An intent understanding method and system

Abstract

The application provides an intention understanding method and system. The method comprises the steps of: performing intention feature extraction on user input data to obtain a task package containing input text, main intention, entity set, main intention confidence and user features. The level of the main intention confidence is determined, if it is high confidence, the corresponding sub-agent is called to process the task package; if it is medium confidence, the task dependency relationship between each sub-intention corresponding to the main intention is determined, then the agent pipeline is determined, and the task package is processed based on the pipeline; if it is low confidence, the sub-agent corresponding to each sub-intention corresponding to the main intention is called to process the task package. When the trigger condition is met, the compliance of the processing result corresponding to the task package is detected. According to the application, the problems of low intention recognition accuracy, low processing efficiency and high output compliance risk existing in the existing intention understanding implementation mode can be solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence and natural language processing, and more particularly relates to an intent understanding method and system. BACKGROUND

[0002] In current intelligent interaction systems, intent understanding is a key technology to achieve efficient human-computer interaction. However, existing intent understanding technologies have many deficiencies. On the one hand, most existing intent understanding systems use a single architecture, which is difficult to handle complex intents and related entity information simultaneously, resulting in low accuracy and completeness of intent recognition. For example, in the customer consultation scenario in the financial field, the customer's intent is often closely related to various financial products, transaction operations and other entities, and the single architecture of the intent understanding system is difficult to accurately extract intent and entity information. On the other hand, existing intent understanding systems lack flexible processing mode switching mechanisms and cannot automatically select the optimal processing mode according to different task requirements and data characteristics, resulting in low system processing efficiency. In addition, in the financial and other fields with extremely high compliance requirements, existing intent understanding systems lack output compliance checking mechanisms for regulatory requirements, and are prone to produce output results that do not comply with regulatory requirements, posing a high compliance risk. SUMMARY

[0003] The purpose of the present application is to solve the problems of low accuracy of intent recognition, low processing efficiency and high risk of output compliance of existing intent understanding implementation.

[0004] According to a first aspect of the present application, an intent understanding method is provided, the method comprising the following steps:

[0005] extracting intent features from user input data to obtain a task package, the task package containing input text, main intent, entity set, main intent confidence and user features;

[0006] determining the level of the main intent confidence, if the main intent confidence is high confidence, calling the sub-agent corresponding to the main intent to process the task package,

[0007] if the main intent confidence is medium confidence, determining the task dependency relationship between each sub-intent corresponding to the main intent, determining an agent pipeline composed of sub-agents corresponding to each sub-intent according to the task dependency relationship, and processing the task package based on the agent pipeline,

[0008] if the main intent confidence is low confidence, respectively calling the sub-agents corresponding to each sub-intent corresponding to the main intent to process the task package;

[0009] According to the domain label of the processing result corresponding to the task package, it is determined whether to trigger domain compliance detection, and if so, the processing result is detected, and the processing result is output if the detection is passed.

[0010] Optionally, the step of performing intent feature extraction on the user input data to obtain a task package comprises:

[0011] Obtaining text corresponding to the user input data, and preprocessing the text;

[0012] Performing joint extraction of intent and entity on the preprocessed text to obtain a triple containing a set of intent labels, an entity list and a confidence vector, and obtaining a confidence distribution containing multiple candidate intents according to the triple;

[0013] Integrating the preprocessed text, the highest confidence intent, the associated entity and the user portrait data to obtain the task package.

[0014] Optionally, the step of obtaining text corresponding to the user input data and preprocessing the text comprises:

[0015] Determining the data form of the user input data, if the user input data is an image, using OCR technology to convert the user input data into text, if the user input data is audio, using ASR technology to convert the user input data into text;

[0016] Cleaning, tokenizing and correcting the text to obtain the standardized input text.

[0017] Optionally, when the main intent confidence is low, the processing result corresponding to the task package is the fusion result of the output results of the respective sub-agents, and the fusion result is obtained based on a DS evidence fusion algorithm.

[0018] Optionally, the step of outputting the processing result if the detection is passed comprises:

[0019] Determining whether the data form of the processing result passed by the detection is consistent with the user input data, if so, directly outputting the processing result, if not, performing data form consistency conversion on the processing result, and outputting the converted processing result.

[0020] According to a second aspect of the present application, an intent understanding system is provided, which comprises a main agent, a plurality of sub-agents and a domain compliance detection module.

