An intelligent question and answer interaction method and system supporting multi-intention coexistence

Abstract

The present disclosure relates to the technical field of artificial intelligence and natural language processing, and provides an intelligent question and answer interaction method and system supporting multiple intents coexistence, wherein the method comprises: acquiring user speech data; predicting the number of user intents contained in the user speech data, and performing intent segmentation on the user speech data to determine an intent semantic segment corresponding to each user potential intent; performing business conflict checking on each intent semantic segment, and determining the execution time sequence relationship between each intent semantic segment when there is no mutual exclusion conflict; and if there is no mutual exclusion conflict, performing a response operation matched with the intent semantic segment by a target intelligent agent corresponding to each intent semantic segment according to the execution time sequence relationship. The technical scheme provided by one or more embodiments of the present disclosure can accurately identify multiple operation intents in a single user dialogue, and can respond to multiple operation intents of a user in a complex business scenario through multiple intelligent agents.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence and natural language processing technology, specifically to an intelligent question-answering interaction method and system that supports the coexistence of multiple intentions. Background Technology

[0002] Natural Language Processing (NLP), a crucial branch of artificial intelligence, aims to enable computers to understand, parse, and generate human language. In human-computer interaction scenarios, intent recognition and dialogue management are core components in building intelligent question-answering systems. Intent recognition aims to automatically detect the user's interaction purpose from their utterances, while dialogue management is responsible for maintaining the dialogue state and deciding on the system's subsequent actions. With the development of deep learning and large-scale pre-trained language models, intelligent question-answering systems have been widely applied in customer service, education, healthcare, finance, and other fields, capable of processing single-turn, single-intent user requests with high accuracy.

[0003] Existing intelligent question-answering methods typically treat user utterances as carriers of a single intent, employing a pipeline-based processing framework of "input-classification-response." This works well for simple, clearly defined questions, but struggles to handle the complex expressions that frequently arise in real-world conversations—users may implicitly contain multiple different operational needs or information query goals within the same sentence. In some practical application areas, such as financial product consultation, e-commerce after-sales service, and government services, a single user's dialogue often carries multiple concurrent interactive intents. For example, in a financial product consultation scenario, a user's statement, "Please check the returns of a certain fund over the past year; if it's suitable, I'll buy it, and at the same time, sell another fund I hold," simultaneously involves three intents: querying, buying, and selling.

[0004] However, most current intelligent question answering systems are still designed based on single intent processing. The few solutions that support multiple intents often parse and respond to each intent independently, failing to fully model the logical relationship between intents and lacking consistent management of shared data states when multiple agents collaborate. This makes it difficult for the system to accurately coordinate multiple sub-tasks in complex business scenarios. Summary of the Invention

[0005] In view of this, one or more embodiments of this disclosure provide an intelligent question-answering interaction method and system that supports the coexistence of multiple intentions, which can accurately identify multiple operation intentions in a single user dialogue, and can coordinately respond to multiple user operation intentions in complex business scenarios through multiple intelligent agents.

[0006] Firstly, this disclosure provides an intelligent question-answering interaction method supporting the coexistence of multiple intents. The method includes: acquiring user speech data in natural language form; predicting the number of user intents contained in the user speech data; segmenting the user speech data according to the number of user intents; determining the semantic segments of intents corresponding to each potential user intent in the user speech data; performing business conflict checks on each semantic segment of intents; determining whether there are mutual exclusion conflicts between the semantic segments of intents; and, if there are no mutual exclusion conflicts, determining the execution sequence relationship between the semantic segments of intents according to the business conflict check results, wherein the sequence relationship includes sequential execution. The execution sequence is defined as follows: if there is no mutual exclusion conflict between the intent semantic segments, then each intent semantic segment is assigned to the corresponding target agent according to the execution timing relationship, so that the target agent executes the business response operation that matches the intent semantic segment; wherein, in the parallel execution phase of the intent semantic segment, the target agent executes the business response operation based on the same multi-version concurrency control snapshot; in the serial execution phase of the intent semantic segment, the target agent in the subsequent execution phase executes the business response operation based on the updated snapshot or the verified pre-output after the pre-operation is submitted and the intermediate state is verified.

[0007] In one embodiment, predicting the number of user intentions contained in the user utterance data includes: inputting the user utterance data into a semantic feature classification unit in a pre-trained intention quantity prediction network to generate classification feature data labeled with classification tags; inputting the classification feature data into a fully connected layer in the intention quantity prediction network, whereby the fully connected layer performs transformation processing on the classification feature data; inputting the transformation result output by the fully connected layer into a Poisson regression function in the intention quantity prediction network, whereby the Poisson regression function is used to determine the intention quantity distribution parameters; and determining the number of user intentions within a preset intention quantity range based on the intention quantity distribution parameters.

[0008] In one implementation, the step of segmenting the user speech data according to the number of user intentions and determining the semantic segments of intentions corresponding to each potential user intention in the user speech data includes: inputting the user speech data into a pre-trained semantic encoding model, so that the semantic encoding model generates feature representation data with contextual semantic information for each word in the user speech data; inputting the feature representation data corresponding to each word and the number of user intentions into a pre-trained conditional random field model, so that the conditional random field model, after masking the start position labels and internal position labels of the associated intention segments with indices greater than the number of user intentions. The constrained sequence is decoded according to a preset label transfer rule, outputting the target position label of each word and the target segment number to which each word belongs; wherein, the structure configuration of the pre-trained conditional random field model includes a preset position label system and the maximum number of intent segment numbers, the preset position label system includes external labels for segments and start position labels and internal position labels set for each intent segment number, and the label constraint rule at least limits the label transfer relationship between adjacent words; according to the target position label and the target segment number of each word, words belonging to the same target segment number are merged to obtain each intent semantic segment.

[0009] In one embodiment, after obtaining each of the intent semantic segments, the method further includes: extracting initial slot information corresponding to each intent semantic segment; performing syntactic structure disambiguation processing on the initial slot information according to preset syntactic dependency rules to generate a first disambiguation result; performing slot-to-slot logical dependency disambiguation processing on the first disambiguation result according to preset logical dependency rules to generate a second disambiguation result; if there are fuzzy slot parameters in the second disambiguation result that do not reach a preset confidence threshold, generating targeted follow-up information for the fuzzy slot parameters and outputting the targeted follow-up information; and updating the fuzzy slot parameters in response to user feedback instructions for the targeted follow-up information to determine the slot filling result of the fuzzy slot parameters.

[0010] In one embodiment, after determining the execution timing relationship between the various intent semantic segments, the method further includes: performing compliance verification on each intent semantic segment according to preset compliance requirements; if any intent semantic segment fails the compliance verification, generating and outputting a compliance exception prompt message to prevent the target intelligent agent from executing the business operation corresponding to the non-compliant semantic segment, and re-determining the processing method of subsequent semantic segments that have a dependency relationship with the non-compliant semantic segment according to the execution timing relationship.

[0011] In one embodiment, the method further includes: if there is a mutual exclusion conflict among the various intent semantic segments, generating clarification consultation information corresponding to the mutual exclusion conflict; outputting the clarification consultation information to obtain conflict resolution instructions from the user regarding the mutual exclusion conflict; revising the intent semantic segments with mutual exclusion conflicts according to the conflict resolution instructions to generate revised semantic segments; replacing the intent semantic segments with mutual exclusion conflicts with the revised semantic segments, and re-checking the business conflict of each intent semantic segment.

