Customer service assistant-based long short-term memory collaborative retrieval and hierarchical storage method

Abstract

Embodiments of the present disclosure disclose a long short-term memory collaborative retrieval and hierarchical storage method based on a customer service assistant, a specific implementation of the method comprising: determining user identification information and query intention feature information; performing matching retrieval on a target user short-term temporary memory library to obtain a matching short-term memory information set; performing long short-term intention completion on the query intention feature information to obtain intention completion information; performing recall on a customer service reply business knowledge base to obtain a query reply information set; performing relevance sorting processing on the query reply information set to obtain a query reply sorting information sequence; performing structured division on current session multi-round question and answer text information to obtain a short-term memory structured information set; and performing short-term memory incremental weight hierarchical storage updating on a user long-term memory storage library. The implementation can realize cross-session communication, improve the accuracy of reply questions and memory retrieval, reduce the retrieval time length, improve the user experience, and reduce the waste of storage resources.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to a long short-term memory collaborative retrieval and hierarchical storage method based on a customer service assistant. Background Technology

[0002] In customer service systems based on customer service assistants (e.g., insurance product applications), the introduction of large language models can assist customer service personnel in responding, improving efficiency and accuracy. However, customer service assistants cannot remember cross-conversation information, and context breaks occur when users switch topics or the conversation is interrupted. The system cannot automatically reconstruct the response, requiring users to repeatedly enter previously asked questions; otherwise, the accuracy of the response is low, necessitating significant effort from customer service personnel, increasing their workload, and reducing user experience. For customer service responses to queries from customer service assistants, the common approach is to directly use the multi-turn question-and-answer text information of the current conversation as short-term memory information for the current query text, serving as context information. Alternatively, the system can retrieve the corresponding long-term memory information set from the statically stored user long-term memory set as context information. Then, based on the current query text information and context information, the customer service response knowledge base is retrieved and reordered to obtain the response information, which is then sent to the response front-end page for customer service personnel to adjust. The current query information and response information are also stored.

[0003] However, in practice, it has been found that when using the above method for collaborative retrieval and hierarchical storage of long and short-term memory based on customer service assistants, the following technical problems often arise: storing the text information of multiple rounds of Q&A in the current session as short-term memory information in its entirety results in wasted memory resources due to the presence of redundant and useless information in the text information of multiple rounds of Q&A in the current session. Treating the entire text information as short-term memory will cause context overflow for the customer service assistant, making it difficult for the customer service assistant to accurately determine key information. Furthermore, using only static long-term memory as context will cause intent drift, leading to irrelevant answers and lower accuracy of retrieved context information. This results in reduced efficiency and accuracy of retrieving customer service response business knowledge base, increased retrieval time, and a decreased user experience.

[0004] The information disclosed in this background section is only intended to enhance the understanding of the background of the present disclosure concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure propose a long short-term memory collaborative retrieval and hierarchical storage method based on a customer service assistant to solve one or more of the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide a method for collaborative retrieval and hierarchical storage of long short-term memory based on a customer service assistant, comprising: in response to a smart reply application based on a customer service assistant receiving current query text information and current session multi-turn question-and-answer text information sent by a target user, determining user identification information and query intent feature information for the aforementioned current query text information; performing a matching retrieval on the target user's short-term temporary memory bank based on the aforementioned user identification information and the aforementioned query intent feature information to obtain a matching short-term memory information set; performing long and short-term intent completion processing on the aforementioned query intent feature information based on the aforementioned matching short-term memory information set and the user's long-term memory bank to obtain intent completion information, and sending the aforementioned intent completion information to... The aforementioned intelligent response application's response front-end page, in response to receiving a selection operation for the aforementioned intent completion information, retrieves the customer service response business knowledge base based on the query intent adjustment information corresponding to the selection operation and the aforementioned user long-term memory repository, obtaining a query response information set; performs relevance sorting processing on the aforementioned query response information set based on the aforementioned query intent adjustment information, obtaining a query response sorting information sequence; sends the aforementioned query response sorting information sequence to the aforementioned response front-end page, and performs structured division of the aforementioned multi-turn question and answer text information of the current conversation, obtaining a short-term memory structured information set; and performs short-term memory incremental weight hierarchical storage update on the aforementioned user long-term memory repository based on the aforementioned short-term memory structured information set.

[0008] Secondly, some embodiments of this disclosure provide a long short-term memory collaborative retrieval and hierarchical storage device based on a customer service assistant, comprising: a determining unit configured to, in response to a smart reply application based on a customer service assistant receiving current query text information and current session multi-turn question-and-answer text information sent by a target user, determine user identification information and query intent feature information for the aforementioned current query text information; a matching retrieval unit configured to, based on the aforementioned user identification information and the aforementioned query intent feature information, perform a matching retrieval on the target user's short-term temporary memory bank to obtain a matching short-term memory information set; and an intent completion unit configured to, based on the aforementioned matching short-term memory information set and the user's long-term memory bank, perform long and short-term intent completion processing on the aforementioned query intent feature information to obtain intent completion information, and send the aforementioned intent completion information to the aforementioned smart reply application. The application includes a response front-end page; a recall unit configured to respond to a selection operation for the aforementioned intent completion information, and to recall the customer service response business knowledge base based on the query intent adjustment information corresponding to the selection operation and the aforementioned user long-term memory repository, thereby obtaining a query response information set; a relevance ranking unit configured to perform relevance ranking processing on the aforementioned query response information set based on the aforementioned query intent adjustment information, thereby obtaining a query response ranking information sequence; a structured partitioning unit configured to send the aforementioned query response ranking information sequence to the aforementioned response front-end page, and to perform structured partitioning on the aforementioned multi-turn question and answer text information of the current session, thereby obtaining a short-term memory structured information set; and a hierarchical storage unit configured to perform short-term memory incremental weight hierarchical storage update on the aforementioned user long-term memory repository based on the aforementioned short-term memory structured information set.

[0009] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, such that when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0010] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described in any implementation of the first aspect.

[0011] The above embodiments of this disclosure have the following beneficial effects: The long short-term memory collaborative retrieval and hierarchical storage method based on customer service assistants in some embodiments of this disclosure can realize cross-session communication, improve the accuracy of answering questions and memory retrieval, reduce retrieval time, improve user experience, and reduce the waste of memory storage resources. Specifically, the reasons for the related reduction in the efficiency and accuracy of retrieving customer service answer business knowledge bases, the increase in retrieval time, and the reduction in user experience are as follows: storing the multi-turn question and answer text information of the current session as short-term memory information in its entirety, since there is redundant and useless information in the multi-turn question and answer text information of the current session, wasting memory resources, and storing the whole as short-term memory will cause context overflow for the customer service assistant, making it difficult for the customer service assistant to accurately determine key information, and relying solely on static long-term memory as context will cause intent drift, resulting in irrelevant answers, reducing the accuracy of retrieved context information, thus reducing the efficiency and accuracy of retrieving customer service answer business knowledge bases, increasing retrieval time, and reducing user experience. Based on this, some embodiments of the long short-term memory collaborative retrieval and hierarchical storage method based on a customer service assistant can first, in response to the intelligent reply application based on the customer service assistant receiving the current query text information and the current session multi-turn question and answer text information sent by the target user, determine the user identification information and query intent feature information for the aforementioned current query text information. Here, the user identification information and query intent feature information facilitate subsequent short-term memory and long-term memory retrieval. Second, based on the aforementioned user identification information and the aforementioned query intent feature information, a matching retrieval is performed on the target user's short-term temporary memory bank to obtain a matching short-term memory information set. Here, short-term memory matching retrieval can achieve cross-session, reduce the occurrence of context breakage problems, and achieve smooth and natural dialogue. Third, based on the aforementioned matching short-term memory information set and the user's long-term memory bank, long and short-term intent completion processing is performed on the aforementioned query intent feature information to obtain intent completion information, and the aforementioned intent completion information is sent to the reply front-end page of the aforementioned intelligent reply application. Here, intent completion through short-term memory and long-term memory can improve the accuracy and comprehensiveness of intent completion information, and long-term memory can extract user preference information to achieve personalized and accurate services, reducing the occurrence of intent deviation and irrelevant answers. Next, in response to the selection operation received regarding the aforementioned intent completion information, the customer service response knowledge base is retrieved based on the query intent adjustment information corresponding to the selection operation and the aforementioned user long-term memory repository, resulting in a query response information set. Here, by manually adjusting the intent completion information again, human collaboration can further reduce the possibility of model illusions by the customer service assistant, improve the accuracy of the intent adjustment information, enhance the accuracy and efficiency of the query response information set, and reduce retrieval time. Subsequently, based on the aforementioned query intent adjustment information, the query response information set is sorted by relevance to obtain a query response ranking sequence.Here, relevance ranking enables the rearrangement and refinement of the query-response information set, improving the semantic relevance, scenario adaptability, and timeliness of the query-response ranking information sequence and the information adjusted based on query intent. This achieves precise responses based on user personalized preferences, enhancing the user experience. Then, the aforementioned query-response ranking information sequence is sent to the aforementioned response front-end page, and the text information of the current session's multi-turn question-and-answer sessions is structurally partitioned to obtain a short-term memory structured information set. This structural partitioning removes redundant and invalid information from the current session's multi-turn question-and-answer text information, improving the quality of the short-term memory structured information set and reducing the memory cost consumed by the subsequent customer service assistant context, thus improving the customer service assistant's reasoning efficiency. Finally, based on the aforementioned short-term memory structured information set, the aforementioned user long-term memory repository is updated with incremental weighted hierarchical storage. This reduces storage resource waste, and hierarchical storage updates improve storage efficiency through incremental information storage. Hierarchical storage effectively distinguishes different types of memory information, enabling precise processing of memory information. Therefore, this customer service assistant-based long and short term memory collaborative retrieval and hierarchical storage method can achieve smooth cross-conversation dialogue and reduce the occurrence of context breakage problems through the collaboration of short-term and long-term memory, avoid irrelevant answers, and improve response efficiency without requiring users to repeat previous content. It can also achieve personalized responses, improve the efficiency and accuracy of retrieving customer service answer business knowledge base, improve the accuracy of answering questions, reduce retrieval time, and improve user experience. The structured division can remove redundant and invalid information, reduce the waste of memory storage resources, and ensure accurate processing of different memories. Attached Figure Description

[0012] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0013] Figure 1 This is a flowchart of some embodiments of the long-term memory collaborative retrieval and multi-level storage method according to the present disclosure; Figure 2 This is a schematic diagram of the structure of some embodiments of the long-term memory collaborative retrieval and multi-level storage device according to the present disclosure; Figure 3 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0016] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0017] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0019] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Figure 1 A flow 100 is shown illustrating some embodiments of the customer service assistant-based long short-term memory collaborative retrieval and hierarchical storage method according to this disclosure. This customer service assistant-based long short-term memory collaborative retrieval and hierarchical storage method includes the following steps: Step 101: In response to the intelligent reply application based on the customer service assistant receiving the current query text information and the current session multi-turn question and answer text information sent by the target user, determine the user identification information and query intent feature information for the current query text information.