[0021] The main agent is configured to:

[0022] performing intent feature extraction on the user input data to obtain a task package, the task package comprising input text, a main intent, an entity set, a main intent confidence, and user features;

[0023] determining a level of the main intent confidence, if the main intent confidence is a high confidence, invoking a sub-agent corresponding to the main intent to process the task package,

[0024] if the main intent confidence is a medium confidence, determining a task dependency relationship between each sub-intent corresponding to the main intent, determining an agent pipeline composed of sub-agents corresponding to the each sub-intent according to the task dependency relationship, and processing the task package based on the agent pipeline,

[0025] if the main intent confidence is a low confidence, respectively invoking sub-agents corresponding to each sub-intent corresponding to the main intent to process the task package;

[0026] determining whether to trigger a domain compliance detection according to a domain label carried by a processing result corresponding to the task package, if yes, invoking the domain compliance detection module to detect the processing result, and outputting the processing result if the detection is passed.

[0027] Optionally, the main agent is further configured to:

[0028] obtain text corresponding to the user input data, and pre-process the text;

[0029] perform joint extraction of intent and entity on the pre-processed text to obtain a triple comprising a set of intent labels, an entity list, and a confidence vector, and obtain a confidence distribution comprising multiple candidate intents according to the triple;

[0030] integrate the pre-processed text, the highest confidence intent, the associated entity, and user portrait data to obtain the task package.

[0031] Optionally, the main agent is further configured to:

[0032] determine a data form of the user input data, if the user input data is an image, convert the user input data into text using an OCR technology, and if the user input data is audio, convert the user input data into text using an ASR technology;

[0033] clean, tokenize, and correct the text to obtain a standardized input text.

[0034] Optionally, when the main intention confidence is a low confidence, the processing result corresponding to the task package is a fusion result of output results of respective sub-agents, and the fusion result is obtained based on a DS evidence fusion algorithm.

[0035] Optionally, the main agent is further configured to:

[0036] determine whether the data form of the processing result detected is consistent with the user input data, if yes, directly output the processing result, and if not, perform data form consistency conversion on the processing result, and output the converted processing result.

[0037] The application has the following beneficial effects:

[0038] The intention understanding method has the following advantages: first, the user input data is subjected to intention feature extraction to obtain a task package, and then the task package is processed by using sub-agents of vertical fields, so that the user intention and related entity information can be extracted more accurately, and the accuracy of intention understanding is significantly improved compared with the existing single-architecture intention understanding system; second, the task package is processed by using different processing modes according to the level of the main intention confidence, so that the optimal processing mode can be selected for tasks of different complexity, unnecessary waste of computing resources is avoided, the processing efficiency of intention understanding is effectively improved, fast response is achieved when processing simple tasks, and accuracy is ensured while not excessively reducing efficiency when processing complex tasks; and third, the processing result of the predetermined field is subjected to field compliance detection, so that the output result of intention understanding in the corresponding application scenario conforms to the relevant regulatory provisions, the compliance risk of the business is reduced, and the reliability and safety of the corresponding intention understanding system in field application are enhanced.

[0039] The intention understanding system belongs to the same general inventive concept as the intention understanding method described above, and has at least the same beneficial effects as the intention understanding method described above, and the beneficial effects will not be repeated here.

[0040] Other features and advantages of the application will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0041] The application can be better understood by reference to the following description taken in conjunction with the accompanying drawings, in which like or similar elements are designated with identical reference numerals in the several figures of the drawing.

[0042] Figure 1 An implementation flowchart of the intention understanding method according to an embodiment of the application is shown;

[0043] Figure 2A structural block diagram of an intention understanding system according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0044] In order for those skilled in the art to more fully understand the technical solutions of the present application, in the following, exemplary embodiments of the present application will be described more fully and in detail with reference to the accompanying drawings. Obviously, the following described one or more embodiments of the present application are only one or more of the specific manners in which the technical solutions of the present application can be implemented, and are not exhaustive. It should be understood that the technical solutions of the present application can be implemented in other manners belonging to one general inventive concept without being limited by the exemplary described embodiments. Based on one or more embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative efforts should belong to the scope of protection of the present application.

