Intelligent media channel inquiry, valuation and negotiation method based on large language model

Abstract

The application relates to the technical field of large language models, and discloses a method for intelligent media channel inquiry, valuation and negotiation based on a large language model, which comprises the following steps: collecting multi-source unstructured media quotation data and performing pretreatment; performing deep semantic analysis on the multi-source unstructured media quotation data by using a large LLM model to obtain a standardized structured quotation data table; constructing a quantitative channel portrait model from multiple dimensions according to the standardized structured quotation data table; constructing and calling an RAG knowledge base to obtain historical transaction bottom prices of the same type of channel and industry market conditions; introducing five valuation factors, calculating the channel quotation by using a multi-factor weighted fusion algorithm, and performing price rationality determination; and generating exclusive inquiry rhetoric, negotiation emphasis, concession negotiation strategies and high-performance-price-ratio alternative channel echelon through the LLM model reasoning. By using the application scheme, the automation degree, the fair pricing and the long-term adaptation capability of the advertising media inquiry business can be comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, specifically to an intelligent media channel inquiry, valuation, and negotiation method based on a large language model. Background Technology

[0002] With the rapid development of diverse advertising media channels such as short video platforms, text-based self-media, information flow influencers, and offline outdoor media, channel inquiry, value assessment, and business negotiation before advertising placement have become routine core business of brand marketing procurement. Media channels within the industry output quotations in various formats, encompassing a large amount of unstructured data such as business chat texts, customized quotation documents, promotional posters, and information displayed on account homepages.

[0003] Currently, the industry's common data parsing methods mainly rely on regular expressions and fixed rule templates, supplemented by basic price comparison tools for simple price comparisons. Meanwhile, artificial intelligence technologies such as LLM (Large Language Model) and RAG (Retrieval-Augmented Generation) are gradually being applied in the advertising and marketing field. General-purpose large language models possess basic text semantic extraction capabilities, while RAG technology can leverage vector storage to achieve historical data retrieval to assist model inference. However, existing technical solutions are all general-purpose designs, lacking customized adaptation for vertical pricing scenarios in advertising media. They lack dedicated structured parsing logic for non-standard media pricing, multi-dimensional channel quantitative valuation models, and a dedicated knowledge accumulation system for pricing and negotiation; they can only achieve basic text processing and simple price sorting functions. The existing technical solutions mainly suffer from the following problems: First, unstructured, non-standard quotations are difficult to parse automatically, resulting in low accuracy in information extraction. Quotations from mainstream media channels (Douyin, Xiaohongshu, Video Accounts, WeChat Official Accounts, outdoor media, and information flow influencers) are all in the form of free text, chat screenshots, and non-standard documents, lacking standardized quotation fields. Traditional technologies rely on regular expressions and fixed rule templates for data extraction, which cannot adapt to the massive amounts of non-standard, conversational, and customized quotation scripts. This leads to problems such as missing or incorrect field extraction, failure to identify implicit clauses, and extremely poor automated parsing capabilities.

[0004] Secondly, channel value lacks a multi-dimensional quantitative model, and price valuation relies on manual experience. Existing inquiry systems only support comparison of fixed list prices and cannot combine multiple factors such as channel account weight, fan verticality, interaction authenticity, audience matching, release schedule popularity, and market premium for comprehensive valuation. New sales personnel cannot quickly grasp inquiry standards, and the industry lacks quantitative and standardized intelligent valuation algorithms, making it impossible to guarantee the fairness of quotations.

[0005] Secondly, the lack of scenario-based negotiation algorithms prevents the automatic generation of negotiation strategies and alternative channels. Traditional price comparison systems only sort prices from highest to lowest and cannot intelligently generate differentiated inquiry scripts, negotiation margins, and alternative channel tiers based on user budgets, product categories, marketing goals, and market conditions, thus failing to achieve intelligent business decision support.

[0006] Finally, business experience cannot be accumulated, and the system lacks self-learning and iterative capabilities. The existing system's data is static and fixed, making it impossible to accumulate industry experience such as the bottom price range of various channels, pricing strategies, discount rules, and hidden consumption terms. It also lacks self-learning and self-optimization capabilities, and as media channel rules iterate, the accuracy of traditional algorithms continues to decline.

[0007] To address these issues, we have developed an intelligent media channel inquiry, valuation, and negotiation method based on a large language model, which solves the aforementioned technical problems. Summary of the Invention

[0008] This invention provides an intelligent media channel inquiry, valuation, and negotiation method based on a large language model. It comprehensively improves the automation, pricing fairness, and long-term adaptability of advertising media inquiry business from four technical dimensions: data analysis, quantitative valuation, intelligent decision-making, and experience iteration.

[0009] Therefore, the present invention provides the following technical solution: A method for intelligent media channel inquiry, valuation, and negotiation based on a large language model, the method comprising: Step 1: Collect multi-source unstructured media pricing data and preprocess the multi-source unstructured media pricing data; Step 2: Use a domain-adjusted LLM large model to perform deep semantic parsing on the preprocessed multi-source unstructured media pricing data to obtain a standardized structured pricing data table; Step 3: Based on the standardized structured quotation data table, construct a quantitative channel profile model from multiple dimensions; Step 4: Build a RAG knowledge base and retrieve the RAG knowledge base to obtain the historical transaction floor price and industry trends of similar channels; Step 5: Introduce five valuation factors, use a multi-factor weighted fusion algorithm to calculate channel quotations, and conduct a price reasonableness assessment; Step 6: Generate customized inquiry scripts, negotiation focus, concession negotiation strategies, and a high-cost-performance alternative channel tier through LLM model reasoning to complete the inquiry analysis and valuation calculation.

