A real-time call recall method and system based on historical call memory

Abstract

This invention discloses a real-time call retrieval method and system based on historical call memory. The method includes: acquiring historical call data of a target customer, extracting call text, performing semantic analysis on the call text to obtain key information corresponding to billing communication; constructing call memory units and multi-type call event indexes based on the key information to store the target customer's call memory; parsing the current-round semantics of the target customer's real-time call, constructing joint query conditions based on the current-round semantics, and outputting the call memory retrieval result based on the joint query conditions; and outputting the call response result based on a preset memory injection mechanism, according to the retrieval result and a dialogue generation model. Therefore, this invention can achieve structured accumulation, accurate retrieval, and controlled injection of key information from historical calls with low response latency, improving the continuity, accuracy, and compliance of billing service communication.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a real-time call recall method and system based on historical call memory. Background Technology

[0002] In billing service scenarios, telephone calls remain the core interaction method due to their high reach and immediacy. With breakthroughs in speech recognition and large language model technologies, AI-powered outbound calling and agent assistance systems are gradually becoming more widespread. A key technical challenge is how to effectively utilize historical information across conversations. Current technologies rely solely on the context window of a single conversation, failing to link historical call records across conversations and time periods. This results in a lack of consistent historical understanding, leading to issues like repeated inquiries and fragmented communication. Secondly, billing service scenarios contain significant amounts of unstructured noise in call text, while billing scenarios require precise extraction of structured facts. Traditional text retrieval methods suffer from insufficient accuracy. Furthermore, billing service calls have strict requirements for real-time performance and compliance, but current full-history retrieval technologies suffer from high response latency and lack effective access control and injection constraint mechanisms to address compliance requirements and identity verification, severely impacting communication efficiency and customer experience. Clearly, existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a real-time call recall method and system based on historical call memory, which can realize the structured accumulation, accurate recall and controlled injection of key information of historical calls with low response latency, thereby improving the continuity, accuracy and compliance of billing service communication.

[0004] To address the aforementioned technical problems, the first aspect of this invention discloses a real-time call recall method based on historical call memory, the method comprising: Obtain the target customer's historical calls and transcribe them into call text. Perform semantic analysis on the call text to obtain key information related to billing communication. Based on the aforementioned key information, a call memory network and a call event index are constructed to store the call memories of target customers; Identify the current-round semantics of the target customer's real-time call, construct joint query conditions based on the current-round semantics, and output the recall result of the call memory based on the joint query conditions; Based on the preset memory injection mechanism, the call response result is output according to the recall result and the dialogue generation model.

[0005] As an optional implementation, in the first aspect of the present invention, obtaining historical call data of a target customer and transcribing it into call text includes: The historical call input speaker differentiation model composes the historical call data into call text, generates role tags for the speakers in the call text, and records the call sequence in the call text using timestamps.

[0006] As an optional implementation, in the first aspect of the present invention, semantic analysis is performed on the call text to obtain key information corresponding to the accounting communication, including: The call text is input into a multi-task joint learning semantic analysis model for semantic analysis and rule verification to obtain the key information of named entities and triples; The named entities are identified and extracted through a pre-trained encoding and decoding layer, and the triples are extracted from the single sentences of the call text and the dialogue context to obtain the intent-slot-value triples. The rule validation is performed by pre-setting the named entities and meanings. Figure Three The normalized domain rules of the tuples are used to verify the consistency between the obtained key information and the call text, and to output structured key information.

[0007] As an optional implementation, in the first aspect of the invention, constructing a call memory network includes: A unique identifier is generated based on the record ID of the historical call and the hash value of the corresponding timestamp of the historical call; Extract a set of key-value pairs of structured facts from the intent-slot-value pairs in the key information triples, and generate fact slots; Based on the fields in the call text that support the structured facts, generate evidence fragments pointing to the timestamp range and text content corresponding to the evidence of that field; The fact slots and evidence fragments are input into a preset summary generation model, and a business summary of the historical call data is output. The time attribute is determined based on the memory construction time, the time of the fact occurrence, and the memory update time; Metadata is generated by extracting confidence scores based on extracted triples from key information, and obtaining customer / agent information sources and related accounting communication task / case IDs based on named entities. The call memory unit for a single historical call corresponding to the target customer is constructed using the unique identifier, fact slot, evidence fragment, business summary, time attribute, and metadata. For each call memory unit, generate a relation pointer based on the logical relationship between any two call memory units; The call memory network of the target customer is constructed using all call memory units and the pointers to the relationships between them.

[0008] As an optional implementation, in the first aspect of the invention, a call event index is constructed to store the call memories of a target customer, including: The text representation of the service summary and fact slot in each of the call memory units is encoded into a high-dimensional dense vector, and the high-dimensional dense vector is used as the first index; Generate a range-filtered and exact-matching inverted index for the key and value corresponding to the fact slot in each of the call memory units, and use the inverted index as the second index; Using the event occurrence time corresponding to the time attribute in each of the call memory units as the primary key, a time-series index supporting time window retrieval is constructed, and the time-series index is used as the third index. By using call memory units as nodes and logical relationship types of relation pointers as edges, a memory graph is constructed from all the relation pointers. A graph index for traversing and querying the memory graph is then constructed, and the graph index is used as the fourth index. Each call memory unit in the call memory network is written into the first index, the second index, the third index, and the fourth index for storage.

[0009] As an optional implementation, in the first aspect of the invention, identifying the current-round semantics of a target customer's real-time call includes: The system acquires the real-time call content of the target customer, identifies the target customer's intent and slot information in the current call based on the semantic analysis model, and generates call status text containing the current call semantics based on the dialogue status tracking module. The dialogue status tracking module is used to maintain the current call status and verify the consistency of the current call semantics in consecutive multiple calls by fusing the intent and slot information.