[0021] In some embodiments, the executing entity (e.g., an electronic device) of the above-described customer service assistant-based long short-term memory collaborative retrieval and hierarchical storage method can, in response to the intelligent response application based on the customer service assistant receiving the current query text information and the current session multi-turn question-and-answer text information sent by the target user, determine the user identification information and query intent feature information for the aforementioned current query text information. The aforementioned customer service assistant can be an intelligent agent used to assist customer service personnel in perceiving the target user's query intent, completing the information input by the target user, and outputting answers to questions to assist customer service personnel. The aforementioned intelligent response application can be an application system that serves the target user in querying information. For example, the aforementioned intelligent response application can be a customer service system for insurance products. The aforementioned target user can be a querying user, i.e., a user who inputs a question and waits for a response from the intelligent response application. The aforementioned current query text information can be a question sent by the target user, waiting for a response from customer service personnel with the assistance of the customer service assistant. The aforementioned current session multi-turn question-and-answer text information can be text information from multiple rounds of question-and-answer interactions between the target user and the aforementioned customer service assistant within the same session as the aforementioned current query text information. For example, the multi-turn Q&A text information in the current conversation could be: "Q: How do I claim insurance if I have a critical illness? A: Hello, if you unfortunately suffer from a critical illness and need to claim, the rules vary slightly depending on the product (critical illness insurance pays the insured amount directly upon diagnosis if the conditions are met, while million-dollar medical insurance reimburses treatment costs). The general process and precautions are as follows… Q: How long is the coverage period (claim period) for this children's accident insurance? A: Most children's accident insurance on the market is a short-term policy with a one-year term, meaning you pay for one year and are covered for one year; there are no lifelong policies. A few term policies may cover up to age 18 / 25, but the risk model changes after adulthood, so it's generally recommended to separate coverage for children during their childhood." The answer is yes, you can switch to adult accident insurance after reaching adulthood. Q: How do I apply? A: Applying for a claim for children's accident insurance is actually simpler than applying for critical illness insurance. The process has four steps: reporting the incident, preparing materials, submitting the claim, and waiting for payment… The aforementioned user identification information can be information used to uniquely identify a user. For example, the aforementioned user identification information can include, but is not limited to, at least one of the following: user's mobile phone number, user's ID card information, user's login account information. The aforementioned query intent characteristic information can be information representing the true intent or core objective of the target user's query, i.e., the semantic information of the current query text information. For example, the aforementioned query intent characteristic information could be "claim application".

[0022] In some optional implementations of certain embodiments, determining the user identification information and query intent feature information for the currently queried text information may include the following steps: The first step, in response to detecting that the target user has successfully logged into the aforementioned intelligent reply application, is to obtain the user identification information for the currently queried text information. This can be achieved by directly extracting the user account information from the login process.

[0023] The second step involves, in response to the detection that the target user is not logged into the aforementioned smart reply application, sending a login reminder message and triggering a login front-end page to prompt the user to log in. The login reminder message can be information used to remind the target user to log in. The login front-end page can be the front-end page on which the user logs into the aforementioned smart reply application.

[0024] The third step involves obtaining user identification information in response to the detected input action of the target user on the login front-end page. This input action could be the target user entering their account password. The information obtained could be the target user's login account information extracted from the input action.

[0025] The fourth step is to perform character correction processing on the current query text information to obtain the corrected query text information. This corrected query text information can be the text information after performing dirty data cleaning, standardizing text formatting, and eliminating interference from symbols, capitalization, and typos.

[0026] The fifth step involves extracting multiple types of keywords from the revised query text information to obtain a query keyword set. The query keywords in this set can be word segments representing the user's query intent. In practice, the executing entity can first use a pre-defined business thesaurus to segment the revised query text information, obtaining a query word segmentation set. This pre-defined business thesaurus can be a database of insurance-related professional terms. For example, it can be the "Industry-Specific Business Terminology Standard." Secondly, stop words are removed from the query word segmentation set to obtain the target query word segmentation set. Stop word removal can be performed by identifying and deleting stop words using a general stop word list and a business-specific stop word list. For example, the general stop word list could be the Harbin Institute of Technology stop word list. The business-specific stop word list can be stop words summarized from historical query information about insurance products. For example, the business-specific stop word list can include, but is not limited to, at least one of the following: help, understand, trouble, ask, want to ask. Then, the target query word segmentation set is segmented and categorized to obtain a target query word segmentation set with word segmentation category annotations, which serves as the query annotation word segmentation set. The word segmentation category annotations can distinguish between nouns (e.g., subject, object), action predicates (e.g., behavior, service), and status words. Status words can include, but are not limited to, at least one of the following: accident, invalid, overdue. Finally, keywords are extracted from the above query annotation word segmentation set to obtain a query keyword set. The keyword extraction can be the extraction of action predicates.

[0027] Step 6: Based on the aforementioned set of query keywords, perform intent mapping and recognition on the revised query text information to obtain query intent feature information. In practice, the executing entity can perform the following intent mapping steps for each query keyword: First, determine the noun groups and state phrases related to the aforementioned query keyword in the revised query text information. Then, in response to determining that neither the aforementioned noun groups nor the aforementioned state phrases are empty, perform text combination on the aforementioned query keyword, the aforementioned noun groups, and the aforementioned state phrases to obtain query intent feature information. Finally, in response to determining that both the aforementioned noun groups and the aforementioned state phrases are empty sets, determine the aforementioned query keyword as query intent feature information.

[0028] Step 102: Based on the user identification information and query intent feature information, perform matching retrieval on the target user's short-term temporary memory database to obtain a matching short-term memory information set.

[0029] In some embodiments, the execution entity can perform matching retrieval on the target user's short-term temporary memory based on the user identification information and the query intent feature information to obtain a matching short-term memory information set. The target user's short-term temporary memory can be a database storing the short-term memory context formed by the multi-turn query text information generated in the previous session for all users. The short-term context can be obtained through the following steps: First, invalid content is removed from the multi-turn query text information generated in the previous session to obtain the previous multi-turn query text information. The invalid content removal can be an operation of removing content without substantial information from the multi-turn query text information generated in the previous session. For example, the content without substantial information may include, but is not limited to, at least one of the following: stop words and interjections. Second, the previous multi-turn query text information is split into dialogues to obtain a turn query text information set. The dialogue splitting can be done by dividing the text information corresponding to the target user's query question and the response text information into independent dialogue units, and adding dialogue time and dialogue turn sequence numbers. Then, the turn query text information set is decomposed into structured fields and stored in the database to obtain the target user's short-term temporary memory. The structured field decomposition described above can be a process of breaking down the query text information for each round into intent fields, slot fields, key entity fields, user preference fields, and interaction node fields. The intent fields can be represented by tags of no more than 10 characters, indicating the user's true intent in the query text information for each round (e.g., life insurance application consultation, claim application, policy inquiry). Fields without a clear intent are marked as "none." The slot fields can be key-value pairs of predefined key fields and corresponding attribute information extracted from the query text information for each round. For example, a slot field could be "Policy Number: P12345678" or "Date of Incident: 2026-04-10." The key entity fields can be entity information extracted from the query text information for each round, and must not overlap with the slot fields. For example, a key entity field could be "Insurance Product: Children's Whole Life Insurance" or "Service Name: Claims Handling Service." The aforementioned user preference field can be a key-value pair format used to record the target user's preference information regarding insurance products in the round-query text information. If there is no preference, it is marked as "None". The aforementioned user preference field could be "User preference: Prefers participating life insurance, inclined towards online claims processing". The aforementioned interaction node field can be a concise statement of no more than 20 characters marking the interaction results in the round-query text information. For example, the aforementioned interaction node field could be the field "User inquires about life insurance premiums, customer service informs of the calculation results".The matching short-term memory information in the aforementioned matching short-term memory information set can be memory information stored in the aforementioned target user's short-term temporary memory bank that has the same query intent as any query intent feature information in the query intent feature information under the same user identification information. The aforementioned matching short-term memory information can be a temporary session cache that includes at least a structured JSON formatted intent field, slot field, key entity field, user preference field, and interaction node field, used by the customer service assistant to maintain the continuity of a single conversation.

[0030] In some optional implementations of certain embodiments, the process of matching and retrieving the target user's short-term temporary memory database based on the user identification information and the query intent feature information to obtain a matching short-term memory information set may include the following steps: The first step is to determine, based on the aforementioned target user short-term temporary memory (STM) database, a set of STM information related to the aforementioned user identification information, as the target STM information set. The target STM information in this set can be information from the STM corresponding to the aforementioned user identification information. In practice, the executing entity can use the aforementioned user identification information as a query condition to filter out STM information related to the aforementioned user identification information from the target user STM database, thus forming the target STM information set.

[0031] The second step involves performing short-term memory matching on the target short-term temporary memory information set based on the aforementioned query intent feature information and the corresponding query keyword set, thereby obtaining the short-term memory information set. The short-term memory information in this set can be the intent of the target short-term temporary memory information, memory information identical to the aforementioned query intent feature information, and where the intersection of the keyword set included in the target short-term temporary memory information and the query keyword set corresponding to the aforementioned query intent feature information is not empty. In practice, the executing entity can utilize Boolean query expressions and vector cosine similarity to perform multi-path recall short-term memory matching on the target short-term temporary memory information set based on the aforementioned query intent feature information and the corresponding query keyword set, thereby obtaining the short-term memory information set.

[0032] The third step involves recalling and filtering the aforementioned short-term memory information set to obtain a filtered short-term memory information set. In practice, the executing entity can filter out short-term memory information sets from the aforementioned short-term memory information set whose Jacarbe similarity to the aforementioned query keyword set is greater than or equal to a preset similarity threshold, and use this as the filtered short-term memory information set. The preset similarity threshold can be a pre-defined critical value for Jacarbe similarity. For example, the preset similarity threshold can be 0.5.

[0033] The fourth step is to determine the memory dimension relevance between each piece of filtered short-term memory information and the query intent feature information in the filtered short-term memory information set, thus obtaining a memory dimension relevance set. This memory dimension relevance includes: intent relevance, keyword Jaccard coefficient, business scenario relevance, and time decay factor relevance. Specifically, the memory dimension relevance can be the degree of correlation between the filtered short-term memory information and the query intent feature information. The intent relevance can be the degree of relevance between the filtered short-term memory information and the query intent feature information in terms of intent. For example, the intent relevance can be obtained by inputting the filtered short-term memory information and the query intent feature information into a cross-encoder to obtain a matching probability value. The keyword Jaccard coefficient can be the Jaccard coefficient between the keywords included in the filtered short-term memory information and the corresponding query keywords in the query intent feature information. The business scenario relevance can be the relevance information used to determine whether the filtered short-term memory information and the query intent feature information belong to the same insurance business scenario. The business scenario relevance can be obtained through the following steps: First, determine the product type or business category in the slot fields included in the filtered short-term memory information as the filtering scenario information. Then, determine the corresponding scenario information for the above-mentioned filtered scenario information and the above-mentioned query intent feature information. If they are the same, the business scenario relevance is 1; otherwise, it is 0. The above-mentioned time decay factor relevance can be the time decay relevance value of the filtered short-term memory information determined by the exponential decay formula. The above-mentioned memory dimension relevance can be the value obtained by weighted summing of the above-mentioned intent relevance, keyword Jaccard coefficient, business scenario relevance, and time decay factor relevance, with the sum of the weight coefficients of the weighted sum being 1. The above-mentioned intent relevance can ensure that the context does not cross domains. The keyword Jaccard coefficient can measure the degree of literal overlap between the filtered short-term memory information and the above-mentioned query intent feature information. The business scenario relevance can forcibly isolate different business lines. The time decay factor relevance determines the sharing degree of subsequent completion through the time difference; the longer the time difference, the easier it is to become invalid.

