AI intelligent outbound call dialogue generation and multi-round interaction system based on deep learning

Abstract

The application discloses an AI intelligent outbound call dialogue generation and multi-round interaction system based on deep learning, which comprises a data acquisition module, a stage anchor point construction module, a double-flow coding module, a stage memory reservation module, an objection bifurcation correction module, a reply generation module and an interaction control module.The data acquisition module is used for collecting interaction data and preprocessing to generate anchor data.The stage anchor point construction module is used for marking stage attribution, migration, promotion depth and fallback state, and generating a stage anchor point sequence.The double-flow coding module is used for inputting round semantic information and the stage anchor point sequence into an improved stage anchor double-flow constraint RetNet.The stage memory reservation module is used for generating a stage memory reservation feature sequence through stage anchor Retention processing.The objection bifurcation correction module is used for detecting objections and correcting the stage anchor point sequence.The reply generation module is used for generating a current round outbound call reply content and a stage closed-loop correction mark.The interaction control module is used for updating the stage anchor point sequence and cyclically interacting, and outputting a result.The application realizes stable control of intelligent outbound call multi-round interaction, and improves objection processing capacity and task completion efficiency.

Description

Technical Field

[0001] This invention relates to the field of dialogue interaction and intelligent outbound calling technology, and in particular to an AI-based intelligent outbound calling dialogue generation and multi-turn interaction system based on deep learning. Background Technology

[0002] With the development of intelligent customer service, speech recognition, natural language processing, and automated outbound calling technologies, outbound calling platforms based on pre-defined scripts have been widely used in scenarios such as notification follow-ups, business confirmation, marketing outreach, service reminders, and result verification. Most existing outbound calling solutions execute interaction processing according to pre-configured business nodes, script templates, and redirection rules, capable of completing basic outbound calling tasks in scenarios where user responses are relatively fixed and business processes are relatively simple. Other solutions incorporate deep learning to semantically recognize user responses and combine multi-turn conversation information to generate reply content, thereby improving the naturalness and continuity of automated interaction.

[0003] Existing technologies still have significant shortcomings in practical applications. On the one hand, existing outbound call interaction processes mostly adopt a linear workflow, which provides a coarse characterization of the business stage to which each interaction round belongs. It is difficult to provide fine-grained representation of the advancement states such as stage maintenance, stage advancement, and stage regression, resulting in unclear stage connections in multi-round interactions and easy deviations between the response content and the current business advancement position. On the other hand, users often express objections such as refusal, questioning, delay, and interruption during outbound calls. Existing solutions mostly stop at keyword triggering or fixed branch jumps in objection responses, lacking constraint correction mechanisms for objection type, objection intensity, and branching paths. It is difficult to effectively rearrange the anchor points of subsequent stages and control the path after an objection occurs. Furthermore, existing multi-round dialogue generation models focus more on text continuity and do not sufficiently integrate outbound call business stage information, advancement depth information, and regression state information. They cannot perform differentiated memory retention processing for different stage transition states, resulting in a lack of clear stage constraints in subsequent response generation and affecting the stability of outbound call task completion.

[0004] Therefore, how to provide an AI-powered intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI-based intelligent outbound call dialogue generation and multi-round interaction system based on deep learning. This invention achieves staged control and multi-round interaction optimization of the outbound call process through stage anchor point construction, objection bifurcation correction, and constraint response generation. It has the advantages of strong interaction continuity, accurate objection handling, and stable task completion.

[0006] According to an embodiment of the present invention, an AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning includes:

[0007] The data acquisition module is used to collect interactive data and preprocess it to generate anchor data.

[0008] The phase anchor point construction module is used to perform phase attribution calibration, phase migration calibration, advance depth calibration and rollback status calibration, and construct a phase anchor point sequence.

[0009] The dual-stream coding module is used to input the round semantic information and stage anchor sequence from the anchored data into the improved stage anchoring dual-stream constraint RetNet, and perform semantic coding and stage coding respectively to generate semantic feature sequence and stage feature sequence.

[0010] The stage memory retention module is used to perform alignment fusion, input the alignment fusion result into the stage anchoring Retention layer, perform same-stage cumulative writing, stage transition enhancement writing and stage backoff suppression writing, and generate stage memory retention feature sequence.

[0011] The objection fork correction module is used to perform objection detection, objection strength determination and objection type classification. When an objection fork is triggered, fork correction is performed on the stage anchor point sequence to generate the constraint update stage anchor point sequence.

[0012] The response generation module is used to input the stage memory retention feature sequence and the constraint update stage anchor point sequence into the response generation layer, generate a candidate response sequence set, perform stage target consistency screening, determine the response content of the current round of outbound calls, and generate stage closed-loop correction markers;

[0013] The interactive control module is used to perform phase maintenance, phase forwarding, and phase recycling on the anchor point sequence of the constraint update phase, generate a new round of phase anchor point sequence, and feed it back to the improved phase anchoring dual-stream constraint RetNet until the new round of phase anchor point sequence reaches the preset termination phase, and output the final outbound call interaction result.

[0014] Optionally, the interactive data in the data acquisition module includes outbound call task identifier, outbound call service type, outbound call target node, user identifier, historical round dialogue content, historical stage progress record, current round user voice data, and current round user text data; preprocessing includes sequentially performing data cleaning, round segmentation, speech transcription, text normalization, semantic segmentation, stage node extraction and association integration processing on the interactive data to generate anchor data; the anchor data includes outbound call task context data, round semantic segment set, and stage node set.

[0015] Optionally, the process of constructing the stage anchor sequence in the stage anchor construction module includes:

[0016] The semantic fragment set of each round is arranged sequentially according to the interaction order to form a round labeling queue;

[0017] A stage position index table is established based on the stage node set to record the position number of each stage node in the business advancement chain;

[0018] For each semantic fragment in the round labeling queue, the corresponding semantic content is matched with each stage node one by one, the stage node with the highest matching result is determined as the target stage node, and written into the stage attribution field to complete the stage attribution labeling.

[0019] Read the target stage node of the semantic fragment in the current round and the target stage node of the semantic fragment in the previous round, and compare their position numbers. If the current position number is greater than the previous position number, mark it as stage forward; if the current position number is equal to the previous position number, mark it as stage hold; if the current position number is less than the previous position number, mark it as stage rollback. Write the stage migration field to complete the stage migration calibration.

[0020] Read the target stage node position number of the semantic fragment in the current round, combine it with the number of valid rounds before the current round and the number of stage forward moves, determine the stage advancement level and round advancement position, write it into the advancement depth field, and complete the advancement depth calibration.

[0021] For the current round semantic segment marked as stage rollback, retrieve the most recent round semantic segment marked as stage forward, determine the rollback start point, rollback end point and rollback span, write them into the rollback status field, and complete the rollback status labeling;

[0022] Combine them in round order to generate a phase anchor sequence.

[0023] Optionally, the process of performing semantic encoding and stage encoding in the dual-stream encoding module includes:

[0024] The round-based semantic information in the anchored data is expanded and arranged in round-based order to form a semantic input sequence;

[0025] The semantic input sequence is segmented, marked with position and round, to form an initial semantic representation sequence. The semantic encoding branch of the improved stage anchored dual-stream constrained RetNet is then input, and context association processing and temporal recursion processing are performed to obtain the semantic feature sequence.

