Intention recognition method and related device

Abstract

An intention recognition method is used for improving the efficiency of recognizing user intentions. In the method, aiming at task data input by a user, a plurality of concern points related to the task data are analyzed through a large language model, the large language model is promoted to identify user intentions corresponding to the task data under the different concern points respectively, and finally candidate intention results obtained under the concern points are synthesized, so that the user intentions corresponding to the task data are identified. And determining a final target intention result. According to the scheme, the roles under different attention angles are set for the large language model by analyzing the multiple attention points related to the task data, so that the large language model recognizes the user intention from different attention angles, the phantom problem of the large language model is solved, and the accuracy of the output target intention result can be ensured; moreover, the general knowledge storage and semantic comprehension capabilities of the existing large language model are utilized to the greatest extent, rules or training data do not need to be provided manually, and the efficiency of realizing intention recognition is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence (AI) technology, and in particular to an intention recognition method and related devices. Background Art

[0002] In human-computer dialogue systems, intent detection (or intent recognition) technology is a key technology for achieving human-computer dialogue. Among them, the main task of intent recognition technology is to identify the user's intention based on the user's input.

[0003] The development of intent recognition technology in human-computer dialogue systems has gone through several important stages. Early intent recognition relied primarily on rule engines, which designed a series of rules to determine user intent. Rule-based intent recognition technology is simple and easy to implement, but its disadvantage is that it requires a large amount of manually designed rules, making it difficult to cover all possible scenarios. With the development of machine learning (ML) technology, statistical-based intent recognition technology has gradually become mainstream. Statistical-based intent recognition technology analyzes large amounts of historical data to learn an intent recognition model, thereby automatically identifying user intent. The advantage of statistical-based intent recognition technology is that it can learn automatically and is applicable to a wide range of scenarios, but the disadvantage is that it requires a large amount of training data to train the intent recognition model.

[0004] In general, current intent recognition technology requires humans to design a large number of rules or provide a large amount of training data for specific intent recognition scenarios, resulting in low efficiency in intent recognition and difficulty in quickly realizing intent recognition in various new scenarios. Summary of the Invention

[0005] The present application provides an intent recognition method that can improve the efficiency of intent recognition.

[0006] In a first aspect, a method for intent recognition is provided for improving the efficiency of intent recognition by utilizing a large language model. In this method, task data input by a user is first obtained. The task data is multimedia data, such as one or more of the following: text, voice, image, or video. Furthermore, the task data indicates a task that the user desires to complete, such as booking a ticket or ordering takeout.

[0007] Then, using the large language model, a set of focus points for the task data is identified. This focus point set includes multiple focus points, each of which indicates the constraints that must be met when executing the task. Multiple focus points can also be understood as indicating different perspectives to consider when executing the task. For example, if the task a user wants to complete is booking a ticket, the multiple focus points for the ticket booking task may include time, convenience, and cost.

[0008] The large language model is used to identify the intent of the task data from each focus point in the focus point set, thereby obtaining multiple candidate intent results corresponding to the focus point set. For example, the large language model can identify the intent of the task data from each focus point separately, thereby obtaining the candidate intent results output by the large language model at each focus point. For each of the multiple focus points, the large language model can output a corresponding candidate intent result, thereby obtaining multiple candidate intent results corresponding one-to-one to the multiple focus points.

[0009] Finally, after obtaining multiple candidate intent results, the multiple candidate intent results can be combined to determine the target intent result. For example, the intent result that appears more frequently among the multiple candidate intent results can be used as the final target intent result.

[0010] In this solution, the large language model analyzes multiple focus points related to the task data input by the user, and prompts the large language model to identify the user intent corresponding to the task data under different focus points. Finally, the candidate intent results obtained under multiple focus points are combined to determine the final target intent result. In this solution, by analyzing multiple focus points related to the task data, different roles are set for the large language model under different focus angles, allowing the large language model to identify user intent from different focus angles. This overcomes the illusion problem inherent in the large language model (i.e., the problem of inaccurate results caused by excessive consideration of factors other than the task when identifying intent), and can ensure the accuracy of the output target intent result. In addition, it maximizes the general knowledge reserve and semantic understanding capabilities of the existing large language model, eliminating the need for manual provision of rules or training data, effectively improving the efficiency of intent recognition.

[0011] In one possible implementation, after obtaining the task data, the intent tree and the task data may be input into the large language model to obtain multiple focus points output by the large language model. Among them, the intent tree may be pre-built according to the current intent recognition scenario, and can indicate the various intentions that the user may express in the current intent recognition scenario. The intent tree includes a plurality of parent intent types, and each parent intent type includes a plurality of sub-class intent types. The large language model is used to identify the target parent intent type to which the task data belongs, and to determine multiple focus points based on the sub-class intent type under the target parent intent type and the semantics of the task data. In other words, the intent tree is used to assist the large language model in identifying the intent that the task data may correspond to, so that the large language model can determine multiple focus points for the task data based on the intent that the task data may correspond to.

[0012] In this solution, by inputting the intent tree in the current intent recognition scenario, the large language model can be assisted to make a good preliminary judgment on the intent of the task data, and then analyze and obtain accurate multiple focus points, which is conducive to improving the accuracy of the focus points output by the large language model.

[0013] In one possible implementation, in the process of determining the target intent result based on multiple candidate intent results, the reliability of the large language model in performing intent recognition can be determined based on the probability of occurrence of each intent result in the multiple candidate intent results. Each of the multiple candidate intent results corresponds to an intent result. Different candidate intent results can correspond to the same intent result or to different intent results. The probability of occurrence of each intent result is the ratio between the number of candidate intent results corresponding to each intent result and the total number of multiple candidate intent results. That is, the more candidate intent results corresponding to the same intent result, the greater the probability of occurrence of such an intent result.

[0014] When the degree of credibility meets the preset conditions, for example, when the degree of credibility reaches a certain height, a target intent result is determined from multiple candidate intent results, and the target intent result corresponds to an intent result with the highest probability of occurrence.

[0015] In this solution, the credibility of the large language model in performing intent recognition is determined based on the multiple candidate intent results output by the large language model, and the final target intent result is determined only when the credibility of the large language model in performing intent recognition is high, thereby ensuring the accuracy of the identified intent result.

[0016] In one possible implementation, in the process of determining a target intent result based on multiple candidate intent results, when the degree of credibility does not meet a preset condition, a negotiation mechanism can be used to adjust some of the multiple candidate intent results, and the target intent result can be determined based on the adjusted candidate intent results. Specifically, when multiple candidate intent results have different intent results, a negotiation mechanism can be used to adjust the focus point based on which the large language model identifies the intent result, thereby adjusting some of the multiple candidate intent results.

[0017] In one possible implementation, the large language model is further used to identify priorities among multiple concerns based on task data.

[0018] In the process of determining a target intent result based on multiple candidate intent results, if the credibility does not meet a preset condition, a target focus point can be first determined from the multiple focus points. The candidate intent result corresponding to the target focus point is different from the candidate intent result corresponding to the first focus point, and the first focus point is the focus point with the highest priority among the multiple focus points.

[0019] Then, the large language model is used to identify the intent of the task data from both the first and target points of interest, obtaining an adjusted candidate intent result corresponding to the target point of interest. In other words, the large language model identifies the intent of the task data from both points of interest.

[0020] Secondly, based on the adjusted candidate intent results and the unadjusted candidate intent results among the multiple candidate intent results, the credibility of the large language model in performing intent recognition is re-determined. That is to say, after adding new focus points and obtaining the adjusted candidate intent results, the distribution of the candidate intent results output by the large language model will change. Therefore, the credibility of the large language model in performing intent recognition can be re-determined based on the adjusted candidate intent results and the unadjusted candidate intent results.

[0021] In addition, when the re-determined degree of credibility meets the preset conditions, the target intention result is determined from the adjusted candidate intention results and the unadjusted candidate intention results.

[0022] In this solution, when setting different roles for the large language model based on different concerns, the priorities between the different roles of the large language model are also set based on the priorities between the concerns. When the intent result output by the role with lower priority is different from the intent result output by the role with higher priority, the role with lower priority can be made to consider both the lower priority concerns and the higher priority concerns at the same time, so as to output the corresponding intent result again, thereby realizing negotiation between the large language models playing different roles. Moreover, the key to the negotiation mechanism is that the large language model with lower role priority needs to pay more attention to the upper-level higher priority concerns, so as to realize the recognition of the intent of the task data while paying attention to both the higher priority concerns and the lower priority concerns.

[0023] In one possible implementation, when the re-determined credibility does not meet the preset conditions, and the adjusted candidate intent result is different from the candidate intent result corresponding to the first focus point, the large language model is used to identify the intent of the task data from the first focus point, the second focus point, and the target focus point at the same time, and the candidate intent result corresponding to the target focus point is obtained after secondary adjustment, where the second focus point is the focus point with the second highest priority among the multiple focus points. That is to say, for the target focus point, if the target focus point and the first focus point are considered at the same time, and the final re-determined credibility still does not meet the conditions, and the adjusted candidate intent result is still different from the candidate intent result corresponding to the first focus point, then the focus points with higher priority that need to be considered (i.e., the second focus point) can be further added, thereby continuing to adjust the intent result output by the large language model under the role corresponding to the target focus point.

[0024] Based on the secondarily adjusted candidate intent results and the remaining unadjusted candidate intent results from the multiple candidate intent results, the credibility of the large language model in performing intent recognition is re-determined. Specifically, after adding the focus points that the large language model needs to consider for a specific role and outputting the corresponding adjusted intent results, the credibility of the large language model in performing intent recognition is again evaluated based on the intent results output by the large language model for different roles, in order to determine whether the final target intent result can be determined.

[0025] In this solution, when the intent results output by the large language models of different roles differ significantly, multiple rounds of negotiation and adjustment are performed between the large language models of different roles. Each round of negotiation and adjustment increases the focus of the large language models with lower role priorities, thereby adjusting the intent results output by these large language models. This ensures that the large language models of different roles ultimately reach a consensus as much as possible, ensuring a highly reliable intent result.