[0010] As an optional implementation, in the first aspect of the present invention, constructing joint query conditions based on the current round semantics and outputting the recall result of the call memory based on the joint query conditions includes: The current round semantics and call status text are input into a pre-trained embedding model to generate a query semantic vector; Extract the filtering fields of intent and slot information from the semantics of the current round, and generate structured filtering conditions; Based on the business rules of accounting communication and the context of the call status text, determine the time range of the query conditions to set the time query window; Based on the causal queries and / or process queries involving consecutive multi-turn calls in the current turn semantics, generate graph query statements with graph traversal conditions; Based on the query semantic vector, structured filtering conditions, time query window, and graph query statement, the call memory network is executed in parallel for multi-path recall to obtain an initial candidate set; For each call memory unit in the initial candidate set, multiple feature scores are calculated. Based on the final score obtained by fusing multiple feature scores, all call memory units are sorted to obtain the recall result. The feature scores include semantic similarity score, time decay score, confidence score, and task relevance score.

[0011] As an optional implementation, in the first aspect of the present invention, based on a preset memory injection mechanism, a call response result is output according to the recall result and the dialogue generation model, including: Define a time decay curve for each type of call memory unit, scan the call memory network according to the time decay curve to determine the validity of each call memory unit, and adjust the sorting weight or perform secondary verification on the call memory units in the recall results according to the reference value of the validity. The call memory network detects first and second memory units that have some identical memories, compares the confidence scores and evidence fragment completeness of the two memory units, and performs a resolution process on the identified conflicting memory units. The conflicting memory units and the call memory units have at least one difference in unique identifier, fact slot, evidence fragment, business summary, time attribute or metadata. The call memory units in the recall results that are greater than the sorting threshold are formatted into fact cards according to a preset card template. The card template is a natural language statement of the recorded facts in the call memory unit. The identity verification status of the target customer is obtained to determine the anonymization level. Based on the anonymization level, the sensitive fields of the fact card are filtered and injected into the dialogue generation model. The filtered fact card and the current call status text are used to generate prompt words and input into the dialogue generation model to output the call response result.

[0012] A second aspect of this invention discloses a real-time call recall system based on historical call memory, the system comprising: The acquisition module is used to acquire the target customer's historical calls and transcribe them into call text, and perform semantic analysis on the call text to obtain key information corresponding to the billing communication. The memory module is used to construct a call memory network and a call event index based on the key information, so as to store the call memories of the target customer; The recall module is used to identify the current-round semantics of the real-time call of the target customer, construct joint query conditions based on the current-round semantics, and output the recall result of the call memory based on the joint query conditions; The output module is used to output the call response result based on the recall result and the dialogue generation model, according to the preset memory injection mechanism.

[0013] A third aspect of the present invention discloses another real-time call recall system based on historical call memory, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the real-time call recall method based on historical call memory disclosed in the first aspect of the present invention.

[0014] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the real-time call recall method based on historical call memory disclosed in the first aspect of the present invention.

[0015] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention extracts key information from transcribed historical call text through semantic analysis, constructs a call memory network and call event index, and improves the relevance of historical call data to overcome the bottleneck of cross-session historical association. By performing semantic analysis on call text, it extracts structured key information, filters out unstructured noise, and eliminates unstructured noise interference. By constructing joint query conditions, it improves the accuracy of recall and reduces the response delay of call replies. Through a dynamic injection mechanism, it forms compliance constraints for historical call data recall, solving the problems of historical information gaps, unstructured noise interference, real-time response delays, and lack of compliance constraints in billing service scenarios. This enables the realization of continuous and consistent historical cognition from single call processing, thereby effectively improving the continuity, accuracy, and compliance of billing service communication. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a real-time call recall method based on historical call memory disclosed in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of a real-time call recall system based on historical call memory disclosed in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of another real-time call recall system based on historical call memory disclosed in an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, 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. 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.

[0021] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] This invention discloses a real-time call recall method and system based on historical call memory. By extracting key information from the transcribed text of historical calls through semantic analysis, a call memory network and call event index are constructed to improve the relevance of historical call data, overcoming the bottleneck of cross-session historical association. Semantic analysis of the call text extracts structured key information and filters out unstructured noise, eliminating unstructured noise interference. By constructing joint query conditions, the accuracy of recall is improved and the response latency of call replies is reduced. A dynamic injection mechanism forms compliance constraints for historical call data recall, solving the problems of historical information gaps, unstructured noise interference, real-time response latency, and lack of compliance constraints in billing service scenarios. This achieves a shift from single-call processing to continuous and consistent historical understanding, thereby effectively improving the continuity, accuracy, and compliance of billing service communication. Detailed explanations follow.

[0024] Example 1 Please see Figure 1 , Figure 1This is a flowchart illustrating a real-time call recall method based on historical call memory, as disclosed in an embodiment of the present invention. Figure 1 The described real-time call recall method based on historical call memory can be applied to data processing systems / data processing devices / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 1 As shown, this real-time call recall method based on historical call memory may include the following operations: 101. Obtain the target customer's historical call history and transcribe it into call text. Perform semantic analysis on the call text to obtain key information corresponding to the billing communication.

[0025] Optionally, the key information can be amount, date, ID number, product name, customer name or business ID, etc., which can be output as entity information, or it can be output as an intent-slot-value triple, containing information such as "objection application - reason: unemployment" or "promise to repay - date: 2023-10-01". It is understood that the above entities and triples need to be obtained by semantic analysis, entity recognition or relation extraction of the key information.

[0026] 102. Based on the key information, construct a call memory network and a call event index to store the call memories of the target customer.

[0027] It should be noted that the call memory network includes several call memory units. Each call memory unit includes at least a structured fact slot, business summary information, original call evidence fragments and timestamps associated with the fact slot, confidence information, and timeliness or decay attributes, which can be used as historical information and effectively utilized for subsequent call response generation.

[0028] 103. Identify the current-round semantics of the target customer's real-time call, construct joint query conditions based on the current-round semantics, and output the recall result of the call memory based on the joint query conditions.

[0029] It should be noted that the joint query conditions are related to the method of constructing the call event index in step 102, and can be generated using the same model. For example, the same embedding model can be used when generating the query vector.

[0030] 104. Based on the preset memory injection mechanism, output the call response result according to the recall result and the dialogue generation model.

[0031] It should be noted that the memory injection mechanism is used to perform compliance checks and sensitive field filtering on the call memory units in the recall results, as well as to filter invalid call memory units. It can gradually unlock the corresponding accounting-related information in the output call response results, thereby minimizing information disclosure and compliance control.