[0026] The target stage node, stage migration field, advance depth field, and rollback status field in the stage anchor point sequence are extracted and combined in round order to form the stage input sequence;

[0027] The stage input sequence is processed by performing stage position mapping, transition state mapping, advance depth mapping, and backtrack state mapping to form an initial stage representation sequence. This initial stage representation sequence is then input into the stage encoding branch of the improved stage-anchored dual-stream constrained RetNet, and stage association processing and temporal recursion processing are performed to obtain the stage feature sequence.

[0028] Optionally, the process of generating the stage memory retention feature sequence in the stage memory retention module includes:

[0029] The alignment and fusion results are input into the stage anchoring Retention layer in round order, and the target stage node, stage transition field, advance depth field and rollback status field corresponding to the current round are read.

[0030] When the target stage node corresponding to the current round is the same as the target stage node corresponding to the previous round, the alignment and fusion result of the current round and the retained result of the previous round are overlaid and updated in the same stage, and the cumulative write in the same stage is performed.

[0031] When the stage migration field corresponding to the current round is marked as stage forward, the feature corresponding to the target stage node in the current round is enhanced and updated according to the advancement depth field, the retention result of the previous round is retained across stages, and the stage jump enhancement write is performed.

[0032] When the stage migration field corresponding to the current round is marked as stage rollback, suppress update is performed on the features located after the rollback start point in the previous round's retained results based on the rollback status field, and recovery update is performed on the features corresponding to the rollback end point in the current round, and stage rollback suppression write is performed;

[0033] The writing results corresponding to each round are arranged in round order to generate a stage memory retention feature sequence.

[0034] Optionally, the process of performing objection detection, objection strength determination, and objection type classification in the objection fork correction module includes:

[0035] Read the feature fragment corresponding to the current round in the stage memory retention feature sequence, read the user response content of the current round, compare the user response content of the current round with the feature fragments corresponding to the historical rounds before the current round in the stage memory retention feature sequence, detect whether there is any objection triggering content in the user response content of the current round that would block, delay, deviate from, or terminate the current stage's progress, and obtain the objection detection result;

[0036] Based on the location, frequency, and consecutive occurrence of the objection trigger content in the current round of user responses, the objection strength level is determined, and the objection strength judgment is completed. Based on the objection detection results and objection strength level, the objection type of the current round of user responses is classified to obtain the objection type corresponding to the current round.

[0037] Optionally, the process of performing fork correction on the stage anchor sequence in the objection fork correction module includes:

[0038] Read the objection type, objection strength level, and target stage node corresponding to the current round from the stage anchor point sequence;

[0039] Determine the direction of the fork correction corresponding to the current round based on the type of objection, and determine the magnitude of the fork correction based on the level of objection intensity.

[0040] Replace the target stage node corresponding to the current round with the fork stage node corresponding to the fork correction direction, and write the fork type flag and fork depth flag at the position corresponding to the current round.

[0041] Search the previous stage anchor point in the previous stage anchor point sequence that corresponds to the current round, and read the target stage node, stage migration field and advance depth field from the previous stage anchor point;

[0042] Based on the bifurcation correction direction, bifurcation correction magnitude, and the advancement depth field in the anchor points of the previous stage, the order of the anchor points after the current round is rearranged and the path constraint is updated. The advancement path of the stage after the bifurcation is consistent with the objection type and objection intensity level corresponding to the current round.

[0043] Arrange the anchor points of each stage after the fork correction is completed in round order to generate the constraint update stage anchor point sequence.

[0044] Optionally, the process of generating a candidate response sequence set in the response generation module includes:

[0045] The memory retains the data corresponding to the current round in the feature sequence of the reading phase and the anchor point sequence of the constraint update phase. Based on the target phase node, fork type marker, fork depth marker and advancement status corresponding to the current round, the response constraint conditions corresponding to the current round are determined.

[0046] Write the response constraints corresponding to the current round into the feature fragments corresponding to the current round in the stage memory retention feature sequence, and perform constraint screening on the feature fragments to obtain candidate response feature fragments;

[0047] Semantic expansion and sequential combination are performed on candidate response feature fragments to form several response paths;

[0048] Perform phase consistency checks and round connection checks on each response path, and retain response paths that meet the check conditions;

[0049] Arrange the retained response paths in the order of response to generate a set of candidate response sequences.

[0050] Optionally, the process of performing phase retention, phase forwarding, and phase reclamation in the interactive control module includes:

[0051] When the closed-loop correction flag indicates that the current round has not completed the current stage objective, the target stage node corresponding to the current round remains unchanged, and the fork type flag, fork depth flag, and progress status corresponding to the current round are maintained during the execution phase.

[0052] When the phase closed-loop correction flag indicates that the current round has completed the current phase objective and meets the conditions for entering the next phase, read the successor phase node of the target phase node in the phase node set corresponding to the current round, replace the successor phase node with the target phase node corresponding to the next round, perform forward update on the corresponding advancement state of the next round, and perform phase forward.

[0053] When the phase closed-loop correction flag indicates that the current round needs to exit the current fork path, read the fork type flag and fork depth flag corresponding to the current round, locate the fork phase node corresponding to the current round, determine the most recent unsuppressed phase node before the fork phase node as the target phase node for recycling, replace the target phase node for recycling with the target phase node corresponding to the next round, and clear the fork type flag and fork depth flag corresponding to the current round, and perform phase recycling.

[0054] Optionally, the process of feeding back to the improved stage-anchored dual-stream constraint RetNet in the interaction control module includes: feeding back the new round of stage anchor sequence and subsequent interaction data to the improved stage-anchored dual-stream constraint RetNet in round order, and continuing to perform semantic encoding, stage encoding, stage memory retention, objection fork correction, response generation, and stage closure correction; after each round of processing is completed, reading the target stage node corresponding to the current round in the new round of stage anchor sequence and comparing it with the preset termination stage, wherein the preset termination stage is the result confirmation stage, the end collection stage, or the termination exit stage; if the target stage node corresponding to the current round is the result confirmation stage and the current stage target has been completed, or the target stage node corresponding to the current round is the end collection stage and there are no new fork stage nodes, or the target stage node corresponding to the current round is the termination exit stage and the interaction has ended, it is determined that the new round of stage anchor sequence has reached the preset termination stage, the feeding back process is stopped, and the final outbound call interaction result is output.

[0055] The beneficial effects of this invention are:

[0056] (1) By constructing a sequence of stage anchor points, and combining stage attribution calibration, stage migration calibration, advancement depth calibration and rollback status calibration, this invention can finely characterize the business advancement position in the outbound call interaction process, so that multi-round interaction no longer stays at the level of simple wording rotation, but can continuously advance around the current business stage, improve the consistency between the response content and the business advancement status, and improve the problems of unclear stage connection and easy deviation of the response from the current business node in the prior art;

[0057] (2) In this invention, a stage anchoring Retention writing mechanism is introduced in the improved stage anchoring dual-stream constraint RetNet. Through the cumulative writing in the same stage, the stage transition enhancement writing, and the stage backoff suppression writing, the interaction features in different stage transition states are treated differently. This enables the previous effective interaction information, the current stage advancement information, and the stage backoff correction information to be retained in the same memory link in an orderly manner, thereby improving the context continuity and stage memory accuracy in the multi-round interaction process and improving the problems of insufficient utilization of historical interaction information and unstable retention of stage features in the prior art.