[0010] Optionally, in step 1, the collected multi-source unstructured media pricing data includes: chat text, pricing documents, poster copy, and channel homepage information. The multi-source unstructured media pricing data is then preprocessed with data cleaning, deduplication, and noise reduction to remove invalid and redundant information.

[0011] Optionally, in step 2, after loading the media inquiry-specific schema field library and the business-specific prompt model, the domain-adjusted LLM model is invoked to perform deep semantic parsing on the preprocessed multi-source unstructured media quotation data. The LLM model extracts key fields such as channel type, number of followers, rate card price, schedule, additional benefits, and hidden charges according to the preset schema field library, and performs error correction operations such as ambiguity cleaning, redundancy removal, and field validation on the key field extraction results. Then, the key field extraction results after error correction enter the data qualification judgment: if the validation fails, the unqualified key field extraction results are returned to the LLM model to perform deep semantic parsing, and the key field extraction is re-executed; if the validation passes, the key field extraction results after error correction are output as the standardized structured quotation data table.

[0012] Optionally, in step 3, based on the standardized structured pricing data table, a quantitative channel profile model is constructed from six dimensions: account verticality, fan audience matching degree, content interaction authenticity, historical campaign ROI, account industry weight, and violation risk, to generate a channel quality score and risk level.

[0013] Optionally, in step 4, when constructing the RAG knowledge base, firstly, the system receives the input data of the new inquiry business in this round and performs query text vectorization encoding on the input inquiry business data; secondly, it performs RAG retrieval based on vector similarity, retrieving historical channel transaction data from the vector database as reference context; it then assembles the RAG retrieval results with the original query text to construct a complete reasoning prompt, assisting the LLM large model in performing accurate reasoning and completing the parsing and valuation calculation of this inquiry; finally, it performs a quality assessment on the reasoning results: if the assessment is excellent, the quotations, valuations, and negotiation cases generated in this business are organized into structured knowledge fragments, vectorized, and then fed back into the vector database; if the assessment is poor, a manual review mechanism is triggered.

[0014] Optionally, in step 5, the five valuation factors include the integrated benchmark base price coefficient, the period popularity fluctuation coefficient, the channel quality premium coefficient, the risk deduction coefficient, and the industry category adaptation coefficient. The integrated benchmark base price coefficient is generated based on the historical transaction base prices of similar channels in the RAG knowledge base and serves as the basis for valuation calculation. When historical data is sufficient, the weighted average of historical transaction base prices is taken. When historical data is insufficient, the publicly listed price of the same category industry is multiplied by a preset industry discount coefficient, with a value range of [0,1]. The aforementioned period popularity fluctuation coefficient is a positive gain coefficient generated by combining the peak and off-peak seasons of the release period and the supply and demand of platform traffic. The higher the popularity, the larger the value of the period popularity fluctuation coefficient, and the value range is [0,1]. The channel quality premium coefficient is generated based on a comprehensive score of channel account verticality, content interaction authenticity, historical ROI and account industry weight. The higher the channel quality, the higher the channel quality premium coefficient, with a value range of [0,1]. The risk deduction coefficient is for violation risk, including account violation records, traffic fraud risk and performance instability risk. A negative correction coefficient is generated by comprehensive scoring. The higher the risk, the larger the risk deduction coefficient, with a value range of [0,1]. The industry category fit coefficient is generated by combining the overlap of advertising promotion category tags, the matching degree of fan audience, and the industry's general pricing rules. The higher the fit, the larger the industry category fit coefficient, with a value range of [0,1].

[0015] Optionally, step 5 includes: Step 51: After collecting the original values ​​of the five major valuation factors, the improved S-curve extreme value compression normalization process is used to eliminate dimensional differences, suppress extreme heat or extreme risk outlier disturbances, and converge all factors to the [0,1] interval. Step 52: The dynamic weights of the five valuation factors are solved using the entropy weight-AHP subjective-objective fusion method. First, an AHP judgment matrix is ​​constructed based on business experience in the advertising industry to solve the subjective experience weights of each of the five valuation factors. Second, the objective data weights of each of the five valuation factors are solved based on the information entropy of historical inquiry sample data. Finally, a balance coefficient λ is introduced, and the dynamic weights of each of the five valuation factors are obtained by fusing the subjective experience weights and objective data weights of each of the five valuation factors. Step 53: Perform differentiated correction on the five valuation factors. For positive gain factors, namely the fusion benchmark floor price coefficient, the period popularity fluctuation coefficient, the channel quality premium coefficient, and the industry category fit coefficient, use power gain correction to amplify the price gain brought by high-quality channels, popular periods, and highly matched categories. For negative risk factors, namely the risk deduction coefficient, set three threshold segments for deduction correction. Step 54: Perform adaptive weighted fusion calculation and complete the price reasonableness judgment; First, substitute the five valuation factors after differentiation correction, the dynamic weights of the five valuation factors, the risk deduction coefficient, and the industry fine-tuning coefficient into the hierarchical fusion valuation model to output the standardized valuation coefficient. Combine this with the historical transaction floor price of the same type of channel obtained from the RAG knowledge base to obtain the fair basic transaction price; Compare and verify the fair basic transaction price with the industry average price of the same period. If the price fluctuation exceeds the reasonable threshold, the process returns to step 53, and the risk deduction coefficient and the channel quality premium coefficient are called back. Factor differentiation correction, adaptive weighted fusion calculation, and price reasonableness judgment are re-executed until the price is in a reasonable range; If the price is in a reasonable range, the fair inquiry range, the maximum negotiation ceiling, and the suggested transaction floor price are output, and risk warnings are marked for channels with abnormal premiums.

