Large model-based long text user portrait updating method and device and electronic equipment

Abstract

The application provides a long text user portrait updating method and device based on a large model and electronic equipment, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining interactive text based on user input writing parameters and a writing agent; collecting the interactive text by recording an agent, and obtaining structured storage user portrait information based on the interactive text; obtaining the change characteristics of the user portrait information by triggering an agent, and triggering a portrait agent if the change characteristics of the user portrait information meet the triggering condition; updating the user portrait according to the user portrait information and the portrait agent; the user portrait comprises at least two layers based on time granularity division; wherein the writing agent, the recording agent, the triggering agent and the portrait agent are all realized based on a large model, which can facilitate long text writing to achieve dynamic balance between short-term task intent and long-term writing preference.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, and electronic device for updating long-text user profiles based on a large model. Background Technology

[0002] With the improvement of large language model capabilities, multi-agent collaborative architectures are gradually being applied to long text writing assistance scenarios. These systems typically consist of multiple agents with specific functions working together to undertake sub-tasks such as content generation, style proofreading, and information retrieval, aiming to provide users with more refined and personalized writing support.

[0003] In personalized writing assistance, the quality of user profiling directly determines the suitability of the system's generated suggestions. User profiling aims to capture users' writing habits, style preferences, and behavioral patterns, providing decision-making support for various writing agents. However, existing multi-agent writing systems generally employ a memory mechanism that processes all interaction information equally in user profiling management, making it difficult to achieve a dynamic balance between short-term task intent and long-term writing preferences. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method, apparatus, and electronic device for updating long-text user profiles based on a large model.

[0005] This invention provides a method for updating long-text user profiles based on a large model, comprising: The interactive text is obtained based on the writing parameters input by the user and the writing agent. The system records the interactive text collected by the intelligent agent and obtains structured user profile information based on the interactive text. The change characteristics of the user profile information are obtained by triggering the intelligent agent. If the change characteristics of the user profile information meet the triggering conditions, the profile intelligent agent is triggered. The user profile is updated based on the user profile information and the profile agent; the user profile includes at least two layers based on time granularity. The writing agent, the recording agent, the triggering agent, and the portrait agent are all implemented based on a large model.

[0006] According to the present invention, a long text user profile update method based on a large model is provided, wherein obtaining structured user profile information based on the interactive text includes: Based on the interactive text, at least one of explicit feedback signals, implicit behavioral signals, and content feature signals is extracted; The user profile information is obtained by obtaining a structured storage based on at least one of the explicit feedback signals, implicit behavioral signals, and content feature signals and a preset format.

[0007] According to the present invention, a long-text user profile update method based on a large model is provided, wherein the change characteristics of the user profile information satisfy triggering conditions, including: The user rejects the writing agent's generation suggestions during consecutive interactions exceeding a first threshold; and / or, Based on the original user profile, the amount of structured user profile information obtained by the recording agent exceeds a second threshold; and / or, Based on the temporal characteristics of the original user profile, the temporal characteristics of the latest structured user profile information obtained by the recording agent exceed a third threshold; and / or, The difference between the semantic distribution of the user profile information and the semantic distribution of the original user profile exceeds a fourth threshold.

[0008] According to the present invention, a long text user profile update method based on a large model is provided. The user profile includes a short-term memory layer and a medium-term summary layer; The step of updating the user profile based on the user profile information and the profile agent includes: During the session, the short-term memory layer of the user profile is updated by the profile agent based on the user profile information. After the session ends, the user profile agent merges the short-term memory layer of the updated user profile and updates the intermediate summary layer of the user profile; wherein, the intermediate summary layer stores the article-level preference features obtained from the merging process.

[0009] According to the present invention, a long text user profile update method based on a large model is provided. The user profile also includes a long-term profile layer; The step of updating the user profile based on the user profile information and the profile agent includes: Based on the intermediate summary layer of the user profile, the long-term profile layer updates the user profile information through incremental summaries by the profile agent.

[0010] The dimensions of the long-term profile layer of the user profile information include at least two of the following: writing style, domain background, modification preferences, instruction style, and short-term task intent.

[0011] According to the present invention, a long text user profile update method based on a large model is provided. Before updating the long-term profile layer of the user profile information through incremental summarization by the profile agent based on the intermediate summary layer of the user profile, the method further includes: Determine the statistical parameters of the incremental information of the intermediate summary layer; The intermediate summary layer based on the user profile, and the long-term profile layer that updates the user profile information through incremental summaries by the profile agent, include: If the incremental information is determined to occur continuously based on the statistical parameters, then the incremental information is used to overwrite the corresponding description of the long-term profile layer of the user profile information; and / or, If it is determined based on the statistical parameters that the incremental information does not appear continuously, then the incremental information is incorporated into the long-term profile layer of the user profile information in a weighted manner.