[0012] In one implementation, assigning each of the intent semantic fragments to a corresponding target agent, so that the target agent executes a business response operation matching the intent semantic fragment, includes: obtaining a three-layer target database, the target database including a user profile layer database, a session layer database, and an intent layer database, the user profile layer database storing cross-session global user data, the session layer database storing user session data generated during the current session interaction, and the intent layer database storing temporary data generated by a single intent semantic fragment during its execution lifecycle; determining the subscription permissions of each target agent to each layer of the target database according to the execution requirements of each target agent; and during the execution of a response operation matching the intent semantic fragment by each target agent, allowing the target agent to perform data read and / or write operations on the target database only within the scope limited by the subscription permissions.

[0013] In one implementation, assigning each intent semantic fragment to a corresponding target agent to enable the target agent to execute a business response operation matching the intent semantic fragment includes: acquiring at least two initial query requests initiated by the target agents during execution; normalizing each initial query request and performing SimHash calculation on the normalized query semantics to generate a semantic fingerprint corresponding to each initial query request; calculating the Hamming distance between any two semantic fingerprints and identifying the initial query requests corresponding to two semantic fingerprints whose Hamming distance is less than a preset distance threshold as candidate duplicate query requests; merging the duplicate query requests into a single merge query request when the operation type, data source, normalization parameters, user or tenant identifier, access permission range, and snapshot version of the candidate duplicate query requests are all consistent; initiating the merge query request to the target database to obtain the merge query result; and feeding back the merge query result to each target agent corresponding to the duplicate query request.

[0014] In one embodiment, the method further includes: acquiring the actual tool call sequence of the target agent during the execution of a response operation; calculating a deviation degree based on the deviation between the actual tool call sequence and the expected tool set; generating an intent omission candidate event if the deviation degree is greater than a preset deviation threshold, or the semantic relevance between the response result output by the target agent and the user speech data is less than a preset relevance threshold, or receiving an execution anomaly flag actively reported by the target agent; inferring the omission intent type and its omission confidence degree by combining the user speech data, the intent semantic fragment, the actual tool call sequence, and the session context of the user speech data when generating the intent omission candidate event; if the omission confidence degree reaches a preset omission threshold, and the inferred intent corresponds to an authorized non-transactional operation, invoking a supplementary agent corresponding to the inferred intent to execute a supplementary response operation; and outputting confirmation information if the inferred intent involves transactions, asset disclosure, or permission expansion, and executing the corresponding operation after obtaining user confirmation and passing identity and permission verification.

[0015] Secondly, this disclosure provides an intelligent question-answering interaction system supporting the coexistence of multiple intents. The system includes: a data acquisition module for acquiring user speech data in natural language form; an intent recognition module for predicting the number of user intents contained in the user speech data, and segmenting the user speech data according to the number of user intents to determine the semantic segments of intents corresponding to each potential user intent; and a conflict detection module for performing business conflict checks on each of the semantic segments of intents to determine whether there are mutual exclusion conflicts between the semantic segments of intents, and, if there are no mutual exclusion conflicts, determining the execution sequence relationship between the semantic segments of intents based on the business conflict check results. The timing relationship includes serial execution and parallel execution; the intent response module is used to assign each intent semantic segment to the corresponding target agent according to the execution timing relationship if there is no mutual exclusion conflict between the various intent semantic segments, so that the target agent executes a business response operation that matches the intent semantic segment; wherein, in the parallel execution phase of the intent semantic segment, the target agent executes the business response operation based on the same multi-version concurrency control snapshot; in the serial execution phase of the intent semantic segment, the target agent in the subsequent execution phase executes the business response operation based on the updated snapshot or the verified pre-output after the pre-operation is submitted and the intermediate state is verified.

[0016] The technical solution provided by one or more embodiments of this disclosure first predicts the number of intents in a user's speech and segments those intents. This allows a user's sentence to be broken down into several independent semantic segments of intents, enabling each potential intent to be identified and processed individually. Compared to traditional methods that only predict a single dominant intent, this segmentation mechanism effectively avoids the problem of losing the user's "other half" needs due to redundancy in intent classification or prioritization. Simultaneously, the entities contained within each semantic segment (such as amount, product code, quantity, etc.) can be naturally assigned to the corresponding intent, clearly defining the scope of use of shared entities across multiple intents and reducing ambiguity in subsequent response execution.

[0017] The technical solutions provided in one or more embodiments of this disclosure, after segmenting into multiple semantic segments of intent, further introduce business conflict checks. This allows for the pre-determination of whether there are mutually exclusive relationships between the intents (e.g., conflicts between "buying" and "selling" for the same product under the same account), thereby preventing business anomalies caused by the simultaneous execution of conflicting intents. Based on the confirmation of no conflicts, the execution sequence relationship (serial or parallel) between the intent segments is determined according to business logic. This provides a clear order basis for the subsequent coordinated execution of multiple agents, enabling dependent operations (such as querying before trading) to be completed sequentially, while independent operations can be processed in parallel, improving execution efficiency.

[0018] This disclosure provides a technical solution through one or more embodiments. During the parallel execution phase of the intent semantic segment, each target agent executes its response operation based on the same multi-version concurrent control snapshot. This ensures that multiple agents see a consistent data state version when accessing shared data (such as user holdings or account balances), avoiding dirty reads or logical errors caused by reading data at different times. During the serial execution phase of the intent semantic segment, subsequent agent operations must wait for the preceding operation to be successfully submitted and verified before proceeding. This avoids the risk of continuing execution based on incomplete or erroneous data, ensuring data consistency and the correctness of the final result. Attached Figure Description

[0019] The features and advantages of the embodiments of this disclosure will be more clearly understood by referring to the accompanying drawings, which are illustrative and should not be construed as limiting the present disclosure in any way. In the drawings: Figure 1 The illustration shows a step diagram of an intelligent question-answering interaction method supporting multiple intents coexisting in one embodiment of this disclosure; Figure 2 A flowchart illustrating an intelligent question-answering interaction method supporting multiple intents coexisting in one embodiment of this disclosure is shown. Figure 3A schematic diagram of the functional modules of an intelligent question-answering interaction system supporting multiple intents coexistence is shown in one embodiment of this disclosure. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0021] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0022] Please see Figure 1 The present disclosure provides an intelligent question-answering interaction method that supports the coexistence of multiple intents, which may include the following steps.

[0023] S1: Obtain user speech data in natural language form.

[0024] In this embodiment, user speech data can be input by the user through various interactive methods such as voice input, keyboard input, touch selection, and gesture drawing. User speech data can be newly input speech content in the current session or speech from subsequent rounds in the historical session record that is yet to be processed.

[0025] In this implementation, user speech data may contain only one user intent or multiple user intents. Different user intents may include different data operation types such as reading, writing, and modifying. For example, "Please check the returns of a certain fund over the past year, and if it's suitable, I'll buy it and sell another fund I hold at the same time," involves three intents: querying, buying, and selling. Different user intents may be concurrent operations (e.g., selling fund A while simultaneously purchasing fund B using the account balance) or may need to be executed sequentially (e.g., selling fund A and then using the received funds to purchase fund B).