[0058] (3) This invention performs structured identification of rejection, questioning, delay and interruption that occur during the outbound call process by objection detection, objection strength determination, objection type classification and bifurcation correction processing. Based on the objection results, the subsequent stage anchor points are reordered and the path constraints are updated. Then, combined with the response constraints, a set of candidate response sequences is generated, so that the response generation process is consistent with the objection processing path, stage goals and rounds, improving the response targeting and outbound call task completion stability in objection scenarios, and improving the problems of rigid objection response and insufficient branch control capability in the prior art. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a module connection diagram of the AI ​​intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning proposed in this invention;

[0061] Figure 2 This is a diagram of the improved stage-anchored dual-stream constraint RetNet structure of the AI ​​intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning proposed in this invention.

[0062] Figure 3 This diagram illustrates the objection bifurcation correction mechanism of the AI ​​intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning proposed in this invention.

[0063] Figure 4 This is a schematic diagram illustrating the process of constructing the stage anchor point sequence of the AI ​​intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figures 1-4 A deep learning-based AI-powered outbound call dialogue generation and multi-turn interaction system, including:

[0066] The data acquisition module is used to collect interactive data and preprocess it to generate anchor data.

[0067] The phase anchor point construction module is used to perform phase attribution calibration, phase migration calibration, advance depth calibration and rollback status calibration, and construct a phase anchor point sequence.

[0068] The dual-stream coding module is used to input the round semantic information and stage anchor sequence from the anchored data into the improved stage anchoring dual-stream constraint RetNet, and perform semantic coding and stage coding respectively to generate semantic feature sequence and stage feature sequence.

[0069] The stage memory retention module is used to perform alignment fusion, input the alignment fusion result into the stage anchoring Retention layer, perform same-stage cumulative writing, stage transition enhancement writing and stage backoff suppression writing, and generate stage memory retention feature sequence.

[0070] The objection fork correction module is used to perform objection detection, objection strength determination and objection type classification. When an objection fork is triggered, fork correction is performed on the stage anchor point sequence to generate the constraint update stage anchor point sequence.

[0071] The response generation module is used to input the stage memory retention feature sequence and the constraint update stage anchor point sequence into the response generation layer, generate a candidate response sequence set, perform stage target consistency screening, determine the response content of the current round of outbound calls, and generate stage closed-loop correction markers;

[0072] The interactive control module is used to perform phase maintenance, phase forwarding, and phase recycling on the anchor point sequence of the constraint update phase, generate a new round of phase anchor point sequence, and feed it back to the improved phase anchoring dual-stream constraint RetNet until the new round of phase anchor point sequence reaches the preset termination phase, and output the final outbound call interaction result.

[0073] In this embodiment, the interactive data in the data acquisition module includes outbound call task identifier, outbound call service type, outbound call target node, user identifier, historical round dialogue content, historical stage progress record, current round user voice data, and current round user text data; preprocessing includes sequentially performing data cleaning, round segmentation, speech transcription, text normalization, semantic segmentation, stage node extraction and association integration on the interactive data to generate anchor data; the anchor data includes outbound call task context data, round semantic segment set, and stage node set.

[0074] In this embodiment, the process of constructing the stage anchor sequence in the stage anchor construction module includes:

[0075] The semantic fragment set of each round is arranged sequentially according to the interaction order to form a round labeling queue;

[0076] A stage position index table is established based on the stage node set to record the position number of each stage node in the business advancement chain;

[0077] For each semantic fragment in the round labeling queue, the corresponding semantic content is matched with each stage node one by one, the stage node with the highest matching result is determined as the target stage node, and written into the stage attribution field to complete the stage attribution labeling.

[0078] Read the target stage node of the semantic fragment in the current round and the target stage node of the semantic fragment in the previous round, and compare their position numbers. If the current position number is greater than the previous position number, mark it as stage forward; if the current position number is equal to the previous position number, mark it as stage hold; if the current position number is less than the previous position number, mark it as stage rollback. Write the stage migration field to complete the stage migration calibration.

[0079] Read the target stage node position number of the semantic fragment in the current round, combine it with the number of valid rounds before the current round and the number of stage forward moves, determine the stage advancement level and round advancement position, write it into the advancement depth field, and complete the advancement depth calibration.

[0080] Specifically, when determining the stage advancement level and round advancement position, firstly, the target stage node position number corresponding to the semantic fragment of the current round is read; secondly, the number of valid rounds that have completed stage affiliation marking before the current round is counted, and the number of valid rounds marked with the stage migration field as stage forward is counted; then, the stage advancement level is determined based on the target stage node position number, and the round advancement position is determined based on the number of historical rounds belonging to the same target stage node before the current round; when the stage migration field is marked as stage forward, the stage advancement level is promoted and the round advancement position is updated; when the stage migration field is marked as stage hold, the stage advancement level remains unchanged and the round advancement position is updated; when the stage migration field is marked as stage rollback, the target stage node position is corrected based on the rollback status marking result, and then the stage advancement level and round advancement position are re-determined.

[0081] For the current round semantic segment marked as stage rollback, retrieve the most recent round semantic segment marked as stage forward, determine the rollback start point, rollback end point and rollback span, write them into the rollback status field, and complete the rollback status labeling;

[0082] The semantic fragments from each round that have completed stage attribution, stage migration, advancement depth, and rollback status are combined in round order to generate a stage anchor sequence.

[0083] In this embodiment, the process of performing semantic coding and stage coding in the dual-stream coding module includes:

[0084] The round-based semantic information in the anchored data is expanded and arranged in round-based order to form a semantic input sequence;

[0085] The semantic input sequence is segmented, marked with position and round, to form an initial semantic representation sequence. The semantic encoding branch of the improved stage anchored dual-stream constrained RetNet is then input, and context association processing and temporal recursion processing are performed to obtain the semantic feature sequence.

[0086] Specifically, when performing word segmentation, position marking, and round marking on the semantic input sequence, the semantic segments of each round are first split into terms in sequence to obtain a term sequence. Then, the position numbers are written according to the order of each term in the semantic segments of the current round, and the round numbers are written according to the order of appearance of each semantic segment in the interaction process to form the initial semantic representation sequence.

[0087] When performing context association processing and temporal recursion processing, the current term representation is first associated with the previous term representation in the same round and the semantic representation in the previous round to form a round context representation; the round context representation of the current round is combined and updated with the recursive state of the previous round according to the round order, and the update result is passed to the next round; after all rounds are processed, the semantic feature sequence is obtained.