[0016] Optionally, step 6 includes: Based on the user's input of the advertising budget, promotion goals, advertising cycle and industry category, combined with real-time market conditions, the system uses an LLM big data model to generate exclusive inquiry scripts, negotiation focus, discount negotiation strategies and a high-cost-performance alternative channel tier. The newly added inquiry and transaction data will be synchronously fed back to the RAG knowledge base, and a quality assessment will be performed on the data: if the assessment is excellent, the quotations, valuations, and negotiation cases generated in this transaction will be organized into structured knowledge fragments, vectorized and encoded, and then fed back into the vector database to update the factor mean, standard deviation, and historical inquiry and transaction sample pool of the five valuation factors, and dynamically update the dynamic weights of the five valuation factors to achieve continuous adaptive optimization of the valuation algorithm; if the assessment is poor, a manual review mechanism will be triggered.

[0017] Optionally, the method further includes, after step 6, outputting a channel price comparison report and a valuation analysis report, and generating channel risk warnings for scenarios such as bundled consumption, high price premiums, and false price quotes during promotional periods.

[0018] A smart media channel pricing, valuation, and negotiation device based on a large language model, the device comprising: The data acquisition and preprocessing unit acquires multi-source unstructured media pricing data and preprocesses the multi-source unstructured media pricing data. The semantic parsing unit uses a domain-adjusted LLM large model to perform deep semantic parsing on the preprocessed multi-source unstructured media pricing data to obtain a standardized structured pricing data table. The channel profile model building unit constructs a quantitative channel profile model from multiple dimensions based on the standardized structured quotation data table. The RAG knowledge base construction unit builds the RAG knowledge base and retrieves the RAG knowledge base to obtain the historical transaction floor price and industry trends of similar channels. The channel pricing calculation unit introduces five valuation factors, uses a multi-factor weighted fusion algorithm to calculate channel prices, and conducts a price reasonableness assessment. The results generation unit uses an LLM large model to generate exclusive inquiry scripts, negotiation focus points, concession negotiation strategies, and a high-cost-performance alternative channel tier, completing the inquiry analysis and valuation calculation.

[0019] This invention provides an intelligent media channel pricing, valuation, and negotiation method based on a large language model. It addresses the shortcomings of existing media channel pricing tools, such as weak non-standard price parsing capabilities, reliance on human experience for valuation, lack of intelligent negotiation decision-making capabilities, and inability to accumulate and iterate industry experience. It integrates four core capabilities: large language model non-standard semantic parsing, multi-factor channel quantitative valuation, scenario-based AI negotiation generation, and RAG knowledge base self-iteration, constructing an end-to-end intelligent media channel pricing algorithm system. On one hand, it achieves automatic structuring of non-standard prices, objective quantification of channel prices, and intelligent negotiation strategy output, significantly reducing reliance on manual intervention and improving pricing efficiency and accuracy. On the other hand, it relies on a vector knowledge base to continuously accumulate and iterate industry pricing experience, effectively adapting to various new media channels and changing pricing rules, ultimately achieving standardized, intelligent, and sustainable optimization of the pricing business. Compared with existing technologies, this invention has the following technical advantages: First, this invention constructs a media inquiry-specific field schema and a vertical scenario-customized prompt. Relying on a finely tuned large language model to complete deep semantic parsing, it automatically identifies information across all dimensions, such as channel prices, schedules, additional benefits, and hidden charges. This enables automated and structured extraction of multi-source non-standard quotation text, solving the problems of poor adaptability, missing or incorrect field extraction, and difficulty in identifying hidden charge clauses in traditional rule extraction schemes. It overcomes the technical shortcomings of existing inquiry tools that rely on regular expressions and fixed templates to extract quotation information, which cannot adapt to unstructured materials such as chat logs, non-standard quotation documents, and colloquial quotations, resulting in poor automated parsing effects. It effectively overcomes the recognition limitations of fixed rules and achieves unified and standardized conversion of various media quotations.

[0020] Secondly, this invention establishes a six-dimensional channel profiling system, integrates five major valuation correction factors to construct a weighted valuation algorithm, and builds a multi-dimensional quantitative valuation model. This eliminates the subjectivity of manual price inquiry and valuation, ensures that the quoted price is fair and quantifiable, and solves the problems of existing systems that only support simple comparison of list prices, cannot comprehensively price based on multiple indicators such as channel quality, schedule popularity, audience matching, and cooperation risks, have inconsistent judgment standards among business personnel, and are difficult for newcomers to learn. It reduces reliance on manual experience, automatically calculates the fair price range for channels, and achieves standardization of media channel price inquiry and valuation.