[0026] In one possible implementation, when determining the reliability of intent recognition performed by a large language model, an uncertainty value for intent recognition performed by the large language model can be determined based on the probability of occurrence of each intent result through information entropy. The uncertainty value is used to indicate the reliability. The preset condition is that the uncertainty value is less than or equal to a target threshold.

[0027] In this scheme, by matching the scattered features of the output results of the large language model with information entropy, the concept of information entropy is introduced to measure the uncertainty of the large language model in performing intention recognition. This can well realize the judgment of the credibility of the large language model in performing intention recognition, which is conducive to finally obtaining a result with higher credibility.

[0028] In one possible implementation, when outputting the target intent result, one or more focus points corresponding to the target intent result may also be displayed, where the candidate intent results corresponding to the one or more focus points are all the same as the target intent result. In other words, how the target intent result was identified is displayed, namely, which focus points were considered to achieve the target intent result.

[0029] In this solution, by displaying one or more focus points corresponding to the target intention result, the logical reasoning path of intention recognition can be clearly displayed, thereby intuitively explaining the reasons for obtaining the intention result and improving the friendliness of human-computer interaction.

[0030] In one possible implementation, multiple focus points are obtained by instructing the large language model to identify requirements expressed in the task data. Generally speaking, in addition to indicating the task to be completed, the task data entered by the user may also express additional user requirements for how to complete the task. Therefore, by instructing the large language model to obtain focus points by identifying the requirements expressed in the task data, the large language model can output more accurate focus points.

[0031] In one possible implementation, the method further includes: executing the task indicated by the task data based on the target intent result and the task data. For example, if the target intent result is to book a flight ticket, an operation of booking a flight ticket for the user may be executed, thereby completing the task desired by the user.

[0032] In a second aspect, an intention recognition device is provided, which includes various modules for executing the intention recognition method in the first aspect or any possible implementation of the first aspect.

[0033] In a third aspect, an intent recognition device is provided, comprising: a processor and a memory; the memory is used to store computer instructions, and when the processor executes the instructions, the intent recognition device executes any of the methods described above.

[0034] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer can execute any of the above methods.

[0035] In a fifth aspect, a computer program product comprising instructions is provided, which, when the instructions are executed on a computer, enables the computer to execute any of the methods described above.

[0036] In the sixth aspect, a chip system is provided, which includes a processor and a communication interface, wherein the communication interface is used to communicate with modules outside the chip shown, and the processor is used to run computer programs or instructions so that the device installed with the chip system can execute any of the methods mentioned above.

[0037] In the seventh aspect, an intent recognition device is provided, comprising a chip system as described in the sixth aspect above, the chip system comprising a processor and a communication interface, the communication interface being used to communicate with modules outside the chip system shown, the processor being used to run computer programs or instructions so that the intent recognition device can execute the method of any of the above aspects.

[0038] In an eighth aspect, an intent recognition device cluster is provided, comprising at least one intent recognition device as described in the third or seventh aspect, wherein any one of the intent recognition devices is configured to execute a computer program or instruction, enabling the intent recognition device cluster to perform any of the aforementioned methods. Alternatively, some or all of the intent recognition devices are configured to execute a computer program or instruction, enabling the intent recognition device cluster to perform any of the aforementioned methods.

[0039] Based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of the intent recognition process provided by this application;

[0041] Figure 2 A schematic diagram of a system architecture 100 provided in this application;

[0042] Figure 3 A flowchart of an intent recognition method provided in this application;

[0043] Figure 4 A schematic diagram of determining multiple focus points through a large language model provided by this application;

[0044] Figure 5 A schematic diagram of the structure of an intent tree provided for this application;

[0045] Figure 6 A schematic diagram of a large language model outputting candidate intent results provided by this application;

[0046] Figure 7 A schematic diagram of a multi-stage implementation flow of an intent recognition method provided in this application;

[0047] Figure 8 A large language model provided by this application performs the process of intent recognition;

[0048] Figure 9 A flowchart of a multi-role negotiation process provided by this application;

[0049] Figure 10 A schematic diagram of the intention result obtained after performing multiple rounds of role negotiation provided by this application;

[0050] Figure 11 A schematic diagram of calculating the uncertainty value of intent recognition performed by a large language model provided in this application;

[0051] Figure 12 A schematic diagram of the structure of an intention recognition device provided in this application;

[0052] Figure 13 A schematic diagram of the structure of a chip provided in this application;

[0053] Figure 14 A schematic diagram of the structure of a computing device 1400 provided in this application;

[0054] Figure 15 A schematic diagram of the structure of a computing device cluster provided in this application;

[0055] Figure 16 A schematic diagram of the structure of another computing device cluster provided in this application;

[0056] Figure 17 A schematic diagram of the structure of a computer-readable storage medium provided in this application. DETAILED DESCRIPTION

[0057] To facilitate understanding, some technical terms involved in this application are first introduced below.

[0058] (1) Human-computer dialogue

[0059] Also known as natural language understanding, this is a key research area within the emerging frontier discipline of artificial intelligence. It studies how to enable computers to understand and use natural languages ​​such as Chinese and English, enabling natural language communication between humans and machines. This allows computers to replace some human mental work, truly extending the human brain.

[0060] During a human-computer dialogue, the computer may request answers to questions, provide parameters, or confirm options. Through this dialogue, the human guides or limits the computer's work and oversees its execution. This approach facilitates incorporating human intention, judgment, and experience into the computer's operational processes, enhancing the flexibility of computer applications and facilitating software development.

[0061] (2) Intent Recognition

[0062] Intent recognition is a method for detecting user intent based on multimedia data such as text or speech. Intent recognition methods are typically used to identify user input text and then match the text with the system's knowledge base to determine the user's intent.

[0063] For example, see Figure 1 , Figure 1 This is a process diagram of intent recognition provided by this application. Figure 1 As shown, the intent recognition module is the execution entity for realizing intent recognition. When the user inputs the text "Help me navigate home", the intent recognition module can analyze the text input by the user to determine that the user's intention is "to navigate a route from the current location to No. XX, XX Road, XX District, XX City".

[0064] (3) Large language model (LLM)

[0065] Large language models are deep learning models trained using large amounts of text data. They can generate natural language text or understand the meaning of text. Large language models can handle a variety of natural language tasks, such as text classification, question-answering, and conversation, and are an important path to artificial intelligence.

[0066] Specifically, large language models are a technology that has emerged in recent years. Because they undergo sophisticated data engineering and training processes, their parameters already incorporate a wealth of existing natural language processing knowledge. This knowledge can already replace humans in many language-related tasks, such as having large language models write code or perform text summarization.

[0067] (4) Information entropy

[0068] Information entropy is a basic concept in information theory, which describes the uncertainty of possible events in the information source.

[0069] In the information source, we should consider not the uncertainty of a single symbol, but the average uncertainty of all possible situations of the information source. If the information source symbol has n values: U1…U i …U n, and the corresponding occurrence probabilities of these source symbols are: P1…P i …P n , and the occurrence of various symbols is independent of each other. In this case, the average uncertainty of the source should be the uncertainty of a single symbol - logP i The statistical average value (E) of can be called information entropy, that is, the information entropy can be expressed as the following formula.

[0070]

[0071] Among them, H(U) represents information entropy, which refers to the uncertainty of the occurrence of each possible event of the information source.

[0072] In human-computer interaction systems, intent recognition is a key technology for achieving this. Its primary task is to identify user intent based on input. Currently, intent recognition techniques primarily include rule-based and statistical approaches.

[0073] Rule-based intent recognition identifies user intent by defining a set of rules. These rules are typically written by experts based on domain knowledge and experience. For example, in a weather app, rules might include "Enter a location, get weather information for that location," "Enter a date, get weather information for that date," and so on. While rule-based intent recognition is simple and easy to understand, suitable for specific domains, it requires a large number of manually written rules, making it difficult to scale to more complex scenarios.

[0074] Statistical intent recognition methods analyze historical data to learn patterns in user input and the corresponding intent. Common methods include: Hidden Markov Model-based methods (predicting user intent by learning transition probabilities and observation probabilities between states), Conditional Random Field-based methods (predicting user intent by learning dependencies between features), and Recurrent Neural Network-based methods (predicting user intent by learning hidden states from sequence data). In summary, statistical intent recognition methods have the advantage of automatically learning rules and are applicable to more complex scenarios. However, they require a large amount of historical data for training, may require manual annotation, and may have poor generalization capabilities in new scenarios.

[0075] In general, current intent recognition technology requires humans to design a large number of rules or provide a large amount of training data for specific intent recognition scenarios, resulting in low efficiency in intent recognition and difficulty in quickly realizing intent recognition in various new scenarios.

[0076] In view of this, the present application provides an intent recognition method, which uses a large language model to analyze multiple focus points related to the task data input by the user, and prompts the large language model to identify the user intent corresponding to the task data under different multiple focus points, and finally integrates the candidate intent results obtained under multiple focus points to determine the final target intent result. In this solution, by analyzing multiple focus points related to the task data, roles under different focus angles are set for the large language model, so that the large language model can recognize user intent from different focus angles, overcome the illusion problem of the large language model itself, and ensure the accuracy of the output target intent result; and, it maximizes the use of the general knowledge storage and semantic understanding capabilities of the existing large language model, without the need for manual provision of rules or training data, effectively improving the efficiency of intent recognition.

[0077] See also Figure 2 , Figure 2 This is a schematic diagram of a system architecture 100 provided in this application. Figure 2 As shown, in this system architecture 100, execution device 110 can be implemented by one or more servers. Optionally, execution device 110 cooperates with other computing devices, such as data storage, routers, and load balancers. Execution device 110 can be deployed at a single physical site or distributed across multiple physical sites. Execution device 110 can use data in data storage system 120 or invoke program code in data storage system 120 to implement the intent recognition method provided herein.

[0078] Users can operate their respective user devices (such as local device 101 and local device 102) to interact with execution device 110. Each local device can represent any computing device, such as a personal computer, a computer workstation, a smart phone, a tablet computer, a laptop computer, and a smart car.

