A voice intent recognition method and system based on an incoming call consultation agent

Abstract

The application relates to the technical field of voice analysis, and discloses a voice intention recognition method and system based on an incoming call consultation Agent, wherein the method comprises the following steps: extracting current round fusion representation data; constructing a current round consultation memory unit; performing gap type driven historical memory recall based on the consultation memory unit; performing entity attribute conflict resolution and evidence fusion; performing hierarchical memory writing and Agent action mapping; compared with the prior art, in a long-term multi-round incoming call consultation scene, especially in the case that a user repeatedly supplements or modifies consultation conditions, the technical problem that accurate historical consultation object backtracking cannot be realized. Due to the consultation memory unit construction, gap type driven recall, conflict resolution and evidence fusion correction, the intention recognition accuracy and response stability of the incoming call consultation Agent in a long dialogue scene are improved.

Description

Technical Field

[0001] This invention relates to the field of voice analysis technology, and in particular to a method and system for voice intent recognition based on a caller ID agent. Background Technology

[0002] Voice-based intelligent customer service and call inquiry systems have been widely applied in scenarios such as education consultation, government services, medical triage, and financial customer service. Their core lies in accurately identifying the current inquiry intent based on the user's real-time voice content and transmitting the identification results to subsequent dialogue management processes. However, in long-duration, multi-round call inquiries, users' expressions are usually not presented in complete, standardized sentence structures, but rather frequently include referential expressions, omitted expressions, comparisons of solutions, and corrections. For example, a user might mention a particular university, major, score range, or matter in a previous round, and then continue to inquire using colloquial expressions such as "that one," "the first solution," "can this be changed," or "I misspoke, it's another one." If the system only classifies the intent based on the current single-round voice text, it is difficult to determine the historical inquiry object corresponding to the above expressions. Existing technology cannot fully meet the needs of cross-round contextual association and historical object tracing in long-dialogue call inquiries. The original invention concept also points out that existing processing methods lack effective cross-round state tracking and contextual association in long-duration, multi-round dialogues, which easily leads to a decrease in the accuracy of identifying referential and omitted expressions.

[0003] Furthermore, existing speech intent recognition methods mostly employ a single-round processing approach of "speech recognition results plus a text classification model," typically treating historical dialogues as ordinary text caches or fixed-window context input models. They fail to structurally maintain the entities, entity attributes, historical intent, emotional tendencies, and evidentiary text fragments within the historical consultation content. For example, in scenarios such as college application, business processing, or pre-sales consultation, users may repeatedly compare multiple options and modify scores, regions, budgets, processing conditions, or personal preferences in subsequent rounds. If the system cannot distinguish between newly added information, historically retained information, and user corrections, it may easily continue to use outdated or corrected historical conditions for intent judgment, causing the recognition results to deviate. Existing technologies cannot fully meet the needs for hierarchical storage, targeted retrieval, and effective updating of historical information based on entities, intents, and emotions.

[0004] Therefore, there is an urgent need for a voice intent recognition method that can still construct consultation memory units, recall historical memories driven by gap types, resolve entity attribute conflicts, and correct intent through multi-evidence fusion in long-term, multi-round call consultation scenarios, especially when users use pronouns, omissions, solution comparisons, conditional corrections, and emotional continuation. This would improve the accuracy of the call consultation agent in retrospecting historical consultation objects, the stability of processing conflicting historical information, and the accuracy of recognizing complex consultation intents. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a voice intent recognition method based on a caller ID agent. This method aims to solve the technical problem in existing technologies where, in long-term, multi-round caller ID scenarios, especially when users repeatedly supplement or modify the ID conditions, it is impossible to accurately trace back historical callers.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a voice intent recognition method based on incoming call consultation agent.

[0007] The aforementioned voice intent recognition method based on caller ID includes: Step S10: Obtain the user's real-time voice stream, and perform a round representation extraction task based on the user's real-time voice stream using a multi-source round representation fusion mechanism, and output the current round fusion representation data; Step S20: Based on the current round fusion representation data, the consultation context memory construction mechanism is used to execute the consultation memory unit construction task, and the consultation memory unit of the current round is output; Step S30: Based on the current round of consultation memory units, a gap-type driven hierarchical recall mechanism is used to execute the historical memory recall task and output the current round of candidate memory set; Step S40: Based on the current round candidate memory set, the intent result correction task is performed using a conflict resolution and evidence fusion correction mechanism, and the current round correction intent data is output; Step S50: Based on the current round of modified intent data, execute the Agent action mapping task using a hierarchical memory closed-loop writing mechanism, and output Agent response control data.

[0008] Preferably, step S10, which involves acquiring the user's real-time speech stream, performing a round representation extraction task based on the user's real-time speech stream using a multi-source round representation fusion mechanism, and outputting the current round fusion representation data, specifically includes: Step S101: Receive the user's real-time voice stream, and perform spectral subtraction and noise reduction, frame segmentation and endpoint detection on the user's real-time voice stream to obtain effective voice segment data; Step S102: Based on the effective speech segment data, generate round text and speech recognition confidence through an end-to-end speech recognition model, and use the BERT model to encode the sentence vectors of the round text to obtain round text representation data; Step S103: Based on the round text representation data, and combined with the sub-language feature vector consisting of speech rate, pause duration, pause frequency and energy change, the Squeeze-and-Excitation feature gating method is used to perform feature weighting to obtain the current round fusion representation data.