[0032] As can be seen, the above-mentioned embodiments of the invention extract key information from the transcribed text of historical calls through semantic analysis, construct a call memory network and a call event index, improve the relevance of historical call data to overcome the bottleneck of cross-session historical association, extract structured key information and filter unstructured noise through semantic analysis of call text, eliminate unstructured noise interference, improve the accuracy of recall and reduce the response delay of call replies by constructing joint query conditions, and form compliance constraints for historical call data recall through a dynamic injection mechanism. This solves the problems of historical information gaps, unstructured noise interference, real-time response delays and lack of compliance constraints in the billing service scenario, thereby realizing the transition from single call processing to continuous and consistent historical cognition, and thus effectively improving the continuity, accuracy and compliance of billing service communication.

[0033] As an optional embodiment, the step of obtaining the target customer's historical call data and transcribing it into call text in the above steps includes: The historical call input speaker differentiation model composes the historical call data into call text, generates role tags for the speakers in the call text, and records the call sequence in the call text using timestamps.

[0034] Optionally, the speaker differentiation model can employ a neural network model based on deep clustering or end-to-end speaker diarization. By fusing acoustic features (such as MFCC, Fbank) and contextual temporal information, it can accurately label each sentence segment of the transcribed text with role tags such as "customer," "agent," or "system," and align them with the timestamps of speech recognition to form a complete dialogue text stream with roles and time sequence.

[0035] As can be seen, through the above optional embodiments, by inputting historical calls to the speaker differentiation model to output call text with speaker labels, the confusion of speakers in mono-channel historical calls is reduced. By generating role labels for speakers in the call text, unstructured noise is filtered to retain structured key information, eliminating unstructured noise interference. By recording the call sequence through timestamps to construct event sequence relationships, the historical information discontinuity rate is reduced to establish a continuous historical cognitive chain. This solves the problems of speaker confusion, unstructured noise interference, and lack of cross-session sequence association in the billing service scenario, and achieves accurate conversion from raw calls to highly structured historical memory, laying a high-quality input foundation for subsequent cross-session retrieval.

[0036] As an optional embodiment, the step above, performing semantic analysis on the call text to obtain key information corresponding to the billing communication, includes: The call text is input into a multi-task joint learning semantic analysis model for semantic analysis and rule verification to obtain the key information of named entities and triples; The named entities are identified and extracted through a pre-trained encoding and decoding layer, and the triples are extracted from the single sentences of the call text and the dialogue context to obtain the intent-slot-value triples. The rule validation is performed by pre-setting the named entities and meanings. Figure Three The normalized domain rules of the tuples are used to verify the consistency between the obtained key information and the call text, and to output structured key information.

[0037] Optionally, the multi-task joint learning model includes a named entity recognition sub-model, an intent and slot joint extraction sub-model, and a rule verification model. The named entity recognition sub-model can be constructed by connecting the encoding layer of a pre-trained language model such as BERT or RoBERTa to a decoding layer. The intent and slot joint extraction sub-model can adopt a joint decoding architecture based on a pointer network or sequence labeling. The rule verification model can adopt a large language model architecture such as LLM.

[0038] Optionally, named entities include, but are not limited to, entities that identify amounts, dates, identification numbers, product names, etc.

[0039] As an example, the extraction of triples first involves identifying the intent entity from named entities, activating the corresponding slot set based on the identified intent entity, and semantically aligning the identified entity with the corresponding slot through hierarchical annotation or pointer mechanism to form preliminary <slot, value> pairs. Then, the content matching the slot is located and extracted from the input text to generate the value. Finally, the identified intent is combined with the aligned slot-value pairs to generate a complete <intent, slot, value> triple.

[0040] As an example, a triple is a triple extracted from a single sentence or dialogue context in the call text, such as "Objection Request - Reason: Unemployment" or "Promise to Repay - Date: 2023-10-01" (Intent, Slot, Value).

[0041] Optionally, the normalization domain rules are configurable preset domain rules, such as date format normalization and monetary unit unification. These rules can be used to calibrate the outputs of the named entity recognition model and the extraction model. At the same time, the preliminary results of the model and the original text fragments are input into an LLM large language model that has been fine-tuned by instructions. The LLM large language model performs consistency verification, conflict detection and information completion to generate the final structured information set.

[0042] As can be seen, through the above optional embodiments, a continuous semantic cognitive chain is constructed by dynamically extracting named entities and triples, which improves the accuracy of semantic association and the coverage of key information. By pre-setting the normalization rules of named entities and triples, the consistency between key information and call text is verified, eliminating semantic ambiguity and compliance risks. Through the semantic analysis and rule verification mechanism of multi-task joint learning, the core problems of unstructured noise interference, inaccurate extraction of key information, and lack of structured semantics in the billing service scenario are solved. It realizes the accurate transformation from raw call text to high-confidence structured key information, providing high-quality input for cross-session recall.

[0043] As an optional embodiment, the step of constructing the call memory network described above includes: A unique identifier is generated based on the record ID of the historical call and the hash value of the corresponding timestamp of the historical call; Extract a set of key-value pairs of structured facts from the intent-slot-value pairs in the key information triples, and generate fact slots; Based on the fields in the call text that support the structured facts, generate evidence fragments pointing to the timestamp range and text content corresponding to the evidence of that field; The fact slots and evidence fragments are input into a preset summary generation model, and a business summary of the historical call data is output. The time attribute is determined based on the memory construction time, the time of the fact occurrence, and the memory update time; Metadata is generated by extracting confidence scores based on extracted triples from key information, and obtaining customer / agent information sources and related accounting communication task / case IDs based on named entities. The call memory unit for a single historical call corresponding to the target customer is constructed using the unique identifier, fact slot, evidence fragment, business summary, time attribute, and metadata. For each call memory unit, generate a relation pointer based on the logical relationship between any two call memory units; The call memory network of the target customer is constructed using all call memory units and the pointers to the relationships between them.

[0044] Optionally, the preset summary generation model can be the BART model or the PEGASUS model, which automatically generates a natural language summary based on evidence fragments and fact slots, making it easy to understand quickly. This application does not impose specific restrictions on the choice of summary generation model.