[0012] According to the present invention, a long text user profile update method based on a large model is provided. The intermediate summary layer based on the user profile, and the long-term profile layer that updates the user profile information through incremental summaries by the profile agent, include: If the incremental information includes the user's display statement, then the corresponding dimension of the long-term profile layer of the user profile information is updated using the incremental information and marked as high confidence.

[0013] The present invention also provides a long text user profile update device based on a large model, comprising: The interactive text determination module is used to obtain interactive text based on the writing parameters input by the user and the writing agent; The profile information determination module is used to collect the interaction text by recording the intelligent agent and obtain structured user profile information based on the interaction text; The profile update triggering module is used to obtain the change characteristics of the user profile information by triggering the intelligent agent. If the change characteristics of the user profile information meet the triggering conditions, the profile intelligent agent is triggered. The user profile update module is used to update the user profile based on the user profile information and the profile agent; the user profile includes at least two layers based on time granularity. The writing agent, the recording agent, the triggering agent, and the portrait agent are all implemented based on a large model.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the long text user profile update method based on the large model as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the long text user profile update method based on a large model as described above.

[0016] This invention provides a method, apparatus, and electronic device for updating user profiles for long texts based on a large model. It obtains interactive text based on user-input writing parameters and a writing agent, records the interactive text collected by the agent, and obtains at least two layers of structured user profile information based on the interactive text. It triggers the agent to obtain the changing characteristics of the user profile information and activates the profile agent when triggering conditions are met. The user profile is updated based on the user profile information and the profile agent. This method can automatically collect signals and store them in a structured, layered manner during the writing interaction process, and proactively trigger profile updates when a valid change in user preferences is detected. This facilitates a dynamic balance between short-term task intent and long-term writing preferences in long text writing. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the long-text user profile update method based on a large model provided by the present invention.

[0019] Figure 2 This is one of the example diagrams illustrating the long text user profile update method based on a large model provided by the present invention.

[0020] Figure 3 This is the second example of the long text user profile update method based on a large model provided by the present invention.

[0021] Figure 4 This is the third example of the long text user profile update method based on a large model provided by the present invention.

[0022] Figure 5 This is a schematic diagram of the structure of the long text user profile update device based on a large model provided by the present invention.

[0023] Figure 6 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] Intelligent writing assistance based on large language models (LLM) has gradually become a trend. However, current intelligent writing methods usually use context-aware prompt word engineering to provide users with functions such as continuation writing, polishing, and summarizing.

[0026] User profiling technology has been widely used in recommendation systems, search engines, and dialogue systems, which can improve the user experience. Based on this, how to combine intelligent writing with user profiling is a technical problem that needs to be solved to improve the user experience of intelligent writing assistance based on large language models (LLM).

[0027] Traditional user profiles are typically constructed through explicit questionnaires, behavioral log statistics, or collaborative filtering, focusing on static modeling of interests and preferences; or by introducing persistent memory modules into dialogue systems to enhance contextual coherence across conversations. However, these techniques all have significant limitations in multi-agent writing assistance scenarios. Specifically: The static profiling approach based on explicit user input requires users to proactively declare their writing style, domain background, and other information through questionnaires or preference settings pages before using the system, and the initialization parameters are configured accordingly. While simple to implement, this approach heavily relies on user cooperation, resulting in low completion rates in practice and failing to reflect the dynamic evolution of user preferences over time.

[0028] Dialogue memory schemes based on fixed-period summaries periodically compress and store dialogue history as long-term memory for subsequent conversations. However, memory updates are typically triggered by fixed rounds or the end of a conversation, resulting in a relatively simple memory granularity. They lack hierarchical modeling of information stability and timeliness, and do not design a dedicated profile dimension system for writing assistance scenarios.

[0029] The existing solutions mentioned above lead to the following problems in practical applications: profile acquisition relies on user input, resulting in low coverage; the memory granularity is too limited to effectively distinguish between the user's short-term task intent in the current article and the long-term writing preferences accumulated across articles, leading to short-term noise interfering with long-term profile modeling, or long-term profiles covering short-term intents; the update mechanism is passive and cannot proactively perceive gradual or abrupt changes in user preferences; the profile dimension design lacks specificity for writing scenarios, making it difficult to effectively support personalized decision-making for writing-related intelligent agents.

[0030] Furthermore, multi-agent collaborative architectures are increasingly being used to solve complex problems. These architectures involve multiple specialized agents each undertaking a sub-task, leading to more refined solutions. Building upon this, this invention provides a method, apparatus, electronic device, and storage medium for updating long-text user profiles based on a large model and multi-agent collaboration. It employs a user profile generation and dynamic management scheme based on automatic collection of interactive information, three-layer hierarchical storage modeling, and multi-condition joint active triggering of updates. This achieves the technical effect of continuously generating multi-dimensional, dynamically adaptive user profiles without requiring active user configuration, and significantly improving the quality of personalized writing assistance.