[0079] Each user's local device can interact with the execution device 110 through a communication network of any communication mechanism / communication standard. The communication network can be a wide area network, a local area network, a point-to-point connection, etc., or any combination thereof.

[0080] In one implementation, the execution device 110 is used to implement the intent recognition method provided in this application to identify the user's intent. Furthermore, when the local device 101 and the local device 102 need to identify the user's intent, the execution device 110 processes the task data input by the user based on the intent recognition method, and after identifying the user's intent, executes the task indicated by the user's intent and returns the corresponding processing results to the local device 101 and the local device 102.

[0081] In another implementation, one or more aspects of the execution device 110 can be implemented by each local device. For example, the local device 101 can provide local data or feedback calculation results to the execution device 110, or execute the intent recognition method provided in this application.

[0082] In general, the intent recognition method provided in this application can be applied to electronic devices, such as the above-mentioned execution device 110, local device 101 or local device 102.

[0083] In actual application scenarios, the intent recognition method provided in this application can be applied to a built-in module in a human-computer dialogue system, or it can be used to implement a separate intent recognition service to provide services for human-computer dialogue.

[0084] See also Figure 3 , Figure 3 This is a flow chart of an intention recognition method provided by this application. Figure 3 As shown, the intention recognition method provided by this application includes the following steps 301-304.

[0085] Step 301 : Acquire task data input by a user. The task data is multimedia data, and the task data is used to indicate a task that the user desires to complete.

[0086] In this application, during the interaction between a user and an electronic device, the electronic device will obtain task data input by the user, and the task data can indicate the task that the user wants the electronic device to complete. The electronic device may be, for example, a cloud server, and the user interacts with the server via a network to send the task data to the server. The electronic device may also be a device used by the user, such as a smartphone, smart car, tablet computer, or laptop computer, and the user sends the task data to the electronic device by inputting text or voice on the electronic device.

[0087] Among them, the task data in this application is multimedia data, such as one or more of the following data: text, voice, image or video. For example, when a user interacts with a ticket booking system with an intent recognition function, the user inputs the above-mentioned task data to the ticket booking system by inputting text; for another example, when a user interacts with a voice assistant with an intent recognition function, the user inputs task data to the voice assistant by inputting voice; for another example, when a user interacts with a digital assistant with an intent recognition function, the user inputs task data to the digital assistant by inputting text + image or text + video. In general, this application does not limit the specific form of task data. For the sake of convenience, the following will take the task data input by the user as text as an example to introduce the intent recognition method of this application.

[0088] In step 302 , a focus set for the task data is identified using a large language model. The focus set includes multiple focus points, each of which is used to indicate a constraint condition that needs to be satisfied when executing the task.

[0089] It is understandable that there are often multiple ways to complete the task that the user desires to complete in the task data. Furthermore, when considering how to complete the task from different perspectives, the conditions that must be met for completing the task are also different, so the ways to complete the task may also often be different.

[0090] For example, if the user's desired task is to book a ticket, the options for completing the task include booking a plane ticket, a train ticket, a ferry ticket, or a bus ticket. Considering arrival time, the specific requirement for completing the task is the earliest arrival time, so the option for completing the task could be booking a plane ticket. Considering convenience, the specific requirement for completing the task is the minimum number of transfers, so the option for completing the task could be booking a plane ticket. Considering cost, the specific requirement for completing the task is the lowest fare, so the option for completing the task could be booking a bus ticket.

[0091] Therefore, in this application, the powerful semantic understanding ability of the large language model can be utilized to identify multiple focus points for task data through the large language model, thereby obtaining a focus point set including multiple focus points. Among them, each of the multiple focus points indicates the constraints that need to be met when performing the task, and the constraints indicated by different focus points are different. Multiple focus points can also be understood as an angle to be considered when performing the task, and different focus points indicate different angles of consideration. When the angle considered when performing the task is fixed, the constraints satisfied when completing the task are usually fixed, so each of the multiple focus points can also be understood as a constraint to be satisfied when determining the task.

[0092] Among them, the constraints that need to be met when executing the task (the angles considered when executing the task) are often closely related to the task itself. Therefore, in this application, a large language model is used to understand the tasks indicated in the task data, and then multiple focus points for the task data are further output.

[0093] For example, see Figure 4 , Figure 4 This application provides a schematic diagram of determining multiple focus points through a large language model. Figure 4As shown, when the task data is "help me book a ticket from Shenzhen to Beijing as soon as possible", the task indicated by the task data is specifically booking a ticket. Therefore, the task data can be combined with the instruction expression "please identify the focus points in the task data" to input into the large language model to instruct the large language model to identify multiple focus points for the task data. The multiple focus points finally output by the large language model may include focus points 1-3, for example. Among them, focus point 1 is the arrival time of the vehicle (that is, the constraint condition when performing the task is to have the shortest possible arrival time), focus point 2 is the convenience of the vehicle (that is, the constraint condition when performing the task is to have the highest possible convenience), and focus point 3 is the cost of the vehicle (that is, the constraint condition when performing the task is to have the lowest possible cost). Obviously, different focus points represent the constraints that need to be met when booking a ticket, or different angles that need to be considered when booking a ticket.

[0094] When the task data is "Help me navigate home", the task indicated by the task data is specifically to navigate a route from the current location to the user's home location. The multiple focus points output by the large language model may include focus points 1 to 4, for example. Focus point 1 is the travel time of the route, focus point 2 is the distance of the route, focus point 3 is the congestion level of the route, and focus point 4 is the highway toll incurred by the route. Different focus points represent different constraints that need to be met when navigating the route.

[0095] Optionally, the aforementioned multiple focus points are obtained by instructing the large language model to identify requirements expressed in the task data. Generally speaking, in addition to indicating the task to be completed, the task data input by the user may also express some additional user requirements for how to complete the task. Therefore, by instructing the large language model to obtain focus points by identifying the requirements expressed in the task data, the large language model can output more accurate focus points.

[0096] Specifically, in this application, the task data and the instruction text may be simultaneously input into the large language model, wherein the instruction text is used to instruct the large language model to recognize the requirements expressed in the task data and output the recognized focus points.

[0097] For example, if the user inputs the task data "Book me a ticket from Shenzhen to Beijing as quickly as possible," the text input to the large language model could be: "Please analyze the semantic needs expressed in the task data 'Book me a ticket from Shenzhen to Beijing as quickly as possible.'" The large language model then outputs the following multiple needs: 1. Time; 2. Destination and departure points; 3. Ticket type.

[0098] Optionally, in order to ensure that the large language model can accurately identify multiple focus points related to the task data, the intent tree and task data can be input into the large language model to obtain multiple focus points output by the large language model. Among them, the intent tree can be pre-built according to the current intent recognition scenario, and can indicate the various intentions that the user may express in the current intent recognition scenario. Specifically, the intent tree includes multiple parent intent types, and each parent intent type includes multiple sub-class intent types. The large language model is used to identify the target parent intent type to which the task data belongs, and determine multiple focus points based on the sub-class intent type under the target parent intent type and the semantics of the task data.

[0099] That is, the intent tree is used to assist the large language model in identifying the intent that the task data may correspond to, thereby enabling the large language model to determine multiple focus points for the task data based on the intent that the task data may correspond to.

[0100] For example, see Figure 5 , Figure 5 This is a schematic diagram of the structure of an intent tree provided by this application. Figure 5 As shown, in the intent tree, booking tickets and ordering takeout are both parent intent types, and each parent intent type includes multiple sub-intent types. Under the parent intent type of booking tickets, there are four sub-intent types: booking plane tickets, booking train tickets, booking boat tickets, and booking bus tickets; under the parent intent type of ordering takeout, there are three sub-intent types: ordering flowers, ordering fruits, and ordering meals. For example, when the task data is "Help me book a ticket from Shenzhen to Beijing as soon as possible", the large language model can identify that the parent intent type of the task data is booking tickets, and determine the four sub-intent types under booking tickets. In this way, based on the semantic information of the task data itself, and the four sub-intent types of booking plane tickets, booking train tickets, booking boat tickets, and booking bus tickets, the following multiple points of interest can be determined: 1. Time; 2. Destination and departure; 3. Ticket type.

[0101] In this solution, by inputting the intent tree in the current intent recognition scenario, the large language model can be assisted to make a good preliminary judgment on the intent of the task data, and then analyze and obtain accurate multiple focus points, which is conducive to improving the accuracy of the focus points output by the large language model.

[0102] In step 303 , the intent of the task data is identified from the focus points in the focus point set through the large language model to obtain multiple candidate intent results.

[0103] In this application, after obtaining a set of focus points including multiple focus points, the large language model can be instructed to identify the intent of the task data based on each focus point individually, thereby obtaining candidate intent results output by the large language model for each focus point. In other words, for each of the multiple focus points, the large language model can output a corresponding candidate intent result, thereby obtaining multiple candidate intent results that correspond one-to-one with the multiple focus points.

[0104] Specifically, in this application, each focus point can be combined with the task data and input into the large language model in sequence, and the large language model can be instructed to identify the intention of the task data based on the input focus point; in this application, all focus points can be combined with the task data and input into the large language model at one time, and the large language model can be instructed to identify the intention of the task data from each focus point respectively.

[0105] For example, see Figure 6 , Figure 6 This is a schematic diagram of a large language model outputting candidate intent results provided by this application. Figure 6 As shown, when the task data is "Help me book a ticket from Shenzhen to Beijing as soon as possible", and the multiple focus points output by the large language model are time, convenience, and cost, the text input into the large language model for the first time is specifically "Please focus on the time requirement. What is the intention of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible?"?" The candidate intent result 1 output by the large language model is "The intention of booking a plane ticket." The text input into the large language model for the second time is specifically "Please focus on the convenience requirement. What is the intention of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible?" The candidate intent result 2 output by the large language model is "The intention of booking a plane ticket." The text input into the large language model for the third time is specifically "Please focus on the cost requirement. What is the intention of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible?" The candidate intent result 3 output by the large language model is "The intention of booking a bus ticket."

