Marketing verbal skill generation method and device

Abstract

The invention relates to a marketing verbal skill generation method and device, and the method comprises the steps: obtaining a marketing verbal skill record with a historical marketing success rate higher than a first threshold value, and converting the marketing verbal skill record into initial text data; performing error vocabulary recognition and correction on the initial text data; sentence fluency detection is conducted on the corrected text data, and high-quality text corpora are screened out; training a marketing verbal skill content selling point keyword recognition model by adopting a feedback human reinforcement learning RLHF framework based on the high-quality text corpus; inputting a to-be-analyzed marketing verbal skill text into the trained selling point keyword recognition model, and outputting package selling point key short words and sentences; and combining the package selling point key short words and sentences and the package basic information into cue words, and generating a target marketing verbal skill based on the cue words and the large language model. Therefore, the defects of dependence on manpower, limited coverage scene and insufficient professional property of a general model in related technologies can be effectively overcome, and the efficiency and success rate of package marketing are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a marketing script generation method and device. BACKGROUND

[0002] In the existing operator package marketing dialogue system, the traditional method usually relies on a pre-processed script library or a recommendation engine based on fixed rules. Such systems have poor flexibility and are difficult to adapt to complex and variable real dialogue scenarios, resulting in monotonous and non-targeted generated scripts.

[0003] With the development of large language model technology, a scheme of directly generating marketing scripts using prompt words has appeared in related technologies. However, since general large models lack professional knowledge in the operator field, the quality of the generated content is strongly related to the number of model parameters, and it is difficult to accurately capture and highlight the core selling points of the package, resulting in insufficient professionalism of the generated scripts and low marketing success rate. SUMMARY

[0004] The present application provides a marketing script generation method and device.

[0005] According to a first aspect of the present application, a marketing script generation method is provided, which comprises: obtaining marketing script recordings with a historical marketing success rate higher than a first threshold value, and converting the marketing script recordings into initial text data; performing error vocabulary recognition and correction on the initial text data to obtain corrected text data; performing sentence fluency detection on the corrected text data to screen out high-quality text corpus; based on the high-quality text corpus, training a marketing script content selling point keyword identification model using a feedback human reinforcement learning (RLHF) framework; inputting the marketing script text to be analyzed into the trained selling point keyword identification model, and outputting package selling point key short sentences; combining the package selling point key short sentences and package basic information into prompt words, and generating target marketing scripts based on the prompt words and a large language model.

[0006] According to a second aspect of the present application, a marketing script generation device is provided, which comprises: a data acquisition module configured to obtain marketing script recordings with a historical marketing success rate higher than a first threshold value, and convert the marketing script recordings into initial text data; a data correction module configured to perform error vocabulary recognition and correction on the initial text data to obtain corrected text data; The data screening module is configured to perform sentence fluency detection on the corrected text data and screen high-quality text corpus; The training module is configured to train a marketing pitch content selling point keyword identification model based on the high-quality text corpus by using a reinforcement learning from human feedback (RLHF) framework. The key short sentence output module is configured to input the marketing pitch text to be analyzed into the trained selling point keyword identification model and output a package selling point key short sentence. The marketing pitch generation module is configured to combine the package selling point key short sentence and package basic information into a prompt word and generate a target marketing pitch based on the prompt word and a large language model.

[0007] According to a third aspect of the embodiments of the present application, an electronic device is provided. The electronic device includes a memory and a processor, the memory has a computer program stored thereon, and the processor implements the method described above when executing the program.

[0008] According to a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided, and the computer readable storage medium has a computer program stored thereon, and the program is executed by a processor to implement the method described above.

[0009] According to a fifth aspect of the embodiments of the present application, a computer program product is provided, and the computer program product includes a computer program, and the computer program is executed by a processor to implement the method described above.