[0009] Preferably, step S20, which involves using the consultation context memory construction mechanism to construct consultation memory units based on the current round's fused representation data and outputting the current round's consultation memory units, specifically includes: Step S201: Based on the current round fusion representation data, use the BERT-CRF entity recognition model to extract the consultation entity and entity attributes to obtain the current round entity attribute data; Step S202: Based on the entity attribute data of the current round, the central theory reference resolution method is used to identify the reference expression, and the frame semantic slot completion method is used to identify the omitted expression, comparison expression and error correction expression to obtain the context gap data of the current round; Step S203: Based on the current round context gap data, construct a current round consultation memory unit that includes the consultation entity, entity attributes, gap type, evidence text fragment, speech recognition confidence, sentiment tendency identifier, and writing round identifier.

[0010] Preferably, step S30, which involves executing a historical memory retrieval task based on a gap-type driven hierarchical retrieval mechanism using the current round of consultation memory units, and outputting the current round of candidate memory sets, specifically includes: Step S301: Based on the gap type in the current round of consultation memory unit, determine the target recall path, wherein when the gap type is an entity gap, the entity attribute index is determined as the first recall path; when the gap type is a solution gap, the semantic vector index is determined as the first recall path; when the gap type is an intent gap, the Markov intent transfer index is determined as the first recall path; when the gap type is an emotion gap, the emotion tendency index is determined as the first recall path. Step S302: Based on the target recall path, perform multi-path recall on the hierarchical memory data, wherein BM25 term matching is used to recall entity-related memories from the entity attribute index, HNSW nearest neighbor retrieval is used to recall semantically related memories from the semantic vector index, Markov intent transfer index is used to recall intent-related memories, and sentiment tendency index is used to recall sentiment-related memories, thereby obtaining a multi-path recall memory set; Step S303: Based on the multi-path recall memory set, in the order of priority of the first recall path, priority of entity attribute consistency, priority of user error correction expression, and priority of writing round proximity, the multi-path recall memory set is subjected to intersection retention, union supplementation, and conflict elimination to obtain the candidate memory set for the current round.

[0011] Preferably, step S40, which involves performing an intent result correction task based on the current round candidate memory set using a conflict resolution and evidence fusion correction mechanism, and outputting the current round corrected intent data, specifically includes: Step S401: Based on the current round candidate memory set, identify numerical conflicts, category conflicts, and temporal conflicts of the entity attributes corresponding to the same consultation entity in the current round candidate memory set to obtain entity attribute conflict data; Step S402: Based on the entity attribute conflict data, according to the order of priority of user error correction expression, priority of speech recognition confidence, and priority of writing round proximity, perform covering, coexistence, deweighting or freezing processing on the candidate memory set of the current round to obtain the conflict resolution memory set; Step S403: Based on the conflict resolution memory set, the Dempster-Shafer evidence fusion method is used to fuse the semantic evidence of the current round, historical intent evidence, and sentiment tendency evidence to obtain the current round's corrected intent data.

[0012] Preferably, step S403, based on the conflict resolution memory set, uses the Dempster-Shafer evidence fusion method to fuse the semantic evidence of the current round, historical intent evidence, and sentiment tendency evidence to obtain the current round's revised intent data, specifically includes: Step S4031: Based on the conflict resolution memory set, generate semantic evidence for the current round, historical intent evidence, and emotional tendency evidence respectively; Step S4032: Based on the semantic evidence of the current round, the historical intent evidence, and the sentiment tendency evidence, the Dempster-Shafer evidence fusion formula is used to calculate the fusion confidence of each candidate intent in the candidate intent set, and the fusion intent confidence data is obtained. The Dempster-Shafer evidence fusion formula is as follows:

[0013] in, Indicates candidate intent fusion confidence, This indicates the semantic evidence for the current round against the candidate intent set. Support Evidence of historical intent is relevant to the candidate intent set. Support Emotional bias evidence for the candidate intention set Support , and Both are subsets of the candidate intent set; Step S4033: Perform low-confidence clarification control based on the fused intent confidence data. When the highest fused confidence is less than a preset confidence threshold, or the difference between the highest fused confidence and the second highest fused confidence is less than a preset interval threshold, write the clarification trigger identifier, clarification object, and candidate intent set into the current round's corrected intent data. When the highest fused confidence is not less than the preset confidence threshold, and the difference between the highest fused confidence and the second highest fused confidence is not less than the preset interval threshold, determine the candidate intent corresponding to the highest fused confidence as the final intent category, and write the final intent category and the corresponding fused confidence into the current round's corrected intent data.

[0014] Preferably, step S50, which involves executing the Agent action mapping task using a hierarchical memory closed-loop writing mechanism based on the current round's modified intent data, and outputting Agent response control data, specifically includes: Step S501: Determine the memory writing strategy for the current round based on the final intent category, fusion confidence, and clarification trigger identifier in the current round correction intent data. The current round memory writing strategy includes at least one of the following: new writing, overwrite writing, concurrent writing, de-weighted writing, and frozen writing. Step S502: Based on the current round memory writing strategy, write the current round consultation memory unit into the entity memory layer, intent memory layer and emotion memory layer, and update the entity attribute index, semantic vector index, intent transfer index and emotion tendency index to obtain the updated hierarchical memory data; Step S503: Based on the updated hierarchical memory data, and combined with the final intent category, fusion confidence, clarification trigger identifier, and clarification object in the current round of corrected intent data, generate Agent response control data. The Agent response control data includes at least one of dialogue action type, clarification question type, entity backtracking pointer, and candidate intent set.