[0106] In general, in this step, different roles can be set for the large language model based on different focus points. The different roles of the large language model represent the starting point of the large language model when identifying intent (that is, the large language model identifies intent from the focus point corresponding to the current role), thereby obtaining the candidate intent results identified under different roles.

[0107] Step 304: Determine the target intent result among multiple candidate intent results.

[0108] After obtaining multiple candidate intent results, the target intent result can be determined by combining the multiple candidate intent results. It is understandable that in this embodiment, each candidate intent result is obtained by identifying task data based on a corresponding focus point, and the intent results obtained by identifying different focus points may be the same, that is, there may be some identical intent results among multiple candidate intent results. Therefore, for example, the intent result that appears more frequently among multiple candidate intent results can be used as the final target intent result.

[0109] For example, in Figure 5 In the example, candidate intent result 1 is the intent to book a plane ticket, candidate intent result 2 is the intent to book a plane ticket, and candidate intent result 3 is the intent to book a bus ticket. Candidate intent results 1 and 2 are the same intent result, and the intent to book a plane ticket appears the most times (twice) among the three candidate intent results. Therefore, the intent to book a plane ticket can be selected as the final target intent result.

[0110] Optionally, after obtaining the target intent result, the task indicated by the task data can be executed based on the target intent result and the task data. For example, if the target intent result is to book a plane ticket, the operation of booking a plane ticket for the user can be executed, thereby completing the task desired by the user.

[0111] In this application, for the task data input by the user, a large language model is used to analyze multiple focus points related to the task data, and the large language model is prompted to identify the user intent corresponding to the task data under different multiple focus points, and finally the candidate intent results obtained under multiple focus points are integrated to determine the final target intent result. In this solution, by analyzing multiple focus points related to the task data, roles under different focus angles are set for the large language model, so that the large language model can identify the user intent from different focus angles, overcome the illusion problem of the large language model itself (that is, when identifying intent, too many factors other than the task are considered, resulting in inaccurate results), and can ensure the accuracy of the output target intent results. It is understandable that since the large language model uses a large amount of corpus for training in the training phase, the large language model will learn a large amount of knowledge from different angles in the same scenario for a large number of scenarios. Therefore, when the large language model outputs the result, it often gives irrelevant answers due to the excessive knowledge learned, that is, it answers questions from other angles that are not very relevant to the input corpus, resulting in illusion problems.

[0112] Furthermore, this solution maximizes the general knowledge and semantic understanding capabilities of existing large language models, eliminating the need for manual rule provision or training data, effectively improving the efficiency of intent recognition. Specifically, existing solutions typically require the pre-design of numerous rules or the preparation of extensive training data, and when expanding into new business scenarios, additional rule design or training data preparation is also required. This solution, however, leverages the inherent knowledge reserves of large language models to achieve intent recognition in a variety of scenarios without the need for rule design or training data preparation, effectively improving the efficiency of intent recognition.

[0113] In this application, the distribution of multiple candidate intent results output by the large language model may vary. For example, the multiple candidate intent results identified by the large language model under different focus points may have small or large differences from each other, that is, the intent results output by the large language model under different roles may have small or large differences.

[0114] When the differences between the multiple candidate intent results output by the large language model are small, it can be considered that the intent results output by the large language model under different roles have small differences, and therefore the credibility of the multiple candidate intent results currently output by the large language model is relatively high; when the differences between the multiple candidate intent results output by the large language model are large, it can be considered that the intent results output by the large language model under different roles have large differences, and therefore the credibility of the multiple candidate intent results currently output by the large language model is not high.

[0115] Based on this, this application can also determine the credibility of the large language model in performing intent recognition based on the multiple candidate intent results output by the large language model, and only when the credibility of the large language model in performing intent recognition is high, determine the final target intent result to ensure the accuracy of the identified intent result.

[0116] Exemplarily, in the process of determining the target intent result based on multiple candidate intent results, the degree of credibility of the large language model in performing intent recognition can be determined based on the probability of occurrence of each intent result. Each of the multiple candidate intent results corresponds to an intent result. Different candidate intent results can correspond to the same intent result or to different intent results. The probability of occurrence of each intent result is the ratio between the number of candidate intent results corresponding to each intent result and the total number of multiple candidate intent results. That is, the more candidate intent results corresponding to the same intent result, the greater the probability of occurrence of such an intent result.

[0117] For example, in Figure 6In the example, candidate intent result 1 and candidate intent result 2 both correspond to the same intent result (booking a plane ticket), while candidate intent result 3 corresponds to another intent result (booking a bus ticket). Therefore, the probability of the intent result "booking a plane ticket" appearing is 2 / 3, and the probability of the intent result "booking a bus ticket" appearing is 1 / 3.

[0118] Specifically, for multiple candidate intent results, since some of the candidate intent results may be identical (i.e., some of the candidate intent results belong to the same type of intent result), the number of different intent results actually corresponding to the multiple candidate intent results is less than or equal to the number of the multiple candidate intent results. By counting the probability of each intent result appearing in the candidate intent results, the probability of each intent result under different focus points can be determined, thereby determining the distribution of different types of intent results.

[0119] Generally speaking, the more dispersed the probability of occurrence of different types of intent results, the lower the credibility of the large language model in performing intent recognition; the more concentrated the probability of occurrence of different types of intent results, the higher the credibility of the large language model in performing intent recognition.

[0120] When the degree of credibility meets the preset conditions, for example, when the degree of credibility reaches a certain height, a target intent result can be determined from multiple candidate intent results. The target intent result corresponds to an intent result with the highest probability of occurrence.

[0121] Optionally, the degree of confidence of the large language model in performing intent recognition can be determined by referring to the concept of information entropy.

[0122] Exemplarily, after determining the probability of occurrence of each intent result based on multiple candidate intent results, the uncertainty value of the large language model performing intent recognition is determined based on the probability of occurrence of each intent result by obtaining information entropy. The uncertainty value can be used to indicate the degree of credibility of the large language model performing intent recognition. The larger the uncertainty value, the higher the uncertainty of the large language model performing intent recognition, and the lower the credibility of the large language model performing intent recognition; the smaller the uncertainty value, the lower the uncertainty of the large language model performing intent recognition, and the higher the credibility of the large language model performing intent recognition. That is, in this application, the uncertainty value is used to measure the credibility of the large language model performing intent recognition, and there is a negative correlation between the uncertainty value and the credibility.

[0123] For example, the uncertainty value of the large language model performing intent recognition can be determined based on the following formula.

[0124]

[0125] Among them, S represents the uncertainty value of the large language model performing intent recognition; C represents all types of intent results; P(c|x) represents the probability of occurrence of intent result c when the input is x.

[0126] In general, when the large language model performs intent recognition, the more unified the output intent results are (that is, the more consistent the opinions of the large language models of different roles are), the lower the uncertainty of the large language model in performing intent recognition, and the higher the credibility of the large language model in performing intent recognition.

[0127] Furthermore, when determining whether the degree of credibility satisfies a preset condition, the preset condition may be that the uncertainty value is less than or equal to a target threshold. The target threshold may be a threshold value determined or adjusted based on actual conditions, such as 0.1 or 0.2, and this application does not impose specific limitations on this.

[0128] In this scheme, by matching the scattered features of the output results of the large language model with information entropy, the concept of information entropy is introduced to measure the uncertainty of the large language model in performing intention recognition. This can well realize the judgment of the credibility of the large language model in performing intention recognition, which is conducive to finally obtaining a result with higher credibility.

[0129] Generally speaking, the requirements raised by users for tasks are divided into different levels of importance, that is, some requirements raised by users may be more important, while others may be less important. Therefore, the focus points obtained by analyzing the requirements expressed in the task data through the large language model can also be divided into priorities. The higher the priority of the focus point, the more the user pays attention to the conditions under the focus point that must be met when completing the task; the lower the priority of the focus point, the less the user pays attention to the conditions under the focus point that must be met when completing the task. In this way, by identifying the priorities between the focus points, when the large language model outputs different candidate intent results for different focus points, the coordination of the candidate intent results can be achieved, so as to improve the credibility of the large language model in performing intent recognition.

[0130] Specifically, when the large language model identifies focus points for task data, it is also used to identify priorities among multiple focus points based on the task data. Thus, when the large language model outputs multiple focus points for task data, it also outputs the priority of each of the multiple focus points.

[0131] In this way, in the process of determining the target intent result based on multiple candidate intent results, when the degree of credibility does not meet the preset conditions, the target focus point can be first determined among multiple focus points. Among them, the candidate intent result corresponding to the target focus point is different from the candidate intent result corresponding to the first focus point, and the first focus point is the focus point with the highest priority among the multiple focus points. In other words, when the degree of credibility does not meet the preset conditions, the focus points whose corresponding candidate intent results are different from the focus point with the highest priority are preferentially determined, so as to facilitate the subsequent adjustment of the candidate intent results corresponding to these focus points.

[0132] Then, the large language model is used to identify the intent of the task data from both the first and target perspectives, yielding an adjusted candidate intent result corresponding to the target perspective. That is, the large language model identifies the intent of the task data from both perspectives simultaneously. For example, if the first perspective is time and the target perspective is the booking type, the text input into the large language model, when combining the first and target perspectives, might be something like, "Please consider the time and booking type requirements. What is the intent of the task 'Book me a ticket from Shenzhen to Beijing as soon as possible'?"

[0133] Secondly, based on the adjusted candidate intent results and the unadjusted candidate intent results among the multiple candidate intent results, the credibility of the large language model in performing intent recognition is re-determined. In other words, after the new focus is considered and the adjusted candidate intent results are obtained, the distribution of the candidate intent results output by the large language model will change. Therefore, the credibility of the large language model in performing intent recognition can be re-determined based on the adjusted candidate intent results and the unadjusted candidate intent results to determine whether the credibility meets the preset conditions.

[0134] Finally, when the re-determined credibility meets the preset conditions, the target intention result is determined from the adjusted candidate intention results and the unadjusted candidate intention results.

