamic elimination condition; Updating the candidate intent based on the determination result includes: removing candidate intents with low confidence.

13. The method according to claim 8, further comprising: If there is at least one candidate intent with a confidence level higher than the confidence level threshold in the processing results of the current round, then the at least one candidate intent is output as the target intent, and the corresponding confidence level is output.

14. An intent recognition device based on a large model, comprising: The candidate intent selection module is used to generate a current query statement based on the scene context and the current dialogue content using a large model, match the current query statement with the intent library, and select candidate intents from the intent library based on the matching results. The target intent acquisition module is used to process the candidate intent and the current query statement in the current round. If the processing result of the current round does not meet the intent recognition requirements, the processing result of the current round is integrated into the scene context to update the current query statement, and the next round of processing is performed according to the updated current query statement until the target intent that meets the intent recognition requirements is obtained.

15. Intelligent agents based on artificial intelligence, including: The input module is used to receive input information; A processing module is configured to execute the method as described in any one of claims 1 to 13 based on the input information received by the input module to obtain output information; An output module is used to output the output information obtained by the processing module.

16. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-13.

17. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-13.

18. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-13.

Images