[0015] The present invention also provides a voice intent recognition system based on a caller ID agent, comprising: The round representation extraction module is used to acquire the user's real-time voice stream, and to perform the round representation extraction task based on the user's real-time voice stream using a multi-source round representation fusion mechanism, and output the current round fusion representation data; The consultation memory construction module is used to perform the consultation memory unit construction task based on the current round fusion representation data using the consultation context memory construction mechanism, and output the current round consultation memory unit; The historical memory retrieval module is used to perform historical memory retrieval tasks based on the current round of consultation memory units using a gap-type driven hierarchical retrieval mechanism, and output the current round of candidate memory set; The intent result correction module is used to perform intent result correction tasks based on the current round candidate memory set using conflict resolution and evidence fusion correction mechanisms, and output the current round corrected intent data; The memory write mapping module is used to perform Agent action mapping tasks by using a hierarchical memory closed-loop write mechanism to correct the intent data according to the current round, and output Agent response control data.

[0016] The present invention also provides a voice intent recognition device based on a caller ID agent. The voice intent recognition device based on a caller ID agent includes: a memory, a processor, and a voice intent recognition program based on a caller ID agent stored in the memory and executable on the processor. When the voice intent recognition program based on a caller ID agent is executed by the processor, it implements the above-described method.

[0017] The present invention also provides a computer program product, the computer program product including a voice intent recognition program based on a caller ID agent, the voice intent recognition program based on a caller ID agent implementing the above method when executed by a processor.

[0018] The beneficial effects of this invention are as follows: 1. This invention uses a multi-source round-based representation fusion mechanism to perform spectral reduction and noise reduction, frame segmentation, and gating weighting of speech signals by combining end-to-end speech recognition and paralinguistic features. This effectively extracts fused representation data containing speech recognition confidence and paralinguistic information, providing high-quality input features for subsequent memory construction.

[0019] 2. This invention constructs a consultation context memory unit and introduces a hierarchical recall mechanism driven by gap type. It can select the optimal index path for multi-path recall based on entity gaps, solution gaps, intention gaps and emotion gaps respectively, which significantly improves the recall accuracy and coverage of historical consultation context-related memories, thereby alleviating the problems of referential resolution and omission completion.

[0020] 3. This invention employs a conflict resolution and evidence fusion correction mechanism. By resolving conflicts in entity attributes in a layered manner and using Dempster-Shafer evidence theory to fuse multi-source evidence, it can automatically perform intent correction and clarification trigger control in the event of information inconsistency, thereby avoiding intent jitter caused by information conflicts and improving the confidence and stability of the output intent.

[0021] 4. This invention does not rely on the novelty of a single algorithm, but rather on the collaborative multi-turn voice processing flow for caller agents: the consultation memory unit provides a structured carrier for historical information, gap type-driven hierarchical recall improves the accuracy of historical object localization, the Markov intent transfer index is used to supplement evidence of intent graph continuation or transfer, conflict resolution prevents invalid conditions before user correction from continuing to affect the judgment, and Dempster-Shafer evidence fusion outputs corrective intent data that can trigger clarification or execution under multi-source uncertain evidence. The combination of these steps can solve the problems of intent drift and historical condition conflicts in long dialogues that are difficult to address with existing single-turn voice classification, fixed window context, and ordinary evidence merging schemes. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the first embodiment of a voice intent recognition method based on a caller ID agent according to the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of a voice intent recognition method based on a caller ID agent according to the present invention. The present invention proposes a first embodiment of a voice intent recognition method based on a caller ID agent.

[0026] In the first embodiment, the voice intent recognition method based on the caller ID agent includes: Step S10: Obtain the user's real-time voice stream, and perform a round representation extraction task based on the user's real-time voice stream using a multi-source round representation fusion mechanism, and output the current round fusion representation data; The multi-source round representation fusion mechanism in this step refers to preprocessing the user's real-time speech stream by performing signal denoising and endpoint detection, then using an end-to-end speech recognition model to generate round text and speech recognition confidence scores, and extracting paralinguistic features such as speech rate, pause duration, pause frequency, and energy changes. These features are then encoded into sentence vectors using a BERT model, and the Squeeze-and-Excitation feature gating method is used to dynamically weight and fuse the features from multiple sources to obtain the current round's fused representation data. This fused representation data directly serves as the input basis for constructing the consultation memory unit in step S20.

[0027] The raw speech signal is transformed into a high-information-density fusion representation while preserving semantic content, speech recognition uncertainty, and paralinguistic cues. This enables subsequent steps to make more accurate judgments by comprehensively utilizing the diverse information in the speech, avoiding potential intent omissions or misjudgments that may result from relying solely on text. This provides the first layer of foundation for maintaining the stability of intent recognition in multi-turn complex dialogues.