[0026] S2: Predict the number of user intentions contained in the user speech data, and segment the user speech data according to the number of user intentions to determine the semantic segments of intentions in the user speech data corresponding to each potential user intention.

[0027] In this implementation, by predicting the number of intents in a user's speech and segmenting those intents, a user's sentence can be broken down into several independent semantic fragments of intent, allowing each user's potential intent to be identified and processed individually. The ingenuity of this design lies in decoupling "identifying the number of user intents" from "identifying the content of user intents," separating "how many intents are in this sentence" from "where each intent begins and ends" into two steps, which reduces computational complexity and improves data processing efficiency.

[0028] In some implementations, predicting the number of user intentions contained in the user utterance data includes: inputting the user utterance data into a semantic feature classification unit in a pre-trained intention quantity prediction network to generate classification feature data labeled with classification tags; inputting the classification feature data into a fully connected layer in the intention quantity prediction network, where the fully connected layer performs transformation processing on the classification feature data; inputting the transformation result output by the fully connected layer into a Poisson regression function in the intention quantity prediction network, using the Poisson regression function to determine the intention quantity distribution parameters; and determining the number of user intentions within a preset intention quantity range based on the intention quantity distribution parameters.

[0029] Specifically, user utterance data can be processed through three stages using a pre-trained intent count prediction network. First, a semantic feature classification unit generates categorical feature data with classification labels (e.g., the [CLS] label in a Transformer network architecture). This categorical feature data captures semantic cues related to intent counts in the user utterance. Then, a fully connected layer performs a nonlinear transformation on this categorical feature data, mapping high-dimensional semantic information to input parameters of a Poisson distribution. Finally, the Poisson regression function outputs a strength parameter characterizing the intent count distribution and calculates the probability of each candidate count within a preset intent count range. The user intent count is determined by the maximum posterior candidate count or a validated discretization rule. Since the Poisson regression outputs distribution parameters rather than naturally non-negative integers, it allows for setting upper limits on intent counts, constraints with a minimum value of 1, and alternative classification or follow-up strategies when confidence is insufficient.

[0030] In some implementations, the step of segmenting the user speech data according to the number of user intentions and determining the semantic segments of intentions corresponding to each potential user intention in the user speech data includes: inputting the user speech data into a pre-trained semantic encoding model, so that the semantic encoding model generates feature representation data with contextual semantic information for each word in the user speech data; inputting the feature representation data corresponding to each word and the number of user intentions into a pre-trained conditional random field model, so that the conditional random field model, after masking the start position labels and internal position labels of the associated intention segments with indices greater than the number of user intentions. The constrained sequence is decoded according to a preset label transfer rule, outputting the target position label of each word and the target segment number to which each word belongs; wherein, the structure configuration of the pre-trained conditional random field model includes a preset position label system and the maximum number of intent segment numbers, the preset position label system includes external labels for segments and start position labels and internal position labels set for each intent segment number, and the label constraint rule at least limits the label transfer relationship between adjacent words; according to the target position label and the target segment number of each word, words belonging to the same target segment number are merged to obtain each intent semantic segment.

[0031] Specifically, a semantic coding model is used to generate feature representation data with contextual semantic information for each word in the user's speech. This ensures that the feature vector representation of each word not only includes its own semantics but also incorporates its contextual role in the entire sentence. The feature representation data of each word is then input into a Conditional Random Field (CRF) model. Under the constraint of the number of user intents, the model performs sequence labeling on each word according to preset labeling rules, outputting the intent category and position label (start, internal, and external) for each word. The core advantage of the CRF model lies in its ability to globally model the label transfer relationships between adjacent words, ensuring that the labeling results of the entire sequence are logically consistent. For example, it avoids invalid labeling sequences such as "the start label of intent A is immediately followed by the internal label of intent B." By merging words labeled with the same intent and arranged consecutively, complete semantic segments of intents can be obtained. This sequence labeling-based segmentation method can accurately characterize the boundary positions of multiple intents in the text, laying a structured foundation for the independent slot filling and business processing of each subsequent intent segment.

[0032] It should be noted that if the predicted user intent is only 1, intent recognition can be performed directly without needing to execute the combined process of the semantic encoding model and the conditional random field model. The combined use of the semantic encoding model and the conditional random field model is mainly for scenarios where multiple intents coexist, enhancing the accuracy of each intent semantic fragment.

[0033] In a practical application example, the semantic encoding model can choose the BERT model, which is responsible for generating high-dimensional vectors rich in contextual semantics for each token (the smallest unit of text processing) in the user's speech data, providing accurate feature representations. The Conditional Random Field (CRF) model, as the decoding layer, is responsible for learning and enforcing global labeling rules for the entire user's speech data (e.g., a "start tag" must be followed immediately by an "internal tag" of the same category), avoiding illogical tag combinations. The combined use of these two models achieves a complementarity between "strong local semantic understanding" and "global optimal path constraints," ensuring that the boundaries of each segmented sub-intent are continuous, complete, and legal, thus perfectly capable of span-level accurate segmentation and entity attribution determination in multi-intent scenarios.

[0034] In a practical application example, a maximum number of intent fragments K is preset, and the fixed label set is {O, Bk, Ik | 1≤k≤K}. For the user utterance data "Help me redeem my fund and subscribe to 30,000", after predicting the number of user intents N=2, labels with k>2 can be masked. Then, according to the order of the first appearance of the fragments, we can get: Help (O)|Me (O)|Redeem (B-1)|Fund (I-1)|And (O)|Subscribe (B-2)|30,000 (I-2). Subsequently, fragment 1 and fragment 2 are classified according to intent, and their business intent types are determined to be redemption and subscription. The fragment number k only indicates the span belonging in the current utterance and is not equivalent to the fixed business intent category.

[0035] In some implementations, after obtaining each of the intent semantic segments, the method further includes: extracting initial slot information corresponding to each intent semantic segment; performing syntactic structure disambiguation processing on the initial slot information according to preset syntactic dependency rules to generate a first disambiguation result; performing slot-to-slot logical dependency disambiguation processing on the first disambiguation result according to preset logical dependency rules to generate a second disambiguation result; if there are fuzzy slot parameters in the second disambiguation result that do not reach a preset confidence threshold, generating targeted follow-up information for the fuzzy slot parameters and outputting the targeted follow-up information; and updating the fuzzy slot parameters in response to user feedback instructions for the targeted follow-up information to determine the slot filling result of the fuzzy slot parameters.

[0036] Specifically, after obtaining the semantic fragments of each intent, the initial slot information corresponding to each fragment (such as key parameters like product name, amount, and quantity) can be extracted. However, this initial slot information may have grammatical ambiguities (e.g., in "redeeming 50%", does the percentage refer to the proportion of the held shares or the proportion of the total assets) or logical dependency ambiguities (e.g., in "reinvesting in this fund", the "this" needs to rely on the preceding context to determine its referent). To address this, the solution first performs syntactic structure disambiguation based on preset grammatical dependency rules to resolve slot comprehension biases caused by unclear sentence component division; then, it performs logical dependency disambiguation between slots based on preset logical dependency rules to resolve semantic ambiguity caused by implicit relationships between slots. If, after the above two rounds of disambiguation, there are still fuzzy slot parameters that do not reach the preset reliability threshold, then targeted follow-up questions are generated and output to the user, and the slot assignment is updated after waiting for user feedback. This mechanism effectively ensures the accurate acquisition of key parameters in semantic segments of each intent in multi-intent scenarios, and handles ambiguity problems that cannot be solved by automatic disambiguation through human-computer interaction, thereby improving the system's fault tolerance for complex natural language expressions.