[0088] The target stage nodes, stage transition fields, advance depth fields, and rollback status fields in the stage anchor point sequence are extracted and combined in round order to form a stage input sequence. During extraction and combination, the target stage nodes, stage transition fields, advance depth fields, and rollback status fields corresponding to each round are read round by round according to the round number. The target stage nodes, stage transition fields, advance depth fields, and rollback status fields corresponding to the same round are concatenated in a preset order to form the stage representation of the current round. The stage representations of each round are arranged in round order to form the stage input sequence.

[0089] The stage input sequence is processed by performing stage position mapping, transition state mapping, advance depth mapping, and backtrack state mapping to form an initial stage representation sequence. This initial stage representation sequence is then input into the stage encoding branch of the improved stage-anchored dual-stream constrained RetNet, and stage association processing and temporal recursion processing are performed to obtain the stage feature sequence.

[0090] In this embodiment, when performing stage position mapping, migration state mapping, advancement depth mapping, and rollback state mapping, the target stage node, stage migration field, advancement depth field, and rollback state field corresponding to the same round are read. Based on the stage position index table, the stage name corresponding to the target stage node is replaced with the corresponding position number to obtain the stage position identifier. Based on the preset migration flag table, the stage forward, stage hold, and stage rollback in the stage migration field are replaced with the corresponding migration state identifiers to obtain the migration state identifiers. The stage advancement level and round advancement position in the advancement depth field are read separately and converted into corresponding level identifiers and... The position identifiers are combined in the order of hierarchy first, then position to obtain the advancement depth identifier. The retreat start point, retreat end point, and retreat span are extracted from the retreat status field. The retreat start point and retreat end point are converted into corresponding position numbers according to the stage position index table, and the retreat span is kept as the corresponding difference identifier. They are combined in the order of start point, end point, and span to obtain the retreat status identifier. Finally, the stage position identifier, migration status identifier, advancement depth identifier, and retreat status identifier corresponding to the same round are concatenated in a preset order to form the stage representation of the current round. The stage representations of each round are arranged in round order to form the initial stage representation sequence.

[0091] In this embodiment, the process of generating the stage memory retention feature sequence in the stage memory retention module includes:

[0092] The alignment and fusion results are input into the stage anchoring Retention layer in round order, and the target stage node, stage transition field, advance depth field and rollback status field corresponding to the current round are read.

[0093] When the target stage node corresponding to the current round is the same as the target stage node corresponding to the previous round, the alignment and fusion result of the current round and the retained result of the previous round are overlaid and updated in the same stage, and the cumulative write in the same stage is performed.

[0094] Specifically, when the target stage node corresponding to the current round is the same as the target stage node corresponding to the previous round, the alignment and fusion result of the current round and the retention result of the previous round are read first. Based on the feature position corresponding to the same stage node, the alignment and fusion result of the current round and the retention result of the previous round are compared at the same position. For the same stage features contained in both the current round and the previous round, they are superimposed and merged according to the one-to-one correspondence of the current position, so that the stage features already retained in the previous round and the stage features newly added in the current round are written into the retention result of the current round. For the same stage features that only appear in the current round but not in the previous round, they are directly written into the retention result of the current round. For the same stage features that only exist in the previous round but do not reappear in the current round, they are kept in their original retention state and continued to be written into the retention result of the current round. After the above processing, the stage features that appear across rounds within the same stage are gradually accumulated in consecutive rounds, forming the accumulated writing result of the same stage.

[0095] When the stage migration field corresponding to the current round is marked as stage forward, the feature corresponding to the target stage node in the current round is enhanced and updated according to the advancement depth field, the retention result of the previous round is retained across stages, and the stage jump enhancement write is performed.

[0096] Specifically, when the stage transition field corresponding to the current round is marked as stage forward, the alignment and fusion results of the current round, the retention results of the previous round, and the advancement depth field corresponding to the current round are read first. Based on the stage advancement level in the advancement depth field, the advancement features corresponding to the target stage node in the current round are selected, and the advancement features are used as the enhancement write objects for the current round. In the retention results of the previous round, the stage features corresponding to the completed stage nodes are selected, the stage features are retained, and they are passed on to the current round. The enhancement write objects of the current round and the stage features passed on from the previous round are combined and updated in round order, so that the current round retains the effective advancement information of the previous stage and highlights the newly added advancement features of the current stage. When the stage advancement level of the current round is higher than that of the previous round, the writing intensity of the features corresponding to the current target stage node in the retention results of the current round is increased, the continuous retention state of the features corresponding to the previous stage is maintained, and the stage transition enhancement write results are formed.

[0097] When the stage migration field corresponding to the current round is marked as stage rollback, the features located after the rollback start point in the previous round's retained results are suppressed and updated according to the rollback status field, and the features corresponding to the rollback end point in the current round are restored and updated. Stage rollback suppression writing is performed.

[0098] Specifically, when the stage migration field corresponding to the current round is marked as stage rollback, the alignment and fusion result of the current round, the retention result of the previous round, and the rollback status field corresponding to the current round are read. The rollback start point, rollback end point, and rollback span are extracted from the rollback status field. Based on the rollback start point, the stage features following the rollback start point are located in the retention result of the previous round. Suppression updates are performed on the located stage features to reduce the write extent of the stage features in the retention result of the current round. Simultaneously, the stage features corresponding to the rollback end point are located in the alignment and fusion result of the current round. The stage features are used as recovery write objects; the recovery write objects are written into the current round's retention result, so that the retention degree of the recovery write objects in the current round's retention result is higher than that of the suppressed subsequent stage features; when the rollback span increases, the range of stage features covered by the suppression update is expanded; when the stage feature corresponding to the rollback endpoint appears repeatedly in the current round, cumulative recovery write is performed on the feature corresponding to the repeatedly appearing rollback endpoint; after processing, the subsequent stage features located after the rollback start point are weakened, while the stage feature corresponding to the rollback endpoint is highlighted again, forming the stage rollback suppression write result;

[0099] The writing results corresponding to each round are arranged in round order to generate a stage memory retention feature sequence.

[0100] In this embodiment, the objection fork correction module performs objection detection, objection strength determination, and objection type classification as follows:

[0101] Read the feature fragment corresponding to the current round in the stage memory retention feature sequence, read the user response content of the current round, compare the user response content of the current round with the feature fragments corresponding to the historical rounds before the current round in the stage memory retention feature sequence, detect whether there is any objection triggering content in the user response content of the current round that would block, delay, deviate from, or terminate the current stage's progress, and obtain the objection detection result;

[0102] Based on the location, frequency, and consecutive occurrence of the objection trigger content in the current round of user responses, the objection strength level is determined, and the objection strength judgment is completed. Based on the objection detection results and objection strength level, the objection type of the current round of user responses is classified to obtain the objection type corresponding to the current round.

[0103] In this implementation, when determining the level of objection intensity, the judgment is based on the number of times the objection trigger content appears in the user response content of the current round, its consecutive occurrence, and the degree of conflict with the semantics of the historical rounds: if it appears infrequently and not consecutively, it is judged as a low-intensity objection; if it appears consecutively or repeatedly, it is judged as a medium-intensity objection; if it appears consecutively and repeatedly and clearly conflicts with the semantics of the historical rounds, it is judged as a high-intensity objection. When classifying objection types, the judgment is based on the way the objection trigger content affects the current stage of progress: if it directly negates the current stage of progress, it is classified as a rejection objection; if it raises doubts about the outbound call content, it is classified as a questioning objection; if it delays the expression of the current processing, it is classified as a delaying objection; if it terminates the current interaction, it is classified as an interruption objection. When multiple judgment conditions are met simultaneously, the objection type with the highest objection intensity level is selected as the classification result.