[0028] Compared to traditional methods that only use speech recognition text as a single input for intent analysis, this step integrates speech recognition confidence and paralinguistic features during the feature extraction stage. This allows the system to provide corrective and weighting cues in advance when speech recognition results are unreliable or when user emotions change through paralinguistic signs. Traditional methods often overlook factors that directly affect intent expression, such as the speaker's speech rate and abnormal pauses. However, this invention, through a gating fusion mechanism, significantly enhances the feature representation capabilities for conversational and weakly expressive call scenarios. For example, in a college application consultation call scenario, a user might express, "My child's score... um... I'd like to see if there are any relatively safe schools in Nanjing." In this case, the end-to-end speech recognition model can obtain the corresponding round of text and speech recognition confidence; the BERT model can extract semantic information such as "score," "Nanjing," and "relatively safe schools"; and paralinguistic features can reflect the pause duration and speech rate changes that occur after the user mentions "score." By weighting the above information using the Squeeze-and-Excitation feature gating method, the system can obtain the current round fusion representation data that simultaneously includes text semantics, speech reliability, and the degree of expression hesitation, thus providing an input basis for subsequent judgments on whether the user has made a solution comparison, risk consultation, or supplementary condition expression.

[0029] Step S20: Based on the current round fusion representation data, the consultation context memory construction mechanism is used to execute the consultation memory unit construction task, and the consultation memory unit of the current round is output; The consultation context memory construction mechanism in this step refers to the following: based on the current round's fused representation data, first, the BERT-CRF model is used to extract consultation entities and their attributes; then, the central theory referencing resolution method is used to parse referencing expressions; and finally, the frame semantic slot completion method is used to identify context gap types such as omissions, comparisons, and corrections. Ultimately, this information, along with evidence text fragments, speech recognition confidence levels, sentiment indicators, and round-of-writing identifiers, is structured and encapsulated into a current round's consultation memory unit. This memory unit serves as the direct query condition for historical memory retrieval in step S30.

[0030] This step transforms unstructured round information into structured memory units tailored to the consultation scenario, explicitly labeling various gaps in the current round. This enables the system to accurately identify "what information the user is missing" and "which information needs to be filled in using historical memory," thereby transforming the need for intent recognition into a precise query of memory, significantly reducing the randomness of subsequent historical memory retrieval.

[0031] Traditional memory construction methods often only retain dialogue text or simple round vectors, lacking structured abstraction of entity relationships, gap types, and emotional tendencies. This makes it difficult to locate specific information that needs to be recalled when dealing with highly abbreviated expressions such as "What about that one? Change it to be the same as it." This invention, through a consultation context memory construction mechanism, classifies gaps into types such as entity gaps, solution gaps, intention gaps, and emotional gaps, achieving fine-grained modeling of information loss in the consultation context and providing clear driving signals for subsequent stratified recall. For example, when a user mentions "computer science major at Nanjing University A" in a previous consultation, but only states "What are the job prospects for this major?" in the current round, "this major" in the current round text is a referential expression. The BERT-CRF entity recognition model can identify entity or attribute cues such as "major" and "job prospects." The central theory referencing resolution method can refer back to "computer science major at Nanjing University A" in the previous round. Then, the frame semantic slot completion method determines that there is an "job prospects gap" in the current round. The current round of consultation memory unit constructed in this way not only records the current expression content, but also records the type of gap that the expression needs to rely on historical memory to fill, providing a basis for subsequent targeted recall of historical memory.

[0032] In this invention, the consultation memory unit is a key data structure connecting speech recognition results and historical memory retrieval. This memory unit is not a typical dialogue log or fixed-window context, but rather a unified encapsulation of the consultation object, attribute conditions, contextual gaps, and evidence sources, enabling the system to distinguish between "content explicitly stated by the user" and "content that needs to be completed from previous rounds." For example, if the user currently only says "Is the second one more reliable?", this text alone is insufficient for intent recognition. However, through the solution gap and the round identifier in the consultation memory unit, the second consultation solution from the previous round can be located, and this solution can be used as the object for subsequent intent judgment. Furthermore, the setting of gap types allows this invention to transform different types of incomplete contextual expressions into different technical processing paths. Entity gaps correspond to the backtracking of specific consultation objects, solution gaps correspond to the location of comparison objects among multiple candidate solutions, intent gaps correspond to the continuation or transfer relationship of the preceding and following intent graphs, and emotion gaps correspond to the continuation of user voice emotion cues such as doubt, denial, or urgency. These gap types, in conjunction with the subsequent hierarchical retrieval path, form a processing mechanism distinct from ordinary context splicing or fixed-window review.

[0033] Step S30: Based on the current round of consultation memory units, a gap-type driven hierarchical recall mechanism is used to execute the historical memory recall task and output the current round of candidate memory set; The gap-type driven hierarchical recall mechanism in this step refers to dynamically determining the target recall path based on the gap type marked in the current round's consultation memory unit: entity gaps prioritize entity attribute indexing and use BM25 term matching for recall; solution gaps prioritize semantic vector indexing and use HNSW nearest neighbor retrieval for recall; intent gaps prioritize Markov intent transfer indexing for recall; and sentiment gaps prioritize sentiment tendency indexing. Then, intersection retention, union supplementation, and conflict removal are performed on the multi-path recall memories, and they are sorted according to the rules of priority for the first recall path, entity attribute consistency, user error correction expression, and proximity in the writing round, outputting the candidate memory set for the current round. This set serves as the evidentiary basis for intent correction in step S40.