[0037] In a practical application example, grammatical dependency rules utilize syntactic analysis, such as determining the direct object of a verb. If the syntax tree explicitly shows that "30,000" is a verb-object relationship for the verb "purchase", then grammatical-level strong binding takes effect directly, without the need for subsequent rules.

[0038] In a practical application example, business logic validation is enabled when syntactic relationships are ambiguous. For instance, if a user's statement is "redeem Fund A and subscribe to Fund B for 30,000 yuan," the syntactic relationship typically binds "30,000 yuan" to "subscribe to Fund B." Even if the user's holdings of Fund A are close to 30,000 yuan and their current account balance is insufficient, this only identifies the temporal relationship that "the subscription funds may depend on the redemption amount," and cannot be used to change "30,000 yuan" to the redemption amount of Fund A. The system should retain the slot for "subscription amount of 30,000 yuan" and generate targeted follow-up information regarding the redemption shares or amount of Fund A; only after obtaining user confirmation should the fund dependency relationship between redemption and subscription be determined.

[0039] In a practical application example, when the syntax and logic fail to accurately determine the user's intent (e.g., the user says, "Handle my fund, the one with 30,000 yuan"), the model abandons automatic guessing and generates a clear and specific rhetorical question to return to the user, such as "Do you want to redeem the 30,000 yuan, or use the 30,000 yuan to subscribe to a new product?". This targeted follow-up question ensures that the model does not make decisions for the user in scenarios with ambiguous intent, thus avoiding the loss of the user's assets.

[0040] S3: Perform a business conflict check on each of the intent semantic segments to determine whether there is a mutual exclusion conflict between the intent semantic segments. If there is no mutual exclusion conflict, determine the execution timing relationship between the intent semantic segments based on the business conflict check results. The timing relationship includes serial execution relationship and parallel execution relationship.

[0041] In this implementation, after segmenting the intent into multiple semantic fragments, a business conflict check is further introduced. This allows for the pre-determination of whether there are mutually exclusive relationships between the intents (e.g., a conflict between "buying" and "selling" the same product under the same account), thereby preventing business anomalies caused by the simultaneous execution of conflicting intents. Based on the confirmation of no conflict, the execution sequence (serial or parallel) between the intent fragments is determined according to business logic. This provides a clear order basis for the subsequent coordinated execution of multiple agents, enabling dependent operations (such as querying before trading) to be completed sequentially, while independent operations can be processed in parallel, improving execution efficiency.

[0042] In some implementations, after determining the execution timing relationship between the various intent semantic segments, the method further includes: performing compliance verification on each intent semantic segment according to preset compliance requirements; if any intent semantic segment fails the compliance verification, generating and outputting a compliance exception prompt message to prevent the target intelligent agent from executing the business operation corresponding to the non-compliant semantic segment, and re-determining the processing method of subsequent semantic segments that have a dependency relationship with the non-compliant semantic segment according to the execution timing relationship.

[0043] Specifically, after determining the execution sequence relationship between various intent fragments, each intent fragment is further validated for compliance based on preset compliance requirements (including but not limited to suitability matching, risk level, position limits, trading hours, and purchase / redemption restrictions) to ensure that the operation corresponding to each intent is legal and permitted at the business rule level. For example, in financial scenarios, compliance validation may include: whether the user has the identity authority to execute the operation, whether the operation meets the time window requirements of regulatory policies, and whether the operation complies with relevant regulations such as anti-money laundering. If any intent fragment fails the compliance validation, it means that the intent cannot be executed under the current business rules. The solution then generates and outputs a compliance anomaly message and terminates the response process of the corresponding target intelligent agent, avoiding risks to users or the platform due to illegal operations. This validation step moves the business compliance check to before actual execution, playing a role in risk interception and access control, reflecting the systematic design of the intelligent question-and-answer interaction method in terms of security.

[0044] In some implementations, the method further includes: if there is a mutual exclusion conflict among the various intent semantic segments, generating clarification consultation information corresponding to the mutual exclusion conflict; outputting the clarification consultation information to obtain conflict resolution instructions from the user regarding the mutual exclusion conflict; revising the intent semantic segments with mutual exclusion conflicts according to the conflict resolution instructions to generate revised semantic segments; replacing the intent semantic segments with mutual exclusion conflicts with the revised semantic segments, and re-checking the business conflict of each intent semantic segment.

[0045] Specifically, when business conflict checks identify mutually exclusive conflicts between various intent semantic fragments (e.g., "buy a fund" and "sell the same fund" cannot be performed simultaneously in the same account at the same time), the solution does not directly reject or discard some intents. Instead, it generates clarification information corresponding to the mutual exclusion conflict and outputs it to the user. This clarification information clearly informs the user of the current conflict point, guiding the user to make a choice or adjustment. After the user provides a conflict resolution instruction, the solution revises the conflicting intent semantic fragments according to the instruction, generates revised semantic fragments, replaces the original conflicting fragments with the revised fragments, and then re-checks the entire intent set for business conflicts. This process forms a closed-loop processing mechanism of "conflict detection—user clarification—intent revision—re-detection," giving the user the final decision-making power in intent conflict scenarios, rather than the system arbitrarily making choices. This design respects the user's true expressed intent while ensuring the feasibility of executing multiple intents, significantly improving the user experience in complex business scenarios.

[0046] S4: If there is no mutual exclusion conflict between the various intent semantic segments, then according to the execution timing relationship, each intent semantic segment is assigned to the corresponding target agent, so that the target agent executes the business response operation that matches the intent semantic segment; wherein, in the parallel execution phase of the intent semantic segment, the target agent executes the business response operation based on the same multi-version concurrent control snapshot; in the serial execution phase of the intent semantic segment, the target agent in the subsequent execution phase executes the business response operation based on the updated snapshot or the verified pre-output after the pre-operation is submitted and the intermediate state is verified.

[0047] In this implementation, target agents within the same parallel execution phase read shared data based on the same Multi-Version Concurrency Control (MVCC) snapshot to obtain a consistent data view. For subsequent execution phases with serial dependencies, an updated snapshot is created or obtained after the preceding operation is committed and passes intermediate state verification, or the verified preceding output is injected into the subsequent agent. MVCC can reduce read / write blocking, but it cannot eliminate concurrent write conflicts; version verification, idempotent keys, conflict detection, and retry or rollback strategies should still be used to address write conflicts.

[0048] In this implementation, MVCC snapshots can be represented by database snapshot identifiers, transaction timestamps, or visible version ranges, without requiring each task to generate a global version number corresponding to historical data. Each target agent in the same parallel phase uses the same snapshot identifier to read data; write operations generate new versions, and the baseline version is verified upon commit. If concurrent updates or changes in business state are detected, the commit is aborted, and the latest snapshot is reread, re-verified, or rolled back. This ensures consistent reading while explicitly controlling concurrent writes and serial dependencies.