[0135] It should be noted that the above embodiment is only an example of the process of re-identifying the intent for a target focus point among multiple focus points. In actual applications, there may be multiple focus points with different candidate intent results from the focus point with the highest priority. For these focus points, the process described in the above embodiment needs to be executed, that is, while considering these focus points separately, the first focus point also needs to be considered, so as to achieve the redetermination of the candidate intent results and the redetermination of the credibility of the large language model in performing intent identification.

[0136] In general, in this solution, when setting different roles for the large language model based on different concerns, the priorities between the different roles of the large language model are also set based on the priorities between the concerns. When the intent result output by the role with lower priority is different from the intent result output by the role with higher priority, the role with lower priority can be made to consider the concerns with lower priority and the concerns with higher priority at the same time, so as to output the corresponding intent result again, and realize the negotiation between the large language models playing different roles. Moreover, the key to the negotiation mechanism is that the large language model with lower role priority needs to pay more attention to the concerns with higher priority in the upper layer, so as to realize the recognition of the intent of the task data while paying attention to the concerns with higher priority and lower priority at the same time.

[0137] That is to say, when there are large differences in the performance of intent recognition by different roles played by the large language model (that is, there are certain differences in the results output by different roles), the negotiation mechanism can achieve communication, reflection and adjustment between the large language models of different roles, so that the large language model can re-execute intent recognition, effectively ensuring the focus on task data during the intent recognition process, and fully and specifically understanding the user's intention.

[0138] Optionally, when the re-determined credibility still does not meet the preset conditions, and the adjusted candidate intent result is different from the candidate intent result corresponding to the first focus point, the large language model can be used to identify the intent of the task data from the first focus point, the second focus point and the target focus point at the same time, and the candidate intent result corresponding to the target focus point after secondary adjustment is obtained, where the second focus point is the focus point with the second highest priority among the multiple focus points. In other words, for the target focus point, if the target focus point and the first focus point are considered at the same time, the final re-determined credibility still does not meet the conditions, and the adjusted candidate intent result is still different from the candidate intent result corresponding to the first focus point, then the focus points with higher priority that need to be considered (i.e., the second focus point) can be continued to be added, thereby continuing to adjust the intent result output by the large language model under the role corresponding to the target focus point.

[0139] Then, based on the secondarily adjusted candidate intent results and the remaining unadjusted candidate intent results from the multiple candidate intent results, the credibility of the large language model in performing intent recognition is re-determined. That is, after adding the focus points that the large language model needs to consider for a specific role and outputting the corresponding adjusted intent results, the credibility of the large language model in performing intent recognition is again evaluated based on the intent results output by the large language model for different roles, in order to determine whether the final target intent result can be determined.

[0140] In general, when the intent results output by large language models in different roles differ significantly, multiple rounds of negotiation and adjustment are performed between the large language models of different roles. Each round of negotiation and adjustment increases the focus of large language models with lower role priorities, thereby adjusting the intent results output by these large language models. This ensures that the large language models of different roles ultimately reach a consensus as much as possible, ensuring a highly reliable intent result.

[0141] Optionally, in order to improve the interpretability of the target intent result of the final output, one or more focus points corresponding to the target intent result can be displayed, where the candidate intent results corresponding to the one or more focus points are the same as the target intent result.

[0142] That is to say, when outputting the target intent result, this application can also show how to identify the target intent result, that is, from which focus points to consider, the target intent result can be obtained. Figure 6 In the example, the target intent is the intent to book a flight, and this intent is recognized when both time and convenience are considered. Therefore, in practical applications, the following text can be displayed: "When considering the task data from the perspectives of 'time' and 'convenience', the intent to book a flight is recognized."

[0143] In this solution, by displaying one or more focus points corresponding to the target intention result, the logical reasoning path of intention recognition can be clearly displayed, thereby intuitively explaining the reasons for obtaining the intention result and improving the friendliness of human-computer interaction.

[0144] Optionally, while displaying one or more focus points corresponding to the target intent result, multiple candidate intent results and the focus points corresponding to each candidate intent result in the multiple candidate intents can also be displayed. That is, each focus point considered by the large language model and the intent results identified under each focus point are displayed, so that users can intuitively understand the entire logical reasoning process of intent recognition performed by the large language model, thereby improving human-computer interaction capabilities. In addition, for scenarios where the semantics of the task data entered by the user are unclear, it can also effectively support multiple rounds of dialogue, providing better support for subsequent intent clarification.

[0145] The above introduces the intention recognition method provided by this application. For ease of understanding, the implementation process of the intention recognition method provided by this application will be described in detail below with reference to specific examples.

[0146] See also Figure 7 , Figure 7 This is a multi-stage implementation flow chart of an intent recognition method provided by this application. Figure 7As shown, the execution process of the intent recognition method provided by this application can be divided into the following multiple stages.

[0147] Phase 1: Identification of focus points.

[0148] In this application, the input for intent recognition is task data entered by the user in the form of multimedia data such as text or voice, for example, "Help me book a ticket from Shenzhen to Beijing as soon as possible." Based on the user input, attention points are first identified using a large language model.

[0149] Specifically, a pre-built intent tree and the task data entered by the user are fed into the large language model, which then outputs the corresponding focus points. First, the large language model matches the parent intent in the intent tree (e.g., booking a ticket) based on the semantics of the task data, thereby identifying multiple sub-intents within the parent intent (e.g., booking a plane ticket, booking a train ticket, booking a bus ticket, etc.). Then, based on the parent intent and the semantics of the task data, the large language model decomposes the multiple focus points specific to the task data (e.g., time, departure and destination, and ticket type).

[0150] For example, see Figure 8 , Figure 8 The process of performing intent recognition on a large language model provided by this application. Figure 8 As shown, when implementing focus decomposition through a large language model, you can enter the prompt "Please decompose the following task focus points based on the intent tree and sort them by importance" into the large language model to instruct the large language model to generate focus points based on the task data and prioritize them according to their importance. For example, the large language model may output the following sorted focus points: O1 (time), O2 (transportation), and O3 (cost).

[0151] Phase 2: Multi-role intention recognition.

[0152] After decomposing multiple focus points through the large language model, different roles can be assigned to the large language model based on these focus points. The different roles of the large language model represent its starting points when determining intent. Each task data entered by a user represents a user demand, and these demands often manifest themselves in a variety of conditions at the semantic level. For example, when booking a ticket, various requirements, such as the arrival time of the transportation, convenience, and cost, are considered. Furthermore, these user demands often have a certain hierarchy and priority, so they can be reflected through multiple prioritized focus points.

[0153] When setting roles for a large language model, each role uniquely corresponds to a focus, and the priority of the role is the same as the priority of the focus to which it corresponds. In this way, based on multiple focus points with priorities, multiple roles with priorities can be set. Figure 8 As shown, for the focus point O1 (time), role R1 can be set, and role R1 is used to focus on time demands; for the focus point O2 (transportation), role R2 can be set, and role R2 is used to focus on transportation demands; for the focus point O3 (cost), role R3 can be set, and role R3 is used to focus on cost demands.

[0154] After setting the corresponding roles for the large language model, the large language model under each role can be used to perform intent recognition on the task data respectively, so as to obtain the intent results output by the large language model under different roles, that is, the corresponding intent results under different focus points.

[0155] Stage three: judgment of intention certainty.

[0156] There may be many situations in the intention results output by the large language model under different roles. It is possible that the intention results output by multiple roles are relatively consistent, or it is possible that the intention results output by multiple roles are quite different. Therefore, this step proposes a measurement method to judge the certainty of the intention results. Specifically, in this step, the certainty of the large language model performing intention recognition can be judged by referring to the information entropy method, and the uncertainty value of the large language model performing intention recognition can be determined based on the probability of occurrence of various intention results. When the uncertainty value is less than or equal to the target threshold, it is considered that the large language model has a high degree of credibility in performing intention recognition and passes the intention certainty judgment; when the uncertainty value is greater than the target threshold, it is considered that the large language model has a low degree of credibility in performing intention recognition and does not pass the intention certainty judgment.

[0157] Stage four: multi-role negotiation.

[0158] In the case that the intent result output by the large language model fails to pass the intent certainty judgment, a multi-role negotiation mechanism is introduced in this step to solve the situation where the divergence of the intent results output by the large language model of different roles does not meet the threshold standard. Specifically, since the priority of each role matches the priority of the focus corresponding to the role, in the process of multi-role negotiation, the role with high priority can adopt a downward communication mechanism. That is, let the role with lower priority consider whether it has not paid attention to the focus of higher priority, thereby increasing the focus of the role with lower priority. In this way, the role with lower priority will often pay attention to the focus corresponding to itself and other focus points with higher priority at the same time. Then, the role with adjusted focus will re-execute intent recognition.

[0159] See also Figure 9 , Figure 9 This is a flowchart of a multi-role negotiation process provided by this application. Figure 9 As shown, multiple rounds of inter-role negotiation can be performed continuously between multiple roles. During round I of inter-role negotiation, if the intent result output by role Ri differs from that output by role R1, role Ri adds a new focus point j, where both i and j have an initial value of 1. This transforms Ri into Ri+j, meaning that role i+j must simultaneously focus on the original focus point i and the newly added focus point j, and then perform intent recognition on the task data. The value of j increases after each round of inter-role negotiation.

[0160] After a round of inter-role negotiation, the intent results output by some roles may change. Therefore, it is necessary to re-evaluate the intent based on the changed intent results and the original unchanged intent results. If the confidence level of intent recognition performed by the large language model meets the preset conditions, the intent result can be output. If the confidence level of intent recognition performed by the large language model still does not meet the preset conditions, the next round of inter-role negotiation will be triggered.

[0161] See also Figure 10 , Figure 10 This is a schematic diagram of the intention result obtained after executing multiple rounds of role negotiation provided by this application. Figure 10 As shown, during the first round of intent recognition, character R1's focus is O1, and the intent result output by character R1 is I(k1); character R2's focus is O2, and the intent result output by character R2 is I(k2); character R3's focus is O3, and the intent result output by character R3 is I(k3). I(k2) is different from I(k1), and I(k3) is also different from I(k1). Therefore, both characters R2 and R3 need to pay attention to the focus point O1 that character R1 pays attention to.