[0045] Embodiment: Figure 1 An implementation flowchart of an intention understanding method according to an embodiment of the present application is shown. Referring to Figure 1 , the intention understanding method according to an embodiment of the present application comprises the following steps:

[0046] Step S100, performing intention feature extraction on user input data to obtain a task package, the task package containing input text, a main intention, an entity set, a main intention confidence and user features;

[0047] Step S200, determining a level of the main intention confidence, if the main intention confidence is high confidence, calling a sub-agent corresponding to the main intention to process the task package,

[0048] if the main intention confidence is medium confidence, determining a task dependency relationship between each sub-intention corresponding to the main intention, determining an agent pipeline composed of sub-agents corresponding to each sub-intention according to the task dependency relationship, and processing the task package based on the agent pipeline,

[0049] if the main intention confidence is low confidence, calling sub-agents corresponding to each sub-intention corresponding to the main intention to process the task package respectively;

[0050] Step S300, judging whether to trigger domain compliance detection according to a domain label carried by a processing result corresponding to the task package, if yes, detecting the processing result, and outputting the processing result in the case of passing the detection.

[0051] Further, in an embodiment of the present application, the step of performing intention feature extraction on user input data to obtain a task package comprises:

[0052] obtaining text corresponding to the user input data, and pre-processing the text;

[0053] Perform joint extraction of intent entities on the preprocessed text to obtain a triple containing an intent label set, an entity list, and a confidence vector. The confidence distribution of multiple candidate intents is obtained based on the triple.

[0054] The preprocessed text, highest confidence intent, related entities and user portrait data are integrated to obtain the task package.

[0055] Furthermore, in an embodiment of the present invention, the steps of obtaining text corresponding to the user input data and preprocessing the text include:

[0056] Determine the data format of the user input data. If the user input data is an image, use OCR technology to convert the user input data into text. If the user input data is an audio, use ASR technology to convert the user input data into text.

[0057] The text is cleaned, segmented and corrected to obtain standardized input text.

[0058] Specifically, in step S100 of the embodiment of the present invention, the user input is converted into structured intent features by extracting intent features from the user's multimodal input. This process mainly includes the following three stages:

[0059] Multimodal normalization processing: Supports multimodal input such as text, voice, and images. It converts multimodal input into a unified text format through ASR / OCR technology, and then performs text cleaning, word segmentation, and error correction to obtain standardized input text.

[0060] Intent-Entity Joint Extraction: This approach uses mainstream NLP models (such as Transformer variants) to extract triples: {intent label set, entity list, confidence vector}. The output includes a confidence distribution of multiple candidate intents, providing a basis for subsequent dynamic routing and processing mode decisions.

[0061] Task package encapsulation: Integrate the original text, the highest confidence intent, related entities, and user profile data to generate a standardized task package: {input text, main image, entity set, main image confidence, user features}.

[0062] Furthermore, in an embodiment of the present invention, when the confidence of the main idea is low, the processing result corresponding to the task package is the fusion result of the output results of the corresponding sub-agents, and the fusion result is obtained based on the DS evidence fusion algorithm.

[0063] Specifically, in step S200 of the embodiment of the present invention, dynamic routing and processing mode decisions are made. Based on a confidence-driven routing strategy, the optimal processing method is dynamically selected according to the intent determinism. The relevant processing methods include the following three:

[0064] Single mode processing (high confidence): When the main intent confidence is higher than the threshold (such as ≥ 0.8), directly call the unique matching sub-agent;

[0065] Sequence processing (medium confidence): When there is a task dependency relationship (such as A task result is B task input), call the sub-agent in the order of dependency;

[0066] Parallel processing (low confidence / ambiguous scenario): Copy the task package and call multiple candidate sub-agents in parallel, and integrate the results through an evidence fusion algorithm.

[0067] Specifically, in the embodiment of the application, after confirming the processing mode corresponding to the task package, whether it is single mode processing, sequence processing or parallel processing, a corresponding vertical domain sub-agent is used for deep analysis. The working mode of the sub-agent is specifically: loading a domain-specific knowledge base (such as a knowledge graph, a rule base, a professional model) to perform secondary analysis on the task package, outputting a domain-specific result (such as a structured recommendation, a work order classification label), including a domain analysis result, a sub-agent processing confidence, a domain label, and completing intent refinement based on domain knowledge. The sub-agent also ensures interface standardization, and the sub-agent interacts with the main agent through a unified API, input and output follow a general data format, supports dynamically expanding new domain sub-agents, and does not need to modify the core routing logic.