[0049] In some implementations, assigning each of the intent semantic fragments to a corresponding target agent, so that the target agent executes a business response operation matching the intent semantic fragment, includes: obtaining a three-layer target database, the target database including a user profile layer database, a session layer database, and an intent layer database, the user profile layer database storing cross-session global user data, the session layer database storing user session data generated during the current session interaction, and the intent layer database storing temporary data generated by a single intent semantic fragment during its execution lifecycle; determining the subscription permissions of each target agent to each layer of the target database according to the execution requirements of each target agent; and during the execution of a response operation matching the intent semantic fragment by each target agent, allowing the target agent to perform data read and / or write operations on the target database only within the scope limited by the subscription permissions.

[0050] Specifically, the three-tiered target database includes a user profile layer (stores global user data across sessions, such as risk preferences and long-term holdings), a session layer (stores user session data generated during the current session interaction, such as historical messages of this conversation), and an intent layer (stores temporary data generated by a single intent semantic fragment during its execution lifecycle, such as intermediate calculation results of a single operation). Based on this, according to the execution needs of each target agent, its subscription permissions to each layer of the database are determined. During execution, the agent is only allowed to perform data read and / or write operations within the scope of its subscription permissions. This design achieves data isolation and access control during concurrent execution of multiple intents. On the one hand, each agent can only access its authorized data layer, preventing data security issues caused by unauthorized access. On the other hand, the intent layer database, as an independent temporary storage space for each intent, ensures that intermediate states generated by different intents during execution do not interfere with each other, improving the isolation and stability of parallel processing of multiple intents.

[0051] In some implementations, assigning each of the intent semantic fragments to a corresponding target agent, so that the target agent executes a business response operation matching the intent semantic fragment, includes: acquiring at least two initial query requests initiated by the target agents during execution; normalizing each initial query request and performing SimHash calculation on the normalized query semantics to generate a semantic fingerprint corresponding to each initial query request; calculating the Hamming distance between any two semantic fingerprints, and identifying the initial query requests corresponding to two semantic fingerprints whose Hamming distance is less than a preset distance threshold as candidate duplicate query requests; merging the duplicate query requests into a single merge query request when the operation type, data source, normalization parameters, user or tenant identifier, access permission range, and snapshot version of the candidate duplicate query requests are all consistent; initiating the merge query request to the target database to obtain the merge query result; and feeding back the merge query result to each of the target agents corresponding to the duplicate query requests.

[0052] Specifically, when at least two target agents initiate initial query requests during execution, the solution performs SimHash (a locality-sensitive hashing algorithm) calculation on each request to generate a corresponding semantic fingerprint. This fingerprint maps the semantic content of the query request to a fixed-length binary code. Then, the Hamming distance between any two semantic fingerprints is calculated. If the Hamming distance is less than a preset distance threshold, the two query requests are determined to be semantically identical or highly similar, and are identified as duplicate query requests. The solution merges these duplicate requests into a single merged query request, initiates a query to the target database, obtains the merged query result, and then feeds it back to the respective target agents. This optimization mechanism is particularly effective in scenarios with concurrent execution of multiple intents—multiple agents may need to query the same information (such as basic data of the same product, holding information of the same account, etc.) to process related intents. By deduplication and merging, redundant database queries are reduced, lowering database access pressure and network overhead, and improving the overall system response speed and resource utilization efficiency.

[0053] In some embodiments, the method further includes: acquiring the actual tool call sequence of the target agent during the execution of the response operation; calculating the deviation degree based on the deviation between the actual tool call sequence and the expected tool set; generating an intent omission candidate event if the deviation degree is greater than a preset deviation threshold, or the semantic relevance between the response result output by the target agent and the user speech data is less than a preset relevance threshold, or receiving an execution anomaly flag actively reported by the target agent; when generating the intent omission candidate event, inferring the omission intent type and its omission confidence degree by combining the user speech data, the intent semantic fragment, the actual tool call sequence, and the session context of the user speech data; if the omission confidence degree reaches a preset omission threshold, and the inferred intent corresponds to an authorized non-transactional operation, invoking a supplementary agent corresponding to the inferred intent to execute a supplementary response operation; if the inferred intent involves a transaction, asset disclosure, or permission expansion, outputting confirmation information, and executing the corresponding operation after obtaining user confirmation and passing identity and permission verification.

[0054] Specifically, considering that tool invocation deviations may also stem from agent planning errors, tool malfunctions, or incomplete configuration of the expected set, deviations cannot be directly equated with missed intents. Therefore, the solution simultaneously monitors three types of abnormal signals during execution: the deviation between the actual tool invocation sequence and the expected set, the semantic relevance of the response result to the user's utterance, and execution anomaly markers actively reported by the agent. When any signal exceeds a preset threshold or triggers a report, a candidate event for missed intent is generated, indicating that the initial intent prediction and segmentation may have missed a reasonable intent implicit in the user's utterance. After the candidate event is generated, the solution combines user utterance data, the current semantic fragment of the intent, the actual tool invocation sequence, and the session context to infer the specific type of the missed intent and calculate the omission confidence. If the confidence of the inferred intent reaches a preset threshold and belongs to a non-transactional authorized operation, the corresponding supplementary agent is invoked to execute a supplementary response; if the inferred intent involves sensitive operations such as transactions, asset disclosures, or permission expansion, the process is switched to user confirmation, and execution can only proceed after identity and permission verification. This mechanism provides a systematic safety net for the interaction process. Even when the initial intent recognition is incomplete, it can drive intent supplementation and diversion processing through abnormal signals in the execution phase. This effectively makes up for the inherent limitations of pre-intention prediction in dealing with implicit expressions and non-explicit needs, and enhances the system's adaptability and robustness to complex dialogue scenarios.

[0055] In some implementations, the results of the operations performed by each target agent are merged and then output in a unified manner.

[0056] Specifically, after each target agent completes its corresponding response operation, the system merges these operation results and provides them to the user in a unified response format. This merged output method integrates the processing results of multiple intents into a clear and organized reply, avoiding information fragmentation or repetitive reading caused by multiple independent responses, thereby improving the coherence of the interaction and the user experience.

[0057] In some implementations, the method further includes: acquiring execution feedback data corresponding to each intent semantic segment, the execution feedback data including at least one of execution status, response latency, and independent feedback submitted by the user for the intent semantic segment; constructing a message-intent two-layer data model, wherein message-level statistical indicators are based on the overall execution result or overall feedback statistics of the entire user utterance data, and intent-level statistical indicators are based on the independent execution data or independent feedback statistics of each intent semantic segment; and performing correlation analysis on the two layers of statistical indicators to locate abnormal intents or performance bottlenecks in the multi-intent execution process. The overall satisfaction level for the entire message must not be directly regarded as the independent satisfaction level for each intent.