[0045] Optionally, the confidence score can be the confidence score obtained by the sub-model when extracting triples, or it can be the validation score output by the rule validation model in the rule validation stage.

[0046] Specifically, relation pointers are used to point to the IDs of other memory units that have a logical connection (such as conflict, premise, or follow-up) with the memory unit, thereby initially constructing the memory network.

[0047] As can be seen, through the above optional embodiments, unique identifiers are generated by recording IDs and timestamp hashes to ensure accurate association across session history records. Key-value pairs are extracted from triples to generate evidence fragments, improving the credibility of key facts. Business summaries are output through a summary generation model, and memory construction time, fact occurrence time, and update time are dynamically marked. A continuous cognitive chain is established in terms of semantic coherence and temporal traceability. Metadata and relation pointers are used to build dynamic associations in the memory network, solving the problems of fragmented historical information, lack of evidence support for key facts, and missing cross-session association logic in the accounting service scenario. This achieves accurate construction from the original call to a high-credibility historical cognitive map, laying a structured memory foundation for real-time cross-session recall.

[0048] As an optional embodiment, the step of constructing a call event index to store the call memories of the target customer in the above steps includes: The text representation of the service summary and fact slot in each of the call memory units is encoded into a high-dimensional dense vector, and the high-dimensional dense vector is used as the first index; Generate a range-filtered and exact-matching inverted index for the key and value corresponding to the fact slot in each of the call memory units, and use the inverted index as the second index; Using the event occurrence time corresponding to the time attribute in each of the call memory units as the primary key, a time-series index supporting time window retrieval is constructed, and the time-series index is used as the third index. By using call memory units as nodes and logical relationship types of relation pointers as edges, a memory graph is constructed from all the relation pointers. A graph index for traversing and querying the memory graph is then constructed, and the graph index is used as the fourth index. Each call memory unit in the call memory network is written into the first index, the second index, the third index, and the fourth index for storage.

[0049] Optionally, the high-dimensional dense vectors can use embedding models that have been pre-trained on domain text (such as customer service dialogues or financial agreements), such as Sentence-BERT or Dense Embedding, to encode the textual representation of the “business summary” and “fact slot” of each memory unit into high-dimensional dense vectors (e.g., 768-dimensional). These vectors are stored in vector databases such as FAISS and Milvus, supporting approximate nearest neighbor (ANN) search.

[0050] It should be noted that inverted indexes need to look up records based on attribute values. Each item includes an attribute value and the address of each record that has that attribute value. The location of the record is determined by the attribute value. For example, the address of the call memory unit is determined based on the key and value corresponding to the fact slot.

[0051] Optionally, time-series indexes can be created using B+ trees or segmented timeline indexes to support efficient retrieval by time window; this application does not impose any restrictions on this.

[0052] It is understood that the first index, second index, third index and fourth index constructed in this embodiment are all call event indexes, and the subsequent joint query conditions can adopt the same construction method as the first index, second index, third index and fourth index, and recall the corresponding call memory units through similarity calculation.

[0053] As can be seen, through the above optional embodiments, the semantic matching accuracy and retrieval efficiency are improved by using vector indexes, the range filtering and precise matching of recall results are achieved by using inverted indexes, compliance and dynamic queries are achieved by using time-series indexes and graph indexes, and cross-session association is achieved by using a multi-type index collaborative architecture. This can reduce the response latency of retrieval and recall, while improving the recall accuracy and cross-session association rate, thereby meeting the real-time and accuracy requirements of billing services for calls. Through the multi-modal index collaborative call event storage mechanism, the problems of low historical information association efficiency, strong unstructured noise interference, slow real-time retrieval response, and lack of compliance constraints in the billing service scenario are solved, enabling historical information to be upgraded from static storage to dynamic intelligent recall, providing a highly reliable cognitive foundation for billing service calls.

[0054] As an optional embodiment, the step of identifying the current-round semantics of the target customer's real-time call in the above steps includes: The system acquires the real-time call content of the target customer, identifies the target customer's intent and slot information in the current call based on the semantic analysis model, and generates call status text containing the current call semantics based on the dialogue status tracking module. The dialogue status tracking module is used to maintain the current call status and verify the consistency of the current call semantics in consecutive multiple calls by fusing the intent and slot information.

[0055] Specifically, during real-time dialogue, semantic parsing is performed on the customer's current speech to identify customer intent and slot information, ensuring consistency between call response and real-time interaction. This step can reuse the slot joint extraction sub-model of the multi-task joint learning model for real-time parsing. At the same time, a dialogue state tracking (DST) module is added. This module maintains a current dialogue state, including filled slots and historical intents. By fusing and updating the parsing results of the current user round with the dialogue state, the true intent and complete information needs of the user in continuous dialogue can be accurately understood.

[0056] As can be seen, through the above optional embodiments, by inputting real-time call content into a semantic analysis model, extracting intent and slots to achieve multi-turn semantic association, by integrating historical memory networks and current-turn semantics through a dialogue state tracking module, verifying the consistency of continuous multi-turns to eliminate communication gaps and repeated queries, by generating call state text containing current-turn semantics as input for joint queries, high-confidence real-time decision-making is achieved, and by real-time semantic recognition and dialogue state dynamic verification mechanism, the problems of gaps between current-turn semantics and historical information, lack of consistency in multi-turn communication, and real-time response delays in billing service scenarios are solved, realizing a transition from isolated single-turn interaction to continuous and consistent dynamic cognition.

[0057] As an optional embodiment, the above steps, including constructing joint query conditions based on the current round semantics and outputting the call memory retrieval results based on the joint query conditions, include: The current round semantics and call status text are input into a pre-trained embedding model to generate a query semantic vector; Extract the filtering fields of intent and slot information from the semantics of the current round, and generate structured filtering conditions; Based on the business rules of accounting communication and the context of the call status text, determine the time range of the query conditions to set the time query window; Based on the causal queries and / or process queries involving consecutive multi-turn calls in the current turn semantics, generate graph query statements with graph traversal conditions; Based on the query semantic vector, structured filtering conditions, time query window, and graph query statement, the call memory network is executed in parallel for multi-path recall to obtain an initial candidate set; For each call memory unit in the initial candidate set, multiple feature scores are calculated. Based on the final score obtained by fusing multiple feature scores, all call memory units are sorted to obtain the recall result. The feature scores include semantic similarity score, time decay score, confidence score, and task relevance score.

