Brand-oriented generative engine optimization perception large language model question word generation method and device, computer equipment and storage medium

Abstract

The application provides a brand-oriented generative engine optimization perception large language model question word generation method and device, computer equipment and storage medium. The method comprises the following steps: obtaining the brand official website page text of a target brand, and extracting the field content of a plurality of core fields from the brand official website page text; performing semantic expansion based on the field content to obtain a semantic expansion vocabulary set; classifying each vocabulary in the vocabulary set under a preset multi semantic dimension classification system, determining the semantic dimension to which each vocabulary belongs, and obtaining the vocabulary set under each semantic dimension; generating a plurality of question words based on the vocabulary set under each semantic dimension through a large language model; and different question words are used to guide the large language model to mention the target brand when generating answers in different guiding modes. The method can generate high-quality large language model question words, so that the large language model can efficiently and naturally mention the target brand when generating answers.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, and storage medium for generating question words for a large language model optimized for brand generative engines. Background Technology

[0002] In brand marketing and generative recommendation optimization scenarios, companies urgently need to use carefully designed question words to encourage large models to naturally mention their own brands during interactions with users, thereby increasing the brand's exposure and influence in AI-generated content.

[0003] Currently, question word generation technologies targeting specific goals (such as increasing brand mention) often rely on rule-based synonym substitution and sentence transformation techniques. These techniques generate variations by superficially replacing basic question words. However, when applied to the specific business objective of brand exposure, rule-based synonym substitution only performs superficial word replacements. While the generated question words may be diverse in form, they struggle to accurately convey the brand's differentiated value proposition and fail to efficiently guide large models to proactively and naturally incorporate brand information when generating answers.

[0004] Therefore, the existing technology suffers from the technical problem of low-quality question words generated by large language models. Summary of the Invention

[0005] Based on this, the purpose of this application is to at least solve one of the above-mentioned technical defects, especially the technical defect of low quality of large language model question words generated in the prior art. This application provides a method, apparatus, computer device and computer-readable storage medium for generating high-quality large language model question words to guide the large language model to efficiently and naturally mention the target brand when generating answers. This is an optimization method for perception of large language model question words for brand generative engines.

[0006] Firstly, this application provides a method for generating question words for a large language model that optimizes perception for brand generative engines. The method includes: Obtain the text of the target brand's official website page, and extract the content of several core fields from the text. Based on the content of each field, semantic expansion is performed to obtain a semantically expanded vocabulary set; Under the pre-defined multi-semantic dimension classification system, each word in the vocabulary set is classified to determine the semantic dimension to which each word belongs, so as to obtain the vocabulary set under each semantic dimension. Multiple question words are generated using a large language model based on vocabulary sets across various semantic dimensions. Different question words are used to guide the large language model to mention the target brand when generating answers in different ways.

[0007] In one exemplary embodiment, the content of several core fields is extracted from the text of a brand's official website page, including: The text on the brand's official website page is divided into multiple text blocks, and the text content of each text block is vectorized to obtain the text block vector of each text block. Determine the similarity between each text block vector and the preset query word vector, and based on the similarity between each text block vector and the preset query vector, determine at least one relevant text block from each text block; The text content of each relevant text block is input into the large language model to identify the field content of each core field.

[0008] In one exemplary embodiment, semantic expansion is performed based on the content of each field to obtain a semantically expanded vocabulary set, including: Using the content of each field as query keywords, multiple related terms are retrieved from multiple semantic databases; By summarizing each core field and each related word, a semantically expanded vocabulary set is obtained.

[0009] In an exemplary embodiment, under a preset multi-semantic dimension classification system, each word in the vocabulary set is classified to determine the semantic dimension to which each word belongs, so as to obtain a vocabulary set under each semantic dimension, including: Each word is vectorized to obtain its word vector; Determine the semantic centroid vector of each semantic dimension in the pre-defined multi-semantic dimension classification system; the semantic centroid vector represents the semantic center of the semantic dimension. Based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension, the semantic dimension to which each word belongs is determined, so as to obtain the word set under each semantic dimension.

[0010] In one exemplary embodiment, the semantic dimension to which each word belongs is determined based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension, including: The initial semantic dimension to which each word belongs is determined based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension. Using a large language model, each word is classified and reasoned through a pre-defined question template. The initial semantic dimension of each word is corrected to determine the semantic dimension of each word.

[0011] In one exemplary embodiment, a large language model generates multiple question words based on vocabulary sets across various semantic dimensions, including: Obtain a pre-constructed user cognition parameter matrix; the user cognition parameter matrix contains user cognition parameters corresponding to various user personas; Select at least one set of user cognitive parameters from the user cognitive parameter matrix, and determine the instruction priority of different semantic dimensions based on the selected user cognitive parameters; The instruction priorities of different semantic dimensions and the vocabulary sets under different semantic dimensions are combined into question word generation instructions and input into the large language model to obtain question words that match the selected user cognitive parameters.

[0012] In one exemplary embodiment, the method further includes: Obtain user action feedback data for multiple question terms, and construct a preference dataset based on the user action feedback data; Fine-tuning of a large language model using a preference dataset.