[0162] Thus, during the second round of intent recognition, character R1's focus and the output intent remain unchanged; character R2's focus is on O2 and O1, and the output intent is I'(k2); character R3's focus is on O3 and O1, and the output intent is I'(k3). While I'(k2) is the same as I(k1), I'(k3) is still different from I(k1). Therefore, character R2 no longer needs to add additional focus points, while R3 needs to continue to focus on O2, the focus point that character R2 is focusing on.

[0163] During the third round of intent recognition, the focus points and output intention results of roles R1 and R2 remain unchanged; role R3's focus points are O3, O1, and O2, and the intention result output by role R3 is I”(k3). Among them, I”(k3) is the same as I(k1). Therefore, role R3 no longer needs to increase the focus points.

[0164] In general, when the credibility of the large language model performing intent recognition does not meet the preset conditions, negotiations are carried out between the large language models of different roles, and after each round of negotiations, the credibility of the large language model performing intent recognition is recalculated to determine whether the credibility has met the preset conditions. If the credibility of the large language model performing intent recognition still does not meet the preset conditions, the next round of negotiations between the large language models of different roles will continue until the credibility of the large language model performing intent recognition meets the preset conditions, or the negotiations between the large language models of different roles have reached a certain number of rounds. Among them, the credibility of the large language model performing intent recognition is measured by the uncertainty value.

[0165] For example, see Figure 11 , Figure 11 This application provides a schematic diagram of calculating the uncertainty value of a large language model performing intent recognition. Figure 11 As shown, for the large language model of role R1, the following prompt words can be input into the large language model: "Please start from the description of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible", focusing on the time requirement, which of the following intentions does this task intention belong to: 1. Book a train ticket, 2. Book a plane ticket, 3. Book a boat ticket, 4. Book a bus ticket", and finally the large language model of role R1 outputs the intention result: the intention of booking a plane ticket.

[0166] For the large language model of role R2, the following prompt word can be input into the large language model: "Please consider the description of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible", focusing on the demands of the destination and departure points. Which of the following intentions does this task intention belong to: 1. Book a train ticket, 2. Book a plane ticket, 3. Book a boat ticket, 4. Book a bus ticket." Finally, the large language model of role R2 outputs the intention result: the intention of booking a plane ticket.

[0167] For the large language model of role R3, the following prompt word can be input into the large language model: "Please start from the description of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible", focusing on the ticket type demand. Which of the following intentions does this task intention belong to: 1. Book a train ticket, 2. Book a plane ticket, 3. Book a boat ticket, 4. Book a bus ticket?" Finally, the large language model of role R3 outputs the intention result: the intention of booking a train ticket.

[0168] At this point, based on the three intent results obtained, we can determine that the probability of "booking a flight ticket" is 2 / 3, and the probability of "booking a train ticket" is 1 / 3. The final calculated uncertainty value for the large language model's intent recognition is 0.276. Given the target threshold of 0.2, 0.276 is greater than 0.2, indicating that the large language model's confidence in intent recognition does not meet the preset requirements, requiring negotiation between the roles.

[0169] After the negotiation between the roles is performed, the focus corresponding to the role R1 can be added to the large language model of role R3. Therefore, the following prompt word can be input into the large language model: "Please start from the description of the task "Help me book a ticket from Shenzhen to Beijing as soon as possible", focusing on the type of ticket and taking into account the time requirement, which of the following intentions does this task intention belong to: 1. Book a train ticket, 2. Book a plane ticket, 3. Book a boat ticket, 4. Book a bus ticket". Finally, the large language model of role R3 outputs the intention result: the intention of booking a plane ticket.

[0170] After the intent result output by character R3 is adjusted, the probability of "intention to book a flight ticket" is determined to be 1, and the final calculated uncertainty value of the large language model's intent recognition is 0. Given a target threshold of 0.2, 0 is less than 0.2, so the large language model's confidence in intent recognition meets the preset conditions, and the final intent result can be output.

[0171] Stage five, demonstrating the explainable intent recognition process.

[0172] Specifically, after obtaining the final intent result, the process of calculating the intent result under which focus points can be displayed to achieve an explainable intent recognition process. For example, Figure 11 In the illustrated embodiment, the following text content can be displayed as an interpretable intent recognition process: "Based on the time requirement, the destination and departure requirements, and the ticket type and time requirements at the same time, it is recognized that the intention is to book a plane ticket."

[0173] It's important to note that if multiple roles have already negotiated a certain number of times and the resulting trustworthiness still doesn't meet the pre-set conditions, this indicates conflicting intent results from different focus points. In this case, the intent result corresponding to the highest-priority focus point can be selected as the final intent result. The intent results for various focus points can then be displayed to explain the entire intent recognition process, effectively supporting multiple rounds of dialogue and providing better support for subsequent intent clarification.

[0174] In general, the intent recognition method based on a large-model multi-role negotiation mechanism provided in this application can effectively address the need to implement human-computer dialogue intent recognition functions in new fields (such as finance) without training corpus and the demand for flexible expansion of intent. In addition, this solution can maximize the general knowledge reserve and semantic understanding capabilities of the large model, while using the negotiation mechanism to constrain the phantom problem of the large model, improving generalization capabilities while ensuring intent recognition accuracy, and also having a certain degree of interpretability.

[0175] The above describes the method provided by the present application in detail. Next, the device provided by the present application for executing the above method will be introduced.

[0176] See also Figure 12 , Figure 12 This is a schematic diagram of the structure of an intention recognition device provided by this application. Figure 12 As shown, the intention recognition device provided by the present application includes: an acquisition module 1201, which is used to obtain task data input by the user, where the task data is multimedia data, and the task data is used to indicate the task that the user expects to complete; a processing module 1202, which is used to identify a set of focus points for the task data through a large language model, where the focus point set includes multiple focus points, and the multiple focus points are used to indicate different angles considered when performing the task; the processing module 1202 is also used to identify the intention of the task data from the focus points in the focus point set through a large language model, and obtain multiple candidate intention results; the processing module 1202 is also used to determine a target intention result among multiple candidate intention results.

[0177] The acquisition module 1201 and the processing module 1202 can be implemented by software or hardware. For example, the implementation of the processing module 1202 will be described below using the processing module 1202 as an example. Similarly, the implementation of the acquisition module 1201 can refer to the implementation of the processing module 1202.

[0178] As an example of a software functional unit, the processing module 1202 may include code running on a computing instance. The computing instance may include at least one of a physical host (computing device), a virtual machine, and a container. Furthermore, the computing instance may be one or more. For example, module A may include code running on multiple hosts / virtual machines / containers. It should be noted that the multiple hosts / virtual machines / containers used to run the code may be distributed in the same region or in different regions. Furthermore, the multiple hosts / virtual machines / containers used to run the code may be distributed in the same availability zone (AZ) or in different AZs, each AZ including one data center or multiple geographically close data centers. Typically, a region may include multiple AZs.

[0179] Similarly, multiple hosts / virtual machines / containers running the code can be distributed within the same virtual private cloud (VPC) or across multiple VPCs. Typically, a VPC is set up within a region. Cross-region communication between two VPCs within the same region, or between VPCs in different regions, requires a communication gateway within each VPC to interconnect the VPCs.

[0180] As an example of a hardware functional unit, the processing module 1202 may include at least one computing device, such as a server. Alternatively, the processing module 1202 may be implemented using a central processing unit (CPU), an application-specific integrated circuit (ASIC), or a programmable logic device (PLD). The PLD may be a complex programmable logical device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), a data processing unit (DPU), a neural network processing unit (NPU), a system on chip (SoC), an offload card, an accelerator card, or any combination thereof.

[0181] The multiple computing devices included in processing module 1202 can be distributed in the same region or in different regions. The multiple computing devices included in processing module 1202 can be distributed in the same AZ or in different AZs. Similarly, the multiple computing devices included in processing module 1202 can be distributed in the same VPC or in multiple VPCs. The multiple computing devices can be any combination of servers, ASICs, PLDs, CPLDs, FPGAs, GALs, DPUs, NPUs, SoCs, offload cards, accelerator cards, and other computing devices.

[0182] In one possible implementation, the processing module is further used to: input the intent tree and the task data into the large language model to obtain the focus set output by the large language model; wherein the intent tree includes multiple parent class intent types, and each parent class intent type includes multiple sub-class intent types, and the large language model is used to identify the target parent class intent type to which the task data belongs, and determine the focus set based on the sub-class intent type under the target parent class intent type and the semantics of the task data.

[0183] In one possible implementation, the processing module 1202 is also used to: determine the credibility of the large language model performing intent recognition based on the occurrence probability of each intent in the multiple candidate intent results, wherein each candidate intent result in the multiple candidate intent results corresponds to an intent result, and the occurrence probability of each intent result is the ratio between the number of candidate intent results corresponding to each intent result and the total number of multiple candidate intent results; when the credibility meets the preset conditions, determine the target intent result from the multiple candidate intent results, and the target intent result corresponds to an intent result with the highest probability of occurrence.

[0184] In one possible implementation, the processing module is further used to: when the credibility does not meet the preset conditions, adopt a negotiation mechanism to adjust some of the multiple candidate intent results, and determine the target intent result based on the adjusted candidate intent results.

[0185] In one possible implementation, the large language model is also used to identify the priority between multiple focus points based on task data; the processing module 1202 is also used to: determine a target focus point among the multiple focus points, the candidate intention result corresponding to the target focus point is different from the candidate intention result corresponding to the first focus point, and the first focus point is the focus point with the highest priority among the multiple focus points; through the large language model, the intention of the task data is identified from the first focus point and the target focus point at the same time, and an adjusted candidate intention result corresponding to the target focus point is obtained; based on the adjusted candidate intention result and the unadjusted candidate intention result among the multiple candidate intention results, the credibility of the large language model in performing intent recognition is re-determined; when the re-determined credibility meets the preset condition, the target intention result is determined from the adjusted candidate intention result and the unadjusted candidate intention result.