[0031] Specifically, the core technical means of this invention include: automatic acquisition of multiple types of signals based on the interaction process, that is, introducing a dedicated recording agent to automatically extract explicit feedback signals, implicit behavioral signals, and content feature signals after each round of interaction and write them into a structured interaction log. This mechanism can eliminate the dependence on user active configuration and achieve zero-threshold profile information acquisition; a multi-dimensional profile dimension system and hierarchical storage structure for writing scenarios, that is, storing profile information in layers according to stability and controlling the impact of recent information on long-term profiles through time decay weights, achieving independent modeling of short-term intentions and long-term preferences; an incremental summary generation and conflict fusion mechanism with the model as the core, that is, the profile agent performs incremental summary merging by dimension and adopts a differentiated fusion strategy for conflicts between new and old information to achieve continuous iterative updates of the profile; and a multi-condition joint active triggering mechanism, that is, the triggering agent judges and activates profile updates according to the priority order of multiple triggering conditions to achieve active and passive perception of changes in user preferences.

[0032] It should be noted that the multi-agent system includes a writing agent, a recording agent, a triggering agent, and a profiling agent. Each agent is understood to possess a certain degree of autonomous decision-making capability. In this embodiment, the agents can form a multi-agent system, with each agent undertaking different sub-tasks. The agents can coordinate with each other through message passing or state sharing to jointly achieve the complex goal of writing long texts.

[0033] Figure 1 This is a flowchart illustrating the long-text user profile update method based on a large model provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps.

[0034] Step 101: Obtain interactive text based on the writing parameters input by the user and the writing agent.

[0035] Writing parameters refer to the input information provided by the user to initiate or guide the writing task. For example, writing parameters may include the writing topic, keywords, target genre, expected length, and initial draft.

[0036] Interactive text refers to the content text generated by the writing agent based on writing parameters, serving as the basis for subsequent interactions. For example, interactive text can be behavioral signals and dialogue content generated by user interaction. It is understood that interactive text is long text, specifically text exceeding a preset threshold in length. The preset threshold can be set according to actual circumstances; this embodiment does not impose any limitations on it.

[0037] It should be noted that the writing agent is an intelligent agent with content generation capabilities, and its specific implementation can be based on any applicable large language model or a dedicated generation model. Users can input writing parameters in a single input or through multi-turn dialogue to gradually clarify them; this embodiment does not impose any limitations on this.

[0038] Step 102: Collect the interactive text by recording the intelligent agent, and obtain structured user profile information based on the interactive text; the user profile information includes at least two layers.

[0039] Among them, the recording agent is the agent responsible for collecting and structurally storing behavioral signals and content features in each round of interaction in real time.

[0040] User profile information, also known as structured interaction logs, consists of at least two layers, which are at least two different levels obtained by dividing the structured user profile information at the time granularity.

[0041] It should be noted that the specific format of structured storage can be designed according to implementation requirements, and this embodiment does not limit it.

[0042] Understandably, by recording intelligent agents to achieve automated data collection and hierarchical structured storage, a data foundation can be provided to distinguish between users' short-term intentions and long-term preferences.

[0043] Step 103: Obtain the change characteristics of the user profile information by triggering the intelligent agent. If the change characteristics of the user profile information meet the triggering condition, then trigger the profile intelligent agent. The triggering agent is responsible for monitoring changes in the recording layer and deciding whether to activate the profile update based on triggering conditions or rules. The recording layer includes at least the structured user profile information obtained by the recording agent.

[0044] It should be noted that the specific measurement method for the changing characteristics and the specific threshold for the triggering conditions can be configured according to the application scenario's requirements for the sensitivity of the profile update; this embodiment does not impose any limitations on this. It is understood that by triggering the intelligent agent to proactively perceive the changing characteristics of user behavior and trigger updates as needed, the response lag problem under a fixed-period update mechanism can be avoided.

[0045] Step 104: Update the user profile based on the user profile information and the profile agent.

[0046] The writing agent, the recording agent, the triggering agent, and the portrait agent are all implemented based on a large model.

[0047] Among them, the profile agent is the agent responsible for performing incremental summarization, generating and maintaining multi-dimensional user profile documents.

[0048] A user profile is a structured description of a user's characteristics, preferences, and behavioral patterns, used to support personalized services. In this embodiment, it specifically refers to a document that continuously models multi-dimensional features such as a user's writing style, domain background, and editing preferences in a writing assistance scenario. Updating a user profile refers to the process of correcting, supplementing, or iterating on the stored user profile description based on newly collected user profile information that meets triggering conditions.

[0049] It should be noted that the updated user profile can be accessed by the writing agent and other agents in the system to generate content that matches the user's personalized preferences. The specific execution method for profile updating is not limited in this embodiment.