[0058] Specifically, a statistical and analytical mechanism for execution feedback is added on top of the basic method, aiming to solve the problem that traditional evaluation at the whole message level cannot distinguish the individual performance of multiple intentions. In actual operation, the system can record the execution results of the agent corresponding to each semantic segment of the intention, including feedback data such as success or failure, duration, and user satisfaction with the overall response. By constructing a two-layer data model of message intentions, the system can perform statistics at two levels: at the message level, it calculates total indicators (such as the weighted average of overall satisfaction) on a whole dialogue basis for macro-evaluating the overall performance of the system; at the intention level, it calculates the success rate, average latency, and user rating for each intention type, thereby accurately reflecting the individual performance of different business operations (such as query, purchase, and sale). Furthermore, the system can cross-compare the data from the two layers. For example, if the overall satisfaction of a certain message is low, it can quickly locate the anomaly of a certain intention subtask (such as a purchase intention timeout or failure), and then optimize the processing logic or resource allocation of that intention accordingly. This design embeds operation and maintenance monitoring capabilities into the interaction process, providing fine-grained data support for continuous system iteration and fault diagnosis, and is especially suitable for complex business scenarios with multiple intentions occurring frequently.

[0059] Please see Figure 2 In response to a real-world financial scenario where a user inputs "Help me redeem the 30,000 yuan from product A, then subscribe to product B, and also check the yield of product C," this disclosure provides an intelligent question-and-answer interaction method that supports the coexistence of multiple intentions, which may include the following process.

[0060] Process 1: Intent boundary segmentation.

[0061] Based on the prediction that the user input contains three user intentions (N=3), we can obtain three semantic fragments of intentions: "redeem 30,000 yuan of product A", "purchase product B", and "query product C". The structured representation of these three semantic fragments of intentions is "I_1(redeem, product=A, amount=30,000) | I_2(purchase, product=B) | I_3(query_return, product=C)".

[0062] Step 2: Conflict pre-check.

[0063] Intent semantic fragments I_1 and I_2 involve different products, but their business operations are financially dependent. Therefore, it is confirmed that I_1 and I_2 are executed sequentially, with I_1 executing before I_2. Compared to I_1 / I_2, intent semantic fragment I_3 involves a pure query operation and does not conflict with I_1 / I_2. Therefore, it is confirmed that I_3 is executed in parallel with I_1. After compliance verification of the three intent semantic fragments, it is confirmed that all three passed the compliance verification.

[0064] Step 3: Shared context pool.

[0065] Load the user profile (including risk level R3, holdings list, etc.) all at once. During the semantic deduplication phase, since both Agent_Redemption and Agent_Purchase require "user holdings information," the query requests for the two agents can be merged into a single query. For this merged query request, create an MVCC snapshot V_1 of the user-related data.

[0066] Step 4: Scheduling and Execution.

[0067] During the parallel execution phase, Agent_Redemption (I_1) and Agent_Yield (I_3) read snapshot V_1. After Agent_Redemption submits a redemption request, Agent_Purchase (I_2) does not directly treat the outstanding redemption amount as an available balance, but enters a dependent waiting state. Once the redemption funds are actually received or reach an available state, a snapshot V_2 is created that updates the user's relevant data. The available balance, the minimum purchase amount for Product B, the transaction time, and the user's authorization are re-verified before the purchase operation is executed or submitted.

[0068] Step 5: Reverse extraction.

[0069] If the execution trajectory of Agent_yield shows a call to "obtain user's holdings information for product C" that exceeds the expected toolset, the system first marks it as a possible agent planning deviation or intent miss candidate, rather than directly inferring that the user wants to disclose holdings profits and losses. The system calculates the confidence level of the candidate intent by combining the original utterance and the conversation context; if the confidence level is insufficient, the system asks the user whether they need to view the profits and losses of product C holdings. Only after obtaining confirmation and passing identity and permission verification can the Agent_holding query be initiated.

[0070] Step 6: Result Synthesis.

[0071] Regarding the redemption of Product A: A redemption application of 30,000 yuan has been submitted, and it is expected to arrive on T+1.

[0072] Regarding the subscription of Product B: Subscription conditions have been recorded; once the redemption funds for Product A are actually received, the available account balance, product transaction status, and user authorization will be re-verified, and the subscription will be submitted when the confirmation or entrustment rules are met.

[0073] Regarding the yield of Product C: The yield over the past 3 months is 2.8%.

[0074] You may also be interested in: You currently hold 1000 units of Product C, with a holding return of +2.1%.

[0075] Step 7: Persistence. Write to a two-layer JSON, and the front end displays 3+1 independent feedback buttons for each intent section (3 potential user intents after the user statement is broken down, and the global user intent corresponding to the overall user statement).

[0076] The technical solution provided by one or more embodiments of this disclosure first predicts the number of intents in a user's speech and segments those intents. This allows a user's sentence to be broken down into several independent semantic segments of intents, enabling each potential intent to be identified and processed individually. Compared to traditional methods that only predict a single dominant intent, this segmentation mechanism effectively avoids the problem of losing the user's "other half" needs due to redundancy in intent classification or prioritization. Simultaneously, the entities contained within each semantic segment (such as amount, product code, quantity, etc.) can be naturally assigned to the corresponding intent, clearly defining the scope of use of shared entities across multiple intents and reducing ambiguity in subsequent response execution.

[0077] The technical solutions provided in one or more embodiments of this disclosure, after segmenting into multiple semantic segments of intent, further introduce business conflict checks. This allows for the pre-determination of whether there are mutually exclusive relationships between the intents (e.g., conflicts between "buying" and "selling" for the same product under the same account), thereby preventing business anomalies caused by the simultaneous execution of conflicting intents. Based on the confirmation of no conflicts, the execution sequence relationship (serial or parallel) between the intent segments is determined according to business logic. This provides a clear order basis for the subsequent coordinated execution of multiple agents, enabling dependent operations (such as querying before trading) to be completed sequentially, while independent operations can be processed in parallel, improving execution efficiency.

[0078] This disclosure provides a technical solution through one or more embodiments. During the parallel execution phase of the intent semantic segment, each target agent executes its response operation based on the same multi-version concurrent control snapshot. This ensures that multiple agents see a consistent data state version when accessing shared data (such as user holdings or account balances), avoiding dirty reads or logical errors caused by reading data at different times. During the serial execution phase of the intent semantic segment, subsequent agent operations must wait for the preceding operation to be successfully submitted and verified before proceeding. This avoids the risk of continuing execution based on incomplete or erroneous data, ensuring data consistency and the correctness of the final result.

[0079] Please see Figure 3 This disclosure also provides an intelligent question-answering interaction system that supports the coexistence of multiple intents, the system comprising: Data acquisition module 100 is used to acquire user speech data in natural language form; The intent recognition module 200 is used to predict the number of user intents contained in the user speech data, and to perform intent segmentation on the user speech data according to the number of user intents, and to determine the intent semantic segment in the user speech data corresponding to each potential user intent. The conflict detection module 300 is used to perform business conflict checks on each of the intent semantic segments, determine whether there is a mutual exclusion conflict between the various intent semantic segments, and, when there is no mutual exclusion conflict, determine the execution timing relationship between the various intent semantic segments based on the business conflict check results. The timing relationship includes serial execution relationship and parallel execution relationship. The intent response module 400 is used to assign each intent semantic segment to the corresponding target intelligent agent according to the execution timing relationship if there is no mutual exclusion conflict between the various intent semantic segments, so that the target intelligent agent executes a business response operation that matches the intent semantic segment; In the parallel execution phase of the intent semantic fragment, the target agent executes the business response operation based on the same multi-version concurrent control snapshot; in the serial execution phase of the intent semantic fragment, the target agent in the subsequent execution phase executes the business response operation based on the updated snapshot or the verified pre-output after the pre-operation is submitted and the intermediate state is verified.