[0021] Third, this invention, based on large-scale model semantic reasoning and combined with real-time market conditions, generates customized negotiation plans and channel tier lists. It automatically generates differentiated pricing strategies according to the specific business scenario, addressing the shortcomings of existing systems that lack intelligent business decision-making support. This overcomes the limitations of traditional price comparison tools, which can only sort prices from highest to lowest and cannot combine advertiser budgets, promotional categories, and marketing objectives to generate negotiation scripts, pricing concessions, and tiered alternative channels. It achieves intelligent decision-making support throughout the entire price inquiry and negotiation process.

[0022] Fourth, we will build a RAG inquiry vector knowledge base to vectorize and persistently store the data from each round of inquiries. Through similarity retrieval to assist model reasoning, we will form a closed-loop iterative system with continuous data feedback and autonomous model optimization. This will enable the automatic accumulation of inquiry business experience and the autonomous iterative optimization of the system, solving the defects of traditional systems such as static data solidification and continuous decay of adaptability. It will also overcome the technical problems of existing tools, such as the inability to retain business data such as channel bottom prices, discount rules, and implicit consumption terms, the lack of self-learning ability, and the continuous decline in recognition and valuation accuracy after media platform rule updates. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a method for intelligent media channel inquiry, valuation and negotiation based on a large language model in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent media channel inquiry, valuation and negotiation system based on a large language model in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the iterative optimization process for constructing a RAG knowledge base in a specific embodiment of the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] like Figure 1 As shown, Figure 1 This is a flowchart of a method for intelligent media channel price inquiry, valuation, and negotiation based on a large language model, according to a specific embodiment of the present invention. The method includes: Step 1: Collect multi-source unstructured media pricing data and preprocess the multi-source unstructured media pricing data.

[0028] The collected multi-source unstructured media pricing data includes chat texts, pricing documents, poster copy, and information published on channel homepages. Data cleaning, deduplication, and noise reduction preprocessing are performed to remove invalid and redundant information, providing a standard data source for model analysis.

[0029] Step 2: Use a domain-adjusted LLM model to perform deep semantic analysis on the preprocessed multi-source unstructured media pricing data to obtain a standardized structured pricing data table.

[0030] Based on the preprocessed multi-source unstructured media pricing data, after loading the media inquiry-specific schema (structured data field template) field library and business-specific prompt model hints, a domain-fine-tuned LLM model is invoked to perform deep semantic parsing on the multi-source unstructured media pricing data. The model automatically extracts key fields such as channel type, number of followers, rate card price, publication schedule, additional benefits, and hidden charges according to the preset schema, and performs ambiguity cleaning, redundancy removal, and field validation and error correction on the extraction results. After execution, the data validity is checked: if the validation fails, it returns to the LLM semantic parsing stage to re-execute field extraction; if the validation passes, a standardized structured pricing data table is output.

[0031] Step 3: Perform LLM semantic analysis based on the standardized structured pricing data table. Based on the LLM semantic analysis capability, construct a quantitative channel profile model from six dimensions: account verticality, fan audience matching degree, content interaction authenticity, historical ROI (Return on Investment), account industry weight, and violation risk. Generate a channel quality score, i.e., channel quality premium coefficient C, and a risk level, i.e. risk deduction coefficient D.

[0032] Step 4: Build a RAG knowledge base and retrieve historical bottom prices and industry trends for similar channels from the RAG knowledge base.

[0033] When constructing the RAG knowledge base, it relies on a vector database to achieve persistent knowledge storage and similar content retrieval, with an iterative optimization process as follows: Figure 3 As shown. The process receives new inquiry data input and performs query text vectorization encoding on the input data. Based on vector similarity, it performs RAG retrieval, retrieving historical channel transaction data from the vector database as contextual reference. The retrieval results are then combined with the original query text to construct a complete inference prompt, assisting the LLM in accurate inference and completing the inquiry parsing and valuation calculation. A quality assessment is performed on the inference results: if the assessment is excellent (field extraction completeness ≥ 95% and valuation deviation ≤ 10%), the quotations, valuations, and negotiation cases generated in this transaction are organized into structured knowledge fragments, vectorized, and then fed back into the vector database; if the assessment is poor, a manual review mechanism is triggered. This process ends, and the returned knowledge data becomes new content in the vector database, participating in subsequent retrieval and inference, achieving continuous iterative expansion of the knowledge base and optimization of model inference accuracy.

[0034] Step 5: Introduce five valuation factors, use a multi-factor weighted fusion algorithm to calculate channel quotations, and conduct a price reasonableness assessment.

[0035] The algorithm designs a multi-factor weighted valuation method that integrates the benchmark base price coefficient, the period popularity fluctuation coefficient, the channel quality premium coefficient, the risk deduction coefficient, and the industry category suitability coefficient. Through weighted calculation, it outputs the fair price range, the maximum negotiation ceiling, and the optimal transaction price for a single channel, thereby achieving standardized intelligent valuation that reduces reliance on human experience.