[0050] Understandably, by dynamically maintaining user profiles through intelligent profiling agents, the characterization of user preferences can be continuously evolved as the user experience progresses.

[0051] The long-text user profile update method based on a large model provided in this invention obtains interactive text based on user-input writing parameters and a writing agent. It records the interactive text collected by the agent and obtains at least two layers of structured user profile information based on the interactive text. It triggers the agent to obtain the change characteristics of the user profile information and triggers the profile agent when the triggering condition is met. The user profile is updated according to the user profile information and the profile agent. It can automatically collect signals and store them in a structured layered manner during the writing interaction process, and actively trigger profile updates when it detects a valid change in user preferences. This makes it easier for long-text writing to achieve a dynamic balance between short-term task intent and long-term writing preferences.

[0052] It should be noted that user profile documents can be organized by user identifier, stored hierarchically in persistent storage on the system side, and support version management. After updating the user profile, the writing agent and the dialogue agent can inject key fields of the current profile into system prompt words or obtain relevant dimensional information as needed through Retrieval-Augmented Generation (RAG) to achieve personalized response generation. Among them, Retrieval-Augmented Generation refers to a method that improves the accuracy and personalization of the generated response by retrieving relevant context from an external knowledge base and injecting prompt words when the large language model generates the response.

[0053] In some embodiments, the multi-agent system further includes a dialogue agent. The multi-agent system comprises a dialogue agent, a writing agent, a recording agent, a triggering agent, and a profiling agent. Each agent undertakes different sub-tasks and coordinates through message passing or state sharing to jointly achieve complex goals. Specifically, the dialogue agent is responsible for interacting with the user using natural language and understanding the writing intent, while the writing agent is responsible for core writing assistance tasks such as content generation, polishing, and continuation of the writing.

[0054] Specifically, such as Figure 2 As shown, users can interact with the dialogue agent through natural language. During the interaction, the dialogue agent calls the user profile from at least the long-term profile layer and sends the understood writing intent to the writing agent. The writing agent sends the generated content to the user. The recording agent records at least the interaction text between the dialogue agent, the writing agent, and the user, and stores the processed user profile information in the target location. The agent is triggered to monitor the user profile information. When the triggering condition is met, the profile agent is triggered to update at least one of the short-term memory layer, the intermediate summary layer, and the long-term profile layer to update the user profile.

[0055] In some embodiments, a vector database, also known as an interaction record, can be used to store the full amount of interactive text. The writing agent retrieves relevant memories on demand through semantic retrieval, eliminating the need for explicit summary generation and further reducing computational overhead.

[0056] Based on the above embodiments, obtaining structured and stored user profile information based on the interactive text includes: Based on the interactive text, at least one of explicit feedback signals, implicit behavioral signals, and content feature signals is extracted; The user profile information is obtained by obtaining a structured storage based on at least one of the explicit feedback signals, implicit behavioral signals, and content feature signals and a preset format.

[0057] Explicit feedback signals refer to clear responses from users to system suggestions, such as acceptance, rejection, partial adoption, or explicit preference statements.

[0058] For example, explicit preference statements can be natural language evaluations such as "too formal" or "this expression is not commensurate," or they can be user-generated modification records of the generated content.

[0059] Implicit behavioral signals refer to preference indications indirectly inferred by analyzing user behavior, such as the number of times the same content is repeatedly generated, the extent of modification of suggested text, and the rate of continued adoption of a certain type of suggestion.

[0060] It is understandable that the number of times the same content is repeatedly generated can reflect the degree of deviation between the generated quality and user expectations, and the continuous adoption rate of a certain type of suggestion can reflect the user's stable preferences.

[0061] It should be noted that the extent of modification to the suggested text can be measured by edit distance or token change rate.

[0062] Among them, content feature signals refer to style or domain features extracted from user input content, such as domain vocabulary distribution, sentence structure, paragraph structure, citation methods, and information density preferences.

[0063] It should be noted that after each round of interaction, the recording agent can extract one or more of the above three types of signals from the interaction text and the user operation sequence.

[0064] Preset formats may include fields such as timestamp, interaction type, original content summary, extracted preference signal tags, article identifier, and session identifier.

[0065] Specifically, the extraction methods for explicit feedback signals may include parsing user interface operation events or performing intent recognition on user natural language evaluations; the calculation methods for implicit behavioral signals may include counting the number of retries and calculating the text edit distance; the extraction methods for content feature signals may include domain vocabulary matching, sentence pattern template recognition, etc., which are not limited in this embodiment.

[0066] Based on any of the above embodiments, the change characteristics of the user profile information satisfy the triggering conditions, including: The user rejects the writing agent's generation suggestions during consecutive interactions exceeding a first threshold; and / or, Based on the original user profile, the amount of structured user profile information obtained by the recording agent exceeds a second threshold; and / or, Based on the temporal characteristics of the original user profile, the temporal characteristics of the latest structured user profile information obtained by the recording agent exceed a third threshold; and / or, The difference between the semantic distribution of the user profile information and the semantic distribution of the original user profile exceeds a fourth threshold.