[0010] The marketing pitch generation method and device provided by the embodiments of the present application can realize an end-to-end automatic and high-quality marketing pitch generation mode by obtaining marketing pitch recordings with a marketing success rate higher than a first threshold value, converting the marketing pitch recordings into initial text data, performing error word identification and correction on the initial text data to obtain corrected text data, performing sentence fluency detection on the corrected text data to screen high-quality text corpus, training a marketing pitch content selling point keyword identification model based on the high-quality text corpus by using a reinforcement learning from human feedback (RLHF) framework, inputting marketing pitch text to be analyzed into the trained selling point keyword identification model to output a package selling point key short sentence, and combining the package selling point key short sentence and package basic information into a prompt word and generating a target marketing pitch based on the prompt word and a large language model. The method effectively overcomes the defects of related technologies, such as dependence on manual work, limited coverage scenarios, and insufficient professional performance of general models, and significantly improves the efficiency and success rate of package marketing. BRIEF DESCRIPTION OF DRAWINGS

[0011] In the following description of exemplary embodiments in conjunction with the accompanying drawings, more details, features and advantages of the present application are disclosed, in which: Figure 1A flowchart of a marketing script generation method provided for an exemplary embodiment of the present application; Figure 2 A flowchart of a marketing script generation method provided for another exemplary embodiment of the present application; Figure 3 A framework diagram of an RLHF learning a reward model from human text input provided for another exemplary embodiment of the present application; Figure 4 A functional module schematic block diagram of a marketing script generation apparatus provided for an exemplary embodiment of the present application; Figure 5 A structural block diagram of an electronic device provided for an exemplary embodiment of the present application; Figure 6 A structural block diagram of a computer system provided for an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0012] Embodiments of the present application will be described in more detail by referring to the attached drawings. Although certain embodiments of the present application are shown in the drawings, it is understood that the present application can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein, but rather, the embodiments are provided so that the present application can be more thoroughly and completely understood. It should be understood that the drawings of the present application and the embodiments are for exemplary purposes only and are not intended to limit the scope of protection of the present application.

[0013] It should be understood that each of the steps recited in the method embodiments of the present application can be executed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit the execution of the steps shown. The scope of the present application is not limited in this respect.

[0014] The term “comprising” and variations thereof as used herein are open-ended, that is, “including but not limited to”. The term “based on” is “based, at least in part, on”. The term “one embodiment” means “at least one embodiment”; the term “another embodiment” means “at least one additional embodiment”; the term “some embodiments” means “at least some embodiments”. Related definitions are given throughout the description. It should be noted that the concepts mentioned “first”, “second” and the like in the present application are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0015] It should be noted that the modification of “one”, “multiple” mentioned in the present application is illustrative and not restrictive, and those skilled in the art should understand that unless otherwise explicitly indicated in the context, it should be understood as “one or more”. Names of messages or information exchanged between multiple devices in embodiments of the present application are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0016] It can be understood that, before using the technical solutions disclosed in the embodiments of the present application, the type, use range, use scenario, etc. of the personal information involved in the present application should be informed to the user and the authorization of the user should be obtained in accordance with relevant laws and regulations.

[0017] For example, in response to receiving an active request of a user, a prompt information is sent to the user to explicitly prompt the user that the operation requested to be performed will require obtaining and using personal information of the user. Thus, the user can autonomously select whether to provide personal information to the electronic device, application program, server or storage medium, etc. software or hardware performing the operation of the technical solutions of the present application according to the prompt information.

[0018] As an optional but non-limiting implementation manner, in response to receiving an active request of a user, the manner of sending a prompt information to the user may, for example, be a pop-up window manner, and the prompt information may be presented in the form of text in the pop-up window. In addition, the pop-up window may also carry a selection control for the user to select "agree" or "disagree" to provide personal information to the electronic device. It can be understood that the above notification and obtaining of user authorization process is only illustrative, and does not limit the implementation manner of the present application, and other manners meeting relevant laws and regulations can also be applied to the implementation manner of the present application.

[0019] Therefore, in order to improve the success rate of package marketing, the embodiments of the present application first provide a marketing script generation method, which automatically extracts accurate package selling points from high-quality historical data, and drives a large model to generate professional and targeted marketing scripts, thereby significantly improving the success rate of marketing. As shown in Figure 1 The method can include the following steps: In step S110, a marketing script recording with a historical marketing success rate higher than a first threshold value is obtained, and the marketing script recording is converted into initial text data.