[0058] It is understandable that the aforementioned query semantic vector, structured filtering conditions, time query window, and the high-dimensional dense vector, inverted index, time-series index, and graph index in the call event index corresponding to the graph query statement can be combined with query conditions to perform multi-way recall based on the call event index, thereby improving the accuracy of recall.

[0059] It should be noted that the business rules for billing communication are time limits corresponding to the business types that the current call can be associated with.

[0060] Specifically, the semantic similarity score is calculated using the vector cosine similarity of the query vector, the time decay score is calculated using the exponential decay function, and the closer the temporal distance record is, the higher the score is. The confidence score is the self-confidence of the call memory unit when it is extracted by the multi-task joint learning model, and the task relevance score is the degree of matching between the intent of the current call and the intent of the call memory unit.

[0061] As can be seen, through the above optional embodiments, unstructured noise interference is reduced and semantic matching accuracy is improved by generating query semantic vectors; accurate historical association is achieved through structured filtering conditions and time windows; cross-session logical association is missing through graph traversal conditions and multi-way recall; and false recall and compliance risks are eliminated through multi-feature fusion scoring and ranking. This solves the problems of cross-session historical association gaps, unstructured noise interference, high real-time response latency, and lack of compliance constraints in the billing service scenario, realizing an intelligent closed loop of dynamic semantic-temporal-relational collaborative recall from static retrieval, and upgrading the billing service from passive waiting to proactive cognition.

[0062] As an optional embodiment, the step above, based on a preset memory injection mechanism, outputs a call response result according to the recall result and the dialogue generation model, including: Define a time decay curve for each type of call memory unit, scan the call memory network according to the time decay curve to determine the validity of each call memory unit, and adjust the sorting weight or perform secondary verification on the call memory units in the recall results according to the reference value of the validity. The call memory network detects first and second memory units that have some identical memories, compares the confidence scores and evidence fragment completeness of the two memory units, and performs a resolution process on the identified conflicting memory units. The conflicting memory units and the call memory units have at least one difference in unique identifier, fact slot, evidence fragment, business summary, time attribute or metadata. The call memory units in the recall results that are greater than the sorting threshold are formatted into fact cards according to a preset card template. The card template is a natural language statement of the recorded facts in the call memory unit. The identity verification status of the target customer is obtained to determine the anonymization level. Based on the anonymization level, the sensitive fields of the fact card are filtered and injected into the dialogue generation model. The filtered fact card and the current call status text are used to generate prompt words and input into the dialogue generation model to output the call response result.

[0063] As an example, the resolution process first defines time decay curves for different types of memory units (e.g., the validity period of a "promised repayment" record is 30 days after the promise date). The system periodically scans the call memory network, and for records that have exceeded their validity period but are not closed, it automatically lowers their weight in the retrieval ranking or marks them as "pending verification." The system periodically or when adding new memory units runs conflict detection rules (e.g., for the same customer and the same outstanding amount, there are two "actual repayment" records with different repayment amounts). After a conflict is detected, the resolution process is initiated, comparing the confidence scores of the conflicting parties, the completeness of the evidence fragments, and the authority of the information source (e.g., agent entry versus customer verbal statement), and checking whether the conflicting records conform to the business process's temporal logic.

[0064] As a further example, for severe conflicts that fail to be resolved automatically, a work order is generated for manual processing, or data from an external authoritative system such as the core accounting system is invoked for calibration.

[0065] Specifically, a gating controller is set up before inputting fact cards into the dialogue generation model (such as GPT, ChatGLM, and other large language models). This controller determines whether to inject information into the call memory unit, which parts of the information to inject, and how to express them, based on the current identity verification status (e.g., whether strong authentication has been passed), the sensitivity level of the information in the memory unit (e.g., bank card numbers and complete ID numbers are highly sensitive), and the current dialogue stage. For example, a repayment reminder record containing the last four digits of the complete card number is only injected after strong authentication is successful. The fact cards, filtered and formatted by the controller, along with the current dialogue history and system instructions, are constructed into structured prompts and input into the large language model to guide it in generating secure, accurate, and natural call responses.

[0066] Furthermore, the corresponding identity verification controls the participation of sensitive fields in the retrieval and injection by setting gating conditions, supporting information security and compliance in accounting communication scenarios. The gating conditions are implemented as a set of configurable policy rules, for example: IF verification level = 'weak verification' THEN only allows the injection of desensitized, non-fund-related memories (such as historical contact record overview); IF verification level = 'strong verification' AND session stage == 'information confirmation' THEN allows the injection of memories containing partially desensitized sensitive information (such as past repayment amounts and dates). It can be understood that the system will mark the fact slot of each memory unit with a sensitivity level label, and the gating controller will dynamically determine which levels of fields in the retrieval index can participate in the query, and which fields can be presented in the final card, based on the real-time verification results and session status.

[0067] As can be seen, through the above optional embodiments, by defining a time decay curve for each type of memory unit, the call memory network is scanned to dynamically adjust the ranking weight of the recall results, reducing the interference of historical memory failure on the recall results. By detecting conflicting units with the same identifier and comparing confidence and evidence fragment completeness, high-confidence memory versions are retained to eliminate logical contradictions in multi-source memories. By formatting the recall results into natural language cards as structured inputs to the dialogue generation model, the accuracy and generation efficiency of semantic decisions are improved. Through dynamic desensitization and compliance constraints of the memory injection mechanism, the risk of sensitive information leakage is resolved. By integrating time decay, conflict resolution, and dynamic desensitization, historical memory is upgraded from a potential source of interference to a compliant decision engine, solving the problems of historical memory obsolescence interference, multi-source memory conflict, sensitive information leakage, and inaccurate real-time decision-making in the accounting service scenario. This achieves a transition from static memory retrieval to dynamic compliant semantic injection, improving subsequent call response results to high-quality outputs with high continuity, high accuracy, and high compliance.