[0013] Secondly, this application provides a large language model question word generation device for optimizing perception in brand generative engines, the device comprising: The acquisition module is used to acquire the text of the target brand's official website page and extract the content of multiple core fields from the text of the brand's official website page; The expansion module is used to semantically expand the vocabulary based on the content of each field, resulting in a semantically expanded vocabulary set. The classification module is used to classify each word in the vocabulary set under a preset multi-semantic dimension classification system, determine the semantic dimension to which each word belongs, and obtain the vocabulary set under each semantic dimension. The generation module is used to generate multiple question words based on the vocabulary set under each semantic dimension using a large language model; different question words are used to guide the large language model to mention the target brand when generating the answer in different ways.

[0014] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0016] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: This application provides a method, apparatus, computer device, and storage medium for generating question words using a large language model for optimizing perception in brand generative engines. The method involves acquiring the text of a target brand's official website page and extracting the content of multiple core fields from the text. Semantic expansion is performed on each field to obtain a semantically expanded vocabulary set. Under a pre-defined multi-semantic dimension classification system, each word in the vocabulary set is classified to determine its semantic dimension, resulting in a vocabulary set for each semantic dimension. A large language model is then used to generate multiple question words based on these semantic dimension vocabulary sets. Different question words are used to guide different behaviors. This approach guides the large language model to mention the target brand when generating answers. By extracting core information from the brand's official website, semantically expanding it, and then categorizing it according to multiple semantic dimensions, a diverse range of question words is generated. This constructs a systematic brand information injection solution, solving the problem that traditional manually written question words are often monotonous and rigid, failing to cover the brand's multi-dimensional communication points. This results in answers generated by the large language model being awkwardly embedded and lacking persuasiveness. The solution generates high-quality, highly guiding, and naturally diverse question words, enabling the large language model to proactively and organically integrate brand information when answering various questions, thereby significantly improving brand exposure and communication depth. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for generating query words from a large language model that optimizes perception for a brand-generating engine, as shown in one embodiment. Figure 2 This is a schematic diagram of a system architecture in one embodiment; Figure 3 This is a flowchart illustrating a method for generating query words for a large language model that optimizes perception for a brand-generating engine, as shown in another embodiment. Figure 4 This is a structural block diagram of a large language model question word generation device for optimizing perception for a brand generative engine in one embodiment. Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Existing rule-based and synonym substitution-based divergent generation techniques primarily rely on pre-defined linguistic rules or statistical probabilities. Their core logic involves manipulating core "seed words" using a thesaurus (such as WordNet) or word embedding space. For example, the EDA (Easy Data Augmentation) algorithm can rapidly generate numerous variant texts by performing synonym substitutions (e.g., replacing "buy" with "get"), random insertions, or word order swaps on the original question.

[0021] Existing Automatic Prompt Engineering (APE) techniques aim to find the "optimal instruction" through model self-iteration. The architecture typically includes two large models: a "generator" and an "evaluator." Based on given input data and the standard answer (Ground Truth), the system automatically generates multiple sets of candidate question words, scores and filters them according to the accuracy of the model's output, and through iterative optimization, ultimately retains the single question word with the highest score.

[0022] Existing reinforcement learning text generation techniques based on general linguistic metrics utilize reinforcement learning (RL) to optimize text generation strategies. The core logic involves setting a reward function based on linguistic statistical features. Positive feedback is given when the generated text satisfies general linguistic metrics (such as fluency, grammatical correctness, and perplexity) or general human preference models. This technique primarily aims to improve the readability and coherence of the model-generated text, making it conform to basic human reading habits.

[0023] The problem with existing technologies is that the core optimization objectives of existing APE and General Reinforcement Learning (GRL) are usually limited to "linguistic metrics" (such as perplexity and fluency) or "general task accuracy." Their reward models are typically built based on general human preferences (such as HH-RLHF). In the optimization scenario of brand generative engines, the optimization objective is not simply sentence fluency, but brand visibility and user adoption intent. Due to the lack of business feedback signals from vertical domains, the model can only learn to "write sentences fluently," but cannot understand higher-order optimization intents such as "getting the large model to mention a certain brand more often" or "conforming to specific industry slang." This application can solve the problems existing in the prior art; please refer to the detailed descriptions of the following embodiments.

[0024] The Generative Engine Optimization (GEO) involved in this application refers to a methodology for optimizing information retrieval within a generative AI-driven ecosystem. This involves collaborative optimization of content strategy, technology adaptation, and user interaction to ensure that information, products, and services of enterprises or individuals are accurately identified and efficiently indexed by generative engines (such as search engines, content platforms, and intelligent recommendation systems with generative capabilities). These products or services then reach target users in the form of "personalized generated content," ultimately achieving optimization in traffic acquisition, user retention, and business conversion.

[0025] In one exemplary embodiment, such as Figure 1 As shown, a method for generating question words from a large language model for optimizing perception in a brand generative engine is provided. Taking the application of this method to a server as an example, the method includes the following steps S102 to S108. Wherein: Step S102: Obtain the text of the target brand's official website page and extract the content of multiple core fields from the text of the official website page.

[0026] The target brand refers to the specific brand entity that needs to be marketed or promoted.

[0027] The brand website page text refers to the raw text data scraped or obtained from the brand's official website, containing content such as brand introduction, product description, brand story, and core values. For example, it could be the raw HTML text of the brand's official website.

[0028] Among them, the core fields refer to the key information categories extracted from the official website text, such as Industry, Category, Brand, and Product, which can provide high-purity seed features for subsequent steps.

[0029] The field content is the specific text corresponding to each core field.

[0030] Optionally, the server first obtains the text of the target brand's official website page. Then, it can use information extraction algorithms or large language models to extract the core fields such as the brand's industry, category, brand, and products from the official website page text.