[0034] This step, through the combination of gap-driven and multi-path indexing, avoids blindly traversing the entire historical memory, significantly reducing recall latency and noise introduction. Selecting the optimal indexing path based on different gap types ensures highly targeted recall for each gap. Furthermore, combining multi-path supplementation and conflict sorting strategies ensures that the recalled candidate memory set possesses both high relevance and logical consistency, directly improving the evidence quality in the intent correction stage.

[0035] Traditional dialogue systems typically employ fixed-length context windows or time-decay-based recall mechanisms, failing to differentiate memory retrieval based on the specific information need at hand. This results in a large amount of irrelevant historical information being used in calculations, while crucial referential and error-correction information may be discarded because it predates the window. This invention, through gap type-driven hierarchical indexing retrieval, accurately achieves on-demand extraction and typified completion. It can locate the most relevant historical memory fragments in complex situations such as error correction for the same entity attribute and intent transfer chains, eliminating the context breakage problem caused by window limitations. For example, when the gap type in the current round of consultation memory units is a "solution gap," and the user states "Which is more stable, the first or the second?", the system should not simply recall all historical content in chronological order. Instead, it should prioritize recalling historical memories previously marked as "Solution 1" and "Solution 2" through semantic vector indexing. When the gap type is an "entity gap," and the user states "Can I apply to that school?", it should prioritize recalling historical entity memories related to "school" through entity attribute indexing. By selecting the recall path based on the gap type, we can reduce the amount of irrelevant rounds of content entering the candidate set, and make the candidate memory set of the current round more focused on covering the historical consultation object that the user actually refers to.

[0036] Step S40: Based on the current round candidate memory set, the intent result correction task is performed using a conflict resolution and evidence fusion correction mechanism, and the current round correction intent data is output; The conflict resolution and evidence fusion correction mechanism in this step refers to first identifying numerical, categorical, and temporal conflicts among the attributes of the same consulting entity in the candidate memory set. Conflicts are then resolved through strategies such as coverage, coexistence, weighting, or freezing to obtain a conflict-resolved memory set. Based on this set, the Dempster-Shafer evidence fusion formula is used to fuse semantic evidence from the current round, historical intent evidence, and sentiment tendency evidence, calculating the fusion confidence of each candidate intent. Clarification is triggered when the confidence does not meet a preset threshold; otherwise, the final intent category is output as the corrected intent data for the current round. This corrected intent data serves as the decision basis for Agent action mapping in step S50.

[0037] This step solidifies intent judgment on two levels: conflict resolution and multi-source evidence fusion. Conflict resolution first purifies the evidence pool, eliminating the interference of contradictory information on reasoning; Dempster-Shafer fusion rigorously integrates the three chains of evidence—semantics, history, and sentiment—quantifying support and outputting fusion confidence. This allows the system to handle user changes of mind and corrections midway through the process, while avoiding forcibly outputting incorrect intent when evidence is insufficient. Instead, it confirms the intent through a clarification mechanism and user interaction, thereby ensuring the reliability of terminal actions.

[0038] Most existing intent correction schemes rely solely on simple rules or single-source classifier confidence for optimization, making it difficult to handle situations where historical information and current semantics directly conflict. For example, if a user initially requests package A but later changes their mind to package B, traditional systems tend to waver between the two. This invention addresses this by first resolving structured conflicts, explicitly prioritizing user error correction expressions and high-confidence strategies to determine coverage or coexistence. Then, it integrates multi-source uncertainties using evidence theory, enabling the system to output a stable and consistent intent even under repeated information updates and fluctuating user emotions.

[0039] For example, if a user previously stated "score approximately 590," but later corrected it to "not 590, but over 580," then there is a numerical conflict in the "score" attribute under the same consultation entity. In this case, the system, prioritizing user corrections, treats the memory corresponding to "over 580" as a valid attribute, while downgrading or freezing the historical memory corresponding to "590." Subsequently, it performs Dempster-Shafer evidence fusion on the semantic evidence from the current round, historical intent evidence, and sentiment bias evidence. If the fusion result shows that the user's current intent is closer to "admission risk assessment," then that final intent is output. If the confidence difference between the two candidate intents, "admission risk assessment" and "school recommendation," is small, a clarification trigger is output so that the Agent can further inquire whether the user wants to assess the risk of the current school or request a new school recommendation.

[0040] The key technical aspect of this step lies in addressing conflicts in historical memory before fusing evidence. In call-in consultation scenarios, users often modify scores, budgets, regions, processing conditions, or preferences in subsequent rounds. If conflicting information is not first covered, coexisted, downweighted, or frozen, old conditions will still participate in intent judgment as high-weight evidence, leading to incorrect agent outputs. This invention purifies the candidate memory set through a conflict resolution mechanism, ensuring that Dempster-Shafer evidence fusion processes evidence that has undergone validity screening. Therefore, Dempster-Shafer evidence fusion in this invention is not a simple vote or weighted average of multiple model classification results, but rather a fusion of semantic evidence from the current round, historical intent evidence, and sentiment bias evidence, combined with the conflict resolution results from previous rounds. When user error correction, speech recognition confidence, and proximity of the writing round indicate that a certain historical condition has expired, the evidence corresponding to that historical condition is downgraded or frozen. When the fusion confidence is insufficient or the difference between the highest confidence and the second highest confidence is too small, the system does not force the output of intent, but triggers clarification control, enabling the Agent to further confirm the user's true needs through follow-up questions.