[0186] In one possible implementation, the processing module 1202 is also used to: when the re-determined degree of credibility does not meet the preset conditions and the adjusted candidate intent result is different from the candidate intent result corresponding to the first focus point, use the large language model to simultaneously identify the intent of the task data from the first focus point, the second focus point and the target focus point, and obtain a second-adjusted candidate intent result corresponding to the target focus point, where the second focus point is the focus point with the second highest priority among the multiple focus points; based on the second-adjusted candidate intent result and the unadjusted candidate intent results among the multiple candidate intent results, redetermine the degree of credibility of the large language model in performing intent recognition.

[0187] In one possible implementation, the processing module 1202 is also used to: determine the uncertainty value of the large language model performing intent recognition based on the probability of occurrence of each intention result by obtaining information entropy, and the uncertainty value is used to indicate the degree of credibility; the preset condition is that the uncertainty value is less than or equal to the target threshold.

[0188] In one possible implementation, the processing module 1202 is further used to: display one or more focus points corresponding to the target intent result, wherein the candidate intent results corresponding to the one or more focus points are the same as the target intent result.

[0189] In one possible implementation, the multiple focus points are obtained by indicating the requirements expressed in the large language model recognition task data.

[0190] In a possible implementation, the processing module 1202 is further configured to: execute the task indicated by the task data based on the target intention result and the task data.

[0191] The intention recognition device provided in the present application can specifically be a chip, and the chip includes: a processing unit and a communication unit. The processing unit can be, for example, a processor, and the communication unit can be, for example, an input / output interface, a pin or a circuit. The processing unit can execute the computer execution instructions stored in the storage unit so that the chip in the electronic device executes the method described in the above embodiment. Optionally, the storage unit is a storage unit in the chip, such as a register, a cache, etc. The storage unit can also be a storage unit located outside the chip in the wireless access device, such as a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM), etc.

[0192] For details, please refer to Figure 13 , Figure 13 This is a schematic diagram of the structure of a chip provided in this application. The chip can be expressed as a neural processing unit (NPU), a graphics processing unit (GPU), or a tensor processing unit (TPU). The following will be introduced using the chip as an NPU 1300 as an example. Specifically, the NPU 1300 is mounted on the main CPU (Host CPU) as a coprocessor, and the Host CPU assigns tasks. The core part of the NPU is the operation circuit 1303, which is controlled by the controller 1304 to extract matrix data from the memory and perform multiplication operations.

[0193] In some implementations, the arithmetic circuit 1303 includes multiple processing units (PEs). In some implementations, the arithmetic circuit 1303 is a two-dimensional systolic array. The arithmetic circuit 1303 can also be a one-dimensional systolic array or other electronic circuitry capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit 1303 is a general-purpose matrix processor.

[0194] For example, assume there are input matrix A, weight matrix B, and output matrix C. The computation circuit retrieves the corresponding data of matrix B from weight memory 1302 and caches it on each PE in the computation circuit. The computation circuit then retrieves the data of matrix A from input memory 1301 and performs a matrix operation on it with matrix B. The partial or final matrix result is stored in accumulator 1308.

[0195] Unified memory 1306 is used to store input and output data. Weight data is directly transferred to weight memory 1302 through the Direct Memory Access Controller (DMAC) 1305. Input data is also transferred to unified memory 1306 through the DMAC.

[0196] BIU stands for Bus Interface Unit 1310 , which is used for interaction between the AXI bus, DMAC, and instruction fetch buffer (IFB) 1309 .

[0197] The bus interface unit 1310 (BIU) is used for the instruction fetch memory 1309 to obtain instructions from the external memory, and is also used for the storage unit access controller 1305 to obtain the original data of the input matrix A or the weight matrix B from the external memory.

[0198] DMAC is mainly used to move input data in the external memory DDR to the unified memory 1306 or to move weight data to the weight memory 1302 or to move input data to the input memory 1301.

[0199] The vector calculation unit 1307 includes multiple operation processing units. When necessary, it further processes the output of the operation circuit 1303, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc. It is mainly used for non-convolutional / fully connected layer network calculations in neural networks, such as batch normalization, pixel-level summation, and upsampling of feature planes.

[0200] In some implementations, the vector calculation unit 1307 can store the processed output vector in the unified memory 1306. For example, the vector calculation unit 1307 can apply a linear function or a nonlinear function to the output of the operation circuit 1303, such as linear interpolation of the feature plane extracted by the convolution layer, or accumulate a vector of values ​​to generate an activation value. In some implementations, the vector calculation unit 1307 generates a normalized value, a pixel-level summed value, or both. In some implementations, the processed output vector can be used as an activation input to the operation circuit 1303, for example, for use in subsequent layers in a neural network.

[0201] An instruction fetch buffer 1309 connected to the controller 1304 is used to store instructions used by the controller 1304;

[0202] Unified memory 1306, input memory 1301, weight memory 1302, and instruction fetch memory 1309 are all on-chip memories. External memories are private to the NPU hardware architecture.

[0203] The processor mentioned in any of the above places can be a general-purpose central processing unit, a microprocessor, an ASIC, or one or more integrated circuits for controlling the execution of the above program.

[0204] This application also provides a computing device 1400, which is used to implement the functions of the above-mentioned intention recognition device. Figure 14 , Figure 14 This is a schematic diagram of the structure of a computing device 1400 provided in this application. Figure 14 As shown, computing device 1400 includes a bus 1402, a processor 1404, a memory 1406, and a communication interface 1408. Processor 1404, memory 1406, and communication interface 1408 communicate with each other via bus 1402. Computing device 1400 can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in computing device 1400.

[0205] The bus 1402 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 14 The bus 1402 may include a path for transmitting information between various components of the computing device 1400 (eg, memory 1406, processor 1404, communication interface 1408).

[0206] The processor 1404 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0207] The memory 1406 may include volatile memory, such as random access memory (RAM). The processor 1404 may also include non-volatile memory, such as read-only memory (ROM), flash memory, a hard disk drive (HDD), or a solid state drive (SSD).

[0208] The memory 1406 stores executable program code, and the processor 1404 executes the executable program code to respectively implement the functions of the aforementioned acquisition module and processing module, thereby implementing the above-mentioned intention recognition method. In other words, the memory 1406 stores instructions for executing the intention recognition method.

[0209] The communication interface 1408 uses a transceiver module such as, but not limited to, a network interface card or a transceiver to implement communication between the computing device 1400 and other devices or a communication network.

[0210] This application also provides a computing device cluster for implementing the functionality of the above-mentioned intent recognition device cluster. The computing device cluster includes at least one computing device. The computing device can be a server, such as a central server, an edge server, or a local server in a local data center. In some embodiments, the computing device can also be a terminal device such as a desktop computer, a laptop computer, or a smartphone.

[0211] See also Figure 15 , Figure 15 This is a schematic diagram of the structure of a computing device cluster provided by this application. Figure 15 As shown, the computing device cluster includes at least one computing device 1400. The memory 1406 in one or more computing devices 1400 in the computing device cluster may store the same instructions for executing the intent recognition method.

[0212] In some possible implementations, the memory 1406 of one or more computing devices 1400 in the computing device cluster may also store some instructions for executing the intent recognition method. In other words, the combination of one or more computing devices 1400 can jointly execute the instructions for executing the intent recognition method.

[0213] It should be noted that the memory 1406 in different computing devices 1400 in the computing device cluster can store different instructions, each for executing a portion of the functions of the data processing apparatus. In other words, the instructions stored in the memory 1406 in different computing devices 1400 can implement the functions of one or more of the aforementioned acquisition module and processing module.

[0214] In some possible implementations, one or more computing devices in a computing device cluster may be connected via a network, which may be a wide area network or a local area network. Figure 16 A possible implementation is shown. Figure 16 This is a schematic diagram of the structure of another computing device cluster provided by this application. Figure 16 As shown, in computing device cluster 1600, two computing devices 1400A and 1400B are connected via a network. Specifically, the connection to the network is achieved through a communication interface within each computing device. In this possible implementation, memory 1406 within computing device 1400A stores instructions for executing the functions of an acquisition module. Simultaneously, memory 1406 within computing device 1400B stores instructions for executing the functions of a processing module.

[0215] It should be understood that Figure 16 The functionality of the computing device 1400A shown in FIG. 14 may also be implemented by multiple computing devices 1400. Similarly, the functionality of the computing device 1400B may also be implemented by multiple computing devices 1400.

[0216] See Figure 17 , Figure 17 This application also provides a computer-readable storage medium. In some embodiments, the above Figure 3 The disclosed methods may be implemented as computer program instructions encoded in a machine-readable format on a computer-readable storage medium or on other non-transitory media or articles of manufacture.

[0217] Figure 17 Schematically illustrates a conceptual partial view of an example computer-readable storage medium including a computer program for executing a computer process on a computing device, arranged in accordance with at least some embodiments presented herein.

[0218] In one embodiment, the computer readable storage medium 1700 is provided using a signal bearing medium 1701. The signal bearing medium 1701 may include one or more program instructions 1702, which when executed by one or more processors may provide the above-mentioned instructions for Figure 3 Describes the functionality or part of the functionality.

[0219] In some examples, the signal bearing medium 1701 may include a computer readable medium 1703 such as, but not limited to, a hard drive, a compact disk (CD), a digital video disk (DVD), a digital tape, memory, ROM or RAM, and the like.

[0220] In some embodiments, the signal-bearing medium 1701 may include a computer-recordable medium 1704, such as, but not limited to, a memory, a read / write (R / W) CD, a R / W DVD, or the like. In some embodiments, the signal-bearing medium 1701 may include a communication medium 1705, such as, but not limited to, a digital and / or analog communication medium (e.g., a fiber optic cable, a waveguide, a wired communication link, a wireless communication link, or the like). Thus, for example, the signal-bearing medium 1701 may be communicated via a wireless form of the communication medium 1705 (e.g., a wireless communication medium conforming to the IEEE 802.X standard or other transmission protocol).