[0067] The first, second, third, and fourth thresholds are all preset configurable parameters. For example, if a user's single modification exceeds a preset threshold, such as a token change rate exceeding 60%, it can be considered that the user rejects the writing agent's generation suggestions.

[0068] It should be noted that the four triggering conditions mentioned above correspond to four types of triggering mechanisms: event triggering, accumulation triggering, time triggering, and drift detection triggering. The first threshold is used to determine whether a user's continuous rejection behavior constitutes a systematic preference bias; the second threshold is used to control the frequency of profile updates based on usage; the third threshold is used to prevent profiles from remaining stagnant for extended periods in low-frequency usage scenarios; and the fourth threshold is used to detect whether significant semantic drift has occurred in user preferences. Each of these thresholds can be configured independently according to actual application needs. When multiple conditions are configured simultaneously, the triggering agent can make judgments according to a preset priority order; this embodiment does not impose any limitations on this.

[0069] For example, the triggering agent can make judgments based on the following rules: if the user refuses to generate suggestions in multiple consecutive rounds of interaction, indicating that the current profile has a systematic bias in its characterization of user preferences, an update is triggered immediately; if the number of newly added structured interaction log entries exceeds a preset number, an update is triggered to ensure that the user profile continues to iterate with usage; if the time since the last profile update exceeds a preset duration, an update is triggered to reduce the risk of long-term stagnation of the user profile in low-frequency usage scenarios; if the triggering agent periodically calculates the difference between the semantic distribution of the current behavior signal and the long-term profile feature description and exceeds a preset threshold, indicating that preference drift has occurred, an update is triggered to achieve proactive perception of gradual changes.

[0070] Preference drift refers to the gradual or abrupt change in user preferences over time. The difference between the semantic distribution of current behavioral signals and the long-term profile feature description can be determined based on the cosine distance of the embedding vectors.

[0071] The priority order of triggering conditions can be set to event triggering take precedence over accumulation triggering, accumulation triggering take precedence over time triggering, and time triggering take precedence over drift detection triggering.

[0072] Understandably, by using a multi-condition triggering mechanism, it is possible to simultaneously cover both sudden events and gradual shifts in user preferences, thereby improving the timeliness and accuracy of profile updates.

[0073] In some embodiments, such as Figure 3 As shown, after a user completes a round of interaction, at least one signal from explicit feedback, implicit behavior, and content features can be collected by the Agent and written into the user profile information. The Agent is then triggered to monitor changes in the user profile information and determine whether the event triggering condition ① is met. If so, the signal merging and deduplication are directly triggered. Multiple signals in the same window are merged into one update task, and the profile Agent is triggered to execute the incremental summary dimension merging + conflict fusion strategy, output the new version profile document update version number and timestamp, and then end. Otherwise, determine if the accumulation trigger condition ② is met. If so, directly trigger signal merging and deduplication, merging multiple signals in the same window into a single update task, and triggering the profile agent to execute incremental summary dimension merging + conflict fusion strategy, outputting the new profile document update version number and timestamp, and then ending. Otherwise, check if condition ③ (time triggering condition) is met. If so, directly trigger signal merging and deduplication, merging multiple signals in the same window into a single update task, and trigger the profile agent to execute incremental summary dimension merging + conflict fusion strategy, outputting the new profile document update version number and timestamp, and then ending the process. Otherwise, determine if condition ④ (drift detection) is met. If so, directly trigger signal merging and deduplication, merging multiple signals in the same window into a single update task, and trigger the profile agent to execute incremental summary dimension merging + conflict fusion strategy, outputting the new profile document update version number and timestamp, and then ending the process. Otherwise, do not update for now, wait for the next round of interaction, and then end.

[0074] In some embodiments, a fixed episode rolling update can also be used, that is, an update is automatically triggered after each article is completed or every fixed number of rounds, which reduces the difficulty of implementation.

[0075] Based on any of the above embodiments, the user profile includes a short-term memory layer and a medium-term summary layer; The step of updating the user profile based on the user profile information and the profile agent includes: During the session, the short-term memory layer of the user profile is updated by the profile agent based on the user profile information. After the session ends, the user profile agent merges the short-term memory layer of the updated user profile and updates the intermediate summary layer of the user profile; wherein, the intermediate summary layer stores the article-level preference features obtained from the merging process. The short-term memory layer stores immediate preference information at the current session level. In some embodiments, short-term task intents can be maintained separately for each long-text writing task.

[0076] Understandably, by introducing a short-term memory layer and a medium-term summary layer, the granularity of user profile information can be further refined, providing structural support for distinguishing between conversational temporary intents and cross-article stable preferences.