[0034] Fifth, based on the aforementioned memory dimension relevance set, the selected short-term memory information set is rearranged to obtain a matching short-term memory information set. In practice, the executing entity can sort the selected short-term memory information from largest to smallest according to the aforementioned memory dimension relevance set, and extract the set of the first preset number of selected short-term memories. The preset number can be a pre-defined value. For example, the preset number can be any integer in [3, 5].

[0035] Step 103: Based on the matching short-term memory information set and the user's long-term memory repository, perform short-term and long-term intent completion processing on the query intent feature information to obtain intent completion information, and send the intent completion information to the response front-end page of the intelligent response application.

[0036] In some embodiments, the execution entity can perform short-term and long-term intent completion processing on the query intent feature information based on the matching short-term memory information set and the user's long-term memory repository to obtain intent completion information, and send the intent completion information to the response front-end page of the intelligent response application. The intent completion information can be information added to the query intent feature information with relevant matching short-term memory information set and determined long-term memory information, making the query intent clearer and more explicit. The response front-end page can be a front-end page in the intelligent response application that displays the interaction information between the target user and customer service. As an example, if the query intent feature information is "claim application," the retrieved matching short-term memory information is "children's accident insurance claim," and the determined long-term memory is "children's accident insurance, claim preference online," then the intent completion information can be "children's accident insurance claim application, claim preference online claim." The user's long-term memory repository can be a vector database used to store key information of the target user's cross-session memory corresponding to the user identification information, to achieve personalization and continuous learning for the target user. The aforementioned key information may include, but is not limited to, at least one of the following: long-term user preference information, key factual information in each session (i.e., entities and relationships), experiential information, and business knowledge. The aforementioned experiential information may include problem-solving information, optimization strategy information, and execution information from historical session sets.

[0037] In some optional implementations of certain embodiments, the above-mentioned long and short term intent completion processing of the query intent feature information based on the above-mentioned matching short-term memory information set and user long-term memory repository to obtain intent completion information may include the following steps: The first step involves reconstructing the query intent feature information using short-term reference resolution based on the aforementioned short-term memory information set, resulting in short-term completed intent information. This short-term completed intent information can be obtained by re-identifying intent information after resolving the pronouns in the current session's multi-turn question-and-answer text information corresponding to the query intent feature information using proximity-based reference resolution. In practice, the executing entity can first, in response to the determination that pronouns exist in the query intent feature information, utilize the proximity-based reference method in the reference resolution algorithm to sequentially select the historical turn dialogue information most recent in time to the current session's multi-turn question-and-answer text information, thereby performing short-term reference resolution on the question text information set corresponding to the query intent feature information, resulting in a reference question text information set. This question text information can be the text information of the question queried by the target user in the current session's multi-turn question-and-answer text information. The historical turn dialogue information can be the dialogue information from the most recent turn in time to the current session's multi-turn question-and-answer text information. Secondly, in response to determining that the aforementioned referential question text information set satisfies the topic interruption condition, at least one topic dialogue information set of referential question text that satisfies the topic interruption condition is determined. The topic interruption condition can be a condition where the target user has not sent a message for more than a preset interruption time threshold, or a condition where the intent set corresponding to the aforementioned referential question text information set changes. The preset interruption time threshold can be a pre-set maximum interruption time, which can be 5 minutes. The topic dialogue information sets in the aforementioned topic dialogue information set can be historical rounds of dialogue information corresponding to the same intent that satisfies the topic interruption condition, with the most recent preset number of rounds. The preset number of rounds can be a pre-set value, which can be any value in [3, 5]. Then, information extraction and reconstruction are performed on the aforementioned topic dialogue information sets to obtain short-term complete intent information. The information extraction can be the extraction of intent information, entity information, completed interaction information, and incomplete interaction information from each topic dialogue information set. The reconstruction can be the reconstruction of intent recognition after reordering the aforementioned intent information, entity information, completed interaction information, and incomplete interaction information in that order. Finally, in response to the determination that there are no pronouns in the above query intent feature information, the question text information set corresponding to the above query intent feature information is determined as the question text information set.

[0038] The second step involves performing long-term memory matching on the aforementioned short-term completion intent information within the user's long-term memory repository to obtain the target long-term memory information set. The target long-term memory information in this set can be long-term memory information stored in the user's long-term memory repository that is similar to the short-term completion intent information. In practice, the executing entity can use the short-term completion intent information as a search key and employ Nearest Neighbor Search (NNS) to perform a similarity search within the user's long-term memory repository to obtain the target long-term memory information set.

[0039] The third step is to determine that the target long-term memory information set is empty, and then determine the above short-term completion intention information as intention completion information.

[0040] The fourth step, in response to the determination that the target long-term memory information set is not empty, involves multi-dimensional information extraction from the target long-term memory information set to obtain a multi-dimensional long-term memory information set. This multi-dimensional long-term memory information set includes one of the following: historical intent information, historical item information, business scenario information, and user preference information. Specifically, the multi-dimensional long-term memory information can be obtained by statistically analyzing the historical dimension tags of the target long-term memory information. The historical intent information can be the query intent feature information corresponding to the target long-term memory information. The historical item information can be relevant information about insurance products that have appeared in the target long-term memory information (e.g., the name, type, insurance content, and services of the insurance product). The business scenario information can be the subdivided scenarios, triggering conditions, and classifications that constrain business occurrences, used to supplement scenario limitations beyond the product itself. For example, the business scenario information can include, but is not limited to, at least one of the following: vehicle accident scenario, premium overdue scenario, and first-time insurance scenario. The business scenario information can be a scenario determined by status words in the query keyword set corresponding to the query intent feature information. The user preference information can be personalized preference information such as the target user's preferences for insurance products and business services.

[0041] The fifth step involves performing information conflict correction processing on the aforementioned multi-dimensional long-term memory information set to obtain a corrected multi-dimensional long-term memory information set. Specifically, the corrected multi-dimensional long-term memory information in this set can be obtained by using a time decay model to correct the conflicting dimensions of the long-term memory information when information conflicts exist within the set.

[0042] Step 6: In response to determining that the aforementioned short-term completed intent information meets the intent scenario classification conditions, based on the aforementioned modified multi-dimensional long-term memory information set, the aforementioned short-term completed intent information is supplemented with memory dimensions to obtain intent completion information. The aforementioned intent scenario classification conditions can be that the intent of the aforementioned short-term completed intent information and the aforementioned modified multi-dimensional long-term memory information set are the same, but the intent of the short-term completed intent information is relatively ambiguous. The aforementioned relatively ambiguous intent can be that the slot fill rate of the short-term completed intent information is less than a preset fill rate threshold. The aforementioned preset fill rate threshold can be a pre-set value, such as 0.6. The slots corresponding to the aforementioned slot fill rate can include: intent information, item information, business scenario information, and user preference information. The aforementioned intent completion information can be obtained by supplementing the aforementioned short-term completed intent information with other dimension information under the same intent. For example, the aforementioned short-term completion intent information could be "claim application," and the revised multi-dimensional long-term memory information set could include: children's accident insurance claims, children's accident insurance, claim application scenarios, and multiple online claim processing preferences. Therefore, the intent completion information could be "online processing of children's accident insurance claims." In practice, the implementing entity can first determine the missing dimensional information set in the aforementioned short-term completion intent information as the short-term memory dimension missing information set. Then, it can filter out the dimensional information set corresponding to the aforementioned short-term memory dimension missing information set from the revised multi-dimensional long-term memory information set as the dimension information set to be completed. Finally, the aforementioned dimension information set to be completed is added to the aforementioned short-term completion intent information to obtain the intent completion information.

[0043] Step 7: In response to the determination that the aforementioned short-term completion intent information does not meet the aforementioned intent scenario classification conditions, the aforementioned short-term completion intent information is completed based on the aforementioned modified multi-dimensional long-term memory information set to obtain intent completion information. In practice, the aforementioned executing entity can use the aforementioned modified multi-dimensional long-term memory information set as the historical matching tag of the aforementioned short-term completion intent information without overwriting the aforementioned short-term completion intent information to obtain intent completion information. For example, the aforementioned short-term completion intent information could be "car insurance claim application," and the aforementioned modified multi-dimensional long-term memory information set could include: accident insurance claim application, accident insurance, claim application scenario. Then, the aforementioned intent completion information could be "car insurance claim application," and adding historical matching tags could be: historical intent information: accident insurance claim application, historical item information: accident insurance.

[0044] Step 104: In response to receiving a selection operation for intent completion information, adjust the information and user long-term memory repository according to the query intent corresponding to the selection operation, recall the customer service response business knowledge base, and obtain the query response information set.

[0045] In some embodiments, the aforementioned execution entity may, in response to receiving a selection operation for the aforementioned intent completion information, recall the customer service response business knowledge base based on the query intent adjustment information corresponding to the selection operation and the aforementioned user long-term memory repository, to obtain a query response information set. The aforementioned selection operation may be a modification, addition, deletion, or addition selection operation performed by customer service personnel on the intent completion information. The aforementioned query intent adjustment information may be information obtained after the customer service personnel perform a selection operation on the aforementioned intent completion information. The aforementioned user long-term memory repository may be a knowledge base used to persistently store information from multiple conversations between the target user corresponding to the user identification information in the aforementioned intelligent response application and the text information between them. The aforementioned persistently stored information may be used to record the user's unique preferences, habits, and identity information to provide personalized services to the target user and ensure the continuity and consistency of the relationship between the intelligent agent and the user. The aforementioned customer service response business knowledge base may be an unstructured database used to store insurance product-related information. For example, the aforementioned customer service response business knowledge base may be a MongoDB database. The aforementioned insurance product-related information may include, but is not limited to, at least one of the following: insurance product terms information, insurance product claims information, and insurance product dividend information. The query response information in the above query response information set can be the response information obtained by the customer service assistant based on the above query intent adjustment information and the large language model inference retrieved by the customer service assistant, which is used to answer the questions raised by the user.

[0046] In addressing the aforementioned technical challenges by employing technical solutions, the application scenario—a dual-index retrieval scenario for enterprise-level customer service response knowledge bases containing large-scale, heterogeneous data related to insurance products—often presents the following technical problems: Due to the long-tail distribution and biases within customer service response knowledge bases (e.g., a large amount of general content for insurance products while high-risk content is located in sparse peripheral areas), existing graph retrieval methods, starting from the central entry point, are prone to getting stuck in local optima for high-frequency content. Furthermore, existing graph retrieval methods typically involve full traversal, leading to long retrieval times, reduced query efficiency, and a poor user experience. Using only one retrieval method also results in lower accuracy of the output query response information, increasing the waste of GPU memory resources. Considering the following requirements for this application scenario: adaptability to large-scale customer service response knowledge bases, high dimensionality, high precision, dual-index retrieval, and multimodal query methods, we have decided to adopt the following solution: In some optional implementations of certain embodiments, the process of retrieving the customer service response business knowledge base based on the query intent adjustment information corresponding to the selection operation and the user long-term memory repository to obtain a query response information set may include the following steps: The first step is to construct a two-level index for the aforementioned customer service response business knowledge base, obtaining the response two-level index information. This response two-level index information can include intent index information and text inverted index information. The intent index information can be an index information where the intent tags obtained from the entry set included in the aforementioned customer service response business knowledge base through a large language model are used as keys, and the product document numbers of the product-related information sets corresponding to the entry set are used as values. The text inverted index information can be an inverted index information formed by using each word segment included in the aforementioned customer service response business knowledge base as keys, and the document frequency and term frequency corresponding to the word segment, as well as the document length and average document length corresponding to the word segment, as values. The document frequency can be the number of documents including the word segment. The term frequency can be the ratio of the number of times the word segment appears in the corresponding product document to the total number of words included in the product document. The document length can be the number of words included in the product document. The average document length can be the average of the document length set of the product document set.