[0020] In the embodiment, the step aims to ensure the quality of training data from the source. Specifically, first, the historical marketing records are obtained from the operator background system, and according to the marketing success rate (such as the ratio of the number of successful businesses to the total number of marketing) and the actual number of successful businesses, all marketers are screened. In a specific implementation, for example, the quartile (Q1) of the number of successful businesses of all marketers can be calculated using the box plot method, and the marketers with the number of successful businesses greater than Q1 are determined as “marketing success rate higher than the first threshold”. Subsequently, the marketing speech recording files of these high-quality marketers are collected, and the voice-to-text model (such as the Whisper Large-v2 model) is used to batch-transcribe them into initial text data. This step effectively filters low-efficiency marketing samples, laying a high-quality data foundation for subsequent model training.

[0021] In step S120, error words in the initial text data are identified and corrected to obtain corrected text data.

[0022] Due to problems such as noisy recording environment and marketer accent, there are often errors and colloquial redundancies in the initial text data obtained by translation. The embodiment can correct and purify the text through an intelligent method. First, a large language model (such as Qianwen 2.0) fine-tuned by marketing speech experts is used to probabilistically label each word in the text, and words with a probability value below a preset second threshold are marked as “potential error words”. Next, a high-frequency word table covering fixed descriptions of packages is constructed, and the comprehensive similarity score of each potential error word with the words in the high-frequency word table is calculated. The score integrates multiple dimensions such as pinyin features (based on pinyin edit distance), character edit distance, and semantic similarity. Finally, the highest-scoring high-frequency word is selected for each potential error word to replace it, resulting in corrected text data with more standardized and accurate word usage.

[0023] In step S130, the corrected text data is subjected to sentence fluency detection to select high-quality text corpus.

[0024] After completing the correction at the word level, this step further controls the corpus quality at the sentence level. Specifically, the N-Gram language model (such as 2-Gram) can be used for fluency detection. First, the probability of each sentence in the corrected text data is calculated. To solve the problem of too large probability value difference between sentences of different lengths, the sentences are grouped by their lengths, and the probabilities of the sentences within the group are normalized to be within a comparable range. Subsequently, the normalized sentence probability is compared with the probability threshold (such as the average value of the probabilities of the sentences in the training set) based on the training corpus statistics, and the sentences that meet the fluency standard are selected, and finally a high-quality text corpus for model training is formed. This step ensures that the corpus learned by the subsequent model meets high standards in terms of language fluency and standardization.

[0025] In step S140, a marketing message content selling point keyword recognition model is trained based on high-quality text corpus using the Reinforcement Learning from Human Feedback (RLHF) framework.

[0026] In this embodiment, this step is used to train a model that can accurately identify the core selling points in the sales pitch. Specifically, it includes: First, prompt words can be written based on the FABE (Features, Advantages, Benefits, Evidence) model. Then, a large language model is used to initially extract key short phrases and words of marketing selling points from high-quality text corpus. After manual annotation and correction, a high-quality training dataset is constructed.

[0027] Then, the RLHF framework is used, combined with the Proximal Policy Optimization (PPO) algorithm, to train the model. This framework may include: Actor Model: This is a keyword recognition model for selling points that needs to be trained. Its strategy is continuously optimized during training.

[0028] The reward model is crucial to this step. Human evaluation feedback text and state-action sequence information are input into a large language model (LLM), which automatically generates a set of states and their corresponding positive or negative labels. Based on these labels, the agent updates the reward probability distribution (modeled as a Beta distribution) corresponding to each state, thereby training a reward model capable of accurately judging the quality of generated content.

[0029] Finally, the trained reward model is used to guide the PPO (Proximal Policy Optimization) algorithm to optimize the strategy of the actor model (i.e., the selling point keyword recognition model), so that its output of selling point keywords is more in line with the preferences of human experts.

[0030] In step S150, the marketing text to be analyzed is input into the trained selling point keyword recognition model, and the model outputs key short phrases of package selling points.