[0031] Step S102 above addresses the "extraction illusion" problem caused by non-core text such as advertisements and navigation bars on brand website pages by performing physically isolated data cleaning. In practical applications, in addition to obtaining the text from the brand website page, it is also possible to obtain the brand name and supplementary brand information. Supplementary brand information can refer to brand-related introductions such as main business directions and products.

[0032] Step S104: Semantic expansion is performed based on the content of each field to obtain a semantically expanded vocabulary set.

[0033] The vocabulary set is a summary of all words obtained after semantic expansion. In this application, the semantically expanded vocabulary set includes the original field content, i.e., the original word roots, as well as related long-tail word roots (including competitor words, scenario words, efficacy words, etc.) obtained by expanding the original word roots.

[0034] Optionally, the server expands the content of each field to obtain a richer vocabulary set.

[0035] Step S106: Under the preset multi-semantic dimension classification system, classify each word in the vocabulary set, determine the semantic dimension to which each word belongs, and obtain the vocabulary set under each semantic dimension.

[0036] The pre-defined multi-semantic dimension classification system refers to a framework system for semantically classifying words, which is pre-constructed based on the core needs of brand communication and user cognitive habits. This system contains multiple independent dimensions with clear semantic orientations.

[0037] For example, it can include nine semantic dimensions: competitors, scenarios, target audience, efficacy, region, product, category, industry, and brand.

[0038] Optionally, the server maps each word in the vocabulary set to a preset multi-semantic dimension classification system, such as classifying reliability into the quality dimension and passion into the emotion dimension, thereby forming a subset of vocabulary organized according to semantic dimensions.

[0039] Step S108: Based on the vocabulary set under each semantic dimension, the large language model generates multiple question words; different question words are used to guide the large language model to mention the target brand when generating the answer in different ways.

[0040] The question word refers to the text instruction used to guide the large language model to generate specific content. In this application, it refers to the text instruction that can guide the large language model to naturally incorporate brand information.

[0041] Optionally, the server calls a large language model to generate diverse question words that can guide the model to naturally mention the brand when answering.

[0042] In the aforementioned method for generating query terms using a large language model to optimize perception for brand generative engines, the process involves obtaining the text of the target brand's official website page and extracting the content of multiple core fields. Semantic expansion is then performed on each field to obtain a semantically expanded vocabulary set. Under a pre-defined multi-semantic dimension classification system, each word in the vocabulary set is categorized to determine its semantic dimension, resulting in a vocabulary set for each semantic dimension. Finally, a large language model generates multiple query terms based on these semantic dimension-specific vocabulary sets. Different query terms are used to guide the large language model in different ways. Mentioning the target brand when generating answers; thus, by extracting core information from the brand's official website and semantically expanding it, then classifying it according to multiple semantic dimensions, and finally generating diverse question words, a systematic brand information injection solution is constructed. This solves the problem that traditional manually written question words are often monotonous and rigid, unable to cover the brand's multi-dimensional communication points, resulting in stiff and unconvincing answers generated by the big language model. It can generate high-quality, highly guiding, and naturally diverse question words, enabling the big language model to actively and organically integrate brand information when answering various questions, thereby significantly improving the brand's exposure and communication depth.

[0043] In one exemplary embodiment, extracting the content of multiple core fields from the text of a brand's official website page includes: dividing the text of the brand's official website page into multiple text blocks, and vectorizing the text content of each text block to obtain a text block vector for each text block; determining the similarity between each text block vector and a preset query term vector, and based on the similarity between each text block vector and the preset query vector, identifying at least one related text block from each text block; and inputting the text content of each related text block into a large language model to identify the content of each core field.

[0044] Among them, text blocks are semantic segments obtained by dividing the original HTML text of the brand's official website into paragraphs, sentences, or fixed lengths.

[0045] Among them, the text block vector is a dense vector converted from a text block by an embedding model (such as BERT) to represent its semantics.

[0046] Among them, the preset query term vector is a vectorized representation of representative words (such as company profile and core products) defined in advance for each core field, which is used to retrieve relevant text blocks.

[0047] Among them, relevant text blocks refer to text fragments that have high semantic similarity to the query terms and may contain core field content.

[0048] Optionally, the server divides the brand's official website page text into multiple text blocks. Then, it uses a text embedding model to convert each text block into a text block vector. At the same time, it prepares one or more preset query term vectors (such as company profile, core products), calculates the cosine similarity between each text block vector and each preset query term vector, and retains the top-K most relevant text blocks as related text blocks. Finally, it concatenates the text content of these related text blocks and uses it as context input to the large language model, and provides instructions for extracting core fields: "Please extract the brand's industry, category, brand, products and other information from the following context text, so as to accurately extract structured field content."

[0049] In this embodiment, text fragments that may contain core fields are quickly located by similarity retrieval between text block vectors and preset query word vectors. Then, a large language model is used for refined extraction, achieving efficient and accurate extraction of core information. This solves the problems of low efficiency and poor accuracy when blindly extracting directly from long web page texts. It utilizes both the speed of vector retrieval and the semantic understanding advantages of LLM, providing high-quality and concise field content input for subsequent semantic expansion, which is the foundation for generating high-quality query terms.

[0050] In one exemplary embodiment, semantic expansion is performed based on the content of each field to obtain a semantically expanded vocabulary set, including: using the content of each field as query keywords to query multiple related words in multiple semantic databases; and summarizing each core field and each related word to obtain a semantically expanded vocabulary set.