[0047] The second step involves retrieving the customer service response business knowledge base based on the aforementioned two-layer index information and query intent adjustment information, thereby obtaining a first query response information sequence. The first query response information in this first query response information set can be response information retrieved from the customer service response business knowledge base that is related to the query intent adjustment information. In practice, the executing entity can first use the query intent adjustment information as the key to perform intent retrieval within the intent index information included in the aforementioned two-layer index information, obtaining a set of product documents identical to the aforementioned query intent adjustment information, which serves as a candidate product document set. Then, using the BM25 (Best Matching 25) algorithm, the document relevance values ​​of the candidate product document set are determined through the text inverted index information included in the aforementioned two-layer index information. The document relevance values ​​in this set can be obtained by inputting the term frequency, inverse document frequency, document length, and average document length of each candidate product document into the BM25 formula. Next, candidate documents with corresponding document relevance values ​​ranked from highest to lowest and above a preset document threshold are selected from the aforementioned candidate product document set and used as the target product document sequence. The preset document threshold can be a pre-defined maximum value among the selected candidate product documents, and can be 50. Finally, a set of product text information related to the aforementioned query intent adjustment information is extracted from the target product document sequence and used as the first query response information sequence.

[0048] The third step involves constructing a proximity graph on the aforementioned customer service response business knowledge base to obtain the response proximity graph. This response proximity graph can be constructed without altering the structure of the customer service response business knowledge base, resulting in a more efficient retrieval system. It organizes the business response data included in the knowledge base into nodes, using Euclidean distances between nodes as connecting edges. The nodes in the response proximity graph can be insurance product-related information for an insurance product, and the connecting edges can represent the Euclidean distance between any two nodes, pruned according to the triangle inequality pruning rules. The Euclidean distance can be calculated by embedding the business response data corresponding to the node into words, normalizing it using the L2 norm, and then calculating the Euclidean distance between nodes. In practice, the execution entity can utilize the NSG (Navigating Spreading-out Graph) algorithm to construct the proximity graph on the aforementioned customer service response business knowledge base to obtain the response proximity graph.

[0049] The fourth step involves extracting nodes from the aforementioned customer service response business knowledge base through hierarchical clustering of the response nearest neighbor graph, resulting in a cluster hub node set. The cluster hub nodes in this set can be the nodes closest to the cluster centers within the clusters obtained from the hierarchical clustering of the customer service response business knowledge base. In practice, the executing entity can first use the business response data included in the customer service response business knowledge base as the target clusters, and then use the K-Means algorithm to perform multiple rounds of hierarchical clustering on the target clusters, i.e., the customer service response business knowledge base, to obtain a customer service response cluster set. The aforementioned multi-round hierarchical clustering process can be as follows: In the first round, the K-Means algorithm is applied to the target clusters to obtain a first set of response clusters. Since the number of first response clusters in the first set is less than a preset number of clusters, a second round of clustering is performed. This involves applying the K-Means algorithm to each first response cluster to obtain a second set of response clusters. The number of second response clusters in the second set is then checked to ensure it is greater than or equal to the preset number of clusters. This process continues until the total number of clusters obtained after multiple rounds is greater than or equal to the preset number, at which point the hierarchical clustering process ends, resulting in a customer service response cluster set. The preset number of clusters can be a pre-defined maximum number of clusters; the specific value can be determined based on specific circumstances and will not be elaborated further. Next, the mean of the set of business response nodes included in each customer service response cluster is determined and normalized using the L2 norm, serving as the cluster center set. Finally, the customer service response node closest to the corresponding cluster center is selected from each customer service response cluster and used as the cluster hub node.

[0050] Fifth, based on the aforementioned clustering hub node set, perform walk sampling encoding on the aforementioned response nearest neighbor graph to obtain a walk subgraph feature vector set. The walk subgraph feature vectors in this set represent the feature vectors of the subgraph obtained by walk sampling the aforementioned response nearest neighbor graph. In practice, the execution entity can first perform the following walk steps for each clustering hub node: First, determine the next walk node for the clustering hub node; second, determine the product of the ratio of the minimum and maximum degrees of nodes in the response nearest neighbor graph and the degree of the next walk node, using this as the neighbor sampling number for the next walk node; third, sample customer service response nodes whose Euclidean distance to the next walk node is half the number of neighbor samples of the maximum and minimum distances, respectively, until the shortest path length from the sampled customer service response node to the aforementioned clustering hub node is less than or equal to the preset number of nodes, thus obtaining a walk local subgraph. The preset number of nodes can be a pre-defined maximum length of the shortest path; its specific value can be determined according to specific circumstances and will not be elaborated here. It should be noted that sampling the maximum and minimum neighbor values ​​by half can balance the exploration of dense and sparse nodes and prevent oversampling. Then, the Graph2Vec algorithm is used to perform graph topological encoding on the obtained multiple walking subgraphs to obtain the feature vector set of the walking subgraphs.

[0051] Step 6: Based on the aforementioned walking subgraph feature vector set, perform topological semantic encoding fusion on the aforementioned clustering hub node set to obtain a node latent feature vector set. The node latent feature vectors in this set can represent the subgraph topological information and query feature information of the aforementioned clustering hub nodes. In practice, the execution entity can first use an embedding model to text-encode the customer service response dataset included in the aforementioned clustering hub node set to obtain a node customer service response feature vector set. Secondly, input the node customer service response feature vector set and the aforementioned walking subgraph feature vector set into a multi-head attention mechanism layer to obtain a node fusion feature vector set. Thirdly, perform L2 norm normalization on the node fusion feature vector set to obtain a normalized fusion feature vector set. Subsequently, input the normalized fusion feature vector set into the node tower model included in the trained customer service recall dual-tower model to obtain a node latent feature vector set. The aforementioned customer service recall dual-tower model also includes a query tower model. The aforementioned customer service recall dual-tower model can be a model that, after representing the input normalized fused feature vector set and query feature vector through a node tower model and a query tower model respectively, outputs a node latent feature vector set and a query projection feature vector respectively, and calculates the cosine similarity between the two in a shared latent space, outputting the cosine similarity value. The aforementioned node tower model can be a model that performs latent space representation on the input normalized fused feature vector set to learn node representations that fuse topology and query features, outputting a node latent feature vector set. The aforementioned node tower model can be a model including a lightweight multilayer perceptron and a ReLU (Linear Rectification function) activation function. The aforementioned query tower model can be a model that performs latent representation on the input query feature vector to learn a query representation aligned with the node representation, outputting a query projection feature vector. The aforementioned query tower model can be a model including a lightweight multilayer perceptron and a ReLU (Linear Rectification function) activation function. The aforementioned customer service recall dual-tower model can be trained offline through the following steps: First, the target long-term memory information set is represented as a textual representation as a sample query vector set. Next, using the cluster hub node set as the entry point, a greedy search is performed on the sample query vector set corresponding to each node in the response nearest neighbor graph to obtain the shortest path hop count set from each sample hub node to the sample query vector set. Subsequently, at least one sample query vector corresponding to a shortest path hop count less than or equal to the minimum shortest path hop count and a preset hop count threshold is designated as the positive sample set, and at least one sample query vector exceeding the preset minimum hop count threshold is designated as the negative sample set. The preset hop count threshold can be a pre-set value, such as 3. The preset minimum hop count threshold can also be a pre-set value, such as 15.Finally, the positive and negative sample sets are input into the InfoNCE loss function to perform comparative learning training on the above customer service recall dual-tower model until the maximum preset number of rounds is reached, and the trained customer service recall dual-tower model is obtained.

[0052] Step 7: Based on the aforementioned latent feature vector set of nodes and the aforementioned response nearest neighbor graph, generate a query customer service response navigation graph for the aforementioned clustering hub node set. This query customer service response navigation graph can be a high-level nearest neighbor graph with adaptive topology and query awareness. The nodes included in the query customer service response navigation graph can be clustering hub nodes, and the connecting edges can be the cosine similarity values ​​between any two clustering hub nodes. In practice, the executing entity can first determine the cosine similarity between any two latent feature vectors of nodes included in the aforementioned latent feature vector set, obtaining a node similarity set. Then, for each node similarity set in the node similarity set, select the node similarity with the largest preset similarity value from the node similarity set to generate a target node similarity set. The preset similarity value can be a pre-set value, and its value can be determined according to specific circumstances, which will not be elaborated here. Finally, the clustered hub node set corresponding to the above target node similarity set is deduplicated to obtain a deduplicated hub node set. Then, connection edges based on node similarity are added to the above deduplicated hub node set to obtain the query customer service response navigation map.

[0053] Step 8: Based on the aforementioned query intent adjustment information, a greedy retrieval and recall process is performed on the aforementioned customer service response navigation map to obtain a second query response information sequence. The second query response information in the aforementioned second query response information set can be response information recalled from the aforementioned customer service response business knowledge base. In practice, the executing entity can first encode the aforementioned query intent adjustment information into text and then perform L2 norm normalization to obtain a query feature vector. Secondly, the aforementioned query feature vector is input into the query tower model included in the trained customer service recall dual-tower model to obtain a query projection feature vector. Thirdly, the cosine similarity between the aforementioned query projection feature vector and the set of latent feature vectors of the nodes included in the aforementioned query customer service response navigation map is determined to obtain a query node similarity set. Next, using a greedy algorithm, starting from the target clustering hub node, a greedy search is performed in the aforementioned query customer service response navigation map to obtain a search hub node set. The aforementioned target clustering hub node can be the clustering hub node corresponding to the query node with the largest similarity value in the aforementioned query node similarity set. Then, using the aforementioned search hub node set as the query entry point, a greedy search algorithm is employed to perform an approximate nearest neighbor search on the aforementioned response nearest neighbor graph. This retrieves responses from all customer service response data in the customer service response business knowledge base, resulting in an initial query response information set. The distance calculations used in the approximate nearest neighbor search are uniformly performed by calculating the Euclidean distance from the feature vectors of nodes after L2 norm normalization following word embedding. Finally, the initial query response information set undergoes secondary filtering to obtain a second query response information sequence. This secondary filtering removes inconsistencies and conflicts between the initial query response information set and the aforementioned query intent adjustment information regarding insurance product type and user preferences.

[0054] Step nine involves fusing and sorting the first query-response information sequence and the second query-response information sequence to obtain a query-response information set, and then sending this set to the response front-end page for display. In practice, the executing entity can first determine the first and second response ranking value sequences for the first and second query-response information sequences, respectively. The first response ranking value in the first response ranking value sequence can be 1 divided by the sum of the order of the first query-response information in the first query-response information sequence and a preset ranking threshold. The second response ranking value in the second response ranking value sequence can be 1 divided by the sum of the order of the second query-response information in the second query-response information sequence and a preset ranking threshold. The preset ranking threshold can be a pre-set value used as a smoothing constant to avoid excessive weighting of top-ranked query-response information; a value of 60 is acceptable. Then, the sum of each first response ranking value in the first response ranking value sequence and the corresponding second response ranking value in the second response ranking value sequence is determined as the target response value set. Specifically, when the first and second response ranking sequences do not have a one-to-one correspondence, the corresponding response ranking value for which there is no one-to-one correspondence can be set to 0. Finally, the maximum number of target response values ​​(preset document values) is selected from the target response value set to form the filtered target response value set. The first and / or second query response information sets corresponding to the filtered target response value set are then determined as the query response information set. The preset document value can be a pre-defined maximum number of query response information items to be selected, and can be 50.