[0068] Specifically, in step S300 of the embodiment of the application, the processing result of the predetermined domain is customized and processed, and the processing result is checked for compliance (such as matching the regulatory rules in the financial domain), filtered for safety (such as desensitizing private data in the medical domain), adapted for business rules (such as inventory checking in the e-commerce domain), etc. according to the domain label of the processing result (such as "finance", "medicine"), and the domain compliance detection supports multiple implementation methods such as rule engine and machine learning model, and outputs the processing result with domain label or correction suggestion.

[0069] Still further, in the embodiment of the application, the step of outputting the processing result through detection includes:

[0070] determining whether the data form of the processing result through detection is consistent with the user input data, if yes, directly outputting the processing result, and if not, performing data form consistency conversion on the processing result and outputting the converted processing result.

[0071] Specifically, in the embodiment of the application, multi-modal output adaptation is implemented on the output of the processing result, corresponding formats (text reply, voice broadcast, graphic-text interaction, etc.) are generated according to the input channel, technical metadata (processing time, routing mode, domain processing state) is attached, etc. Process, finally generate the final response that users can interact.

[0072] Correspondingly, based on the intent understanding method of the embodiments of the present application, the embodiments of the present application further propose an intent understanding system.

[0073] Figure 2 The structural block diagram of the intent understanding system of the embodiments of the present application is shown. Referring to Figure 2 The intent understanding system of the embodiments of the present application comprises a main intelligent agent, a plurality of sub-intelligent agents and a domain compliance detection module;

[0074] The main intelligent agent is used for:

[0075] performing intent feature extraction on the user input data to obtain a task package, the task package comprising input text, a main intent, an entity set, a main intent confidence and user features;

[0076] determining the level of the main intent confidence, if the main intent confidence is a high confidence, calling the sub-intelligent agent corresponding to the main intent to process the task package,

[0077] if the main intent confidence is a medium confidence, determining the task dependency relationship between each sub-intent corresponding to the main intent, determining an intelligent agent pipeline composed of the sub-intelligent agents corresponding to each sub-intent according to the task dependency relationship, and processing the task package based on the intelligent agent pipeline,

[0078] if the main intent confidence is a low confidence, calling the sub-intelligent agent corresponding to each sub-intent corresponding to the main intent to process the task package respectively;

[0079] determining whether to trigger domain compliance detection according to the domain label carried by the processing result corresponding to the task package, if yes, calling the domain compliance detection module to detect the processing result, and outputting the processing result in the case of passing the detection.

[0080] Further, in the embodiments of the present application, the main intelligent agent is further used for:

[0081] obtaining the text corresponding to the user input data, and pre-processing the text;

[0082] performing joint extraction of intent and entity on the pre-processed text to obtain a triple comprising an intent label set, an entity list and a confidence vector, and obtaining a confidence distribution comprising multiple candidate intents according to the triple;

[0083] integrating the pre-processed text, the highest confidence intent, the associated entity and the user portrait data to obtain the task package.

[0084] Still further, in the embodiments of the present application, the main intelligent agent is further used for:

[0085] judging a data form of the user input data, if the user input data is an image, adopting an OCR technology to convert the user input data into text, if the user input data is audio, adopting an ASR technology to convert the user input data into text;

[0086] cleaning, segmenting and correcting the text to obtain the standardized input text.

[0087] Further, in the embodiment of the present application, the main intelligent agent is further used for:

[0088] judging whether the data form of the processing result detected is consistent with the user input data, if yes, directly outputting the processing result, if not, performing data form consistency conversion on the processing result, and outputting the converted processing result.

[0089] The intent understanding system of the embodiment of the present application is described in more detail based on a specific example as follows:

[0090] (I) System architecture building

[0091] The main intelligent agent is deployed, and an intent entity joint extraction model is configured, which can adopt an existing natural language processing model, such as a model based on Transformer, for preliminary processing of the text input by the user, extracting the intent and related entity, and calculating the confidence of the intent.

[0092] According to different vertical fields (such as finance, customer service, etc.), corresponding vertical field sub-intelligent agents are deployed, and each sub-intelligent agent is optimized for a specific field and has professional knowledge and processing ability in the field. For example, when applied in the financial field, a post-processing rule engine is deployed, and pre-configuration of compliance verification rules based on financial regulatory requirements is performed, such as rules for prohibiting recommendation of financial products that do not meet the risk tolerance of the user.