[0080] For a detailed explanation of each module, please refer to the relevant content in the method embodiment above, which will not be repeated here.

[0081] In one embodiment, the intent recognition module 200 is specifically configured to: input the user utterance data into a semantic feature classification unit in a pre-trained intent quantity prediction network to generate classification feature data labeled with classification tags; input the classification feature data into a fully connected layer in the intent quantity prediction network, where the fully connected layer performs transformation processing on the classification feature data; input the transformation result output by the fully connected layer into a Poisson regression function in the intent quantity prediction network, using the Poisson regression function to determine intent quantity distribution parameters; and determine the number of user intents within a preset intent quantity range based on the intent quantity distribution parameters.

[0082] In one embodiment, the intent recognition module 200 is further configured to: input the user utterance data into a pre-trained semantic coding model, so that the semantic coding model generates feature representation data with contextual semantic information for each word in the user utterance data; input the feature representation data corresponding to each word and the number of user intents into a pre-trained conditional random field model, so that the conditional random field model, after masking the start position labels and internal position labels of the associated intent fragment numbers greater than the number of user intents, performs constrained sequence decoding according to a preset label transfer rule, and outputs the target position label of each word and the target fragment number to which each word belongs; wherein, the structural configuration of the pre-trained conditional random field model includes a preset position label system and the maximum number of intent fragment numbers, the preset position label system includes external labels of fragments and start position labels and internal position labels set for each intent fragment number respectively, and the label constraint rule at least limits the label transfer relationship between adjacent words; and merge words belonging to the same target fragment number according to the target position label and the target fragment number of each word to obtain each intent semantic fragment.

[0083] In one embodiment, after obtaining each of the intent semantic segments, the intent recognition module 200 is further configured to: extract initial slot information corresponding to each intent semantic segment; perform syntactic structure disambiguation processing on the initial slot information according to preset syntactic dependency rules to generate a first disambiguation result; perform slot logical dependency disambiguation processing on the first disambiguation result according to preset logical dependency rules to generate a second disambiguation result; if there is a fuzzy slot parameter in the second disambiguation result that does not reach a preset confidence threshold, generate targeted follow-up information for the fuzzy slot parameter and output the targeted follow-up information; and, in response to the user's feedback instruction on the targeted follow-up information, assign and update the fuzzy slot parameter to determine the slot filling result of the fuzzy slot parameter.

[0084] In one embodiment, the conflict detection module 300 is further configured to: perform compliance verification on each of the intent semantic segments according to preset compliance requirements; if any of the intent semantic segments fails the compliance verification, generate and output a compliance exception prompt message to prevent the target intelligent agent from executing the business operation corresponding to the non-compliant semantic segment, and re-determine the processing method of subsequent semantic segments that have a dependency relationship with the non-compliant semantic segment according to the execution timing relationship.

[0085] In one embodiment, the conflict detection module 300 is further configured to: generate clarification consultation information corresponding to the mutual exclusion conflict if there is a mutual exclusion conflict among the various intent semantic segments; output the clarification consultation information to obtain the user's conflict resolution instruction for the mutual exclusion conflict; revise the intent semantic segments with mutual exclusion conflicts according to the conflict resolution instruction to generate revised semantic segments; replace the intent semantic segments with mutual exclusion conflicts with the revised semantic segments, and re-check the business conflict of each intent semantic segment.

[0086] In one embodiment, the intent response module 400 is specifically configured to: acquire a three-tier target database, the target database including a user profile layer database, a session layer database, and an intent layer database, wherein the user profile layer database stores global user data across sessions, the session layer database stores user session data generated during the current session interaction, and the intent layer database stores temporary data generated by a single intent semantic fragment during its execution lifecycle; determine the subscription permissions of each target agent to each layer of the target database according to the execution requirements of each target agent; and during the execution of a response operation matching the intent semantic fragment by each target agent, allow the target agent to perform data reading and / or writing operations on the target database only within the scope limited by the subscription permissions.

[0087] In one embodiment, the intent response module 400 is further configured to: acquire at least two initial query requests initiated by the target agents during execution; normalize each initial query request and perform SimHash calculation on the normalized query semantics to generate a semantic fingerprint corresponding to each initial query request; calculate the Hamming distance between any two semantic fingerprints and determine the initial query requests corresponding to two semantic fingerprints whose Hamming distance is less than a preset distance threshold as candidate duplicate query requests; merge the duplicate query requests into a single merge query request when the operation type, data source, normalization parameters, user or tenant identifier, access permission range, and snapshot version of the candidate duplicate query requests are all consistent; initiate the merge query request to the target database to obtain the merge query result; and feed back the merge query result to each target agent corresponding to the duplicate query request.

[0088] In one embodiment, the system further includes a reverse extraction module, which is used to: obtain the actual tool call sequence of the target agent during the execution of the response operation; calculate the deviation degree based on the deviation between the actual tool call sequence and the expected tool set; generate an intent omission candidate event if the deviation degree is greater than a preset deviation threshold, or the semantic relevance between the response result output by the target agent and the user speech data is less than a preset relevance threshold, or an execution anomaly flag actively reported by the target agent is received; when generating the intent omission candidate event, infer the omission intent type and its omission confidence degree by combining the user speech data, the intent semantic fragment, the actual tool call sequence and the session context of the user speech data; if the omission confidence degree reaches a preset omission threshold and the inferred intent corresponds to an authorized non-transactional operation, call the supplementary agent corresponding to the inferred intent to execute a supplementary response operation; if the inferred intent involves a transaction, asset disclosure or permission expansion, output confirmation information and execute the corresponding operation after obtaining user confirmation and passing identity and permission verification.

[0089] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing this application, the functions of each module can be implemented in one or more software and / or hardware.

[0090] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0091] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0092] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0093] Although embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A smart question-answering interaction method supporting the coexistence of multiple intents, characterized in that, The method includes: Acquire user speech data in natural language format; Predict the number of user intentions contained in the user speech data, and segment the user speech data according to the number of user intentions to determine the semantic segments of intentions in the user speech data corresponding to each potential user intention. A business conflict check is performed on each of the intent semantic segments to determine whether there is a mutual exclusion conflict between the intent semantic segments. If there is no mutual exclusion conflict, the execution timing relationship between the intent semantic segments is determined based on the business conflict check results. The timing relationship includes serial execution relationship and parallel execution relationship. If there is no mutual exclusion conflict between the various intent semantic segments, then according to the execution timing relationship, each intent semantic segment is assigned to the corresponding target intelligent agent, so that the target intelligent agent executes the business response operation that matches the intent semantic segment; In the parallel execution phase of the intent semantic fragment, the target agent executes the business response operation based on the same multi-version concurrent control snapshot; In the serial execution phase of the intent semantic fragment, the target agent in the subsequent execution phase, after submitting the pre-operation and passing the intermediate state verification, executes the business response operation based on the updated snapshot or the verified pre-output.