[0051] Among them, semantic databases refer to external semantic association databases, which may store semantic relationships between words (such as synonyms, near-synonyms, hyponyms, associations, etc.), such as synonym dictionaries, WordNet, concept graphs, etc.

[0052] Among them, related words are words that are semantically related to the query keywords and are obtained by querying the semantic database.

[0053] Optionally, the server uses the content of each field obtained as an index anchor point, and accesses multiple external semantic association databases through a wide area network data aggregation interface to aggregate a massive number of related long-tail keywords (including competitor keywords, scenario keywords, efficacy keywords, etc.), expanding the "small sample official definition" into a "massive market semantic cloud", that is, forming a semantically expanded vocabulary set.

[0054] In this embodiment, by utilizing multiple semantic databases to expand the core vocabulary of the brand, the semantic expression related to the brand is greatly enriched. This solves the problem that relying solely on the original official website text has a limited vocabulary and cannot cover the diverse language habits and expression scenarios of users. The semantically expanded vocabulary set contains more potential user search terms and natural expressions, providing sufficient corpus for the subsequent generation of diverse question words. This allows the generated question words to be integrated into brand information in a way that is closer to user expression, improving the naturalness and coverage of brand mentions.

[0055] In an exemplary embodiment, under a preset multi-semantic dimension classification system, each word in the vocabulary set is classified to determine the semantic dimension to which each word belongs, so as to obtain a vocabulary set under each semantic dimension. This includes: vectorizing each word to obtain a word vector; determining the semantic centroid vector of each semantic dimension in the preset multi-semantic dimension classification system; the semantic centroid vector represents the semantic center of the semantic dimension; and determining the semantic dimension to which each word belongs based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension, so as to obtain a vocabulary set under each semantic dimension.

[0056] Among them, the word vector is the vector representation obtained by transforming words through word embedding models (such as Word2Vec, GloVe, BERT).

[0057] The semantic centroid vector is a vector obtained by averaging the vectors of representative words in each semantic dimension. It is used to represent the core semantic direction or semantic center of that semantic dimension.

[0058] For example, for the "scene" dimension, representative scene words such as "office," "outdoor travel," and "family gathering" can be collected in advance. The semantic centroid vector of the "scene" dimension is obtained by averaging the vectors of these words. In this application, the semantic centroid vector will be updated in real time according to the new words added to the semantic dimension.

[0059] Optionally, the server first generates a word vector for each word in the vocabulary set using a pre-trained word vector model. At the same time, for each preset semantic dimension, it obtains the semantic centroid vector of that dimension and calculates the similarity (such as cosine similarity) between each word vector and each semantic centroid vector. For each word, the dimension with the highest similarity is selected as its semantic dimension. Finally, the words of the same dimension are aggregated together to form a vocabulary set under each semantic dimension.

[0060] In this embodiment, the automatic and refined dimensional classification of words is achieved through vectorization and centroid calculation. The semantic centroid can accurately grasp the core semantics of each dimension, so that words can be accurately classified into the dimension that best reflects their brand value. This semantic space-based classification provides a scientific and reasonable structural basis for the subsequent generation of different question words, ensuring that the question words of each dimension can focus on the core connotation of that dimension, thereby enhancing the accuracy of brand information transmission.

[0061] In an exemplary embodiment, determining the semantic dimension to which each word belongs based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension includes: determining the initial semantic dimension to which each word belongs based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension; using a large language model to perform classification reasoning on each word through a preset question template, correcting the initial semantic dimension to which each word belongs, and determining the semantic dimension to which each word belongs.

[0062] The initial semantic dimension is a preliminary classification result obtained based on vector similarity calculation.

[0063] The preset question template is a fixed instruction format used to guide the large language model to perform semantic dimension classification. For example, "Please determine which dimension the following word '{word}' is most suitable to be classified into, and briefly explain the reason."

[0064] Optionally, the server first performs vector similarity calculation to obtain an initial semantic dimension for each word. Considering that the vector method may have biases or ambiguities, a large language model is introduced for secondary verification. Each word is substituted into a preset question template to construct a specific classification request, which is then submitted to the large language model. Based on its deep understanding of the semantics of the words, the large language model outputs a dimension judgment and its reasoning. The output of the large language model is compared with the initial semantic dimension. If they match, the judgment is adopted; if they do not match, further analysis is performed to determine the final corrected semantic dimension.

[0065] In this embodiment, the accuracy of word classification is improved through a dual mechanism of initial vector judgment and large-scale model refinement. This ensures that each word is ultimately assigned to a dimension that truly matches its semantic meaning, laying a solid and reliable foundation for the subsequent generation of question words and further enhancing the quality and guiding effect of the question words.

[0066] In an exemplary embodiment, a large language model generates multiple question words based on vocabulary sets under various semantic dimensions, including: obtaining a pre-constructed user cognitive parameter matrix; the user cognitive parameter matrix contains user cognitive parameters corresponding to various user personas; selecting at least one set of user cognitive parameters from the user cognitive parameter matrix, and determining the instruction priority of different semantic dimensions based on the selected user cognitive parameters; combining the instruction priorities of different semantic dimensions and the vocabulary sets under different semantic dimensions into question word generation instructions and inputting them into the large language model to obtain question words that match the selected user cognitive parameters.