[0041] Step S50: Based on the current round of modified intent data, execute the Agent action mapping task using a hierarchical memory closed-loop writing mechanism, and output Agent response control data.

[0042] In this step, the hierarchical memory closed-loop writing mechanism refers to determining the appropriate memory writing strategy (such as adding, overwriting, coexisting, downweighting, or freezing) based on the final intent category, fusion confidence, and clarification trigger identifier in the current round's revised intent data. The current round's consultation memory unit is then written into the entity memory layer, intent memory layer, and emotion memory layer, with the corresponding indexes for each layer updated synchronously. Subsequently, the revised intent data and the updated hierarchical memory are combined to generate Agent response control data, including dialogue action types, clarification question types, and entity backtracking pointers, which are used to directly control the intelligent agent to perform corresponding operations.

[0043] The technical advantage of this step lies in forming a complete closed loop from perception to memory update to action mapping. By strategically writing back the confirmed intent and consultation information to hierarchical memory and updating the index, the system can continuously accumulate dialogue experience, making subsequent memory retrieval increasingly accurate. Simultaneously, agent response control data is directly mapped to specific clarification scripts, business processing actions, or information submission operations, significantly shortening the path from intent recognition to business process execution.

[0044] Traditional methods often lack fine-grained control over memory writing. Simply appending historical information can easily lead to memory bloat and index pollution, while directly discarding unconfirmed or conflicting information can result in the loss of critical clues. This invention uses a hierarchical closed-loop memory writing system, which can select overwrite, coexist, de-weight, or freeze writing as needed. This ensures both the activity and accuracy of memory while preventing the solidification of erroneous information. At the same time, it tightly integrates agent action mapping with correction intent, enabling automatic triggering of clarification intent and direct execution intent, significantly improving the system's call handling capabilities and self-learning efficiency.

[0045] For example, when the current round's corrected intent data output in step S40 indicates that the user's final intent is "admission risk assessment," and the fusion confidence level meets the preset requirements, the system writes the confirmed scores, target regions, target universities, and majors from the current round into the entity memory layer, writes "admission risk assessment" into the intent memory layer, and updates the corresponding indexes simultaneously. Subsequently, the Agent response control data can instruct the Agent to directly generate risk analysis scripts. If step S40 outputs a clarification trigger flag, the Agent response control data can instruct the Agent to generate clarification questions, such as "Do you want to compare the admission stability of the two options, or do you want to re-recommend a more stable school?" This avoids the system directly outputting incorrect results when the intent is unclear.

[0046] It should be understood that while Markov intent transfer and Dempster-Shafer evidence fusion can be used for state prediction and uncertainty fusion respectively, existing technologies typically use them as standalone intent prediction models or tools for merging results from multiple classifiers. This fails to address unique challenges in call consultation scenarios, such as cross-round referential tracing, object localization for scheme comparison, user error correction information coverage, low-confidence speech recognition results participating in inference, and continuous historical memory writing pollution. Therefore, the technical improvement of this invention is not a simple superposition of the aforementioned algorithms, but rather embeds them into a continuous processing flow of consultation memory units, gap-type driven recall, conflict resolution, and closed-loop writing. This ensures that historical intent transfer serves only as one piece of evidence for recall and correction, and outputs response control data executable by the agent through evidence fusion after conflict resolution.

[0047] This invention processes each round of incoming call voice messages not merely as text to be classified, but converts them into consultation memory units that include consultation entities, entity attributes, gap types, evidence text fragments, voice recognition confidence, sentiment indicators, and round-of-writing identifiers. Then, different historical memory retrieval paths are selected based on the gap type exposed in the current round. Subsequently, conflicting entity attributes in the retrieval results are covered, coexisted, downweighted, or frozen. Finally, evidence such as current semantics, historical intent, and sentiment tendencies are fused to obtain a control result that can both identify the current intent and guide the agent to clarify, retrospectively analyze, or execute actions.

[0048] Example 2: Furthermore, the present invention provides a voice intent recognition system based on a caller ID agent, employing a voice intent recognition method based on a caller ID agent as described in the above embodiments, which can solve the technical problem of voice intent recognition based on a caller ID agent. The beneficial effects of the voice intent recognition system based on a caller ID agent provided by the present invention are the same as those of the voice intent recognition method based on a caller ID agent provided in the above embodiments, and other technical features of the voice intent recognition system based on a caller ID agent are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0049] Example 3: This invention provides a voice intent recognition device based on a caller ID agent. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to perform the voice intent recognition method based on a caller ID agent described in Example 1. The voice intent recognition device based on a caller ID agent in this invention can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This voice intent recognition device based on a caller ID agent is merely an example and should not limit the functionality or scope of the invention. A voice intent recognition device based on a caller ID agent may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of the voice intent recognition device based on a caller ID agent. The processing unit, the read-only memory, and the random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including, for example, magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the voice intent recognition device based on a caller ID agent to communicate wirelessly or wiredly with other devices to exchange data. Although a voice intent recognition device based on a caller ID agent with various systems is shown in the figures, it should be understood that it is not required to implement or possess all the systems shown. It can be implemented alternatively or with more or fewer systems.