[0077] In some embodiments, the profiling agent may invoke a Large Language Model (LLM) to merge the short-term memory layer of the updated user profile and update the medium-term summary layer of the user profile.

[0078] In other embodiments, rule-based signal extraction methods can be used, such as counting the frequency and acceptance rate of labels in each dimension, thereby further reducing computational overhead.

[0079] In some embodiments, the user profile further includes a long-term profile layer; The step of updating the user profile based on the user profile information and the profile agent includes: Based on the intermediate summary layer of the user profile, the long-term profile layer updates the user profile information through incremental summaries by the profile agent; The dimensions of the long-term profile layer of the user profile information include at least two of the following: writing style, domain background, modification preferences, instruction style, and short-term task intent.

[0080] Incremental summarization refers to a method that processes new content and merges it with existing summaries to generate an updated summary. It can be understood that incremental summarization can reduce the computational overhead of updating user profiles.

[0081] Specifically, the writing style dimension may include formality, sentence length preference, rhetorical style, etc.; the domain background dimension may include professional domain tags, high-frequency domain vocabulary, knowledge depth preference, etc.; the revision preference dimension may include the historical acceptance rate, rejection rate, and revision magnitude distribution of different types of suggestions; the instruction style dimension may include the user's habitual instruction expression patterns; and the short-term task intent refers to the conversational information such as the topic of the current article, target audience, and expected style.

[0082] It should be noted that the profile agent can call the Large Language Model (LLM) to merge the newly added content from the intermediate summary layer into the corresponding dimension of the original user profile for fusion processing, so as to update the user profile.

[0083] Among them, the large language model is a language generation model based on the Transformer architecture and pre-trained on a large-scale corpus.

[0084] It should be noted that the user profile agent can invoke a large language model to fuse the original user profile and the newly added content from the intermediate summary layer from each dimension of the user profile. Different dimensions can have their own independent data structures and update logic. The five dimensions mentioned above are exemplary dimension combinations designed for writing assistance scenarios. In practical applications, dimension definitions can be added, deleted, or adjusted as needed; this embodiment does not impose any limitations on this.

[0085] In some embodiments, such as Figure 4 As shown, the immediate intent and temporary preferences of the user in the current session are obtained by inputting at least one of the interaction signals extracted from the interactive text, including explicit feedback signals, implicit behavioral signals, and content feature signals. These are then stored in the short-term memory layer (Session-level), such as the current article topic and target audience, the style of the immediate instructions in this session, and temporary preference statements. After the session ends, the records in the short-term memory layer are automatically merged and summarized into the intermediate summary layer. The intermediate summary layer (Article-level) summarizes the interaction records of several recent articles, retaining the preference features at the article level, such as the recent suggestion acceptance rate distribution, recent common modification types and magnitudes, and recent domain vocabulary and style tendencies. This serves as an input buffer for the long-term layer update, and incremental summary updates are performed on the long-term profile layer through accumulation triggering and other methods. The long-term profile layer (User-level) stores the stable user preference features accumulated across articles, such as writing style: formality / sentence length / rhetoric, domain background: professional tags / high-frequency words, modification preferences: acceptance rate / rejection rate / magnitude, instruction style: command / description / question, etc. These features are organized by dimension into a structured profile document for the writing agent to call. The structured profile document can be in JSON format or contain natural language descriptions.

[0086] In some embodiments, the influence of the short-term memory layer, intermediate summary layer, and long-term profile layer on the long-term profile can be controlled by time decay weights. The information flow between the three layers can be unidirectional, that is, from the short-term memory layer to the intermediate summary layer, and then to the long-term profile layer.

[0087] Understandably, by designing a multi-dimensional profiling system, it is possible to depict users' writing preferences from multiple perspectives, providing richer personalized decision-making basis for the writing intelligence agent.

[0088] Based on any of the above embodiments, before updating the long-term profile layer of the user profile information through the incremental summary of the profile agent based on the intermediate summary layer of the user profile, the method further includes: Determine the statistical parameters of the incremental information of the intermediate summary layer; The intermediate summary layer based on the user profile, and the long-term profile layer that updates the user profile information through incremental summaries by the profile agent, include: If the incremental information is determined to occur continuously based on the statistical parameters, then the incremental information is used to overwrite the corresponding description of the long-term profile layer of the user profile information; and / or, If it is determined based on the statistical parameters that the incremental information does not appear continuously, then the incremental information is incorporated into the long-term profile layer of the user profile information in a weighted manner.

[0089] The statistical parameters refer to statistical quantities used to characterize the frequency or continuity of specific information in the incremental summary.

[0090] It should be noted that continuous occurrence usually refers to the appearance of the same type of behavioral signal in multiple consecutive sessions or articles, which can reflect the stable drift of user preferences; non-continuous occurrence usually refers to occasional signals, such as a single occurrence, which has a relatively low confidence level.