[0067] In this user cognition parameter matrix, each element represents a user parameter corresponding to a user persona. The user cognition parameter matrix summarizes user parameters with different cognitive granularities and consumer psychology (e.g., industry expert / beginner, price-sensitive / quality-oriented). In the two-dimensional coordinate system corresponding to the user cognition parameter matrix of this application, the horizontal axis represents cognitive granularity, involving five cognitive granularities from Level 1 (beginner) to Level 5 (expert), and the vertical axis represents consumer psychology, involving two consumer psychology types: Type A (price-sensitive) and Type B (quality-oriented).

[0068] Among them, instruction priority is the different levels of importance assigned to different semantic dimensions (such as quality, price, and environmental protection) based on the characteristics of the current user persona, which is used to guide the large language model on which dimensions to focus on when generating question words.

[0069] The question word generation instruction is a comprehensive prompt that combines priority, vocabulary set, and generation requirements.

[0070] Optionally, the server first loads the user cognition parameter matrix, selects a coordinate point from the matrix (e.g., expert and price-sensitive), and calls the corresponding persona description template library. Then, based on the "expert + price-sensitive" persona parameters, it adds weight to the quality semantic dimension, i.e., it determines that instructions for the quality semantic dimension have higher priority and instructions for other language dimensions have lower priority. Next, it integrates the priority information and the corresponding vocabulary set for each dimension into a structured question word generation instruction and submits it to the large language model. The large language model then generates multiple question words that conform to the persona characteristics and can naturally integrate brand information.

[0071] In this embodiment, by introducing a user cognitive parameter matrix and dynamic priority calculation, highly personalized and precise targeting of question word generation is achieved. It can intelligently adjust the weight of each semantic dimension according to the psychological characteristics and concerns of the target group, thereby generating question words that are most likely to resonate with the group.

[0072] In the question word generation process of this application, K-Means clustering and the maximum boundary correlation algorithm (MMR) are also introduced to calculate the cosine similarity between newly generated question words and historical question words in the vector space in real time. This forces question words generated in different batches to maintain the maximum semantic distance, eliminates pseudo-diversity results with highly repetitive semantics, and ensures that the seed bank can cover the breadth of the long-tail semantic space.

[0073] To facilitate understanding by those skilled in the art, the following uses a structured vector after lexical classification processing as an example to illustrate the specific method for generating question terms in this application. The structured vector is "{"Related Root Popularity": 13125, "Related Root": How is the evaluation of coffee brand A?, "Original Category": Category, "Original Brand": Coffee brand A, "Original Root": Coffee, "Related Word Classification": Brand Word}". In the process of generating question terms, firstly, based on the constructed user cognition parameter matrix, the horizontal axis (cognition granularity) includes Level 1 (beginner) to Level 5 (expert), and the vertical axis (consumer psychology) includes Type_A (price-sensitive) - Type_B (quality-oriented). Then, a coordinate point is randomly selected from the user cognition parameter matrix (e.g., expert + price-sensitive), and the corresponding persona description template library is called. Then, based on the randomly selected parameters, the question term is generated... Structured features are "differentiated packaging." For example, an expert persona assigns high weight to the original word root (coffee), guiding the question to focus on technical details; a novice persona assigns high weight to related word categories (brand words), guiding the question to focus on social evaluation. Then, the persona of "industry expert + quality orientation" is injected to generate dynamic question words such as "You are now a coffee quality appraiser with 10 years of experience (persona injection). Please combine the brand 'A Coffee Brand' (structured feature) and focus on professional perspectives such as the roasting consistency of coffee beans and the expression of regional flavors (quality orientation weight) to provide an in-depth analysis of the question 'How would you rate A Coffee Brand?'" These question words can then be used together with the user's original question to input into the large language model.

[0074] In one exemplary embodiment, the method further includes: acquiring user operation feedback data for multiple question words, and constructing a preference dataset based on the user operation feedback data; and fine-tuning a large language model using the preference dataset.

[0075] User action feedback data includes both explicit and implicit feedback. Explicit feedback refers to the user's "activation" action, which is considered a strong positive validation by the system, directly confirming the semantic accuracy and industry depth of the query terms. Implicit feedback refers to the user's "deletion" action, which is considered a negative validation, marking it as invalid or low-quality content.

[0076] The preference dataset is a set of training samples built based on user operation feedback data to represent users' preferences for different question words.

[0077] Optionally, the server presents the generated multiple question words to the user in the form of a list. The user, as the final decision-maker, performs activation or deactivation operations on the question words in the list. The server cleans and structures the user interaction data in real time, constructs "Prompt Preference Pairs", marks the question words activated by the user as Golden Samples, and marks the question words deactivated by the user as Rejected Samples. This forms a high-quality preference dataset with clear industry characteristics and user aesthetic preferences. Based on the preference dataset, supervised fine-tuning or preference alignment of the large language model is triggered periodically.

[0078] In this embodiment, real feedback data is used to fine-tune the large language model. The model learns "what the user selected" and "what the user discarded" and automatically adjusts the weights of its internal parameters. As the frequency of user use increases, the generated question words will become more and more in line with industry jargon and questioning habits, achieving a low-cost adaptive upgrade from "cold start" to "personalized expert".

[0079] Figure 2 A system architecture diagram is provided, such as Figure 2 The system architecture presented in this application adopts a dual-closed-loop architecture, decoupling and deploying each logical unit through microservice design to ensure the stability and scalability of the system when processing massive amounts of brand data. The system supports cloud-native distributed deployment, with core components divided into four independently scalable clusters that communicate asynchronously via high-throughput message queues.