[0055] The above-described technical solution and its related content, as an inventive point of this disclosure, solve the technical problem of "long retrieval time, reduced query efficiency, reduced user experience, low accuracy of query response information, and increased waste of video memory resources." Factors contributing to long retrieval time, reduced query efficiency, reduced user experience, low accuracy of query response information, and increased waste of video memory resources are often as follows: Customer service response knowledge bases often exhibit long-tail distribution and bias (e.g., insurance products have a lot of general content while high-risk content is located in sparse peripheral areas). Existing graph retrieval methods, starting from the central entry point, are prone to getting trapped in local optima of high-frequency content. Furthermore, existing graph retrieval methods typically involve full traversal retrieval, leading to long retrieval times, reduced query efficiency, and a poor user experience. Using a single retrieval method also results in low accuracy of the output query response information, increasing waste of video memory resources. Solving these factors can achieve the effects of reducing retrieval time, improving query efficiency, enhancing user experience, increasing the accuracy of query response information, and reducing waste of video memory resources. To achieve this effect, this disclosure firstly uses the dual-level index information of the response to perform intent index recall, obtaining a first query response information sequence. The intent index included in the dual-level index information of the response can be used for coarse screening, and the text inverted index information provided provides prior knowledge, which can reduce the missed recall due to vector offset in word segmentation matching. Intent index recall can avoid full traversal and improve the recall accuracy based on keywords. Secondly, the construction of the response nearest neighbor graph can remove redundant long edges in the above customer service response business knowledge base, ensure the global connectivity of the graph, prevent the search from getting stuck in local optima, and reduce the degree of each node to facilitate subsequent improvement of the efficiency of approximate nearest neighbor search. Thirdly, hierarchical clustering node extraction can divide massive unordered nodes into multiple clusters, effectively reducing the average maximum distance of any query vector from the cluster hub node, which can reduce the problem of search path explosion caused by the entry point falling in the wrong cluster, and improve the visibility of long tail and sparse data, avoiding getting stuck in local optima. Next, wandering sampling encoding can balance the sampling bias between high-degree nodes (cluster hub nodes) and low-degree nodes (edge ​​nodes), and the conversion of graph structure into feature vectors allows nodes to automatically perceive their own topological connectivity. Subsequently, the fusion of topological semantic encoding of the node tower model based on the dual-tower model and the generation of the query customer service reply navigation graph can construct potential vectors with textual semantics and business topology, forcing the model to learn the optimal entry mapping rather than content similarity mapping; the query customer service reply navigation graph can perform a greedy search during retrieval to quickly determine the optimal underlying search entry point without traversing the entire graph, breaking the original spatial distance limitations and making it highly relevant in business logic.Finally, the greedy retrieval recall can align query intent information to the hub space, achieving optimal physical search through fixed entry points and local depth-first traversal. First, determining the global route in the query customer service response navigation graph, then descending to the response nearest neighbor graph for greedy diffusion shortens the search path and reduces search complexity, improving search stability and efficiency. Secondary filtering ensures the security of the knowledge base and improves the semantic and business relevance of the recalled response information. Finally, fusion ranking eliminates the differences between multi-path recalls, achieving complementary advantages. This allows the query response information set to combine the dual advantages of exact matching and semantic generalization, improving the quality of the query response information set and reducing the number of times it is sent to and displayed on the front-end page, thus reducing the waste of GPU memory and transmission resources.

[0056] Step 105: Adjust the information according to the query intent, sort the query response information set by relevance, and obtain the query response sorting information sequence.

[0057] In some embodiments, the execution entity can perform relevance sorting on the query response information set based on the query intent adjustment information to obtain a query response sorting information sequence. This query response sorting information sequence can be a sequence ordered from largest to smallest based on the relevance between each query response and the query intent adjustment information corresponding to the selection operation. In practice, the execution entity can input the query response information set and the query intent adjustment information into a cross-encoder for fine-tuning to obtain the query response sorting information sequence. The cross-encoder can be an encoder that matches and sorts the input query response information set and the query intent adjustment information. For example, the cross-encoder can be a CohereRerank model.

[0058] In addressing the technical challenges mentioned above, the application scenario—high-concurrency, real-time online customer service for enterprise-level insurance products involving large-scale, heterogeneous data during holidays—often presents the following technical issues: Due to the dramatic increase in user inquiries during holidays, the timeliness of customer service is crucial. Furthermore, relying solely on a single semantic dimension and user long short-term memory based on large language processing leads to low accuracy and long query times, reducing user experience. Repeated sorting further wastes memory and transmission resources. Considering the following requirements for this application scenario: adaptability to high concurrency, low latency for timely responses to real-time requests, adaptability to personalization and high timeliness, and adaptability to parallel determination of multi-dimensional sorting, we have decided to adopt the following solution: In some optional implementations of certain embodiments, the process of adjusting the query response information set according to the query intent information to obtain a query response ranking information sequence may include the following steps: The first step is to determine the semantic relevance set of question-answer pairs between the aforementioned query intent adjustment information and the aforementioned query response information set. Specifically, the semantic relevance of the question-answer pairs in this set can be considered as the degree of semantic matching between the query intent adjustment information and each query response information in the question-answer pair. In practice, the executing entity can input the query intent adjustment information and each query response information into a cross-encoder to obtain the semantic relevance set of question-answer pairs.

[0059] The second step involves determining the user preference matching degree set of the aforementioned query intent adjustment information and the aforementioned query response information set based on the target user's matching short-term memory information set and the target long-term memory information set. The user preference matching degree set can be the degree of matching between user preferences in the query response information and preferences in the target user's short-term and long-term memories. In practice, the executing entity can first, in response to the determination that the matching short-term memory information set and the target long-term memory information set do not meet the preset memory data volume condition, determine the shared group preference information under the aforementioned query intent adjustment information. The preset memory data volume condition can be a condition that the data volume of the matching short-term memory information set and the target long-term memory information set is less than or equal to the preset memory data volume. The preset memory data volume can be the minimum data volume used to distinguish whether the target user is a new user, a user with little interaction, or a user with a user preference profile (i.e., a user who does not need a cold start). The value can be determined according to specific circumstances and will not be elaborated here. The shared group preference information can be the preference information of commonly used groups under the intent corresponding to the aforementioned query intent adjustment information. Next, the aforementioned query response information set is input into the FastText model for preference information classification to obtain a query preference classification information set. Next, the cosine similarity set of the aforementioned query preference classification information set and the aforementioned shared group preference information set is determined as the user preference matching degree set. Then, in response to determining that the matching short-term memory information set and the target long-term memory information set meet the preset memory data volume condition, the matching short-term memory information set and the target long-term memory information set are text-encoded and then weighted and summed to obtain the user preference feature vector. Afterwards, the cosine similarity set of the aforementioned user preference feature vector and the query preference feature vector set corresponding to the aforementioned query preference classification information set is determined as the user preference matching degree set. Here, the aforementioned query preference feature vector set can be a feature vector obtained by text encoding the query preference classification information set.

[0060] The third step is to determine the timeliness fit set and the structured field Jacquard similarity set for the aforementioned query intent adjustment information and query response information set. The timeliness fit set can be the timeliness value of the timestamp of the query response information. The structured field Jacquard similarity set can be the average of the Jacquard similarities of the dialogue intent information, dialogue slot information set, and dialogue entity information set obtained after structured partitioning of the aforementioned query intent adjustment information and query response information. In practice, the executing entity can first use a business scenario timeliness mapping table to determine the time decay function set and time decay scale set of the aforementioned query response information set based on the business scenario category information set. The business scenario timeliness mapping table can be a pre-defined table used to record the mapping relationship between business scenario category information, time decay functions, and time decay parameters. For example, the records included in the aforementioned business scenario timeliness mapping table could be: "If the business scenario category information is policy change / rate adjustment, then the time decay function can be an exponential decay function with a time decay scale of any integer between [3 days, 7 days]; if the business scenario category information is product change / new insurance product release, then the time decay function can be a Gaussian decay function with a time decay scale of any integer between [14 days, 30 days]; if the business scenario category information is claims process scenario, then the time decay function can be a linear decay function with a time decay scale of any integer between [90 days, 180 days]." Secondly, through the entity activity mapping table, based on the word segmentation information set included in the aforementioned query intent adjustment information and the activity level of the aforementioned target user, the dynamic decay factor set in the aforementioned time decay function is determined. The aforementioned dynamic decay factor can be any value between [0, 1] used to control the steepness of the curve. The aforementioned time decay scale can be a value used to control the degree of decay. The aforementioned entity activity mapping table can be a table used to record the mapping relationship between the word segmentation information set, user activity, and dynamic decay factor. The entity activity mapping table mentioned above includes records where: "For entities containing time-sensitive words (e.g., latest, new policy, update, etc.) and frequently active users, the dynamic decay factor can be increased from the base decay factor [0.2, 0.5]; for entities containing time-sensitive words and moderately active users, the dynamic decay factor can be the original decay factor; for entities containing time-sensitive words and low-frequency active users, the dynamic decay factor can be decreased from the original decay factor [0.2, 0.4]." The base decay factor can be a preset decay factor with a value of 0.7. Frequently active users are those who log in to the smart reply application more than 5 times per week. Moderately active users are those who log in 1-4 times per week. Low-frequency active users are those who have not logged in for the past 30 days.Then, the time difference between the current timestamp and the timestamp stored in the query-response information set is input into the corresponding time decay function set based on the dynamic decay factor set and the time decay scale set to obtain the response time decay value set. Finally, the corresponding product of the above response time decay value set and the above question-answer pair semantic relevance set is determined as the timeliness fit set.

[0061] The fourth step involves determining the normalized sum of the aforementioned question-and-answer pair semantic relevance set, user preference matching set, timeliness suitability set, and structured field Jacquard similarity set to obtain the response dimension value set. The response dimension values ​​in this set can be obtained by summing the Z-scores of the question-and-answer pair semantic relevance and the corresponding user preference matching, timeliness suitability, and structured field Jacquard similarity, respectively.

[0062] Fifth, filter at least one query response from the above query response information set whose corresponding response dimension value is greater than or equal to a preset response relevance threshold. The preset response relevance threshold can be a pre-defined set of the top 20 response dimension values ​​from the above response dimension value set, ranked from largest to smallest.

[0063] Step 6: Utilize a response uncertainty prediction model to determine the response uncertainty value of the at least one query response information and the query intent adjustment information. The response uncertainty value can represent the uncertainty of the entire sequence after sorting the question-answer pair set composed of the at least one query response information and the query intent adjustment information. The response uncertainty prediction model can be a deep neural network model that determines the uncertainty of the input at least one query response information and query intent adjustment information and outputs the uncertainty value. For example, the response uncertainty prediction model can be Monte Carlo Dropout. In practice, the executing entity can repeatedly input the at least one query response information into the response uncertainty prediction model to obtain a set of response statistical variances. Next, the variance of each response statistical variance group in the set of response statistical variances is determined as the response variance, resulting in a response variance set. Then, the response variance set is input into a Sigmoid function for normalization to obtain the response uncertainty value.