[0050] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for voice intent recognition based on a caller ID agent. The computer program product provided by this invention can solve the technical problem of voice intent recognition based on a caller ID agent. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the voice intent recognition method based on a caller ID agent provided in the above embodiments, and will not be repeated here.

[0051] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0052] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0053] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for recognizing voice intent based on a caller ID agent, characterized in that, The methods include: Step S10: Obtain the user's real-time voice stream, and perform a round representation extraction task based on the user's real-time voice stream using a multi-source round representation fusion mechanism, and output the current round fusion representation data; Step S20: Based on the current round fusion representation data, the consultation context memory construction mechanism is used to execute the consultation memory unit construction task, and the consultation memory unit of the current round is output; Step S30: Based on the current round of consultation memory units, a gap-type driven hierarchical recall mechanism is used to execute the historical memory recall task and output the current round of candidate memory set; Step S40: Based on the current round's candidate memory set, perform the intent result correction task using a conflict resolution and evidence fusion correction mechanism, and output the current round's corrected intent data; wherein, the step of performing the intent result correction task using the conflict resolution and evidence fusion correction mechanism based on the current round's candidate memory set and outputting the current round's corrected intent data specifically includes: Based on the current round candidate memory set, numerical conflicts, category conflicts, and temporal conflicts are identified for the entity attributes corresponding to the same consultation entity in the current round candidate memory set to obtain entity attribute conflict data. Based on the entity attribute conflict data, the candidate memory set of the current round is subjected to covering, coexistence, deweighting or freezing processing in the order of priority of user error correction expression, priority of speech recognition confidence and priority of writing round proximity, so as to obtain the conflict resolution memory set. Based on the conflict resolution memory set, the Dempster-Shafer evidence fusion method is used to fuse the semantic evidence of the current round, historical intent evidence, and emotional tendency evidence to obtain the current round's revised intent data. Step S50: Based on the current round of modified intent data, execute the Agent action mapping task using a hierarchical memory closed-loop writing mechanism, and output Agent response control data.

2. The voice intent recognition method based on caller ID agent as described in claim 1, characterized in that, Step S10, which involves acquiring the user's real-time speech stream, performing a round representation extraction task based on the user's real-time speech stream using a multi-source round representation fusion mechanism, and outputting the current round fusion representation data, specifically includes: Step S101: Receive the user's real-time voice stream, and perform spectral subtraction and noise reduction, frame segmentation and endpoint detection on the user's real-time voice stream to obtain effective voice segment data; Step S102: Based on the effective speech segment data, generate round text and speech recognition confidence through an end-to-end speech recognition model, and use the BERT model to encode the sentence vectors of the round text to obtain round text representation data; Step S103: Based on the round text representation data, and combined with the sub-language feature vector consisting of speech rate, pause duration, pause frequency and energy change, the Squeeze-and-Excitation feature gating method is used to perform feature weighting to obtain the current round fusion representation data.

3. The voice intent recognition method based on caller ID agent as described in claim 1, characterized in that, Step S20, which involves constructing consultation memory units based on the current round's fused representation data using a consultation context memory construction mechanism and outputting the current round's consultation memory units, specifically includes: Step S201: Based on the current round fusion representation data, use the BERT-CRF entity recognition model to extract the consultation entity and entity attributes to obtain the current round entity attribute data; Step S202: Based on the entity attribute data of the current round, the central theory reference resolution method is used to identify the reference expression, and the frame semantic slot completion method is used to identify the omitted expression, comparison expression and error correction expression to obtain the context gap data of the current round; Step S203: Based on the current round context gap data, construct a current round consultation memory unit that includes the consultation entity, entity attributes, gap type, evidence text fragment, speech recognition confidence, sentiment tendency identifier, and writing round identifier.

4. The voice intent recognition method based on caller ID agent as described in claim 1, characterized in that, Step S30, which involves executing a historical memory retrieval task based on a gap-type driven hierarchical retrieval mechanism using the current round of consultation memory units, and outputting the current round of candidate memory sets, specifically includes: Step S301: Based on the gap type in the current round of consultation memory unit, determine the target recall path, wherein when the gap type is an entity gap, the entity attribute index is determined as the first recall path; when the gap type is a solution gap, the semantic vector index is determined as the first recall path; when the gap type is an intent gap, the Markov intent transfer index is determined as the first recall path; when the gap type is an emotion gap, the emotion tendency index is determined as the first recall path. Step S302: Based on the target recall path, perform multi-path recall on the hierarchical memory data, wherein BM25 term matching is used to recall entity-related memories from the entity attribute index, HNSW nearest neighbor retrieval is used to recall semantically related memories from the semantic vector index, Markov intent transfer index is used to recall intent-related memories, and sentiment tendency index is used to recall sentiment-related memories, thereby obtaining a multi-path recall memory set; Step S303: Based on the multi-path recall memory set, in the order of priority of the first recall path, priority of entity attribute consistency, priority of user error correction expression, and priority of writing round proximity, the multi-path recall memory set is subjected to intersection retention, union supplementation, and conflict elimination to obtain the candidate memory set for the current round.