[0091] It is understandable that using incremental summary information to cover the corresponding description of the original user profile can quickly respond to the real migration of user preferences. By integrating the incremental summary information into the long-term profile layer of user profile information in a weighted manner to obtain the user profile, it is possible to avoid excessive disturbance to the long-term profile caused by single noise.

[0092] It should be noted that the specific calculation method for the weighting weights may include fixed weight coefficients, time decay functions, or dynamic weights based on signal strength, etc., and this embodiment does not limit this.

[0093] Understandably, by distinguishing between continuous and sporadic signals and adopting differentiated fusion strategies, it is possible to balance sensitivity and stability during the image update process.

[0094] Based on any of the above embodiments, the intermediate summary layer based on the user profile, and the long-term profile layer that updates the user profile information through incremental summaries by the profile agent, include: If the incremental information includes the user's display statement, then the corresponding dimension of the long-term profile layer of the user profile information is updated using the incremental information and marked as high confidence.

[0095] Explicit declarations are user preferences or modification instructions that are clearly expressed through language or actions.

[0096] It should be noted that since explicit statements have a high degree of certainty regarding user intent, they can be directly used to update user profiles without the need for frequency accumulation and can be given a high confidence label, thus distinguishing them from preference features inferred from implicit behavior.

[0097] It is understandable that by granting explicit declarations priority update rights in this embodiment, it is possible to ensure that user-expressed preferences are adopted in a timely and accurate manner.

[0098] In some embodiments, the user profile type is a user profile document, which contains natural language descriptions and structured fields for various dimensions, and records version number and update timestamp, thereby providing support for version backtracking.

[0099] For example, in some embodiments, after receiving an update instruction, the profile agent can read the summary content added since the last profile update in the intermediate summary layer, as well as the current version of the long-term profile document; using the large language model as the core processing module, it extracts and summarizes the incremental information one by one according to the profile dimension, generates a fused new description, and outputs a new version of the user profile document.

[0100] In summary, the long-text user profile update method based on a large model provided by this invention can eliminate the dependence on user-initiated configuration and lower the threshold for obtaining user profiles by automatically collecting data on the interaction process of the intelligent agent; through a hierarchical storage structure, it can achieve independent modeling of short-term intentions and long-term preferences, avoiding mutual interference between information at different time granularities; through a multi-condition joint active triggering mechanism, it can promptly perceive sudden changes and gradual shifts in user preferences, ensuring the timeliness of the profile; and through a model-driven incremental summarization and conflict fusion strategy, it can achieve continuous iteration of the profile.

[0101] Therefore, this invention can continuously generate accurate, multi-dimensional, and dynamically adaptive user profiles without requiring active user participation, significantly improving the personalization capabilities and long-term service quality of multi-agent writing assistance systems.

[0102] The following describes the long text user profile update device based on a large model provided by the present invention. The long text user profile update device based on a large model described below can be referred to in correspondence with the long text user profile update method based on a large model described above.

[0103] Figure 5 This is a schematic diagram of the structure of the long text user profile update device based on a large model provided by the present invention, as shown below. Figure 5 As shown, the device includes: The interactive text determination module 510 is used to obtain interactive text based on the writing parameters input by the user and the writing agent. The profile information determination module 520 is used to collect the interaction text by recording the intelligent agent and obtain structured user profile information based on the interaction text; The profile update triggering module 530 is used to obtain the change characteristics of the user profile information by triggering the intelligent agent. If the change characteristics of the user profile information meet the triggering conditions, the profile intelligent agent is triggered. User profile update module 540 is used to update the user profile based on the user profile information and the profile intelligent agent; the user profile includes at least two layers based on time granularity. The writing agent, the recording agent, the triggering agent, and the portrait agent are all implemented based on a large model.

[0104] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a long-text user profile update method based on a large model. This method includes: obtaining interactive text based on user-input writing parameters and a writing agent; collecting the interactive text through a recording agent and obtaining structured user profile information based on the interactive text; obtaining the change characteristics of the user profile information through a triggering agent; if the change characteristics of the user profile information meet the triggering condition, triggering the profile agent; updating the user profile according to the user profile information and the profile agent; the user profile includes at least two layers based on time granularity; wherein the writing agent, the recording agent, the triggering agent, and the profile agent are all implemented based on a large model.

[0105] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the long text user profile update method based on a large model provided by the above methods. The method includes: obtaining interactive text based on writing parameters input by the user and a writing agent; collecting the interactive text through a recording agent and obtaining structured user profile information based on the interactive text; obtaining the change characteristics of the user profile information through a triggering agent; if the change characteristics of the user profile information meet the triggering condition, triggering the profile agent; updating the user profile according to the user profile information and the profile agent; the user profile includes at least two layers based on time granularity; wherein the writing agent, the recording agent, the triggering agent, and the profile agent are all implemented based on a large model.