[0031] Once the model is trained, it can be put into practical application. When marketing is needed, a marketing script to be analyzed (e.g., a transcript of a real-time customer service conversation or a historical script record) can be input into the trained keyword recognition model. The model will automatically analyze the text content and output extracted key phrases and keywords for the package's selling points, such as "gigabit broadband," "family sharing," and "monthly discount."

[0032] In step S160, key short phrases highlighting the package's selling points and basic package information are combined into prompt words, and target marketing scripts are generated based on the prompt words and the large language model.

[0033] Finally, the embodiment combines the key short phrases of the package's selling points output in step S150 with the basic information of the package (such as name, price, and contents) to construct a structured prompt. This prompt is then input into a general or finely tuned large language model to generate a complete, fluent, and precise marketing script that highlights the core selling points of the package, which can be used directly or as a reference by frontline marketing personnel.

[0034] Through the technical solutions described in the above embodiments, this application embodiment realizes an end-to-end automated, high-quality marketing script generation method, which effectively overcomes the shortcomings of related technologies such as reliance on manual labor, limited coverage of scenarios, and insufficient professionalism of general models, and significantly improves the efficiency and success rate of package marketing.

[0035] Based on the above embodiments, in another embodiment provided in this application, by combining intelligent recognition of a large language model with multi-dimensional similarity matching, high-precision location and correction of erroneous words in translated text can be achieved. Therefore, when identifying and correcting erroneous words in the initial text data, the above step S120 may further include the following steps: Step S121: Use a fine-tuned large language model to perform probability labeling on the words in the initial text data, and mark words with probabilities below a second threshold as potential error words.

[0036] In this embodiment, the core of this step lies in locating potential erroneous words. First, the initial text data is processed using a large language model (e.g., Qianwen 2.0) that has been supervised and fine-tuned by marketing script experts. Based on learned high-quality script language patterns, this model generates a probability score for each word in the text, which reflects the reasonableness of the word's appearance in a given context.

[0037] Subsequently, a probability threshold (i.e., a second threshold) is set to filter out low-probability words. In one specific implementation, this threshold can be calculated by taking the mean and standard deviation of the probabilities of all words in each batch of text data, and using the value of "mean minus three times the standard deviation" as a dynamic threshold. Of course, other statistical methods such as box plots or clustering can also be used to determine this threshold in the embodiments. Finally, all words with probability values ​​lower than the second threshold are marked as "potentially incorrect words" and included in the subsequent processing list.

[0038] Step S122: Calculate the comprehensive score of the potential erroneous words and the words in the preset high-frequency word list in terms of pinyin features, edit distance and semantic similarity.

[0039] In this embodiment, this step is used to find the most suitable replacement for each potentially erroneous word. First, a preset high-frequency word list needs to be constructed. This word list is derived from statistics on a large number of historically successful package marketing scripts and includes core words and common expressions in the fixed descriptions of each package.

[0040] Then, for each potentially misspelled word, its similarity to each candidate word in the high-frequency word list is calculated. To comprehensively address various situations such as translation errors due to inaccurate pronunciation and synonym substitution, this embodiment innovatively adopts a multi-dimensional comprehensive scoring mechanism, specifically including: Pinyin Features: This dimension converts words into their pinyin form and calculates the edit distance between pinyin syllables. It is specifically designed to address transcription errors caused by similar pronunciations.

[0041] Edit distance: Directly calculates the edit distance between the original word and the candidate word at the character level, used to capture spelling errors or errors with similar glyphs.

[0042] Semantic similarity: Words are transformed into vectors using semantic embedding models (such as Word2Vec or BERT), and the cosine similarity between vectors is calculated. This dimension ensures that the replacement words are semantically consistent with the original context.

[0043] Finally, the scores from the three dimensions are combined into a single comprehensive score using a weighted average method. The weights are dynamically calculated based on the accuracy of each dimension on historical validation samples and the difference coefficient of its score distribution, thereby ensuring the optimization of the matching strategy.

[0044] Step S123: Select the high-frequency words with the highest comprehensive score to replace the potential erroneous words.