[0068] In one specific implementation scheme, a real-time call recall system based on historical call memory is provided according to the technical solution of the present invention, which addresses the following shortcomings of the prior art: (1) Most existing systems only perform context modeling within a single session or a limited number of multi-turn dialogue windows, making it difficult to effectively utilize historical call records across sessions and time periods. When the same customer repeatedly communicates about billing issues, expresses objections, or makes commitments in multiple billing calls on different dates, the system cannot form a continuous and consistent historical understanding, which can easily lead to problems such as repeated inquiries and fragmented communication content, thereby affecting communication efficiency and customer experience; (2) Some systems attempt to introduce retrieval enhancement generation or dialogue memory mechanisms, but usually treat historical call texts as ordinary text for vectorized retrieval. Since call texts contain a large number of small talk, interruptions, slips of the tongue and invalid rounds, and the accounting communication scenario is highly dependent on structured information such as commitment dates, installment plans, reasons for objections and identity verification results, the above methods are prone to problems such as inaccurate recall, omission of key information or inconsistent generated content in practical applications.

[0069] (3) In addition, accounting communication calls have strict requirements for real-time performance. They usually need to complete understanding, decision-making and response generation within 500ms in a single round of dialogue, which makes it difficult to accept high-latency solutions that repeatedly retrieve and summarize the entire historical call data. At the same time, financial accounting scenarios have high requirements for privacy protection and compliance. Sensitive historical information must not be injected into the generation model before identity verification is completed, and existing general dialogue memory solutions usually lack effective access control and injection constraint mechanisms.

[0070] To address the aforementioned shortcomings, this recall system proposes a mechanism for utilizing historical interaction memory, which involves first performing structured data storage, then joint recall, and finally controlled injection. Specifically, instead of directly using the original historical call text as the retrieval object, it constructs historical calls as call memory units. During real-time processing, it dynamically constructs joint retrieval queries based on the semantics of the current round, recalls relevant call memory units from multiple types of indexes, and finally injects them in a controlled manner into the dialogue generation model in the form of fact cards, outputting the next call response result. This mechanism enables the structured storage, accurate recall, and controlled injection of key information from historical calls while maintaining low latency, thereby improving the continuity, accuracy, and compliance of the accounting communication system.

[0071] Typically, in an intelligent billing service system, call records between the customer and the service provider are used for subsequent billing communication, service improvement, and audit analysis. To illustrate the effectiveness of this technical solution, let's take a complex billing negotiation involving multiple communications as an example.

[0072] The recall system's working method provides the following example in complex billing negotiation scenarios: S1: In an initial call, Customer A calls the service hotline to inquire about installment plans for their credit card bill. Customer Service Representative B engages in brief pleasantries with Customer A before getting to the point. The representative records Customer A's current preferred installment plan and the date they commit to making installment payments—this information is automatically extracted and used to generate a call memory unit. This unit includes structured fact slots such as "preferred installment plan," "committed payment date," associated fragments of original verbal evidence, and the timestamp and confidence level at which this information was created. S2: Several days later, customer A called again to inquire about changes to the installment payment terms from the previous call. The real-time dialogue engine obtains customer A's semantic information by parsing the semantic information steps of the current round in the real-time dialogue, identifies that the customer wants to modify the installment plan, and dynamically calls the joint query conditions based on the new semantic information. Combining semantic filtering and installment-related task constraints, the system retrieves and recalls the previously recorded call memory units. S3: Through sorting mechanisms, such as task consistency priority, to ensure that the most relevant information is presented first, the system quickly selects the unit containing the initial installment intention from multiple call memory units, and generates a response for customer A again through the model generation system; S4: When generating a response, the system verifies the customer's identity through an identity verification mechanism before allowing the memory unit containing sensitive information to be injected into the prompts of the real-time dialogue generation model. This process ensures information compliance and privacy protection, thereby generating real-time call response results.

[0073] Through the above process, the system completes the entire process from historical retrieval to real-time dialogue generation within a 500ms latency limit, effectively avoiding duplicate queries and inconsistent responses. This embodiment demonstrates that, through the innovative method of structured memory units and dynamic joint retrieval, the system achieves efficient utilization of historical information, enhancing information consistency and decision support in continuous dialogue.

[0074] In summary, the technical solutions disclosed in the embodiments of the present invention have the following advantages: (1) The structured storage of key information from historical calls is achieved through the call memory unit, which significantly reduces noise interference; (2) Adopt a joint retrieval method that combines semantics, structure, and task constraints to improve the accuracy of historical information retrieval; (3) By combining offline construction with online lightweight recall, the requirements for low latency in real-time calls are met; (4) Introduce evidence fragments and confidence information to enhance the interpretability and auditability of the generated results; (5) By using identity verification gating and desensitization injection mechanisms, compliance risks in financial accounting communication scenarios can be effectively reduced.

[0075] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a real-time call recall system based on historical call memory, as disclosed in an embodiment of the present invention. Figure 2 The described real-time call recall system based on historical call memory can be applied to data processing systems / data processing equipment / data processing servers (wherein, the server includes local processing servers or cloud processing servers). For example... Figure 2As shown, the real-time call recall system based on historical call memory may include: The acquisition module 201 is used to acquire the target customer's historical calls and transcribe them into call text, and perform semantic analysis on the call text to obtain key information corresponding to the billing communication. The memory module 202 is used to construct a call memory network and a call event index based on the key information, so as to store the call memories of the target customer; The recall module 203 is used to identify the current-round semantics of the real-time call of the target customer, construct joint query conditions based on the current-round semantics, and output the recall result of the call memory based on the joint query conditions; The output module 204 is used to output the call response result based on the recall result and the dialogue generation model, according to the preset memory injection mechanism.

[0076] As an optional implementation, acquiring the target customer's historical call data and transcribing it into call text includes: The historical call input speaker differentiation model composes the historical call data into call text, generates role tags for the speakers in the call text, and records the call sequence in the call text using timestamps.