[0036] The five valuation factors specifically include: Benchmark Base Price Coefficient A: Generated based on historical transaction base prices from the same platform and product category channels in the RAG knowledge base, serving as the basis for valuation calculations. When historical data is sufficient, the weighted average of historical transaction base prices is used (the more recent the time, the higher the weight); when historical data is insufficient (cold start scenario), the publicly listed price of the same product category is multiplied by a preset industry discount coefficient. Value range: [0,1].

[0037] Release Date Popularity Coefficient B: This coefficient is a positive gain factor generated by combining the peak and off-peak seasons of the release period with the supply and demand of traffic on the platform. The higher the popularity, the larger the coefficient value. The value range is [0,1].

[0038] Channel quality premium coefficient C: This is generated based on a comprehensive score of account verticality, content interaction authenticity, historical campaign ROI, and account industry weight. The higher the channel quality, the higher the premium. The value range is [0,1].

[0039] Risk deduction coefficient D: A negative correction coefficient generated based on a comprehensive score of violation risk (account violation records, volume fraud risk, and unstable performance risk). The higher the risk, the greater the deduction intensity. Value range: [0,1].

[0040] Industry category fit coefficient E: This coefficient is generated by combining the overlap of advertising category tags, the matching degree of fan audience, and the industry's general pricing rules. The higher the fit, the larger the coefficient. The value range is [0,1].

[0041] The multi-factor weighted fusion algorithm specifically includes: Step 51: Factor nonlinear normalization preprocessing.

[0042] After collecting the original values ​​of the five major factors, an improved S-curve extreme value compression and normalization process was used to eliminate dimensional differences and suppress outlier disturbances caused by extreme heat / extreme risk, thus converging all factors to the [0,1] interval and outputting the normalized factors. A', B', C', D', E' .

[0043] The general normalization formula is:

[0044] in: For the original factor value of type i, The normalized factor values ​​(stable convergence to [0,1]); This is the historical mean of the factor. The factor standard deviation is used to achieve adaptive calibration based on historical data, with a minimum value not lower than 0.001; α is a compression coefficient (ranging from 1.5 to 2.0, preferably 1.8), used to adjust the steepness of the curve, weaken extreme value disturbances, and enhance the factor discrimination in the middle interval. The optimal α is selected based on the fact that it makes the skewness coefficient of the normalized factor distribution closest to 0 on the validation set (i.e., closest to the symmetry of the standard normal distribution).

[0045] Step 52: Entropy weight - AHP (Analytic Hierarchy Process) dynamic weight solution combining subjective and objective factors.

[0046] Abandoning fixed manual weights, a two-layer weight calculation logic is integrated: Step 1a: Construct an AHP judgment matrix based on business experience in the advertising placement industry, and solve for the subjective experience weights. 、 、 、 、 ; Step 1b: Solve for the objective data weights based on the information entropy of historical inquiry sample data. 、 、 、 、 ; Step 1c: Introduce a balance coefficient λ (range 0.4~0.8, preferably 0.6) to fuse and obtain real-time dynamic weights. ,satisfy: .in, The optimal criteria are: minimizing the root mean square error (RMSE) between the valuation result under the fusion weight and the actual transaction price on the validation set. and Normalization was performed before fusion to ensure that the weight sum is always 1 after fusion, x1∈{A1, B1, C1, D1,E1}, x2∈{A2,B2, C2, D2,E2}.

[0047] Step 53: Factor differential nonlinear correction.

[0048] Differential correction logic is applied to different types of factors: Positive gain factors A', B', C', E': Employ power-law gain correction to amplify the price gains brought by high-quality channels, popular sales periods, and highly compatible product categories. , Where β = 1.2 (gain exponent), k ∈ {A, B, C, E}.

[0049] Negative risk factor D': Set three thresholds for segmented deduction correction: Low-risk interval (D'∈[0,0.3)): ,in =0.3 (slight deduction) Medium-risk interval (D'∈[0.3,0.7)): ,in =0.8 (linear gradient) High-risk interval (D'∈[0.7,1]): ,in =2.0 (Strong threshold circuit breaker) Step 54: Adaptive weighted fusion calculation + price reasonableness determination.

[0050] The corrected factors, dynamic weights, risk deduction coefficients, and industry fine-tuning coefficients are substituted into the stratified fusion valuation model to output standardized valuation coefficients. These coefficients are then converted into a fair basic transaction price using the channel benchmark floor price retrieved from RAG. The calculation results are compared with the industry average price for the same period. If the price fluctuation exceeds the reasonable threshold of ±15%, the algorithm is re-executed, automatically adjusting the risk deduction coefficient and channel premium coefficient, and iterating the entire algorithm calculation again. If the price is within a reasonable range, the fair inquiry range, maximum negotiation ceiling, and suggested transaction floor price are output, and channels with abnormal premiums are marked with risk warnings.

[0051] The formula for the tiered fusion valuation model is as follows:

[0052] in: - The final standardized valuation coefficient can be converted into the actual valuation price by combining the benchmark floor price; - , , , , The values ​​of the five major factors after nonlinear correction; - Risk penalty coefficient (range 0.5~2.0), adaptively configured according to industry scenario; - Industry fine-tuning coefficient (range 0.8~1.2), dynamically calibrated based on category characteristics.

[0053] Step 6: Generate the inquiry script and perform adaptive algorithm iteration.