[0104] In this embodiment, the process of performing fork correction on the stage anchor sequence in the objection fork correction module includes:

[0105] Read the objection type, objection strength level, and target stage node corresponding to the current round from the stage anchor point sequence;

[0106] Determine the direction of the fork correction corresponding to the current round based on the type of objection, and determine the magnitude of the fork correction based on the level of objection intensity.

[0107] Replace the target stage node corresponding to the current round with the fork stage node corresponding to the fork correction direction, and write the fork type flag and fork depth flag at the position corresponding to the current round.

[0108] Search the previous stage anchor point in the previous stage anchor point sequence that corresponds to the current round, and read the target stage node, stage migration field and advance depth field from the previous stage anchor point;

[0109] Based on the bifurcation correction direction, bifurcation correction magnitude, and the advancement depth field in the anchor points of the previous stage, the order of the anchor points after the current round is rearranged and the path constraint is updated. The advancement path of the stage after the bifurcation is consistent with the objection type and objection intensity level corresponding to the current round.

[0110] Arrange the anchor points of each stage after the fork correction is completed in round order to generate the constraint update stage anchor point sequence.

[0111] In this implementation, when determining the fork correction direction corresponding to the current round based on the objection type, the objection type and the target stage node corresponding to the current round are read. According to the preset fork correction rules, the objection type is matched with the jumpable stage nodes in the stage node set. When the objection type is a rejection objection, the fork correction direction is determined to be from the current target stage node to the objection processing stage node. When the objection type is a challenge objection, the fork correction direction is determined to be from the current target stage node to the explanation confirmation stage node. When the objection type is a delay objection, the fork correction direction is determined to be from the current target stage node to the delayed confirmation stage node. When the objection type is an interruption objection, the fork correction direction is determined to be from the current target stage node to the end and recycling stage node. If an objection fork marker already exists before the current round, the fork stage node corresponding to the most recent objection fork is read first, and the current fork correction direction is determined to be from the fork stage node to the next fork stage node corresponding to the current objection type.

[0112] When determining the fork correction magnitude based on the objection intensity level, the objection intensity level corresponding to the current round and the position number of the current target stage node in the stage position index table are read; the stage position difference between the current target stage node and the fork stage node corresponding to the fork correction direction is calculated; when the objection intensity level is low, the single-stage offset in the stage position difference is determined as the fork correction magnitude, so that the current round deviates from the current target stage node by only one stage position; when the objection intensity level is medium, the multi-stage offset in the stage position difference is determined as the fork correction magnitude, so that the current round crosses the intermediate stage between the current target stage node and the fork stage node; when the objection intensity level is high, the current target stage node is directly replaced with the fork stage node corresponding to the fork correction direction, all stages after the current target stage node are marked as constraint suppression state, and the overall offset corresponding to the direct replacement is determined as the fork correction magnitude; when objection triggering content has appeared continuously before the current round, an offset of one level is added to the original fork correction magnitude to expand the fork correction range of the current round.

[0113] When performing order rearrangement and path constraint updates on anchor points after the current round, the following steps are taken: First, read the fork stage node, fork correction direction, fork correction magnitude, and target stage nodes corresponding to anchor points after the current round. Second, using the fork stage node as the new starting anchor point, reorder the target stage nodes after the current round according to the fork correction direction. Third, delete the corresponding stage connection relationships for target stage nodes located before the fork stage node and not matching the current objection type. Fourth, retain the original connection order for target stage nodes located after the fork stage node and consistent with the fork correction direction. Fifth, for target stage nodes that conflict with the fork correction direction... The segment node is adjusted to a position following the fork stage node; when the fork correction magnitude is a single-stage offset, only the first target stage node after the current round is replaced; when the fork correction magnitude is a multi-stage offset, the intermediate stage node is crossed and the target stage node after the crossing is moved forward and connected; when the fork correction magnitude is an overall offset, all subsequent stages after the current target stage node are set to a constraint-suppressed state, and only the subsequent path corresponding to the fork stage node is retained; the rearranged stage anchor points are re-established to reconnect the preceding and following relationships, and the updated stage connection order, stage suppression state, and stage retention state are written into each stage anchor point to generate a constraint update stage anchor point sequence.

[0114] In this embodiment, the process of generating a candidate response sequence set in the response generation module includes:

[0115] The memory retains the data corresponding to the current round in the feature sequence of the reading phase and the anchor point sequence of the constraint update phase. Based on the target phase node, fork type marker, fork depth marker and advancement status corresponding to the current round, the response constraint conditions corresponding to the current round are determined.

[0116] When determining the response constraints, read the target stage node, fork type marker, fork depth marker, and advancement status corresponding to the current round; determine the target stage node as the stage theme, the fork type marker as the response direction, the fork depth marker as the deflection magnitude, and the advancement status as the stage position; and use the stage theme, response direction, deflection magnitude, and stage position together as the response constraints.

[0117] Write the response constraints corresponding to the current round into the feature fragments corresponding to the current round in the stage memory retention feature sequence, and perform constraint screening on the feature fragments to obtain candidate response feature fragments;

[0118] When performing constraint screening, feature sub-fragments corresponding to the current target stage node are retained based on the stage theme; feature sub-fragments consistent with the current fork type label are retained based on the response direction; the retention range of feature sub-fragments is controlled based on the deflection magnitude, where only feature sub-fragments within the current stage are retained when the deflection magnitude is small, and feature sub-fragments of the current stage and adjacent stages are retained when the deflection magnitude is large; the retained feature sub-fragments are arranged sequentially according to the stage position to form candidate response feature fragments.

[0119] Semantic expansion and sequential combination are performed on candidate response feature fragments to form several response paths;

[0120] When performing semantic expansion and sequential combination on candidate response feature fragments, semantic fragments related to the same stage are added according to the stage theme, and corresponding guiding semantic fragments are added according to the response direction to form expanded response fragments; then, the expanded response fragments are sorted according to the stage position and spliced ​​together in order of sorting results to form a response path.

[0121] Perform phase consistency checks and round connection checks on each response path, and retain response paths that meet the check conditions;

[0122] During the execution of phase consistency checks and round continuity checks, each response fragment in the response path is read one by one, and the phase information corresponding to each response fragment is compared with the current target phase node. When each response fragment is within the phase range allowed by the current target phase node, the phase consistency check is passed. The first response fragment of each response path is compared with the current round user response content and historical response content. When the first response fragment can inherit the current round user response content and subsequent response fragments maintain continuity, the round continuity check is passed. Finally, response paths that pass both checks are retained.

[0123] Arrange the retained response paths in the order of response to generate a set of candidate response sequences.