2. The method according to claim 1, characterized in that, The prediction of the number of user intents contained in the user utterance data includes: The user utterance data is input into the semantic feature classification unit in a pre-trained intent quantity prediction network to generate classification feature data labeled with classification tags; The classification feature data is input into the fully connected layer of the intent quantity prediction network, and the fully connected layer transforms the classification feature data. The transformation result output by the fully connected layer is input into the Poisson regression function in the intent quantity prediction network, and the intent quantity distribution parameters are determined by the Poisson regression function. The number of user intents is determined within a preset range based on the intent quantity distribution parameters.

3. The method according to claim 1 or 2, characterized in that, The step of segmenting the user speech data according to the number of user intentions, and determining the semantic segment of the intention corresponding to each user's potential intention in the user speech data, includes: The user speech data is input into a pre-trained semantic coding model so that the semantic coding model generates feature representation data with contextual semantic information for each word in the user speech data. The feature representation data corresponding to each word and the number of user intents are input into a pre-trained Conditional Random Field (CRF) model. The CRF model, after masking the start position labels and internal position labels of the associated intent fragment numbers greater than the number of user intents, performs constrained sequence decoding according to preset label transfer rules, outputting the target position label for each word and the target fragment number to which each word belongs. The pre-trained CRF model's structural configuration includes a preset position label system and a maximum number of intent fragment numbers. The preset position label system includes external fragment labels and start position labels and internal position labels set for each intent fragment number. The label constraint rules at least limit the label transfer relationship between adjacent words. Based on the target location label and target segment number of each word, words belonging to the same target segment number are merged to obtain each intention semantic segment.

4. The method according to claim 3, characterized in that, After obtaining each of the stated intent semantic fragments, the method further includes: Extract the initial slot information corresponding to each of the aforementioned intent semantic segments; According to the preset grammatical dependency rules, the initial slot information is subjected to syntactic structure disambiguation processing to generate the first disambiguation result; According to the preset logical dependency rules, the first disambiguation result is processed for inter-slot logical dependency disambiguation to generate the second disambiguation result. If there are fuzzy slot parameters in the second disambiguation result that do not reach the preset confidence threshold, then targeted follow-up information for the fuzzy slot parameters is generated and the targeted follow-up information is output. In response to the user's feedback instruction regarding the targeted follow-up information, the fuzzy slot parameter is updated and the slot filling result of the fuzzy slot parameter is determined.

5. The method according to claim 1, characterized in that, After determining the execution timing relationship between the various intent semantic segments, the method further includes: According to the preset compliance requirements, each of the stated intent semantic segments is subject to compliance verification; If any of the stated intent semantic segments fails the compliance check, a compliance error message is generated and output to prevent the target agent from executing the business operation corresponding to the non-compliant semantic segment. Based on the execution sequence relationship, the processing method for subsequent semantic segments that are dependent on the non-compliant semantic segment is re-determined.

6. The method according to claim 1, characterized in that, The method further includes: If there is a mutual exclusion conflict among the various intentional semantic segments, then clarification consultation information corresponding to the mutual exclusion conflict is generated; Output the clarification consultation information to obtain conflict resolution instructions from the user regarding the mutual exclusion conflict feedback; According to the conflict resolution instructions, the semantic segments of intentions that have mutual exclusion conflicts are revised to generate revised semantic segments; Replace the mutually exclusive semantic fragments with the revised semantic fragments, and re-check the business conflict of each of the intention semantic fragments.

7. The method according to claim 1, characterized in that, Assigning each of the intent semantic fragments to a corresponding target agent, so that the target agent executes a service response operation matching the intent semantic fragment, includes: Obtain the target database with a three-tier architecture, which includes a user profile layer database, a session layer database, and an intent layer database. The user profile layer database stores global user data across sessions, the session layer database stores user session data generated during the current session interaction, and the intent layer database stores temporary data generated by a single intent semantic fragment during its execution lifecycle. Based on the execution requirements of each target intelligent agent, determine the subscription permissions of each target intelligent agent to each layer of the database in the target database; During the execution of a response operation that matches the intent semantic fragment by each target agent, the target agent is only allowed to perform data reading and / or writing operations on the target database within the scope limited by the subscription permissions.

8. The method according to claim 1 or 7, characterized in that, Assigning each of the intent semantic fragments to a corresponding target agent, so that the target agent executes a service response operation matching the intent semantic fragment, includes: Obtain at least two initial query requests initiated by the target agents during execution; Each initial query request is normalized, and the normalized query semantics are calculated using SimHash to generate a semantic fingerprint corresponding to each initial query request. Calculate the Hamming distance between any two semantic fingerprints, and identify the initial query requests corresponding to two semantic fingerprints whose Hamming distance is less than a preset distance threshold as candidate duplicate query requests; When the operation type, data source, normalization parameters, user or tenant identifier, access permission scope, and snapshot version of the candidate duplicate query requests are all the same, the duplicate query requests will be merged into a single merge query request. Initiate the merge query request to the target database to obtain the merge query results; The merged query results are then fed back to the respective target agents corresponding to the duplicate query requests.

9. The method according to claim 1, characterized in that, The method further includes: Obtain the actual tool call sequence of the target agent during the execution of the response operation; Calculate the deviation degree based on the deviation between the actual tool call sequence and the expected tool set; If the deviation is greater than a preset deviation threshold, or the semantic relevance between the response output by the target agent and the user speech data is less than a preset relevance threshold, or an execution anomaly flag is received by the target agent, then an intent miss detection candidate event is generated. When generating the intent missed candidate events, the type of missed intent and its missing confidence are inferred by combining the user speech data, the semantic fragments of the intent, the actual tool call sequence and the session context of the user speech data. If the omission confidence level reaches the preset omission threshold, and the inferred intent corresponds to an authorized non-transactional operation, then the supplementary agent corresponding to the inferred intent is invoked to perform a supplementary response operation. If the inferred intent involves a transaction, asset disclosure, or expansion of permissions, a confirmation message is output, and the corresponding operation is executed after obtaining user confirmation and passing identity and permission verification.

10. An intelligent question-answering interaction system supporting the coexistence of multiple intents, characterized in that, The system includes: The data acquisition module is used to acquire user speech data in natural language form; The intent recognition module is used to predict the number of user intents contained in the user speech data, and to segment the user speech data according to the number of user intents, and to determine the semantic segment of intent in the user speech data corresponding to each potential user intent. The conflict detection module is used to perform business conflict checks on each of the intent semantic segments, determine whether there is a mutual exclusion conflict between the intent semantic segments, and, if there is no mutual exclusion conflict, determine the execution timing relationship between the intent semantic segments based on the business conflict check results. The timing relationship includes serial execution relationship and parallel execution relationship. The intent response module is used to assign each intent semantic segment to the corresponding target agent according to the execution timing relationship if there is no mutual exclusion conflict between the various intent semantic segments, so that the target agent executes a business response operation that matches the intent semantic segment; In the parallel execution phase of the intent semantic fragment, the target agent executes the business response operation based on the same multi-version concurrent control snapshot; In the serial execution phase of the intent semantic fragment, the target agent in the subsequent execution phase, after submitting the pre-operation and passing the intermediate state verification, executes the business response operation based on the updated snapshot or the verified pre-output.

Images