[0064] Step 7: In response to the aforementioned response uncertainty value being greater than or equal to a preset uncertainty threshold, the aforementioned query intent adjustment information, the aforementioned at least one query response information, the target long-term memory information set, and the aforementioned matching short-term memory information set are input into the response information reordering model to obtain a query response ranking information sequence, and the aforementioned query response ranking information sequence is sent to the response front-end page for display. The aforementioned response information reordering model can be a large language model that uses the input aforementioned query intent adjustment information, the aforementioned at least one query response information, the target long-term memory information set, and the aforementioned matching short-term memory information set to form response information reordering prompts for reordering, thereby obtaining a query response ranking information sequence. For example, the aforementioned response information reordering model can be a deepseek model or a thousand-question model. The aforementioned response information reordering prompts can be prompts used to guide the response information reordering model to re-rank the input at least one query response information and output the ranked sequence. The aforementioned preset uncertainty threshold can be a predetermined critical value for whether to reorder the aforementioned at least one query response information. For example, the aforementioned preset uncertainty threshold can be 0.7.

[0065] Step 8: In response to the fact that the uncertainty value of the above-mentioned response is less than a preset uncertainty threshold, the at least one query response information is sorted according to the above-mentioned response dimension value set to obtain a query response sorting information sequence, and the above-mentioned query response sorting information sequence is sent to the response front-end page for display. In practice, the above-mentioned executing entity can sort the above-mentioned at least one query response information according to the above-mentioned response dimension value set from largest to smallest to obtain a query response sorting information sequence.

[0066] The above-described technical solution and its related content, as an inventive point of this disclosure, solve the technical problem of "low sorting accuracy and long query time, reducing user experience and increasing waste of video memory and transmission resources." Factors leading to low sorting accuracy, long query time, reduced user experience, and increased waste of video memory and transmission resources are often as follows: During holidays, the number of user inquiries to customer service increases dramatically, placing higher demands on the timeliness of customer service. Furthermore, sorting the query response information set based solely on a single semantic dimension and user long short-term memory (LSTM) of a large language leads to low sorting accuracy and long query time, reducing user experience and requiring repeated sorting, thus increasing waste of video memory and transmission resources. Solving these factors can improve sorting accuracy, shorten query time, improve user experience, and reduce waste of video memory and transmission resources. To achieve this effect, this disclosure first employs a cross-encoder for deep semantic interaction computation, which can accurately capture subtle semantic relationships between query intent adjustment information and query response information sets, reducing noise problems caused by essentially unrelated vector similarities in vector retrieval. Secondly, a user preference matching set is determined. User categories are distinguished by the amount of short-term and long-term memory information. Introducing shared group preference information during cold starts can improve the experience for new and inactive users. For active users, the short-term and long-term memory information is encoded and fused, and then similarity is calculated with query preference classification information. This allows for precise matching of personalized user needs, achieving a unique experience and preventing irrelevant answers. Thirdly, a timeliness fit set and a structured field Jacquard similarity set are determined. Timeliness fit is dynamically determined based on different business scenarios and user activity levels, ensuring high weight for highly time-sensitive information and improving timeliness in ranking. Structured field Jacquard similarity enforces consistency with the current scenario, effectively preventing invalid recall and ranking across business scenarios. Finally, normalization, summation, and filtering are performed. Normalization and summation remove the influence of differences in units of measurement. Determining the number of answer dimensions from multiple perspectives improves accuracy, while filtering removes inaccurate query responses, reducing subsequent computational load. Then, using a response uncertainty prediction model, the response uncertainty value is determined, quantifying the model's uncertainty level regarding the current ranking result, which facilitates subsequent re-ranking. Finally, when the response uncertainty value is greater than or equal to a preset uncertainty threshold, a re-ranking based on the response information re-ranking model is performed, triggering a list-style re-ranking of the large model. Utilizing the global semantic understanding and reasoning capabilities of the large model for re-ranking can effectively solve the logical inconsistency problem in complex queries, reduce the illusion problem of large language models, and improve the quality and accuracy of the query response ranking information sequence.When the uncertainty value of the response is less than the preset uncertainty threshold, the responses are sorted according to the set of values ​​of the response dimension, which improves the timeliness and real-time performance of the response. This can greatly reduce computational latency and cost. The combination of the two achieves a balance between accuracy and timeliness / efficiency, reduces the number of repeated sorting, and reduces the waste of video memory and transmission resources.

[0067] Step 106: Send the query response sorting information sequence to the response front-end page, and perform structured division of the multi-turn question and answer text information of the current session to obtain a short-term memory structured information set.

[0068] In some embodiments, the executing entity may send the query-response sorting information sequence to the response front-end page and perform structured partitioning of the current session's multi-turn question-and-answer text information to obtain a short-term memory structured information set. The short-term memory structured information in the short-term memory structured information set may be JSON-formatted structured information composed of intent fields, slot fields, key entity fields, user preference fields, and interaction node fields in that order.

[0069] In some optional implementations of certain embodiments, the above-mentioned structural partitioning of the multi-turn question-and-answer text information of the current session to obtain a short-term memory structured information set may include the following steps: The first step involves filtering the current session's multi-turn Q&A text information to obtain the target session's multi-turn dialogue information. This target session's multi-turn dialogue information can be dialogue information where invalid information (i.e., content lacking substantive information) has been removed from the current session's multi-turn Q&A text information, retaining only the business semantic information. In practice, the executing entity can extract information from the current session's multi-turn Q&A text information using a pre-set business invalid term list and a pre-set business semantic term list, obtaining a set of business invalid information and a set of business semantic information. Then, the business invalid information set is removed from the current session's multi-turn Q&A text information set, while the business semantic information set is retained, to obtain the target session's multi-turn dialogue information. The pre-set business invalid term list can be a pre-defined term list composed of invalid information appearing in insurance product customer service, and is updated periodically. The aforementioned pre-defined invalid business term list may include: polite expressions (e.g., "Hello," "Thank you for your hard work," "You're welcome," etc.), modal particles, purely interrogative phrases without business focus (e.g., "Are you there?", "Can you hear me?", etc.), emotional expression phrases without business entities (e.g., "Okay," "Sure," etc.), purely interactive instructions without requests (e.g., "Please wait a moment," "Transfer to human operator," etc.), and tags. The aforementioned pre-defined business semantic term list may be composed of word segments containing business semantic information, business entities, and business requirement information, and may be updated periodically.

[0070] The second step is to perform dialogue segmentation processing on the multi-turn dialogue information of the target session to obtain a dialogue segmentation fragment information set. The dialogue segmentation fragment information in the above dialogue segmentation fragment information set can be fragment information grouped together by the target user's query text information and the reply text information provided by the customer service assistant, and the above dialogue segmentation fragment information includes a dialogue timestamp and a dialogue turn sequence number.

[0071] Third, for each dialogue segment information in the above dialogue segment information set, perform the following extraction steps: Sub-step 1 involves performing intent recognition on the aforementioned dialogue segment information to obtain dialogue intent information. This dialogue intent information may include query intent feature information.

[0072] Sub-step 2 involves filling the aforementioned dialogue fragment information with slots to obtain a dialogue slot information set. The dialogue slot information in this set can be key-value pairs formed from a set of values ​​extracted from the aforementioned dialogue fragment information. This preset slot information set can be a pre-defined set of entity information. For example, the preset slot information set may include, but is not limited to, at least one of the following: policy number, policyholder gender, date of incident, and premium amount.

[0073] Sub-step 3 involves performing dialogue entity recognition on the aforementioned dialogue segment information to obtain a dialogue entity information set. The dialogue entity information in this set can be the entity information from the aforementioned dialogue segment information. In practice, the executing entity can utilize a named entity recognition model to perform dialogue entity recognition on the aforementioned dialogue segment information to obtain the dialogue entity information set. This named entity recognition model can be a Lattice LSTM (Long Short-Term Memory) model.

[0074] Sub-step 4 involves extracting user preference information from the aforementioned dialogue fragments to obtain a user preference information set. This user preference information can be information representing user behavioral preferences. For example, the user preference information set may include, but is not limited to, at least one of the following: preference information for participating life insurance, preference information for online claims. In practice, the executing entity can utilize a large language model to extract user preference information from the aforementioned dialogue fragments to obtain the user preference information set.

[0075] Sub-step 5 involves extracting interactive response information from the aforementioned dialogue fragments to obtain user interactive response information. This user interactive response information can be the interactive information received in response to a query text message. For example, the user interactive response information could be "The user inquired about life insurance premiums, and customer service informed them of the calculation results," or "The user applied for a claim, and customer service confirmed the policy number." In practice, the executing entity can input the aforementioned dialogue fragments into a large language model for summary extraction to obtain the user interactive response information.

[0076] Sub-step 6 involves structuring the aforementioned dialogue intent information, dialogue slot information set, dialogue entity information set, user preference information set, and user interaction response information to obtain short-term memory structured information. This short-term memory structured information can be converted into JSON format structured information in the order of deduplicated dialogue intent information, dialogue slot information set, dialogue entity information set, user preference information set, and user interaction response information. In practice, the executing entity can first deduplicate the aforementioned dialogue intent information, dialogue slot information set, dialogue entity information set, user preference information set, and user interaction response information to obtain deduplicated dialogue intent information, deduplicated dialogue slot information set, deduplicated dialogue entity information set, deduplicated user preference information set, and deduplicated user interaction response information. Then, the deduplicated dialogue intent information, deduplicated dialogue slot information set, deduplicated dialogue entity information set, deduplicated user preference information set, and deduplicated user interaction response information are converted into JSON format in that order to obtain short-term memory structured information.

[0077] Step 107: Based on the short-term memory structured information set, perform incremental weighted hierarchical storage updates for the user's long-term memory repository.

[0078] In some embodiments, the aforementioned executing entity can perform incremental weighted hierarchical storage updates on the aforementioned user long-term memory repository based on the aforementioned short-term memory structured information set. Specifically, this incremental weighted hierarchical storage update determines the memory timeliness weight (i.e., the time difference from the current time) of each short-term memory structured information in the aforementioned short-term memory structured information set, and classifies it accordingly into permanent memory, time-sensitive memory, and forgotten memory. These are then stored in three layers—permanent memory, time-sensitive memory, and forgotten memory—for incremental storage updates, avoiding full overwrite and preserving historical traces. The aforementioned permanent memory layer can be a storage layer for storing information that is retained long-term by the target user, has no expiration time, is related to the user's long-term needs, and will not be deleted upon the end of the session. The information included in the aforementioned permanent memory layer may include, but is not limited to, at least one of the following: user preference information, user basic information, historical claims, and historical insurance information. The aforementioned user basic information may include, but is not limited to, at least one of the following: mobile phone number and frequently used beneficiary information. The aforementioned time-sensitive memory layer can be a storage layer for storing information with specific time limits and related to the user's recent needs. The aforementioned time-sensitive memory layer may include information such as: recent policies (e.g., expiry date: 6 months after the policy expires), temporary request information (e.g., expiry date: 1 month after the request is completed), and recent claims inquiry information (e.g., expiry date: 2 months after the inquiry is completed). The aforementioned forgotten memory layer may be a storage layer used to store expired, invalid, or worthless information that needs to be periodically cleared. The information stored in the aforementioned forgotten memory layer may include, but is not limited to, at least one of the following: invalid casual chat records, and query information for expired policies.