[0077] As an optional implementation, semantic analysis of the call text is performed to obtain key information corresponding to the billing communication, including: The call text is input into a multi-task joint learning semantic analysis model for semantic analysis and rule verification to obtain the key information of named entities and triples; The named entities are identified and extracted through a pre-trained encoding and decoding layer, and the triples are extracted from the single sentences of the call text and the dialogue context to obtain the intent-slot-value triples. The rule validation is performed by pre-setting the named entities and meanings. Figure Three The normalized domain rules of the tuples are used to verify the consistency between the obtained key information and the call text, and to output structured key information.

[0078] As an optional implementation, a call memory network is constructed, including: A unique identifier is generated based on the record ID of the historical call and the hash value of the corresponding timestamp of the historical call; Extract a set of key-value pairs of structured facts from the intent-slot-value pairs in the key information triples, and generate fact slots; Based on the fields in the call text that support the structured facts, generate evidence fragments pointing to the timestamp range and text content corresponding to the evidence of that field; The fact slots and evidence fragments are input into a preset summary generation model, and a business summary of the historical call data is output. The time attribute is determined based on the memory construction time, the time of the fact occurrence, and the memory update time; Metadata is generated by extracting confidence scores based on extracted triples from key information, and obtaining customer / agent information sources and related accounting communication task / case IDs based on named entities. The call memory unit for a single historical call corresponding to the target customer is constructed using the unique identifier, fact slot, evidence fragment, business summary, time attribute, and metadata. For each call memory unit, generate a relation pointer based on the logical relationship between any two call memory units; The call memory network of the target customer is constructed using all call memory units and the pointers to the relationships between them.

[0079] As an optional implementation, a call event index is constructed to store the call memories of the target customer, including: The text representation of the service summary and fact slot in each of the call memory units is encoded into a high-dimensional dense vector, and the high-dimensional dense vector is used as the first index; Generate a range-filtered and exact-matching inverted index for the key and value corresponding to the fact slot in each of the call memory units, and use the inverted index as the second index; Using the event occurrence time corresponding to the time attribute in each of the call memory units as the primary key, a time-series index supporting time window retrieval is constructed, and the time-series index is used as the third index. By using call memory units as nodes and logical relationship types of relation pointers as edges, a memory graph is constructed from all the relation pointers. A graph index for traversing and querying the memory graph is then constructed, and the graph index is used as the fourth index. Each call memory unit in the call memory network is written into the first index, the second index, the third index, and the fourth index for storage.

[0080] As an optional implementation, identifying the current-round semantics of the target customer's real-time call includes: The system acquires the real-time call content of the target customer, identifies the target customer's intent and slot information in the current call based on the semantic analysis model, and generates call status text containing the current call semantics based on the dialogue status tracking module. The dialogue status tracking module is used to maintain the current call status and verify the consistency of the current call semantics in consecutive multiple calls by fusing the intent and slot information.

[0081] As an optional implementation, a joint query condition is constructed based on the current round semantics, and the recall result of the call memory is output based on the joint query condition, including: The current round semantics and call status text are input into a pre-trained embedding model to generate a query semantic vector; Extract the filtering fields of intent and slot information from the semantics of the current round, and generate structured filtering conditions; Based on the business rules of accounting communication and the context of the call status text, determine the time range of the query conditions to set the time query window; Based on the causal queries and / or process queries involving consecutive multi-turn calls in the current turn semantics, generate graph query statements with graph traversal conditions; Based on the query semantic vector, structured filtering conditions, time query window, and graph query statement, the call memory network is executed in parallel for multi-path recall to obtain an initial candidate set; For each call memory unit in the initial candidate set, multiple feature scores are calculated. Based on the final score obtained by fusing multiple feature scores, all call memory units are sorted to obtain the recall result. The feature scores include semantic similarity score, time decay score, confidence score, and task relevance score.

[0082] As an optional implementation, based on a preset memory injection mechanism, and according to the recall result and the dialogue generation model, the call response result is output, including: Define a time decay curve for each type of call memory unit, scan the call memory network according to the time decay curve to determine the validity of each call memory unit, and adjust the sorting weight or perform secondary verification on the call memory units in the recall results according to the reference value of the validity. The call memory network detects first and second memory units that have some identical memories, compares the confidence scores and evidence fragment completeness of the two memory units, and performs a resolution process on the identified conflicting memory units. The conflicting memory units and the call memory units have at least one difference in unique identifier, fact slot, evidence fragment, business summary, time attribute or metadata. The call memory units in the recall results that are greater than the sorting threshold are formatted into fact cards according to a preset card template. The card template is a natural language statement of the recorded facts in the call memory unit. The identity verification status of the target customer is obtained to determine the anonymization level. Based on the anonymization level, the sensitive fields of the fact card are filtered and injected into the dialogue generation model. The filtered fact card and the current call status text are used to generate prompt words and input into the dialogue generation model to output the call response result.

[0083] Example 3 Please see Figure 3 , Figure 3 This is another real-time call recall system based on historical call memory disclosed in the embodiments of the present invention. Figure 3 The described real-time call recall system based on historical call memory is applied in a data processing system / data processing equipment / data processing server (wherein, the server includes a local processing server or a cloud processing server). For example... Figure 3 As shown, the real-time call recall system based on historical call memory may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the real-time call recall method based on historical call memory described in Embodiment 1.

[0084] Example 4 This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps of the real-time call recall method based on historical call memory described in Embodiment 1.

[0085] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the real-time call recall method based on historical call memory described in Embodiment 1.

[0086] The foregoing has described specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily have to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0087] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0088] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0089] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure One One or more processes and / or boxes Figure One A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure One One or more processes and / or boxes Figure One The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure One One or more processes and / or boxes Figure One The steps of the function specified in one or more boxes.

[0093] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0094] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0095] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0096] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0097] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0098] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0099] Finally, it should be noted that the real-time call recall method and system based on historical call memory disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time call recall method based on historical call memory, characterized in that, The method includes: Obtain the target customer's historical calls and transcribe them into call text. Perform semantic analysis on the call text to obtain key information related to billing communication. Based on the aforementioned key information, a call memory network and a call event index are constructed to store the call memories of target customers; Identify the current-round semantics of the target customer's real-time call, construct joint query conditions based on the current-round semantics, and output the recall result of the call memory based on the joint query conditions; Based on the preset memory injection mechanism, the call response result is output according to the recall result and the dialogue generation model.