[0124] In this embodiment, the process of maintaining, advancing, and reclaiming the execution phase in the interactive control module includes:

[0125] When the closed-loop correction flag indicates that the current round has not completed the current stage objective, the target stage node corresponding to the current round remains unchanged, and the fork type flag, fork depth flag, and progress status corresponding to the current round are maintained during the execution phase.

[0126] When the phase closed-loop correction flag indicates that the current round has completed the current phase objective and meets the conditions for entering the next phase, read the successor phase node of the target phase node corresponding to the current round in the phase node set, replace the successor phase node with the target phase node corresponding to the next round, perform forward update on the corresponding advancement state of the next round, and perform phase forward.

[0127] When the phase closed-loop correction flag indicates that the current round needs to exit the current fork path, read the fork type flag and fork depth flag corresponding to the current round, locate the fork phase node corresponding to the current round, determine the most recent unsuppressed phase node before the fork phase node as the target phase node for recycling, replace the target phase node for recycling with the target phase node corresponding to the next round, and clear the fork type flag and fork depth flag corresponding to the current round, and perform phase recycling.

[0128] In this embodiment, the process of feeding back the interactive control module to the improved stage-anchored dual-stream constraint RetNet includes: feeding back the new round of stage anchor sequence and subsequent interactive data to the improved stage-anchored dual-stream constraint RetNet in round-by-round order, and continuing to perform semantic encoding, stage encoding, stage memory retention, objection bifurcation correction, response generation, and stage closure correction; after each round of processing is completed, reading the target stage node corresponding to the current round in the new round of stage anchor sequence and comparing it with the preset termination stage, wherein the preset termination stage is the result confirmation stage, the end collection stage, or the termination exit stage; when the target stage node corresponding to the current round is in the result confirmation stage and the current stage target has been... If the current round's target stage node is either complete, or the target stage node is in the end-of-cycle phase and there are no new branching stage nodes, or the target stage node is in the termination-exit phase and the interaction has ended, then the new round's stage anchor sequence is determined to have reached the preset termination stage. Feedback processing is stopped, and the final outbound call interaction result is output. The final outbound call interaction result is used to characterize whether the current outbound call task has achieved the preset business objective, the termination stage type of the current interaction, and the processing conclusion corresponding to the current user response. Specifically, it may include at least one of the following: task completion status, result confirmation status, end-of-cycle status, termination-exit status, user acceptance conclusion, user rejection conclusion, delayed processing conclusion, and follow-up status.

[0129] Example 1: To verify the feasibility of this invention in practice, it was applied to a financial service-related telephone outreach scenario. The objective was to remind existing users of expired services, verify their identity, assess their willingness to participate, and guide them through subsequent processing. During the scenario, traditional outbound calling methods rely heavily on fixed script trees for interaction. When user responses deviate from the preset path, the outbound calling process is prone to issues such as distorted stage judgments, abrupt script transitions, increased repeated follow-up questions, and ineffective objection handling. Especially when users provide feedback such as "temporarily inconvenient," "doesn't understand the notification," "already handled," or "will contact you later," conventional solutions often only trigger simple branch jumps, making it difficult to accurately identify the current stage of the interaction or adjust subsequent scripts based on the strength of objections. This leads to prolonged multi-round conversations, decreased task completion rates, and increased user call interruption rates. This example addresses these problems by deploying the system of this invention in an actual outbound calling process, performing staged modeling, objection branch correction, and constraint response generation for the multi-round interaction process.

[0130] In practical applications, the data acquisition module first receives interactive data, which consists of outbound call task identifier, outbound call business type, outbound call target node, user identifier, historical rounds of dialogue content, historical stage progress records, current round user voice data, and current round user text data. After acquisition, the system performs data cleaning, round segmentation, speech-to-text transcription, text normalization, semantic segmentation, stage node extraction, and association integration on the interactive data to form anchor data. After the anchor data enters the stage anchor point construction module, the system arranges the round semantic segments according to the interaction order and establishes a stage position index table based on the stage node set. For each... In a round of user replies, the system does not directly determine whether to continue broadcasting. Instead, it first determines whether the current reply belongs to the identity verification stage, explanation confirmation stage, intention judgment stage, objection handling stage, result confirmation stage, or end and recovery stage. Then, it combines the position in the previous round to determine whether the current state is stage forward, stage hold, or stage rollback. For replies that are repeatedly confirmed within the same stage, the system identifies whether the current round is at the initial stage, intermediate stage, or near completion stage by using the depth of advancement calibration. In the case where the user suddenly returns from a later stage to a previous stage, the system clarifies the rollback start point, rollback end point, and rollback span by using the rollback status calibration.

[0131] After receiving the round semantic information and the stage anchor sequence, the dual-stream encoding module encodes the user's expressions and stage progress status in each round. The semantic encoding branch focuses on extracting the semantics of the current round, the semantics of previous rounds, and the continuity between consecutive rounds. The stage encoding branch focuses on extracting the target stage node, stage transition field, progress depth field, and fallback state field, enabling the model to not only "understand" what the user said but also "know" where the current call is at. The stage memory retention module inputs the alignment and fusion results into the stage anchoring Retention layer. When the current round and the previous round are in the same stage, it performs cumulative writing within the same stage, gradually superimposing the effective information that appears consecutively within the stage. When the current round enters the next stage, it performs stage transition enhancement writing, highlighting the features of the new stage while retaining the confirmed information of the previous stage. When the current round falls back to the previous stage, it performs stage fallback suppression writing, weakening the subsequent features after the fallback starting point and restoring the core features corresponding to the fallback endpoint. After this processing, the system does not retain all historical dialogues equally, but selectively strengthens, continues, or suppresses different features according to stage changes.

[0132] When a user expresses a clear objection, the objection fork correction module reads the memory retention features of the current stage and the user's response content, detects whether there is any triggering content that would block, delay, deviate from, or terminate the current stage's progress, and determines the objection strength level based on the frequency of occurrence, continuity, and degree of conflict with historical progress semantics. It then categorizes the objection as a rejection objection, a questioning objection, a delaying objection, or an interruption objection. After the determination, the system no longer mechanically advances along the original path, but instead determines the fork correction direction based on the objection type, the fork correction magnitude based on the objection strength level, and performs a reordering of the stage anchor point sequence and updates the path constraints. For example, with "Contact me later," the system will guide the current path to the delayed confirmation stage; with "I didn't understand this notification," the system will guide the current path to the explanation confirmation stage; and with "It's not necessary to continue now," the system will guide the current path to the end and recycling stage. After receiving the feature sequence from the stage memory and the anchor sequence from the constraint update stage, the response generation module first determines the response constraints, then performs constraint filtering, semantic expansion, and sequential combination on the feature fragments to form multiple candidate response paths. Subsequently, it performs stage goal consistency checks and round connection checks, retaining only the response paths that can carry the semantics of the current user and conform to the current stage goal, and outputs the response content for the current round of outbound calls. The interaction control module performs stage maintenance, stage forwarding, and stage retrieval based on the stage closed-loop correction flag, so that the new round of stage anchor sequence is re-inputted into the improved stage anchoring dual-stream constraint RetNet until the result confirmation stage, end retrieval stage, or termination exit stage meets the termination conditions, forming the final outbound call interaction result.