[0079] In some optional implementations of certain embodiments, the above-mentioned incremental weighted hierarchical storage update of the user's long-term memory repository based on the above-mentioned short-term memory structured information set may include the following steps: The first step is to perform the following weighted summation step for each short-term memory structured information in the above short-term memory structured information set: Sub-step 1: Based on the dialogue turn information of the aforementioned short-term memory structured information, determine the structured turn timeliness weight value of the aforementioned short-term memory structured information. The structured turn timeliness weight value can represent the importance of the aforementioned short-term memory structured information in terms of timeliness. The dialogue turn information can be the dialogue turn number of the aforementioned short-term memory structured information. In practice, the executing entity can first determine the dialogue turn information that terminates the short-term structured information as the current turn information, wherein the terminated short-term structured information can be the short-term memory structured information located at the termination position in the aforementioned short-term memory structured information set. Then, determine the difference between the current turn information and the dialogue turn information as the turn difference value. Finally, determine the structured turn timeliness weight value through a turn timeliness mapping table. The turn timeliness mapping table can be a table used to record the mapping relationship between the turn difference value and the structured turn timeliness weight value. For example, the above round timeliness mapping table can record the following: "When the round difference is 1, the structured round timeliness weight value can be 3; when the round difference is any value in [2, 3], the structured round timeliness weight value can be 2; when the round difference is any value in [4, 5], the structured round timeliness weight value can be 1; when the round difference is greater than 5, the structured round timeliness weight value can be 0."

[0080] Sub-step 2 involves determining the intent relevance weight values ​​for the aforementioned short-term memory structured information and the aforementioned short-term completion intent information. The intent relevance weight value can represent the importance of the aforementioned short-term memory structured information in terms of intent. In practice, the executing entity can first extract the intent feature information corresponding to the aforementioned termination short-term structured information and the aforementioned short-term completion intent information from the aforementioned query intent feature information, using them as termination intent feature information and target intent feature information. Then, the cosine similarity between the aforementioned termination intent feature information and the aforementioned target intent feature information is determined as the intent similarity. Finally, the intent relevance weight value is determined through a similarity weight mapping table. This similarity weight mapping table can be a table used to record the mapping relationship between the aforementioned intent similarity and intent relevance weight values. For example, the content recorded in the similarity weight mapping table could be: "When the aforementioned intent similarity is greater than or equal to a first preset similarity threshold, the intent relevance weight value can be 3; when the aforementioned intent similarity is less than the first preset similarity threshold but greater than or equal to a second preset similarity threshold, the intent relevance weight value can be 2; when the intent similarity is less than the second preset similarity threshold, the intent relevance weight value can be 0." Both the first preset similarity threshold and the second preset similarity threshold can be preset values. For example, the values ​​of the first preset similarity threshold and the second preset similarity threshold can be 0.6 and 0.3, respectively.

[0081] Sub-step 3 involves determining the business importance weight values ​​of the aforementioned short-term memory structured information. These business importance weight values ​​characterize the degree of importance of the aforementioned short-term memory structured information. In practice, the executing entity can first determine the Jacquard similarity groups of the terminated short-term structured information and the current query text information in the dialogue slot information set, dialogue entity information set, and user preference information set, respectively. Then, using a Jacquard importance mapping table, the business importance weight values ​​are determined based on the aforementioned Jacquard similarity groups. The Jacquard importance mapping table can be a table used to record the mapping relationship between each Jacquard similarity and the aforementioned business importance weight values. For example, the content recorded in the aforementioned Jacarbec importance mapping table could be: "When the Jacarbec similarity corresponding to both the dialogue entity information set and the user preference information set is greater than or equal to a preset Jacarbec similarity threshold, the business importance weight value can be 4; when the Jacarbec similarity corresponding to both the dialogue entity information set and the user preference information set is less than a first preset Jacarbec similarity threshold, and the Jacarbec similarity corresponding to the user preference information set is greater than or equal to a second preset Jacarbec similarity threshold, the business importance weight value can be 2; in other cases, the business importance weight value can be 0." The aforementioned first and second preset Jacarbec similarity thresholds can both be pre-set values, taking values ​​of 0.65 and 0.4, respectively.

[0082] Sub-step 4: Determine the sum of the above-mentioned structured round timeliness weight value, the above-mentioned intent relevance weight value, and the above-mentioned business importance weight value to obtain short-term memory storage level information.

[0083] The second step involves matching the aforementioned short-term memory structured information set with the long-term memory information set included in the user's long-term memory repository to obtain a memory matching result set. The memory matching results in this set characterize whether the aforementioned short-term memory structured information has been stored. This matching can be a precise match using hash values.

[0084] The third step involves storing the aforementioned short-term memory structured information set into the user's long-term memory repository according to the obtained short-term memory storage level information and the memory matching result set.

[0085] In some optional implementations of certain embodiments, the above-mentioned storage of the short-term memory structured information set into the user's long-term memory repository according to the obtained multiple short-term memory storage level information and the above-mentioned memory matching result set, in a time-series hierarchical manner, includes: The first step involves filtering the short-term memory structured information set based on the aforementioned multiple short-term memory storage level information, resulting in a filtered short-term memory structured information set. In practice, the executing entity can first remove short-term memory structured information from the aforementioned short-term memory structured information set where any one of the following weight values—business importance weight, structured round timeliness weight, and intent relevance weight—is 0, thus creating a target short-term memory structured information set. Then, at least one target short-term memory structured information is selected from the target short-term memory structured information set whose corresponding target short-term memory storage level information is less than a preset level value. This preset level value can be a pre-defined threshold value used to determine whether short-term memory structured information should be removed, and can be 3. Finally, the at least one short-term memory structured information is removed from the target short-term memory structured information set, resulting in the filtered short-term memory structured information set.

[0086] The second step involves, in response to the determination that at least one piece of structured short-term memory information satisfying long-term storage conditions exists in the aforementioned set of filtered short-term memory structured information, at least one memory matching result corresponding to the aforementioned at least one piece of structured short-term memory information is selected from the aforementioned memory matching result set. The aforementioned long-term storage conditions can be that the filtered short-term memory structured information is structured information that needs to be retained long-term, has no expiration time, is related to the target user's long-term needs, and will not be deleted upon the end of the session. For example, the filtered short-term memory structured information satisfying the long-term storage conditions can include, but is not limited to, at least one of the following: user basic preference information, user basic information (e.g., user's mobile phone number, user account, frequently used beneficiary information), and historical claims / insurance information. The aforementioned historical claims / insurance information could be "Insured for children's accident insurance in March 2025, with one claim made."

[0087] Third, in response to determining that at least one of the above memory matching results contains a filtered memory matching result set whose representation has been stored, the filtered short-term memory structured information set corresponding to the above-mentioned memory matching result set whose representation has been stored is stored in a time-series chain to the user's long-term memory repository. The time-series chained storage can be a storage method that stores the filtered short-term memory structured information set corresponding to the above-mentioned memory matching result set whose representation has been stored in a chain at previously stored locations, rather than overwriting existing data.

[0088] Fourth, in response to determining that at least one of the above memory matching results contains a set of memory matching results whose representations have not been stored, the filtered short-term memory structured information set corresponding to the memory matching result set whose representations have not been stored is subjected to privacy encryption processing to obtain an encrypted long-term memory information set. The encrypted long-term memory information in the encrypted long-term memory information set can be obtained by privacy desensitizing the filtered short-term memory structured information corresponding to the memory matching result set whose representations have not been stored and encrypting it using a national cryptographic algorithm.

[0089] The fifth step involves storing the aforementioned encrypted long-term memory information set into the permanent memory layer included in the aforementioned user long-term memory repository, and updating it periodically as needed. The aforementioned permanent memory layer can be a level that stores memories that will not be deleted when the session ends. The aforementioned periodic updates can be updates made on demand by the target user when modifying their basic user information during the session.

[0090] Step 6: In response to determining that a structured information set satisfying the time-sensitive memory storage conditions exists within the aforementioned filtered short-term memory structured information set, the filtered short-term memory structured information set satisfying the time-sensitive memory storage conditions is stored in the time-sensitive memory layer of the aforementioned user long-term memory repository, and periodically updated over a first time period. The aforementioned time-sensitive memory storage conditions may include the filtered short-term memory structured information having a preset expiration date, existing in multiple rounds of question-and-answer text information in the current session but not having a complete match in the user's long-term memory repository, and being stored in the forgetting memory layer after the expiration date has passed. The aforementioned time-sensitive memory layer may be a hierarchy storing expiration date markers. For example, the expiration date marker may be a marker adding a 6-month validity period to the policy's expiration date. The memory information included in the aforementioned time-sensitive memory layer may include, but is not limited to, at least one of the following: policy information with an expiration date, demand information with an expiration date (e.g., deleted 1 month after the demand is completed), and claims consultation record information with an expiration date (e.g., expired 2 months after the consultation). The aforementioned periodic updates in the first time period could be periodic updates that are automatically checked daily by the user's long-term memory repository.

[0091] Step 7: Perform an expiration check on the set of time-sensitive long-term memory information included in the time-sensitive memory layer of the aforementioned user long-term memory repository to obtain an expiration check result information set. The expiration check result information in this set can be information on whether the time-sensitive long-term memory information included in the aforementioned time-sensitive memory layer has exceeded the time specified for its expiration.

[0092] Step 8 involves storing the filtered short-term memory structured information set corresponding to at least one expired validity period detection result in the forgotten memory layer of the aforementioned user long-term memory repository, and performing periodic memory cleanup in the second time period. The forgotten memory layer can be a storage layer for storing expired and invalid information that requires periodic cleanup. The memory information included in the forgotten memory layer can include: expired memory information and duplicate or redundant memory information. The second time period can be weekly.

[0093] Further reference Figure 2 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a long short-term memory collaborative retrieval and hierarchical storage device based on a customer service assistant. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this customer service assistant-based long short-term memory collaborative retrieval and hierarchical storage device can be specifically applied to various electronic devices.

[0094] like Figure 2As shown, a long short-term memory collaborative retrieval and hierarchical storage device 200 based on a customer service assistant includes: a determination unit 201, a matching retrieval unit 202, an intent completion unit 203, a recall unit 204, a relevance ranking unit 205, a structured partitioning unit 206, and a hierarchical storage unit 207. The determination unit 201 is configured to: in response to the intelligent reply application based on the customer service assistant receiving current query text information and current session multi-turn question-and-answer text information sent by the target user, determine the user identification information and query intent feature information for the aforementioned current query text information. The matching retrieval unit 202 is configured to: perform matching retrieval on the target user's short-term temporary memory bank based on the aforementioned user identification information and the aforementioned query intent feature information to obtain a matching short-term memory information set. The intent completion unit 203 is configured to: perform long and short-term intent completion processing on the aforementioned query intent feature information based on the aforementioned matching short-term memory information set and the user's long-term memory bank to obtain intent completion information, and send the intent completion information to the reply front-end page of the intelligent reply application. Recall unit 204 is configured to: in response to receiving a selection operation for the aforementioned intent completion information, recall the customer service response business knowledge base based on the query intent adjustment information corresponding to the selection operation and the aforementioned user long-term memory repository, to obtain a query response information set. Relevance ranking unit 205 is configured to: perform relevance ranking processing on the aforementioned query response information based on the aforementioned query intent adjustment information, to obtain a query response ranking information sequence. Structured partitioning unit 206 is configured to: send the aforementioned query response ranking information sequence to the aforementioned response front-end page, and perform structured partitioning on the aforementioned multi-turn question-and-answer text information of the current session, to obtain a short-term memory structured information set. Hierarchical storage unit 207 is configured to: perform short-term memory incremental weight hierarchical storage update on the aforementioned user long-term memory repository based on the aforementioned short-term memory structured information set.