[0054] Based on the user's input of the advertising budget, promotion goals, advertising cycle, and industry category, combined with real-time market conditions, LLM reasoning automatically generates exclusive inquiry scripts, negotiation focus, discount negotiation strategies, and a tier of high-cost-performance alternative channels; The newly added inquiry and transaction data will be synchronously fed back into the RAG knowledge base, automatically updating the factor mean, standard deviation, and entropy weight calculation samples, and dynamically updating the weights of each factor to achieve continuous adaptive optimization of the valuation algorithm. The entropy weight calculation samples consist of all historical inquiry and transaction cases stored in the RAG knowledge base. Each sample includes: the original values ​​of the five valuation factors and the actual transaction price from the channel. A quality assessment will be performed on this data: if the assessment is excellent, the quotations, valuations, and negotiation cases generated in this transaction will be organized into structured knowledge fragments, vectorized, and fed back into the vector database to update the factor mean, standard deviation, and historical inquiry and transaction sample pool of the five valuation factors, and dynamically update the dynamic weights of the five valuation factors to achieve continuous adaptive optimization of the valuation algorithm; if the assessment is poor, a manual review mechanism will be triggered.

[0055] Step 7: Output channel price comparison reports and valuation analysis reports, and generate channel risk warnings for scenarios such as bundled consumption, high price premiums, and false price quotes during promotional periods.

[0056] This invention revolves around a five-factor adaptive weighted valuation algorithm based on entropy weight-AHP (objective and subjective dynamic weighting) and the integration of a RAG knowledge base to achieve autonomous model iteration. It establishes an intelligent pricing algorithm system for media channels, achieving significant technological advancements in information processing efficiency, pricing objectivity, business intelligence, and long-term adaptability. Specifically, these advancements are reflected in the following aspects: (1) Achieve automated and high-precision parsing of non-standard quotation texts, breaking free from the limitations of traditional rule extraction.

[0057] Traditional quotation tools rely on regular expressions and fixed templates to extract quotation information, which cannot adapt to conversational quotation content such as free chat text and non-standard documents. This makes it prone to omissions and errors in field extraction, and makes it difficult to identify hidden charges. This invention utilizes a customized schema field library and scenario-specific Prompt calls to LLM to complete semantic parsing. Combined with ambiguity removal and field validation mechanisms, it automatically converts various unstructured quotations into standardized data tables, replacing manual data entry and significantly improving the processing efficiency and data integrity of the pre-quotation stage.

[0058] (2) To achieve quantitative and objective calculation of channel inquiry prices and eliminate subjective bias caused by manual valuation.

[0059] Traditional price inquiries rely solely on simple comparisons of list prices, with pricing heavily dependent on the individual experience of sales personnel and lacking standardized quantitative criteria. This invention employs an entropy-weighted AHP (Advanced Personal Hierarchy Process) scheme to dynamically allocate weights across five key factors. It uses S-shaped extreme value normalization to suppress data disturbances caused by extreme periods and high-risk accounts, implements nonlinear segmented correction for positive gain factors and negative risk factors, and adds a risk circuit breaker mechanism. By relying on a standardized mathematical model to calculate fair prices, it unifies pricing standards, ensures objective and fair price inquiry results, and significantly lowers the learning curve for new sales personnel.

[0060] (3) Relying on the RAG knowledge base to form a self-learning closed loop, the valuation algorithm can be continuously iterated and optimized.

[0061] Traditional systems store business data statically, failing to accumulate bottom-line price information and pricing experience. As media channel rules are updated, the accuracy of the algorithm continuously declines. This invention vectorizes the results of each round of price inquiries and stores them in the RAG knowledge base, continuously expanding historical transaction samples and updating factor distribution and model weight parameters. This enables the valuation model to iterate autonomously and adapt to the ever-changing pricing formats and market conditions of new media channels in the long term.

[0062] (4) Based on the valuation results, the scenario-based negotiation plan is automatically output to realize intelligent business auxiliary decision-making throughout the whole process.

[0063] Traditional price comparison systems can only sort prices and cannot provide negotiation references. This invention combines the fair price range output by the model with large-scale model inference to automatically generate negotiation space, inquiry scripts, and tiered alternative channels, upgrading simple price comparison into complete intelligent inquiry decision support, and further improving the overall intelligence level of the inquiry business.

[0064] Accordingly, embodiments of the present invention also provide an intelligent media channel inquiry, valuation, and negotiation system based on a large language model, such as... Figure 2 The image shown is a schematic diagram of the system's structure. This intelligent media channel pricing, valuation, and negotiation system based on a large language model includes the following modules: The data acquisition and preprocessing unit 201 acquires multi-source unstructured media pricing data and preprocesses the multi-source unstructured media pricing data. Semantic parsing unit 202 uses a domain-fine-tuned LLM large model to perform deep semantic parsing on the preprocessed multi-source unstructured media pricing data to obtain a standardized structured pricing data table; Channel profile model building unit 203 constructs a quantitative channel profile model from multiple dimensions based on the standardized structured quotation data table. RAG knowledge base construction unit 204 constructs a RAG knowledge base and retrieves the RAG knowledge base to obtain historical transaction prices and industry trends for similar channels; The channel pricing calculation unit 205 introduces five valuation factors, uses a multi-factor weighted fusion algorithm to calculate channel pricing, and conducts a price reasonableness assessment. Result generation unit 206 uses LLM large model reasoning to generate exclusive inquiry scripts, negotiation focus, concession negotiation strategies, and a high-cost-performance alternative channel tier, completing the inquiry analysis and valuation calculation, and evaluating the overall result quality.