[0107] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the long text user profile update method based on a large model provided by the above methods. The method includes: obtaining interactive text based on writing parameters input by the user and a writing agent; collecting the interactive text through a recording agent and obtaining structured user profile information based on the interactive text; obtaining the change characteristics of the user profile information through a triggering agent, and triggering the profile agent if the change characteristics of the user profile information meet the triggering condition; updating the user profile according to the user profile information and the profile agent; the user profile includes at least two layers based on time granularity; wherein the writing agent, the recording agent, the triggering agent, and the profile agent are all implemented based on a large model.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for updating long-text user profiles based on a large model, characterized in that, include: The interactive text is obtained based on the writing parameters input by the user and the writing agent. The system records the interactive text collected by the intelligent agent and obtains structured user profile information based on the interactive text. The change characteristics of the user profile information are obtained by triggering the intelligent agent. If the change characteristics of the user profile information meet the triggering conditions, the profile intelligent agent is triggered. The user profile is updated based on the user profile information and the profile agent; the user profile includes at least two layers based on time granularity. The writing agent, the recording agent, the triggering agent, and the portrait agent are all implemented based on a large model.

2. The method for updating long-text user profiles based on a large model according to claim 1, characterized in that, The structured user profile information obtained based on the interactive text includes: Based on the interactive text, at least one of explicit feedback signals, implicit behavioral signals, and content feature signals is extracted; The user profile information is obtained by obtaining a structured storage based on at least one of the explicit feedback signals, implicit behavioral signals, and content feature signals and a preset format.

3. The method for updating long-text user profiles based on a large model according to claim 1, characterized in that, The changes in the user profile information satisfy the triggering conditions, including: The user rejects the writing agent's generation suggestions during consecutive interactions exceeding a first threshold; and / or, Based on the original user profile, the amount of structured user profile information obtained by the recording agent exceeds a second threshold; and / or, Based on the temporal characteristics of the original user profile, the temporal characteristics of the latest structured user profile information obtained by the recording agent exceed a third threshold; and / or, The difference between the semantic distribution of the user profile information and the semantic distribution of the original user profile exceeds a fourth threshold.

4. The method for updating long-text user profiles based on a large model according to claim 1, characterized in that, The user profile includes a short-term memory layer and a medium-term summary layer; The step of updating the user profile based on the user profile information and the profile agent includes: During the session, the short-term memory layer of the user profile is updated by the profile agent based on the user profile information. After the session ends, the user profile agent merges the short-term memory layer of the updated user profile and updates the intermediate summary layer of the user profile; wherein, the intermediate summary layer stores the article-level preference features obtained from the merging process.

5. The method for updating long-text user profiles based on a large model according to claim 4, characterized in that, The user profile also includes a long-term profile layer; The step of updating the user profile based on the user profile information and the profile agent includes: Based on the intermediate summary layer of the user profile, the long-term profile layer updates the user profile information through incremental summaries by the profile agent. The dimensions of the long-term profile layer of the user profile information include at least two of the following: writing style, domain background, modification preferences, instruction style, and short-term task intent.

6. The method for updating long-text user profiles based on a large model according to claim 5, characterized in that, Before updating the long-term profile layer of the user profile information through incremental summarization by the profile agent based on the intermediate summary layer of the user profile, the method further includes: Determine the statistical parameters of the incremental information of the intermediate summary layer; The intermediate summary layer based on the user profile, and the long-term profile layer that updates the user profile information through incremental summaries by the profile agent, include: If the incremental information is determined to occur continuously based on the statistical parameters, then the incremental information is used to overwrite the corresponding description of the long-term profile layer of the user profile information; and / or, If it is determined based on the statistical parameters that the incremental information does not appear continuously, then the incremental information is incorporated into the long-term profile layer of the user profile information in a weighted manner.

7. The method for updating long-text user profiles based on a large model according to claim 5, characterized in that, The intermediate summary layer based on the user profile, and the long-term profile layer that updates the user profile information through incremental summaries by the profile agent, include: If the incremental information includes the user's display statement, then the corresponding dimension of the long-term profile layer of the user profile information is updated using the incremental information and marked as high confidence.

8. A device for updating long-text user profiles based on a large model, characterized in that, include: The interactive text determination module is used to obtain interactive text based on the writing parameters input by the user and the writing agent; The profile information determination module is used to collect the interaction text by recording the intelligent agent and obtain structured user profile information based on the interaction text; The profile update triggering module is used to obtain the change characteristics of the user profile information by triggering the intelligent agent. If the change characteristics of the user profile information meet the triggering conditions, the profile intelligent agent is triggered. The user profile update module is used to update the user profile based on the user profile information and the profile agent; the user profile includes at least two layers based on time granularity. The writing agent, the recording agent, the triggering agent, and the portrait agent are all implemented based on a large model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the long text user profile update method based on a large model as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the long text user profile update method based on a large model as described in any one of claims 1 to 7.

Images