[0080] The intelligent parsing and vectorization layer deploys distributed data cleaning and vectorization nodes, semantic filtering service nodes, and entity extraction and inference service nodes. The distributed data cleaning and vectorization nodes integrate a block-based vectorization unit, responsible for physically isolating and slicing the original text of the brand's official website. The semantic filtering service node runs a semantic retrieval filtering unit, calculating similarity based on preset high-confidence query vectors and removing noisy paragraphs. The entity extraction and inference service node runs a core entity extraction unit, using a large model to accurately parse four core fields: industry, category, brand, and product, providing clean seed features for downstream applications.

[0081] The semantic computation and generation layer deploys dynamic feature computation nodes, parameterized generation and inference nodes, and vector deduplication service nodes. The dynamic feature computation nodes carry multi-source semantic mapping units and dynamic centroid classification units, are responsible for accessing wide area network data interfaces, executing online learning mechanisms, and periodically calculating the distribution positions and cluster centroids of semantic anchor points in the vector space. The parameterized generation and inference nodes run multi-dimensional cognitive parameter injection units, fusing structured features with user cognitive parameters and inputting them into the large language model. The vector deduplication service nodes run vector space semantic deduplication units, performing MMR algorithms or K-Means clustering in memory to remove redundant features from the generated results.

[0082] The dynamic knowledge and feature storage layer deploys a high-dimensional vector and metadata hybrid storage cluster as the underlying data support of the system, replacing the traditional single database architecture. Specifically, the vector feature storage area stores text block vectors output by the intelligent parsing and vectorization layer, as well as dynamic centroid vectors constructed by the semantic computation and generation layer. The structured metadata area stores brand attributes, product associations, and long-tail keyword tags parsed by the semantic computation and generation layer.

[0083] The interactive feedback and model evolution layer deploys an interactive feedback gateway and an automated fine-tuning pipeline. The interactive feedback gateway interfaces with the front-end interface, runs interactive probe activation and annotation logic, and captures the user's "activate" and "discard" commands in real time. The automated fine-tuning pipeline integrates a preference dataset automatic construction and generation module for model-oriented fine-tuning. It is responsible for converting accumulated high-quality and rejection samples into training data, periodically triggering supervised fine-tuning or direct preference optimization tasks, and realizing targeted updates and adaptive upgrades of model weights.

[0084] The information flow within the system strictly follows a "dual closed-loop" mechanism, consisting of a forward data flow and a reverse feedback flow, achieving a fully automated closed loop of "parsing-building-generating-evolution".

[0085] The forward data flow is the core business process, specifically: original information of the brand to be analyzed (brand name, official website URL, supplementary information) → distributed data cleaning and vectorization node (block vectorization) → semantic filtering service node (noise removal) → entity extraction and reasoning service node (four-dimensional analysis of industry, category, brand, and product) → dynamic feature calculation node (multi-source semantic mapping and dynamic centroid classification) → parameterized generation and reasoning node (cognitive parameter injection and question word generation) → vector deduplication service node (redundant feature removal) → output layer (structured probe problem).

[0086] The reverse feedback flow is a self-evolving process, specifically: front-end probe matrix interacts with data → interactive feedback gateway (captures "activate" and "discard" instructions) → automatic construction module of preference dataset (labels high-quality samples and rejection sample pairs) → automated fine-tuning pipeline (triggers supervised fine-tuning or direct preference optimization training tasks) → parameterized generation of inference nodes (model weight redirection update) → improvement of semantic alignment capability of core generation layer → optimization of forward data flow generation accuracy.

[0087] Compared with the prior art, the technical solution of this application has the following significant advantages: First, based on physical isolation cleaning and geometric prior guidance, this invention addresses the "input illusion" and "concept drift" problems that traditional RAG or EDA methods easily encounter when processing multi-source heterogeneous web page data. It eliminates advertising and navigation bar noise through a semantic retrieval filtering unit, ensuring that only Top-K core paragraphs are input into the large model, avoiding misattribution caused by contextual pollution in general models. Utilizing "geometric prior retrieval" and "strong / weak consensus update" mechanisms, it solves the problem of misjudging new words due to knowledge lag in pre-trained models, enabling the system to automatically update centroid vectors as new market concepts emerge, ensuring the timeliness and accuracy of brand classification. Overall, through block vectorization, physical isolation cleaning, and a dynamic centroid mechanism guided by geometric priors, it effectively achieves source blocking of noise and adaptive feature alignment.

[0088] Secondly, the system boasts groundbreaking semantic space expansion and multi-dimensional cognitive differentiation generation capabilities. It goes beyond limited official website data with external long-tail keywords (competitors, scenarios, efficacy), expanding brand characteristics into a massive market semantic cloud and significantly broadening the scope of perception. By introducing a dynamic cognitive matrix (expert / novice, price / quality, etc.) and combining it with the MMR algorithm, the system forces different batches of query terms to maintain maximum semantic distance within the vector space. This ensures that the generated probe matrix not only has broad coverage but also exhibits extremely high cognitive granularity, effectively simulating the complex and ever-changing user search intent in the real world. Overall, by utilizing multi-source semantic mapping combined with multi-dimensional cognitive parameter injection and a vector space semantic deduplication mechanism, the system solves the serious "semantic collapse" and "homogenization" problems inherent in traditional APE (Automatic Prompt Engineering) technology-generated Prompts.