[0093] (II) Processing flow implementation

[0094] Taking user intent processing in an intelligent investment advisor master control system as an example:

[0095] S1 the user inputs a query text such as "I want to know about low-risk investment products", the main intelligent agent receives the text, identifies the intent as "knowing about investment products" and the entity as "low risk" through the intent entity joint extraction model, and calculates the confidence of the intent as 0.8 (assuming that the confidence threshold of the single prompt mode is 0.7).

[0096] S2 since the confidence 0.8 is higher than the single prompt mode threshold, the main intelligent agent adopts the single prompt mode to directly assign the task to the investment product sub-intelligent agent in the financial field.

[0097] The S3 investment product sub-agent receives the task, uses its internal professional knowledge and models to further analyze the user's intention, and screens out the investment product information that meets the requirements in combination with the entity "low risk".

[0098] The S4 sub-agent processing result is returned to the main agent, and the main agent transmits the result to the post-processing rule engine (financial field). The rule engine checks the compliance according to the regulatory requirements, for example, checks whether the recommended investment product is within the scope allowed by the regulation and whether it meets the user's basic information (such as age, risk level, etc.).

[0099] S5 The result of the compliance check is fed back to the user by the main agent; if it is not compliant, it will be adjusted or prompted that the relevant information cannot be provided.

[0100] When encountering more complex user intentions, such as problems involving both investment product understanding and account management, the confidence of the main agent in extracting the intention may be 0.5 (assuming that the sequence mode threshold is 0.4-0.7, and the parallel mode threshold is lower than 0.4), at which point it is determined whether sequence processing is needed, i.e., first processing the investment product sub-intention, and then processing the account management sub-intention; if the confidence is lower and multiple sub-intentions can be processed independently, the parallel mode is used to improve efficiency.

[0101] In the customer service intention recognition scenario, the processing flow is similar. The main agent assigns the corresponding sub-agents in the customer service field according to the customer's problem, such as the order query sub-agent, the complaint suggestion sub-agent, etc. The sub-agent processing result does not need to go through the post-processing rule engine of the financial field, but is directly output to the customer after the regular compliance check (such as language specifications).

[0102] Through the above implementation, the intention understanding system can efficiently and accurately process user intentions in different scenarios, meet the special needs of the financial, medical and other fields, and achieve the goal of hierarchical multi-agent intention understanding.

[0103] The intention understanding system of the embodiment of the application does not rely on specific domain knowledge and achieves generalization ability through the following mechanisms:

[0104] Interface standardization: define a unified task package format (JSON / Protobuf) and interaction protocol (HTTP / gRPC) to support mixed deployment of sub-agents with different technology stacks;

[0105] Processing mode abstraction: encapsulate domain-specific processing (such as financial compliance) as a pluggable post-processing module, and the core routing logic remains general;

[0106] Model replaceability: Intent extraction, evidence fusion, and other links support multiple algorithm implementations (such as replacing BiLSTM with T5 model), adapting to different computing power and accuracy requirements;

[0107] Through the above generalization design, the embodiment of the application constructs a layered architecture of "general framework + domain plug-in", which not only guarantees the efficiency of the core processing flow, but also supports quick adaptation to diversified industry scenarios.

[0108] The intent understanding system of the embodiment of the application has the following beneficial effects:

[0109] Accuracy improvement: Based on the cascaded agent architecture of intent entity joint extraction, through the cooperative work of the main agent and the vertical field sub-agent, the user intent and related entity information can be more accurately extracted, and compared with the traditional single architecture model, the accuracy of intent understanding is significantly improved.

[0110] Efficiency optimization: The dynamic routing mechanism automatically switches the processing mode according to the confidence threshold, so that the system can select the optimal processing mode for tasks of different complexity, avoiding unnecessary waste of computing resources and effectively improving the processing efficiency of the system. When processing simple tasks, it responds quickly, and when processing complex tasks, it ensures accuracy without reducing too much efficiency.

[0111] Compliance guarantee: The post-processing rule engine checks the output results based on regulatory requirements to ensure that the output results of intent understanding in multi-field (financial, medical, etc.) application scenarios meet the relevant regulatory requirements, reduce the compliance risk of business, and enhance the reliability and security of the system in field applications.

[0112] Although one or more embodiments of the application have been described above, it should be understood by those skilled in the art that the application can be implemented in any other form without departing from the spirit and scope of the application. Therefore, the above-described embodiments are illustrative rather than limiting, and many modifications and substitutions are obvious to those skilled in the art without departing from the spirit and scope of the application as defined by the appended claims.