[0221] The one or more program instructions 1702 may be, for example, computer-executable instructions or logic-implemented instructions. In some examples, the computing device may be configured to provide various operations, functions, or actions in response to the program instructions 1702 communicated to the computing device via one or more of computer-readable media 1703, computer-recordable media 1704, and / or communication media 1705.

[0222] It should also be noted that the device embodiments described above are merely illustrative, in which the units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines.

[0223] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course can also be implemented by special hardware including application-specific integrated circuits, special CPUs, special memories, special components, etc. In general, all functions performed by computer programs can be easily implemented with corresponding hardware, and the specific hardware structures used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits, etc. However, for the present application, software program implementation is a better implementation method in most cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer's floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., and includes a number of instructions to enable a computer device (which can be a personal computer, training equipment, or network equipment, etc.) to execute the methods of each embodiment of the present application.

[0224] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0225] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, training equipment or data center to another website, computer, training equipment or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a training equipment, data center, etc. that includes one or more available media integrations. Available media can be magnetic media (e.g., floppy disk, hard disk, tape), optical media (e.g., DVD), or semiconductor media (e.g., solid-state drive (SSD)).

Claims

1. A method for identifying intention, characterized in that: include: Acquire task data input by a user, wherein the task data is multimedia data and is used to indicate a task that the user desires to complete; Identifying, by means of a large language model, a set of focus points for the task data, the set of focus points comprising a plurality of focus points, each of the plurality of focus points being used to indicate a constraint condition to be satisfied when executing the task; Identifying the intent of the task data from the focus points in the focus point set using the large language model to obtain a plurality of candidate intent results; Among the multiple candidate intent results, determine the target intent result.

2. The method according to claim 1, characterized in that The identifying of a set of focus points for the task data using a large language model includes: Inputting the intent tree and the task data into the large language model to obtain the focus set output by the large language model; Among them, the intent tree includes multiple parent intent types, and each parent intent type includes multiple sub-class intent types. The large language model is used to identify the target parent intent type to which the task data belongs, and determine the focus set based on the sub-class intent type under the target parent intent type and the semantics of the task data.

3. The method according to claim 1 or 2, characterized in that Determining a target intent result based on the multiple candidate intent results includes: Determining the credibility of the large language model in performing intent recognition based on the occurrence probability of each intent in the multiple candidate intent results, wherein each candidate intent result in the multiple candidate intent results corresponds to an intent result, and the occurrence probability of each intent result is the ratio between the number of candidate intent results corresponding to each intent result and the total number of the multiple candidate intent results; When the credibility meets the preset conditions, the target intention result is determined from the multiple candidate intention results, and the target intention result corresponds to an intention result with the highest probability of occurrence.

4. The method according to claim 3, characterized in that The determining of a target intent result based on the plurality of candidate intent results further includes: When the credibility does not meet the preset conditions, a negotiation mechanism is used to adjust some of the multiple candidate intent results, and the target intent result is determined based on the adjusted candidate intent results.

5. The method according to claim 4, characterized in that The large language model is further used to identify priorities among the multiple concerns based on the task data; The adopting a negotiation mechanism to adjust some of the plurality of candidate intent results, and determining the target intent result based on the adjusted candidate intent results, includes: Determining a target point of interest among the multiple points of interest, the candidate intent result corresponding to the target point of interest being different from the candidate intent result corresponding to the first point of interest, the first point of interest being the point of interest with the highest priority among the multiple points of interest; Using the large language model, simultaneously identifying the intent of the task data from the first focus and the target focus, and obtaining an adjusted candidate intent result corresponding to the target focus; Re-determining the credibility of the large language model in performing intent recognition based on the adjusted candidate intent result and the unadjusted candidate intent results among the multiple candidate intent results; When the re-determined degree of credibility meets the preset condition, the target intention result is determined from the adjusted candidate intention results and the unadjusted candidate intention results.

6. The method according to claim 5, characterized in that The determining of a target intent result based on the plurality of candidate intent results further includes: When the re-determined credibility does not meet the preset condition, and the adjusted candidate intent result is different from the candidate intent result corresponding to the first focus point, using the large language model to simultaneously identify the intent of the task data from the first focus point, the second focus point, and the target focus point, to obtain a second-adjusted candidate intent result corresponding to the target focus point, where the second focus point is the focus point with the second highest priority among the multiple focus points; Based on the secondarily adjusted candidate intent results and the unadjusted candidate intent results among the multiple candidate intent results, the credibility of the large language model in performing intent recognition is re-determined.

7. The method according to any one of claims 3 to 6, characterized in that: Determining the credibility of the large language model in performing intent recognition based on the occurrence probability of each intent in the multiple candidate intent results includes: Based on the occurrence probability of each intent in the multiple candidate intent results, determining an uncertainty value of intent recognition performed by the large language model by using information entropy, wherein the uncertainty value is used to indicate the degree of credibility; The preset condition is that the uncertainty value is less than or equal to a target threshold.

8. The method according to any one of claims 1 to 7, characterized in that The method further comprises: Display one or more focus points corresponding to the target intent result, wherein the candidate intent results corresponding to the one or more focus points are the same as the target intent result.

9. The method according to any one of claims 1 to 8, characterized in that The plurality of focus points are obtained by instructing the large language model to recognize requirements expressed in the task data.

10. The method according to any one of claims 1 to 9, characterized in that: The method further comprises: Based on the target intention result and the task data, the task indicated by the task data is performed.

11. An intention recognition device, characterized in that: include: An acquisition module is used to acquire task data input by a user, wherein the task data is multimedia data and is used to indicate a task that the user desires to complete; a processing module, configured to identify a set of focus points for the task data using a large language model, wherein the set of focus points includes a plurality of focus points, each of which is used to indicate a constraint condition to be satisfied when executing the task; The processing module is further configured to identify the intent of the task data from the focus points in the focus point set using the large language model to obtain a plurality of candidate intent results; The processing module is also used to determine the target intention result among the multiple candidate intention results.

12. The device according to claim 11, characterized in that The processing module is further configured to: Inputting the intent tree and the task data into the large language model to obtain the focus set output by the large language model; Among them, the intent tree includes multiple parent intent types, and each parent intent type includes multiple sub-class intent types. The large language model is used to identify the target parent intent type to which the task data belongs, and determine the focus set based on the sub-class intent type under the target parent intent type and the semantics of the task data.

13. The device according to claim 11 or 12, characterized in that The processing module is further configured to: Determining the credibility of the large language model in performing intent recognition based on the occurrence probability of each intent in the multiple candidate intent results, wherein each candidate intent result in the multiple candidate intent results corresponds to an intent result, and the occurrence probability of each intent result is the ratio between the number of candidate intent results corresponding to each intent result and the total number of the multiple candidate intent results; When the credibility meets the preset conditions, the target intention result is determined from the multiple candidate intention results, and the target intention result corresponds to an intention result with the highest probability of occurrence.

14. The device according to claim 13, characterized in that The processing module is further configured to: When the credibility does not meet the preset conditions, a negotiation mechanism is used to adjust some of the multiple candidate intent results, and the target intent result is determined based on the adjusted candidate intent results.

15. The device according to claim 14, characterized in that The large language model is further used to identify priorities among the multiple concerns based on the task data; The processing module is further configured to: Determining a target point of interest among the multiple points of interest, the candidate intent result corresponding to the target point of interest being different from the candidate intent result corresponding to the first point of interest, the first point of interest being the point of interest with the highest priority among the multiple points of interest; Using the large language model, simultaneously identifying the intent of the task data from the first focus and the target focus, and obtaining an adjusted candidate intent result corresponding to the target focus; Re-determining the credibility of the large language model in performing intent recognition based on the adjusted candidate intent result and the unadjusted candidate intent results among the multiple candidate intent results; When the re-determined degree of credibility meets the preset condition, the target intention result is determined from the adjusted candidate intention results and the unadjusted candidate intention results.

16. The device according to claim 15, characterized in that The processing module is further configured to: When the re-determined credibility does not meet the preset condition, and the adjusted candidate intent result is different from the candidate intent result corresponding to the first focus point, using the large language model to simultaneously identify the intent of the task data from the first focus point, the second focus point, and the target focus point, to obtain a second-adjusted candidate intent result corresponding to the target focus point, where the second focus point is the focus point with the second highest priority among the multiple focus points; Based on the secondarily adjusted candidate intent results and the unadjusted candidate intent results among the multiple candidate intent results, the credibility of the large language model in performing intent recognition is re-determined.

17. The device according to any one of claims 13 to 16, characterized in that The processing module is further configured to: Based on the occurrence probability of each intent in the multiple candidate intent results, determining an uncertainty value of intent recognition performed by the large language model by using information entropy, wherein the uncertainty value is used to indicate the degree of credibility; The preset condition is that the uncertainty value is less than or equal to a target threshold.

18. The device according to any one of claims 11 to 15, characterized in that The method further comprises: Display one or more focus points corresponding to the target intent result, wherein the candidate intent results corresponding to the one or more focus points are the same as the target intent result.

19. The device according to any one of claims 11 to 16, characterized in that The plurality of focus points are obtained by instructing the large language model to recognize requirements expressed in the task data.

20. An intention recognition device, characterized in that It includes a memory and a processor; the memory stores code, and the processor is configured to execute the code. When the code is executed, the intention recognition device executes the method according to any one of claims 1 to 10.

21. A chip system, characterized in that: The chip system includes a processor and a communication interface, the communication interface is used to communicate with a module outside the chip system, and the processor is used to execute the method according to any one of claims 1 to 10.

22. A computer storage medium, characterized in that The computer storage medium stores instructions, which, when executed by a computer, cause the computer to implement the method according to any one of claims 1 to 10.

23. A computer program product, characterized in that The computer program product stores instructions, which, when executed by a computer, cause the computer to implement the method according to any one of claims 1 to 10.

Images