[0133] To verify the practical effect of the present invention in the above-mentioned scenarios, this embodiment selects several consecutive batches of outbound call samples for comparative verification, with a total sample size of 12,000 calls. Among them, the sample size using the system of the present invention is 6,000 calls, and the sample size using the conventional fixed-script outbound call scheme is 6,000 calls. The samples maintain consistency in user type distribution, task type distribution, and initial intention state distribution, and the results are shown in Table 1 below.

[0134] Table 1: Comparison of Interaction Effects of Multi-Round Outbound Calls

[0135] Indicator Item Standard fixed script This invention system Increase Total sample size (general) 6000 6000 — Average number of effective dialogue rounds (rounds) 8.7 6.2 Decrease of 28.7% Average one-way interaction time (seconds) 98.4 74.6 Decrease of 24.2% Stage identification accuracy (%) 84.1 95.3 An increase of 11.2 percentage points Objection identification accuracy (%) 78.6 93.8 An increase of 15.2 percentage points Consistency rate of responses with the current stage (%) 81.9 96.1 An increase of 14.2 percentage points Success rate of continuing the process after objection (%) 51.4 74.9 An increase of 23.5 percentage points Percentage of paths that revert to effective paths after a phase rollback (%) 46.8 72.5 An increase of 25.7 percentage points Early disconnection rate (%) 18.7 8.9 A decrease of 9.8 percentage points Task completion rate (%) 63.5 79.2 An increase of 15.7 percentage points Success rate of result confirmation (%) 58.3 76.8 An increase of 18.5 percentage points Percentage of manual secondary connection (%) 22.6 11.7 A decrease of 10.9 percentage points Accuracy rate of identifying procrastination-related objections (%) 79.8 94.6 An increase of 14.8 percentage points Success rate of reverting to the main path after objections raised (%) 49.7 71.3 An increase of 21.6 percentage points

[0136] As shown in Table 1, the present invention significantly improves the overall performance in multi-round outbound call scenarios. Firstly, regarding interaction efficiency, the average number of effective dialogue rounds decreased from 8.7 to 6.2, and the average single-call interaction time decreased from 98.4 seconds to 74.6 seconds. This indicates that the present invention, through stage anchor point construction, stage memory retention, and constraint response generation, can more quickly locate the current stage of the interaction, output response content that better matches the current business progress, reduce invalid round-trip confirmations and repeated follow-up questions, and make the outbound call process more efficient.

[0137] In terms of interaction understanding and stage control capabilities, the stage identification accuracy improved from 84.1% to 95.3%, the objection identification accuracy improved from 78.6% to 93.8%, and the consistency rate between the response and the current stage improved from 81.9% to 96.1%. This result demonstrates that the present invention does not simply perform text matching on user statements, but rather integrates round semantic information and stage anchor sequence into an improved stage-anchored dual-stream constrained RetNet for processing. Combined with the stage anchoring retention writing mechanism, it differentiates and preserves information under different stage transition states. Therefore, it exhibits stronger stability and accuracy in the three key stages of stage judgment, objection detection, and response generation.

[0138] In terms of objection handling and path recovery capabilities, the advantages of this invention are more prominent. The success rate of continuing to advance after objection increases from 51.4% to 74.9%, the proportion of returning to an effective path after a stage rollback increases from 46.8% to 72.5%, and the premature hang-up rate decreases from 18.7% to 8.9%. This indicates that when users refuse, question, delay, or interrupt their expression, this invention can rearrange the order of subsequent stage anchor points and update path constraints through objection intensity judgment, objection type classification, and bifurcation correction. Combined with stage maintenance, stage forwarding, and stage recovery, it completes closed-loop control, enabling the dialogue to return to an effective advancement path even after deviations occur. This significantly improves the problems of rigid objection handling and high interaction interruption rates in conventional fixed-script solutions.

[0139] From the final business results, the task completion rate increased from 63.5% to 79.2%, the result confirmation success rate increased from 58.3% to 76.8%, and the proportion of manual secondary takeover decreased from 22.6% to 11.7%. The accuracy rate of identifying procrastination-type objections reached 94.6%, and the success rate of returning to the main path after questioning-type objections reached 71.3%, further demonstrating that this invention not only improves basic dialogue capabilities but also enhances the ability to successfully complete outbound call tasks. Combined with the entire set of data, it can be seen that this invention has achieved good technical results in outbound call stage identification, objection bifurcation handling, response generation, and interactive closed-loop control, improving the continuity, targeting, and task completion stability of multi-round outbound call interactions.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based AI intelligent outbound call dialogue generation and multi-turn interaction system, characterized in that, include: The data acquisition module is used to collect interactive data and preprocess it to generate anchor data. The phase anchor point construction module is used to perform phase attribution calibration, phase migration calibration, advance depth calibration and rollback status calibration, and construct a phase anchor point sequence. The dual-stream coding module is used to input the round semantic information and stage anchor sequence from the anchored data into the improved stage anchoring dual-stream constraint RetNet, and perform semantic coding and stage coding respectively to generate semantic feature sequence and stage feature sequence. The stage memory retention module is used to perform alignment fusion, input the alignment fusion result into the stage anchoring Retention layer, perform same-stage cumulative writing, stage transition enhancement writing and stage backoff suppression writing, and generate stage memory retention feature sequence. The objection fork correction module is used to perform objection detection, objection strength determination and objection type classification. When an objection fork is triggered, fork correction is performed on the stage anchor point sequence to generate the constraint update stage anchor point sequence. The response generation module is used to input the stage memory retention feature sequence and the constraint update stage anchor point sequence into the response generation layer, generate a candidate response sequence set, perform stage target consistency screening, determine the response content of the current round of outbound calls, and generate stage closed-loop correction markers; The interactive control module is used to perform phase maintenance, phase forwarding, and phase recycling on the anchor point sequence of the constraint update phase, generate a new round of phase anchor point sequence, and feed it back to the improved phase anchoring dual-stream constraint RetNet until the new round of phase anchor point sequence reaches the preset termination phase, and output the final outbound call interaction result.

2. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The interactive data in the data acquisition module includes outbound call task identifier, outbound call service type, outbound call target node, user identifier, historical round dialogue content, historical stage progress record, current round user voice data, and current round user text data; preprocessing includes sequentially performing data cleaning, round segmentation, speech transcription, text normalization, semantic segmentation, stage node extraction and association integration on the interactive data to generate anchor data; the anchor data includes outbound call task context data, round semantic segment set, and stage node set.

3. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of constructing the stage anchor point sequence in the stage anchor point construction module includes: The semantic fragment set of each round is arranged sequentially according to the interaction order to form a round labeling queue; A stage position index table is established based on the stage node set to record the position number of each stage node in the business advancement chain; For each semantic fragment in the round labeling queue, the corresponding semantic content is matched with each stage node one by one, the stage node with the highest matching result is determined as the target stage node, and written into the stage attribution field to complete the stage attribution labeling. Read the target stage node of the semantic fragment in the current round and the target stage node of the semantic fragment in the previous round, and compare their position numbers. If the current position number is greater than the previous position number, mark it as stage forward; if the current position number is equal to the previous position number, mark it as stage hold; if the current position number is less than the previous position number, mark it as stage rollback. Write the stage migration field to complete the stage migration calibration. Read the target stage node position number of the semantic fragment in the current round, combine it with the number of valid rounds before the current round and the number of stage forward moves, determine the stage advancement level and round advancement position, write it into the advancement depth field, and complete the advancement depth calibration. For the current round semantic segment marked as stage rollback, retrieve the most recent round semantic segment marked as stage forward, determine the rollback start point, rollback end point and rollback span, write them into the rollback status field, and complete the rollback status labeling; Combine them in round order to generate a phase anchor sequence.

4. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of performing semantic coding and stage coding in the dual-stream coding module includes: The round-based semantic information in the anchored data is expanded and arranged in round-based order to form a semantic input sequence; The semantic input sequence is segmented, marked with position and round, to form an initial semantic representation sequence. The semantic encoding branch of the improved stage anchored dual-stream constrained RetNet is then input, and context association processing and temporal recursion processing are performed to obtain the semantic feature sequence. The target stage node, stage migration field, advance depth field, and rollback status field in the stage anchor point sequence are extracted and combined in round order to form the stage input sequence; The stage input sequence is processed by performing stage position mapping, transition state mapping, advance depth mapping, and backtrack state mapping to form an initial stage representation sequence. This initial sequence is then input into the stage encoding branch of the improved stage-anchored dual-stream constrained RetNet, and stage association processing and temporal recursion processing are performed to obtain the stage feature sequence.

5. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of generating the stage memory retention feature sequence in the stage memory retention module includes: The alignment and fusion results are input into the stage anchoring Retention layer in round order, and the target stage node, stage transition field, advance depth field and rollback status field corresponding to the current round are read. When the target stage node corresponding to the current round is the same as the target stage node corresponding to the previous round, the alignment and fusion result of the current round and the retained result of the previous round are overlaid and updated in the same stage, and the cumulative write in the same stage is performed. When the stage migration field corresponding to the current round is marked as stage forward, the feature corresponding to the target stage node in the current round is enhanced and updated according to the advancement depth field, the retention result of the previous round is retained across stages, and the stage jump enhancement write is performed. When the phase migration field corresponding to the current round is marked as phase rollback, the features located after the rollback start point in the previous round's retained results are suppressed and updated according to the rollback status field, and the features corresponding to the rollback end point in the current round are restored and updated. The phase rollback suppression write is then performed. The writing results corresponding to each round are arranged in round order to generate a stage memory retention feature sequence.

6. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The objection fork correction module includes the following processes for objection detection, objection strength determination, and objection type classification: Read the feature fragment corresponding to the current round in the stage memory retention feature sequence, read the user response content of the current round, compare the user response content of the current round with the feature fragments corresponding to the historical rounds before the current round in the stage memory retention feature sequence, detect whether there is any objection triggering content in the user response content of the current round that would block, delay, deviate from, or terminate the current stage's progress, and obtain the objection detection result; Based on the location, frequency, and consecutive occurrence of the objection trigger content in the current round of user responses, the objection strength level is determined, and the objection strength judgment is completed. Based on the objection detection results and objection strength level, the objection type of the current round of user responses is classified to obtain the objection type corresponding to the current round.

7. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of performing fork correction on the stage anchor sequence in the objection fork correction module includes: Read the objection type, objection strength level, and target stage node corresponding to the current round from the stage anchor point sequence; Determine the direction of the fork correction corresponding to the current round based on the type of objection, and determine the magnitude of the fork correction based on the level of objection intensity. Replace the target stage node corresponding to the current round with the fork stage node corresponding to the fork correction direction, and write the fork type flag and fork depth flag at the position corresponding to the current round. Search the previous stage anchor point in the previous stage anchor point sequence that corresponds to the current round, and read the target stage node, stage migration field and advance depth field from the previous stage anchor point; Based on the bifurcation correction direction, bifurcation correction magnitude, and the advancement depth field in the anchor points of the previous stage, the order of the anchor points after the current round is rearranged and the path constraint is updated. The advancement path of the stage after the bifurcation is consistent with the objection type and objection intensity level corresponding to the current round. Arrange the anchor points of each stage after the fork correction is completed in round order to generate the constraint update stage anchor point sequence.

8. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of generating a candidate response sequence set in the response generation module includes: The memory retains the data corresponding to the current round in the feature sequence of the reading phase and the anchor point sequence of the constraint update phase. Based on the target phase node, fork type marker, fork depth marker and advancement status corresponding to the current round, the response constraint conditions corresponding to the current round are determined. Write the response constraints corresponding to the current round into the feature fragments corresponding to the current round in the stage memory retention feature sequence, and perform constraint screening on the feature fragments to obtain candidate response feature fragments; Semantic expansion and sequential combination are performed on candidate response feature fragments to form several response paths; Perform phase consistency checks and round connection checks on each response path, and retain response paths that meet the check conditions; Arrange the retained response paths in the order of response to generate a set of candidate response sequences.

9. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of executing phase retention, phase forwarding, and phase recycling in the interactive control module includes: When the closed-loop correction flag indicates that the current round has not completed the current stage objective, the target stage node corresponding to the current round remains unchanged, and the fork type flag, fork depth flag, and progress status corresponding to the current round are maintained during the execution phase. When the phase closed-loop correction flag indicates that the current round has completed the current phase objective and meets the conditions for entering the next phase, read the successor phase node of the target phase node corresponding to the current round in the phase node set, replace the successor phase node with the target phase node corresponding to the next round, perform forward update on the corresponding advancement state of the next round, and perform phase forward. When the phase closed-loop correction flag indicates that the current round needs to exit the current fork path, read the fork type flag and fork depth flag corresponding to the current round, locate the fork phase node corresponding to the current round, determine the most recent unsuppressed phase node before the fork phase node as the target phase node for recycling, replace the target phase node for recycling with the target phase node corresponding to the next round, and clear the fork type flag and fork depth flag corresponding to the current round, and perform phase recycling.

10. The AI-based intelligent outbound call dialogue generation and multi-turn interaction system based on deep learning according to claim 1, characterized in that, The process of feeding back to the improved stage-anchored dual-stream constraint RetNet in the interaction control module includes: feeding back the new round of stage anchor sequence and subsequent interaction data to the improved stage-anchored dual-stream constraint RetNet in round order, and continuing to perform semantic encoding, stage encoding, stage memory retention, objection fork correction, response generation, and stage closure correction; after each round of processing is completed, reading the target stage node corresponding to the current round in the new round of stage anchor sequence and comparing it with the preset termination stage, wherein the preset termination stage is the result confirmation stage, the end collection stage, or the termination exit stage; if the target stage node corresponding to the current round is the result confirmation stage and the current stage target has been completed, or the target stage node corresponding to the current round is the end collection stage and there are no new fork stage nodes, or the target stage node corresponding to the current round is the termination exit stage and the interaction has ended, it is determined that the new round of stage anchor sequence has reached the preset termination stage, the feeding back process is stopped, and the final outbound call interaction result is output.

Images