Claims

1. A method for understanding intention, characterized in that: include: Performing intention feature extraction on user input data to obtain a task package, wherein the task package includes input text, main image, entity set, main image confidence and user features; Determine the confidence level of the main idea, and if the confidence level of the main idea is high, call the sub-agent corresponding to the main idea to process the task package, If the confidence of the main intention is medium, determine the task dependency relationship between the sub-intentions corresponding to the main intention, determine the agent pipeline composed of sub-agents corresponding to the sub-intentions according to the task dependency relationship, and process the task package based on the agent pipeline. If the confidence of the main intention is low, the sub-agents corresponding to the sub-intentions corresponding to the main intention are called to process the task package respectively; Determine whether to trigger domain compliance detection based on the domain tag of the processing result corresponding to the task package. If so, detect the processing result and output the processing result if it passes the detection.

2. The intention understanding method according to claim 1, characterized in that: The step of extracting intention features from user input data to obtain a task package includes: Obtaining text corresponding to the user input data and preprocessing the text; Perform joint extraction of intent entities on the preprocessed text to obtain a triple containing an intent label set, an entity list, and a confidence vector, and obtain a confidence distribution containing multiple candidate intents based on the triple; The preprocessed text, the highest confidence intent, the associated entities and the user portrait data are integrated to obtain the task package.

3. The intention understanding method according to claim 2, characterized in that: The step of obtaining text corresponding to the user input data and preprocessing the text includes: Determine the data format of the user input data; if the user input data is an image, convert the user input data into text using an OCR technology; if the user input data is an audio, convert the user input data into text using an ASR technology; The text is cleaned, segmented and corrected to obtain the standardized input text.

4. The intention understanding method according to claim 1, characterized in that: When the confidence of the main idea is low, the processing result corresponding to the task package is the fusion result of the output results of the corresponding sub-agents, and the fusion result is obtained based on the DS evidence fusion algorithm.

5. The intention understanding method according to claim 1, characterized in that: The step of outputting the processing result when the detection is passed comprises: Determine whether the data format of the processing result detected is consistent with the user input data. If so, directly output the processing result. If not, perform data format consistency conversion on the processing result and output the converted processing result.

6. An intention understanding system, characterized in that: It includes a main agent, several sub-agents and a domain compliance detection module; The main agent is used to: Performing intention feature extraction on user input data to obtain a task package, wherein the task package includes input text, main image, entity set, main image confidence and user features; Determine the confidence level of the main idea, and if the confidence level of the main idea is high, call the sub-agent corresponding to the main idea to process the task package, If the confidence of the main intention is medium, determine the task dependency relationship between the sub-intentions corresponding to the main intention, determine the agent pipeline composed of sub-agents corresponding to the sub-intentions according to the task dependency relationship, and process the task package based on the agent pipeline. If the confidence of the main intention is low, the sub-agents corresponding to the sub-intentions corresponding to the main intention are called to process the task package respectively; Determine whether to trigger domain compliance detection based on the domain tag of the processing result corresponding to the task package. If so, call the domain compliance detection module to detect the processing result, and output the processing result if it passes the detection.

7. The intention understanding system according to claim 6, characterized in that: The master agent is further configured to: Obtaining text corresponding to the user input data and preprocessing the text; Perform joint extraction of intent entities on the preprocessed text to obtain a triple containing an intent label set, an entity list, and a confidence vector, and obtain a confidence distribution containing multiple candidate intents based on the triple; The preprocessed text, the highest confidence intent, the associated entities and the user portrait data are integrated to obtain the task package.

8. The intention understanding system according to claim 7, characterized in that: The master agent is further configured to: Determine the data format of the user input data; if the user input data is an image, convert the user input data into text using an OCR technology; if the user input data is an audio, convert the user input data into text using an ASR technology; The text is cleaned, segmented and corrected to obtain the standardized input text.

9. The intention understanding system according to claim 6, characterized in that: When the confidence of the main idea is low, the processing result corresponding to the task package is the fusion result of the output results of the corresponding sub-agents, and the fusion result is obtained based on the DS evidence fusion algorithm.

10. The intention understanding system according to claim 6, characterized in that: The master agent is further configured to: Determine whether the data format of the processing result detected is consistent with the user input data. If so, directly output the processing result. If not, perform data format consistency conversion on the processing result and output the converted processing result.

Images