2. The real-time call recall method based on historical call memory according to claim 1, characterized in that, Obtain historical call data of the target customer and transcribe it into call text, including: The historical call input speaker differentiation model composes the historical call data into call text, generates role tags for the speakers in the call text, and records the call sequence in the call text using timestamps.

3. The real-time call recall method based on historical call memory according to claim 1, characterized in that, Semantic analysis of the call text yielded key information related to the billing communication, including: The call text is input into a multi-task joint learning semantic analysis model for semantic analysis and rule verification to obtain the key information of named entities and triples; The named entities are identified and extracted through a pre-trained encoding and decoding layer, and the triples are extracted from the single sentences of the call text and the dialogue context to obtain the intent-slot-value triples. The rule verification uses preset normalized domain rules for the named entity and intent triples to verify the consistency between the obtained key information and the call text, and outputs structured key information.

4. The real-time call recall method based on historical call memory according to claim 1, characterized in that, Constructing a call memory network includes: A unique identifier is generated based on the record ID of the historical call and the hash value of the corresponding timestamp of the historical call; Extract a set of key-value pairs of structured facts from the intent-slot-value pairs in the key information triples, and generate fact slots; Based on the fields in the call text that support the structured facts, generate evidence fragments pointing to the timestamp range and text content corresponding to the evidence of that field; The fact slots and evidence fragments are input into a preset summary generation model, and a business summary of the historical call data is output. The time attribute is determined based on the memory construction time, the time of the fact occurrence, and the memory update time; Metadata is generated by extracting confidence scores based on extracted triples from key information, and obtaining customer / agent information sources and related accounting communication task / case IDs based on named entities. The call memory unit for a single historical call corresponding to the target customer is constructed using the unique identifier, fact slot, evidence fragment, business summary, time attribute, and metadata. For each call memory unit, generate a relation pointer based on the logical relationship between any two call memory units; The call memory network of the target customer is constructed using all call memory units and the pointers to the relationships between them.

5. The real-time call recall method based on historical call memory according to claim 4, characterized in that, Build a call event index to store the call memories of the target customer, including: The text representation of the service summary and fact slot in each of the call memory units is encoded into a high-dimensional dense vector, and the high-dimensional dense vector is used as the first index; Generate a range-filtered and exact-matching inverted index for the key and value corresponding to the fact slot in each of the call memory units, and use the inverted index as the second index; Using the event occurrence time corresponding to the time attribute in each of the call memory units as the primary key, a time-series index supporting time window retrieval is constructed, and the time-series index is used as the third index. By using call memory units as nodes and logical relationship types of relation pointers as edges, a memory graph is constructed from all the relation pointers. A graph index for traversing and querying the memory graph is then constructed, and the graph index is used as the fourth index. Each call memory unit in the call memory network is written into the first index, the second index, the third index, and the fourth index for storage.

6. The real-time call recall method based on historical call memory according to claim 3, characterized in that, Identify the current-round semantics of a target customer's real-time call, including: The system acquires the real-time call content of the target customer, identifies the target customer's intent and slot information in the current call based on the semantic analysis model, and generates call status text containing the current call semantics based on the dialogue status tracking module. The dialogue status tracking module is used to maintain the current call status and verify the consistency of the current call semantics in consecutive multiple calls by fusing the intent and slot information.

7. The real-time call recall method based on historical call memory according to claim 1, characterized in that, Based on the current round semantics, construct joint query conditions, and output the call memory recall results based on the joint query conditions, including: The current round semantics and call status text are input into a pre-trained embedding model to generate a query semantic vector; Extract the filtering fields of intent and slot information from the semantics of the current round, and generate structured filtering conditions; Based on the business rules of accounting communication and the context of the call status text, determine the time range of the query conditions to set the time query window; Based on the causal queries and / or process queries involving consecutive multi-turn calls in the current turn semantics, generate graph query statements with graph traversal conditions; Based on the query semantic vector, structured filtering conditions, time query window, and graph query statement, the call memory network is executed in parallel for multi-path recall to obtain an initial candidate set; For each call memory unit in the initial candidate set, multiple feature scores are calculated. Based on the final score obtained by fusing multiple feature scores, all call memory units are sorted to obtain the recall result. The feature scores include semantic similarity score, time decay score, confidence score, and task relevance score.

8. The real-time call recall method based on historical call memory according to claim 1, characterized in that, Based on a preset memory injection mechanism, and according to the recall results and the dialogue generation model, the call response results are output, including: Define a time decay curve for each type of call memory unit, scan the call memory network according to the time decay curve to determine the validity of each call memory unit, and adjust the sorting weight or perform secondary verification on the call memory units in the recall results according to the reference value of the validity. The call memory network detects first and second memory units that have some identical memories, compares the confidence scores and evidence fragment completeness of the two memory units, and performs a resolution process on the identified conflicting memory units. The conflicting memory units and the call memory units have at least one difference in unique identifier, fact slot, evidence fragment, business summary, time attribute or metadata. The call memory units in the recall results that are greater than the sorting threshold are formatted into fact cards according to a preset card template. The card template is a natural language statement of the recorded facts in the call memory unit. The identity verification status of the target customer is obtained to determine the anonymization level. Based on the anonymization level, the sensitive fields of the fact card are filtered and injected into the dialogue generation model. The filtered fact card and the current call status text are used to generate prompt words and input into the dialogue generation model to output the call response result.

9. A real-time call recall system based on historical call memory, characterized in that, The system includes: The acquisition module is used to acquire the target customer's historical calls and transcribe them into call text, and perform semantic analysis on the call text to obtain key information corresponding to the billing communication. The memory module is used to construct a call memory network and a call event index based on the key information, so as to store the call memories of the target customer; The recall module is used to identify the current-round semantics of the real-time call of the target customer, construct joint query conditions based on the current-round semantics, and output the recall result of the call memory based on the joint query conditions; The output module is used to output the call response result based on the recall result and the dialogue generation model, according to the preset memory injection mechanism.

10. A real-time call recall system based on historical call memory, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the real-time call recall method based on historical call memory as described in any one of claims 1-8.

Images