[0065] This invention differs from traditional manual price inquiry and rule-based price comparison systems by employing an LLM closed-loop semantic extraction mechanism for non-standard advertising quotes. It also differs from the single-text extraction method of general large models by binding a structured schema template specific to the advertising media and using business-specific prompts to drive deep semantic parsing within the LLM. The extracted results undergo ambiguity cleaning and field validation; unqualified data is automatically fed back into the large model for secondary parsing, forming an extraction-validation-feedback rerun closed loop. This addresses the shortcomings of general LLM systems that directly process non-standard verbal quotes, such as missing fields, unidentified hidden charges, and low accuracy in single-sample extraction. This invention also proposes a multi-factor fusion dynamic valuation algorithm for media channels, and innovatively proposes a five-factor adaptive weighted valuation algorithm based on the dynamic weighting of entropy weight-AHP subjective and objective fusion. It suppresses abnormal pricing interference through S-shaped extreme value normalization, distinguishes between positive gain and negative risk factors for nonlinear segmented correction, and achieves accurate quantification of channel prices with a risk circuit breaker mechanism, which is different from the simple valuation method of linear multiplication with fixed coefficients. At the same time, it relies on RAG historical transaction data to continuously update factor distribution and weights, realizes the autonomous iteration of valuation model, and solves the technical defects of existing valuation algorithms such as fixed weights, valuation distortion in extreme scenarios, single risk deduction, and inability to adapt to industry changes.

[0066] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0067] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.

[0069] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and systems of the present invention, and are merely some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent media channel inquiry, valuation, and negotiation based on a large language model, characterized in that, The method includes: Step 1: Collect multi-source unstructured media pricing data and preprocess the multi-source unstructured media pricing data; Step 2: Use a domain-adjusted LLM large model to perform deep semantic parsing on the preprocessed multi-source unstructured media pricing data to obtain a standardized structured pricing data table; Step 3: Based on the standardized structured quotation data table, construct a quantitative channel profile model from multiple dimensions; Step 4: Build a RAG knowledge base and retrieve the RAG knowledge base to obtain the historical transaction floor price and industry trends of similar channels; Step 5: Introduce five valuation factors, use a multi-factor weighted fusion algorithm to calculate channel quotations, and conduct a price reasonableness assessment; Step 6: Generate customized inquiry scripts, negotiation focus, concession negotiation strategies, and a high-cost-performance alternative channel tier through LLM model reasoning to complete the inquiry analysis and valuation calculation.

2. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 1, characterized in that, In step 1, the collected multi-source unstructured media pricing data includes: chat text, pricing documents, poster copy, and channel homepage information. The multi-source unstructured media pricing data is then preprocessed with data cleaning, deduplication, and noise reduction to remove invalid and redundant information.

3. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 1, characterized in that, In step 2, after loading the media inquiry-specific schema field library and the business-specific prompt model, the domain-adjusted LLM model is invoked to perform deep semantic parsing on the preprocessed multi-source unstructured media quotation data. The LLM model extracts key fields such as channel type, number of followers, rate card price, schedule, additional benefits, and hidden charges according to the preset schema field library, and performs error correction operations such as ambiguity cleaning, redundancy removal, and field validation on the key field extraction results. Then, the key field extraction results after error correction enter the data qualification judgment: if the validation fails, the unqualified key field extraction results are returned to the LLM model to perform deep semantic parsing, and the key field extraction is re-executed. If the verification is successful, the key fields extracted after the error correction operation will be output as the standardized structured quotation data table.

4. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 1, characterized in that, In step 3, based on the standardized structured pricing data table, a quantitative channel profile model is constructed from six dimensions: account verticality, fan audience matching degree, content interaction authenticity, historical campaign ROI, account industry weight, and violation risk, and a channel quality score and risk level are generated.

5. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 4, characterized in that, In step 4, when constructing the RAG knowledge base, firstly, the system receives the input data of the new inquiry business in this round and performs query text vectorization encoding on the input inquiry business data; secondly, it performs RAG retrieval based on vector similarity, retrieving historical channel transaction data from the vector database as reference context; it then assembles the RAG retrieval results with the original query text to construct a complete inference prompt, assisting the LLM large model in performing accurate inference and completing the parsing and valuation calculation of this inquiry; finally, it performs a quality assessment on the inference results: if the assessment is excellent, the quotations, valuations, and negotiation cases generated in this business are organized into structured knowledge fragments, vectorized, and then fed back into the vector database; if the assessment is poor, a manual review mechanism is triggered.

6. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 4, characterized in that, In step 5, the five valuation factors include the integrated benchmark base price coefficient, the period popularity fluctuation coefficient, the channel quality premium coefficient, the risk deduction coefficient, and the industry category adaptation coefficient. The integrated benchmark base price coefficient is generated based on the historical transaction base prices of similar channels in the RAG knowledge base, and serves as the basis for valuation calculation. When historical data is sufficient, the weighted average of historical transaction base prices is taken. When historical data is insufficient, the publicly listed price of the same category industry is multiplied by a preset industry discount coefficient, and the value range is [0,1]. The aforementioned period popularity fluctuation coefficient is a positive gain coefficient generated by combining the peak and off-peak seasons of the release period and the supply and demand of platform traffic. The higher the popularity, the larger the value of the period popularity fluctuation coefficient, and the value range is [0,1]. The channel quality premium coefficient is generated based on a comprehensive score of channel account verticality, content interaction authenticity, historical ROI and account industry weight. The higher the channel quality, the higher the channel quality premium coefficient, with a value range of [0,1]. The risk deduction coefficient is for violation risk, including account violation records, traffic fraud risk and performance instability risk. A negative correction coefficient is generated by comprehensive scoring. The higher the risk, the larger the risk deduction coefficient, with a value range of [0,1]. The industry category fit coefficient is generated by combining the overlap of advertising promotion category tags, the matching degree of fan audience, and the industry's general pricing rules. The higher the fit, the larger the industry category fit coefficient, with a value range of [0,1].

7. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 6, characterized in that, Step 5 includes: Step 51: After collecting the original values ​​of the five major valuation factors, the improved S-curve extreme value compression normalization process is used to eliminate dimensional differences, suppress extreme heat or extreme risk outlier disturbances, and converge all factors to the [0,1] interval. Step 52: The dynamic weights of the five valuation factors are solved using the entropy weight-AHP subjective-objective fusion method. First, an AHP judgment matrix is ​​constructed based on business experience in the advertising industry to solve the subjective experience weights of each of the five valuation factors. Second, the objective data weights of each of the five valuation factors are solved based on the information entropy of historical inquiry sample data. Finally, a balance coefficient λ is introduced, and the dynamic weights of each of the five valuation factors are obtained by fusing the subjective experience weights and objective data weights of each of the five valuation factors. Step 53: Perform differentiated correction on the five valuation factors. For positive gain factors, namely the fusion benchmark floor price coefficient, the period popularity fluctuation coefficient, the channel quality premium coefficient, and the industry category fit coefficient, use power gain correction to amplify the price gain brought by high-quality channels, popular periods, and highly matched categories. For negative risk factors, namely the risk deduction coefficient, set three threshold segments for deduction correction. Step 54: Perform adaptive weighted fusion calculation and complete the price reasonableness judgment; First, substitute the five valuation factors after differentiation correction, the dynamic weights of the five valuation factors, the risk deduction coefficient, and the industry fine-tuning coefficient into the hierarchical fusion valuation model to output the standardized valuation coefficient. Combine this with the historical transaction floor price of the same type of channel obtained from the RAG knowledge base to obtain the fair basic transaction price; Compare and verify the fair basic transaction price with the industry average price of the same period. If the price fluctuation exceeds the reasonable threshold, the process returns to step 53, and the risk deduction coefficient and the channel quality premium coefficient are called back. Factor differentiation correction, adaptive weighted fusion calculation, and price reasonableness judgment are re-executed until the price is in a reasonable range; If the price is in a reasonable range, the fair inquiry range, the maximum negotiation ceiling, and the suggested transaction floor price are output, and risk warnings are marked for channels with abnormal premiums.

8. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 6, characterized in that, Step 6 includes: Based on the user's input of the advertising budget, promotion goals, advertising cycle and industry category, combined with real-time market conditions, the system uses an LLM big data model to generate exclusive inquiry scripts, negotiation focus, discount negotiation strategies and a high-cost-performance alternative channel tier. The newly added inquiry and transaction data will be synchronously fed back to the RAG knowledge base, and a quality assessment will be performed on the data: if the assessment is excellent, the quotations, valuations, and negotiation cases generated in this transaction will be organized into structured knowledge fragments, vectorized and encoded, and then fed back into the vector database to update the factor mean, standard deviation, and historical inquiry and transaction sample pool of the five valuation factors, and dynamically update the dynamic weights of the five valuation factors to achieve continuous adaptive optimization of the valuation algorithm; if the assessment is poor, a manual review mechanism will be triggered.

9. The intelligent media channel inquiry, valuation, and negotiation method based on a large language model according to claim 1, characterized in that, The method also includes, after step 6, outputting a channel price comparison report and a valuation analysis report, and generating channel risk warnings for scenarios such as bundled consumption, high price premiums, and false price quotes during promotional periods.

10. A smart media channel inquiry, valuation, and negotiation device based on a large language model, characterized in that, The device includes: The data acquisition and preprocessing unit acquires multi-source unstructured media pricing data and preprocesses the multi-source unstructured media pricing data. The semantic parsing unit uses a domain-adjusted LLM large model to perform deep semantic parsing on the preprocessed multi-source unstructured media pricing data to obtain a standardized structured pricing data table. The channel profile model building unit constructs a quantitative channel profile model from multiple dimensions based on the standardized structured quotation data table. The RAG knowledge base construction unit builds the RAG knowledge base and retrieves the RAG knowledge base to obtain the historical transaction floor price and industry trends of similar channels. The channel pricing calculation unit introduces five valuation factors, uses a multi-factor weighted fusion algorithm to calculate channel prices, and conducts a price reasonableness assessment. The results generation unit uses an LLM large model to generate exclusive inquiry scripts, negotiation focus points, concession negotiation strategies, and a high-cost-performance alternative channel tier, completing the inquiry analysis and valuation calculation.

Images