[0095] It is understandable that the units and references recorded in the customer service assistant-based long short-term memory collaborative retrieval and hierarchical storage device 200 are related. Figure 1 The steps described in the method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the long short-term memory collaborative retrieval and hierarchical storage device 200 based on the customer service assistant and the units contained therein, and will not be repeated here.

[0096] The following is for reference. Figure 3 It shows a schematic diagram of the structure of an electronic device (e.g., an electronic device) 300 suitable for implementing some embodiments of the present disclosure. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0097] like Figure 3As shown, the electronic device 300 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0098] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0099] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 309, or installed from storage device 308, or installed from ROM 302. When the computer program is executed by processing device 301, it performs the functions defined in the methods of some embodiments of this disclosure.

[0100] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0101] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0102] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: in response to a smart reply application based on a customer service assistant receiving current query text information and current session multi-turn question-and-answer text information sent by a target user, determine user identification information and query intent feature information for the aforementioned current query text information; perform matching retrieval on the target user's short-term temporary memory bank based on the aforementioned user identification information and the aforementioned query intent feature information to obtain a matching short-term memory information set; perform long and short-term intent completion processing on the aforementioned query intent feature information based on the aforementioned matching short-term memory information set and the user's long-term memory bank to obtain intent completion information; and send the aforementioned intent completion information... The information is sent to the response front-end page of the aforementioned intelligent response application; in response to receiving the selection operation for the aforementioned intent completion information, the customer service response business knowledge base is recalled based on the query intent adjustment information corresponding to the selection operation and the aforementioned user long-term memory repository to obtain a query response information set; based on the aforementioned query intent adjustment information, the aforementioned query response information set is sorted by relevance to obtain a query response sorting information sequence; the aforementioned query response sorting information sequence is sent to the aforementioned response front-end page, and the aforementioned multi-turn question and answer text information of the current conversation is structurally divided to obtain a short-term memory structured information set; based on the aforementioned short-term memory structured information set, the aforementioned user long-term memory repository is updated by incremental weighted storage of short-term memory.

[0103] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0104] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0105] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a determination unit, a matching retrieval unit, an intent completion unit, a recall unit, a relevance ranking unit, a structured partitioning unit, and a hierarchical storage unit. The names of these units do not necessarily limit the unit itself; for example, the determination unit may also be described as "a unit that, in response to a smart reply application based on a customer service assistant receiving current query text information and current session multi-turn question-and-answer text information sent by a target user, determines user identification information and query intent feature information for the aforementioned current query text information."

[0106] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0107] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for collaborative retrieval and hierarchical storage of long short-term memory based on customer service assistants, comprising: In response to the intelligent reply application based on the customer service assistant receiving the current query text information and the current session multi-turn question and answer text information sent by the target user, the user identification information and query intent feature information for the current query text information are determined; Based on the user identification information and the query intent feature information, a matching retrieval is performed on the target user's short-term temporary memory database to obtain a matching short-term memory information set; Based on the matching short-term memory information set and the user's long-term memory repository, the query intent feature information is processed to complete the short-term and long-term intents, resulting in intent completion information. The intent completion information is then sent to the response front-end page of the intelligent response application. In response to receiving a selection operation for the intent completion information, the customer service response business knowledge base is retrieved based on the query intent adjustment information corresponding to the selection operation and the user long-term memory repository to obtain a query response information set; Based on the query intent adjustment information, the query response information set is sorted by relevance to obtain a query response sorting information sequence. The query response sorting information sequence is sent to the response front-end page, and the multi-turn question and answer text information of the current session is structured to obtain a short-term memory structured information set; Based on the short-term memory structured information set, the user's long-term memory repository is updated by performing incremental weighted storage of short-term memory.

2. The method according to claim 1, wherein, The determination of user identification information and query intent feature information for the currently queried text information includes: In response to detecting that a target user has successfully logged into the smart reply application, the user identification information for the currently queried text information is obtained; In response to detecting that the target user is not logged into the smart reply application, a login reminder message is sent, and a login front-end page is triggered to prompt the user to log in; In response to detecting an input operation by a target user on the login front-end page, user identification information is obtained; The current query text information is subjected to character correction processing to obtain the corrected query text information; The modified query text information is subjected to multi-type keyword extraction to obtain a query keyword set; Based on the set of query keywords, intent mapping and recognition are performed on the corrected query text information to obtain query intent feature information.

3. The method according to claim 1, wherein, The step of matching and retrieving the target user's short-term temporary memory database based on the user identification information and the query intent feature information to obtain a matching short-term memory information set includes: Based on the target user's short-term temporary memory database, a short-term temporary memory information set related to the user identification information is determined as the target short-term temporary memory information set; Based on the query intent feature information and the query keyword set corresponding to the query intent feature information, short-term memory matching is performed on the target short-term temporary memory information set to obtain the short-term memory information set. The short-term memory information set is subjected to recall and screening to obtain a filtered short-term memory information set; The memory dimension relevance between each piece of filtered short-term memory information in the filtered short-term memory information set and the query intent feature information is determined to obtain a memory dimension relevance set, wherein the memory dimension relevance includes: intent relevance, keyword Jaccard coefficient, business scenario relevance and time decay factor relevance; Based on the memory dimension relevance set, the filtered short-term memory information set is rearranged and selected to obtain a matching short-term memory information set.

4. The method according to claim 1, wherein, The step of performing short-term and long-term intent completion processing on the query intent feature information based on the matching short-term memory information set and the user's long-term memory repository to obtain intent completion information includes: Based on the matching short-term memory information set, the query intent feature information is reconstructed by short-term referential resolution and completion to obtain short-term completed intent information. Based on the short-term completion intent information, long-term memory matching is performed on the user's long-term memory repository to obtain the target long-term memory information set; In response to determining that the target long-term memory information set is an empty set, the short-term completion intention information is determined as intention completion information; In response to determining that the target long-term memory information set is not an empty set, multi-dimensional information extraction is performed on the target long-term memory information set to obtain a multi-dimensional long-term memory information set. The multi-dimensional long-term memory information in the multi-dimensional long-term memory information set includes one of the following: historical intent information, historical item information, business scenario information, and user preference information. The multidimensional long-term memory information set is subjected to information conflict correction processing to obtain a corrected multidimensional long-term memory information set. In response to determining that the short-term completion intent information meets the intent scenario classification conditions, the short-term completion intent information is supplemented with memory dimensions according to the modified multi-dimensional long-term memory information set to obtain intent completion information; In response to determining that the short-term completion intent information does not meet the intent scenario classification conditions, intent completion is performed on the short-term completion intent information according to the modified multi-dimensional long-term memory information set to obtain intent completion information.

5. The method according to claim 1, wherein, The structured partitioning of the multi-turn question-and-answer text information of the current session yields a short-term memory structured information set, including: Information filtering is performed on the current session's multi-turn question-and-answer text information to obtain the target session's multi-turn dialogue information; The multi-turn dialogue information of the target session is processed by dialogue segmentation to obtain a set of dialogue segment information. For each dialogue segment information in the dialogue segment information set, perform the following extraction steps: Intent recognition is performed on the segmented dialogue information to obtain dialogue intent information; The dialogue segment information is filled with slots to obtain a dialogue slot information set; Dialogue entity recognition is performed on the segmented dialogue information to obtain a dialogue entity information set; User preference information is extracted from the segmented dialogue information to obtain a user preference information set; Interactive response information is extracted from the segmented dialogue information to obtain user interactive response information; The dialogue intent information, the dialogue slot information set, the dialogue entity information set, the user preference information set, and the user interaction response information are structured to obtain short-term memory structured information.

6. The method according to claim 1, wherein, The step of performing incremental weighted hierarchical storage updates on the user's long-term memory repository based on the short-term memory structured information set includes: For each short-term memory structured information in the aforementioned short-term memory structured information set, perform the following weighted summation step: Based on the dialogue turn information of the short-term memory structured information, determine the structured turn timeliness weight value of the short-term memory structured information; Determine the intent relevance weight values ​​of the short-term memory structured information and the short-term completion intent information; Determine the business importance weight value of the short-term memory structured information; The sum of the structured round timeliness weight value, the intent relevance weight value, and the business importance weight value is determined to obtain short-term memory storage level information; The short-term memory structured information set and the long-term memory information set included in the user's long-term memory repository are matched to obtain a memory matching result set; Based on the obtained short-term memory storage level information and the memory matching result set, the short-term memory structured information set is stored in the user's long-term memory repository in a time-series hierarchical manner.

7. The method according to claim 6, wherein, The step of storing the short-term memory structured information set into the user's long-term memory repository in a time-series hierarchical manner, based on the obtained multiple short-term memory storage level information and the memory matching result set, includes: Based on the multiple short-term memory storage level information, the short-term memory structured information set is subjected to memory filtering processing to obtain a filtered short-term memory structured information set; In response to determining that at least one piece of structured short-term memory information that satisfies the long-term storage conditions exists in the set of structured short-term memory information, at least one memory matching result corresponding to the at least one piece of structured short-term memory information is selected from the set of memory matching results. In response to determining that there exists a set of filtered memory matching results in the at least one memory matching result, the set of filtered short-term memory structured information corresponding to the set of memory matching results in which the characteristics are stored is stored in a time-chain manner to the user's long-term memory repository. In response to determining that there is a set of memory matching results that represent that the memory matching results have not been stored, the filtered short-term memory structured information set corresponding to the set of memory matching results that represent that the memory matching results have not been stored is subjected to privacy encryption processing to obtain an encrypted long-term memory information set; The encrypted long-term memory information set is stored in the permanent memory layer included in the user's long-term memory repository and updated periodically as needed; In response to determining that there exists a structured information set that meets the time-sensitive memory storage conditions in the filtered short-term memory structured information set, the filtered short-term memory structured information set that meets the time-sensitive memory storage conditions is stored in the time-sensitive memory layer in the user's long-term memory repository, and is periodically updated in the first time period. The validity period of the time-sensitive long-term memory information set included in the time-sensitive memory layer of the user's long-term memory repository is checked to obtain the validity period check result information set; The structured short-term memory information set corresponding to at least one expiration date detection result that has exceeded its expiration date is stored in the forgotten memory layer of the user's long-term memory repository, and periodic memory clearing is performed in the second time period.

8. An apparatus for a long short-term memory collaborative retrieval and hierarchical storage method based on a customer service assistant, comprising: The determining unit is configured to, in response to a smart reply application based on a customer service assistant receiving current query text information and current session multi-turn question and answer text information sent by a target user, determine user identification information and query intent feature information for the current query text information; The matching retrieval unit is configured to perform a matching retrieval on the target user's short-term temporary memory based on the user identification information and the query intent feature information to obtain a matching short-term memory information set; The intent completion unit is configured to perform short-term and long-term intent completion processing on the query intent feature information based on the matching short-term memory information set and the user's long-term memory repository to obtain intent completion information, and to send the intent completion information to the response front-end page of the intelligent response application. The recall unit is configured to, in response to receiving a selection operation for the intent completion information, recall the customer service response business knowledge base according to the query intent adjustment information corresponding to the selection operation and the user long-term memory repository, to obtain a query response information set. The relevance ranking unit is configured to perform relevance ranking processing on the query response information set according to the query intent adjustment information to obtain a query response ranking information sequence; The structured partitioning unit is configured to send the query response sorting information sequence to the response front-end page, and to perform structured partitioning of the current session's multi-turn question and answer text information to obtain a short-term memory structured information set; The hierarchical storage unit is configured to perform incremental weighted hierarchical storage updates of the user's long-term memory repository based on the short-term memory structured information set.

9. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images