[0089] Third, the system leverages user feedback data to drive preference alignment and model self-evolution. By capturing users' explicit actions (activation / deactivation) of Prompts, a preference dataset containing Golden / Rejected samples is automatically constructed. Based on this dataset, Supervised Fine-Tuning (SFT) or Preference Alignment (DPO) is periodically triggered, enabling the general-purpose model to quickly learn industry-specific jargon and users' questioning habits. As usage frequency increases, the system automatically evolves from a "cold start" state into a "personalized expert" that understands the business and pain points, solving the pain point of general-purpose models being difficult to implement in vertical fields. Overall, by introducing "interactive probe activation" and "SFT / DPO targeted fine-tuning pipeline" into the question word optimization closed loop, this system establishes an evolutionary mechanism based on user business preferences, unlike existing technologies that only focus on language fluency.

[0090] In another embodiment, such as Figure 3 As shown, a method for generating question words from a large language model for optimizing perception in a brand generative engine is provided. Taking the application of this method to a server as an example, the method includes the following steps: Step S302: Obtain the text of the target brand's official website page and extract the content of multiple core fields from the text of the official website page.

[0091] Step S304: Semantic expansion is performed based on the content of each field to obtain a semantically expanded vocabulary set.

[0092] Step S306: Under the preset multi-semantic dimension classification system, classify each word in the vocabulary set, determine the semantic dimension to which each word belongs, and obtain the vocabulary set under each semantic dimension.

[0093] Step S308: Obtain a pre-constructed user cognition parameter matrix; the user cognition parameter matrix contains user cognition parameters corresponding to various user personas.

[0094] Step S310: Select at least one set of user cognitive parameters from the user cognitive parameter matrix, and determine the instruction priority of different semantic dimensions based on the selected user cognitive parameters.

[0095] Step S312: Combine the instruction priorities of different semantic dimensions and the vocabulary sets under different semantic dimensions into question word generation instructions and input them into the large language model to obtain question words that match the selected user cognitive parameters; different question words are used to guide the large language model to mention the target brand when generating answers in different ways.

[0096] It should be noted that the specific limitations of the above steps can be found in the above description of the specific limitations of a method for generating question words for a large language model that optimizes perception for a brand generative engine.

[0097] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0098] The following describes the large language model question word generation device for brand generative engine optimization perception provided in the embodiments of this application. The large language model question word generation device for brand generative engine optimization perception has the same inventive concept as the above-described large language model question word generation method for brand generative engine optimization perception. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the large language model question word generation device for brand generative engine optimization perception provided below can be referred to the limitations of the large language model question word generation method for brand generative engine optimization perception described above. The large language model question word generation device for brand generative engine optimization perception described below and the large language model question word generation method for brand generative engine optimization perception described above can be referred to each other, and will not be repeated here.

[0099] In one exemplary embodiment, Figure 4 This application provides a schematic diagram of the structure of a large language model question word generation device for optimizing perception in a brand generative engine, as shown in the embodiments of this application. Figure 4 As shown, the large language model question word generation device for optimizing perception in a brand generative engine includes: an acquisition module 402, an expansion module 404, a classification module 406, and a generation module 408, wherein: The acquisition module 402 is used to acquire the text of the target brand's official website page and extract the content of multiple core fields from the text of the brand's official website page; The expansion module 404 is used to perform semantic expansion based on the content of each field to obtain a semantically expanded vocabulary set. The classification module 406 is used to classify each word in the vocabulary set under a preset multi-semantic dimension classification system, determine the semantic dimension to which each word belongs, and obtain the vocabulary set under each semantic dimension. The generation module 408 is used to generate multiple question words based on the vocabulary set under each semantic dimension through the large language model; different question words are used to guide the large language model to mention the target brand when generating the answer in different ways.

[0100] In an exemplary embodiment, the acquisition module 402 is specifically used to segment the text of the brand official website page into multiple text blocks, and to vectorize the text content of each text block to obtain the text block vector of each text block; to determine the similarity between each text block vector and the preset query word vector, and to determine at least one related text block from each text block based on the similarity between each text block vector and the preset query vector; and to input the text content of each related text block into the large language model to identify the field content of each core field.

[0101] In an exemplary embodiment, the expansion module 404 is specifically used to use the content of each field as query keywords to query multiple related words in multiple semantic databases; and to summarize each core field and each related word to obtain a semantically expanded word set.

[0102] In an exemplary embodiment, the classification module 406 is specifically used to perform vectorization processing on each word to obtain the word vector of each word; determine the semantic centroid vector of each semantic dimension in the preset multi-semantic dimension classification system; the semantic centroid vector represents the semantic center of the semantic dimension; and determine the semantic dimension to which each word belongs based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension, so as to obtain the word set under each semantic dimension.

[0103] In an exemplary embodiment, the classification module 406 is specifically used to determine the initial semantic dimension to which each word belongs based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension; to perform classification reasoning on each word using a large language model through a preset question template, to correct the initial semantic dimension to which each word belongs, and to determine the semantic dimension to which each word belongs.

[0104] In an exemplary embodiment, the generation module 408 is specifically used to obtain a pre-constructed user cognitive parameter matrix; the user cognitive parameter matrix contains user cognitive parameters corresponding to various user personas; select at least one set of user cognitive parameters from the user cognitive parameter matrix, and determine the instruction priority of different semantic dimensions based on the selected user cognitive parameters; combine the instruction priorities of different semantic dimensions and the vocabulary sets under different semantic dimensions into question word generation instructions and input them into the large language model to obtain question words that match the selected user cognitive parameters.