5. The voice intent recognition method based on caller ID agent as described in claim 1, characterized in that, Step S403, based on the conflict resolution memory set, uses the Dempster-Shafer evidence fusion method to fuse the semantic evidence of the current round, historical intent evidence, and sentiment tendency evidence to obtain the current round's revised intent data. This step specifically includes: Step S4031: Based on the conflict resolution memory set, generate semantic evidence for the current round, historical intent evidence, and emotional tendency evidence respectively; Step S4032: Based on the semantic evidence of the current round, the historical intent evidence, and the sentiment tendency evidence, the Dempster-Shafer evidence fusion formula is used to calculate the fusion confidence of each candidate intent in the candidate intent set, and the fusion intent confidence data is obtained. The Dempster-Shafer evidence fusion formula is as follows: in, Indicates candidate intent fusion confidence, This indicates the semantic evidence for the current round against the candidate intent set. Support Evidence of historical intent is relevant to the candidate intent set. Support Emotional bias evidence for the candidate intention set Support , and Both are subsets of the candidate intent set; Step S4033: Perform low-confidence clarification control based on the fused intent confidence data. When the highest fused confidence is less than a preset confidence threshold, or the difference between the highest fused confidence and the second highest fused confidence is less than a preset interval threshold, write the clarification trigger identifier, clarification object, and candidate intent set into the current round's corrected intent data. When the highest fused confidence is not less than the preset confidence threshold, and the difference between the highest fused confidence and the second highest fused confidence is not less than the preset interval threshold, determine the candidate intent corresponding to the highest fused confidence as the final intent category, and write the final intent category and the corresponding fused confidence into the current round's corrected intent data.

6. The voice intent recognition method based on caller ID agent as described in claim 5, characterized in that, Step S50, which involves executing the Agent action mapping task using a hierarchical memory closed-loop writing mechanism based on the current round's modified intent data, and outputting Agent response control data, specifically includes: Step S501: Determine the memory writing strategy for the current round based on the final intent category, fusion confidence, and clarification trigger identifier in the current round correction intent data. The current round memory writing strategy includes at least one of the following: new writing, overwrite writing, concurrent writing, de-weighted writing, and frozen writing. Step S502: Based on the current round memory writing strategy, write the current round consultation memory unit into the entity memory layer, intent memory layer and emotion memory layer, and update the entity attribute index, semantic vector index, intent transfer index and emotion tendency index to obtain the updated hierarchical memory data; Step S503: Based on the updated hierarchical memory data, and combined with the final intent category, fusion confidence, clarification trigger identifier, and clarification object in the current round of corrected intent data, generate Agent response control data. The Agent response control data includes at least one of dialogue action type, clarification question type, entity backtracking pointer, and candidate intent set.

7. A voice intent recognition system based on a caller ID agent, applied to the voice intent recognition method based on a caller ID agent as described in any one of claims 1 to 6, characterized in that, The voice intent recognition system includes: The round representation extraction module is used to acquire the user's real-time voice stream, and to perform the round representation extraction task based on the user's real-time voice stream using a multi-source round representation fusion mechanism, and output the current round fusion representation data; The consultation memory construction module is used to perform the consultation memory unit construction task based on the current round fusion representation data using the consultation context memory construction mechanism, and output the current round consultation memory unit; The historical memory retrieval module is used to perform a historical memory retrieval task based on the current round of consultation memory units using a gap-type driven hierarchical retrieval mechanism, and output the current round of candidate memory set; The intent result correction module is used to perform an intent result correction task based on the current round's candidate memory set using a conflict resolution and evidence fusion correction mechanism, and output the current round's corrected intent data; wherein, the step of performing the intent result correction task based on the current round's candidate memory set using a conflict resolution and evidence fusion correction mechanism and outputting the current round's corrected intent data specifically includes: Based on the current round candidate memory set, numerical conflicts, category conflicts, and temporal conflicts are identified for the entity attributes corresponding to the same consultation entity in the current round candidate memory set to obtain entity attribute conflict data. Based on the entity attribute conflict data, the candidate memory set of the current round is subjected to covering, coexistence, deweighting or freezing processing in the order of priority of user error correction expression, priority of speech recognition confidence and priority of writing round proximity, so as to obtain the conflict resolution memory set. Based on the conflict resolution memory set, the Dempster-Shafer evidence fusion method is used to fuse the semantic evidence of the current round, historical intent evidence, and emotional tendency evidence to obtain the current round's revised intent data. The memory write mapping module is used to perform Agent action mapping tasks by using a hierarchical memory closed-loop write mechanism to correct the intent data according to the current round, and output Agent response control data.

8. A voice intent recognition device based on a caller ID agent, characterized in that, The voice intent recognition device based on caller ID includes: a memory, a processor, and a voice intent recognition program based on caller ID stored in the memory and executable on the processor. When the voice intent recognition program based on caller ID is executed by the processor, it implements a voice intent recognition method based on caller ID according to any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product includes a voice intent recognition program based on a caller ID agent, which, when executed by a processor, implements a voice intent recognition method based on a caller ID agent according to any one of claims 1 to 6.

Images