[0105] In one exemplary embodiment, the apparatus further includes: a fine-tuning module, configured to acquire user operation feedback data for multiple question words, and construct a preference dataset based on the user operation feedback data; and fine-tune a large language model using the preference dataset.

[0106] The modules in the aforementioned large language model question word generation device optimized for brand generative engines can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0107] In one exemplary embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the brand generative engine-optimized perception large language model question word generation methods described above.

[0108] In one exemplary embodiment, this application also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the brand generative engine-oriented perception-based large language model question word generation methods described above.

[0109] In one exemplary embodiment, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the brand generative engine-optimized perception large language model question word generation methods described in the above embodiments.

[0110] Indicatively, such as Figure 5 As shown, Figure 5 This is a schematic diagram of the internal structure of a computer device 500 provided in an embodiment of this application. The computer device 500 can be provided as a server. (Refer to...) Figure 5 The computer device 500 includes a processing component 502, which further includes one or more processors, and memory resources represented by memory 501 for storing instructions, such as application programs, that can be executed by the processing component 502. The application programs stored in memory 501 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 502 is configured to execute instructions to perform the brand generative engine-optimized perception large language model question word generation method of any of the above embodiments.

[0111] The computer device 500 may also include a power supply component 503 configured to perform power management of the computer device 500, a wired or wireless network interface 504 configured to connect the computer device 500 to a network, and an input / output (I / O) interface 505. The computer device 500 may operate on an operating system stored in memory 501, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0112] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0113] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0114] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0115] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0116] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating question words for a large language model that optimizes perception for brand generative engines, characterized in that, The method includes: Obtain the text of the target brand's official website page, and extract the content of multiple core fields from the text of the official website page; Based on the content of each field, semantic expansion is performed to obtain a semantically expanded vocabulary set; Under the preset multi-semantic dimension classification system, each word in the vocabulary set is classified to determine the semantic dimension to which each word belongs, so as to obtain the vocabulary set under each semantic dimension. Multiple question words are generated using a large language model based on the vocabulary sets under each semantic dimension; different question words are used to guide the large language model to mention the target brand when generating answers in different ways.

2. The method according to claim 1, characterized in that, The field content extracted from the text of the brand's official website page includes: The text of the brand's official website page is divided into multiple text blocks, and the text content of each text block is vectorized to obtain the text block vector of each text block. Determine the similarity between each text block vector and the preset query word vector, and based on the similarity between each text block vector and the preset query vector, determine at least one related text block from each text block; The text content of each of the relevant text blocks is input into the large language model to identify the field content of each of the core fields.

3. The method according to claim 1, characterized in that, The semantic expansion based on the content of each field, resulting in a semantically expanded vocabulary set, includes: Using the content of each field as query keywords, multiple related terms are retrieved from multiple semantic databases; By summarizing the core fields and related words, a semantically expanded vocabulary set is obtained.

4. The method according to claim 1, characterized in that, The step of classifying each word in the vocabulary set under a preset multi-semantic dimension classification system, determining the semantic dimension to which each word belongs, and obtaining a vocabulary set under each semantic dimension includes: Each of the aforementioned words is vectorized to obtain a word vector for each of the aforementioned words; Determine the semantic centroid vector of each semantic dimension in the preset multi-semantic dimension classification system; the semantic centroid vector represents the semantic center of the semantic dimension; Based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension, the semantic dimension to which each word belongs is determined, so as to obtain the word set under each semantic dimension.

5. The method according to claim 4, characterized in that, The step of determining the semantic dimension to which each word belongs based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension includes: The initial semantic dimension to which each word belongs is determined based on the similarity between the word vector of each word and the semantic centroid vector of each semantic dimension. The large language model is used to classify and reason about each word using a preset question template, and the initial semantic dimension of each word is corrected to determine the semantic dimension of each word.

6. The method according to claim 1, characterized in that, The process involves generating multiple question words based on the vocabulary sets under each semantic dimension using a large language model, including: Obtain a pre-constructed user cognition parameter matrix; the user cognition parameter matrix contains user cognition parameters corresponding to various user personas; Select at least one set of user cognitive parameters from the user cognitive parameter matrix, and determine the instruction priority of different semantic dimensions based on the selected user cognitive parameters; The instruction priorities of different semantic dimensions and the vocabulary sets under different semantic dimensions are combined into a question word generation instruction and input into the large language model to obtain question words that match the selected user cognitive parameters.

7. The method according to claim 1, characterized in that, The method further includes: Obtain user operation feedback data for the multiple question words, and construct a preference dataset based on the user operation feedback data; The large language model is fine-tuned using the preference dataset.

8. A large language model question word generation device for optimizing perception in brand generative engines, characterized in that, The device includes: The acquisition module is used to acquire the text of the target brand's official website page and extract the content of multiple core fields from the text of the official website page. An expansion module is used to semantically expand the vocabulary based on the content of each field to obtain a semantically expanded vocabulary set. The classification module is used to classify each word in the vocabulary set under a preset multi-semantic dimension classification system, determine the semantic dimension to which each word belongs, and obtain the vocabulary set under each semantic dimension. The generation module is used to generate multiple question words based on the vocabulary sets under each semantic dimension using a large language model; different question words are used to guide the large language model to mention the target brand when generating answers in different ways.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images