[0045] In this embodiment, this step completes the final correction operation. For each potential erroneous word, after the calculation in step S122, the system selects the word with the highest comprehensive score from all its corresponding candidate high-frequency words as the best replacement candidate word. Subsequently, this candidate word is automatically used to replace the potential erroneous words in the text, thereby generating corrected text data with more accurate and standardized wording.

[0046] The solution provided in this embodiment can efficiently and accurately clean up speech-to-text translations, significantly improving the overall quality of the corpus in subsequent processes and providing reliable data support for model training and key information extraction.

[0047] Based on the above embodiments, in another embodiment provided in this application, the specific implementation process of sentence fluency detection (i.e., step S130) on the corrected text data is further refined. This method effectively solves the problem of excessive probability differences caused by varying sentence lengths by introducing length-based normalization, thereby achieving accurate and fair evaluation of sentence fluency. Therefore, when performing sentence fluency detection on the corrected text data, step S130 may further include the following steps: Step S131: Calculate the probability of sentences in the corrected text data using an N-Gram language model.

[0048] In this embodiment, this step is used to establish an objective and quantifiable fluency benchmark for each sentence to be detected. Specifically, an N-Gram language model (preferably N=2, i.e., a Bigram model) can be used to calculate the probability of the sentence. For a given sentence... S = w 1, w 2,..., w m Its probability P ( S The calculation can be performed using the chain rule and the Markov assumption, and the formula is as follows: (1) in, Indicates the preceding word In the case of occurrence, the current word The conditional probability of occurrence is calculated, and this probability value is derived statistically from a large-scale training corpus. This step allows for the calculation of an initial probability value for each sentence in the corrected text data. P ( S ).

[0049] Step S132: Group the sentences according to their length and normalize the probability of sentences in each group.

[0050] In this embodiment, this step is used to eliminate the absolute influence of sentence length on the probability value. In practice, the probability value of long sentences is usually much lower than that of short sentences, but this does not necessarily mean that long sentences are not fluent. Directly comparing the original probabilities of sentences of different lengths can lead to judgment bias.

[0051] To address this issue, this embodiment employs a grouping normalization strategy: Grouping by length: All sentences in the corrected text data are grouped according to the number of words (or characters) they contain, and sentences of the same or similar length are grouped together.

[0052] Intra-group normalization: The probability values ​​P(S) of all sentences within the same length group are normalized, for example, by scaling them to the [0,1] interval. A preferred method is to calculate the Z-score (standard score) of the sentence probabilities within the group, or to use the maximum-minimum normalization method. After processing, the fluency of each sentence has a relatively comparable standard score within the group.

[0053] This step ensures that the fluency evaluation criteria are consistent and fair across different length groups.

[0054] Step S133: Compare the normalized sentence probability with a preset probability threshold to determine whether the fluency of the sentences in the corrected text data meets the standard.

[0055] In this embodiment, this step can perform the final screening decision based on the normalized result. The preset probability threshold can be pre-calculated based on a high-quality training corpus. In a specific implementation, this threshold can be set as a statistic of the normalized probabilities of all sentences in the training corpus, such as the mean or median.

[0056] Then, the normalized probability of each sentence obtained in step S132 is compared with the preset threshold: If the normalized probability of a sentence is higher than or equal to the threshold, the sentence is deemed to meet the fluency standard and is selected as part of the high-quality text corpus.

[0057] If the normalized probability of a sentence is lower than the threshold, the sentence is deemed to be unfluent and is removed from the subsequent training corpus.

[0058] In the special case where no sentence of the same length can be found in the training corpus, the system can downgrade the processing and directly compare the original probability of the sentence with the global probability threshold of the entire training sample.

[0059] The solution provided in this embodiment can overcome the sensitivity of traditional fluency detection methods to length factors, thereby more scientifically and accurately selecting high-quality speech texts with fluent language, and providing key data quality assurance for subsequent model training.

[0060] Based on the above embodiments, in another embodiment provided in this application, the key step of training the reward model used for guidance optimization during the training of the marketing message content selling point keyword recognition model using the RLHF framework is refined. This method innovatively utilizes a large language model to transform human text feedback into quantifiable training signals, effectively reducing the reliance of traditional RLHF on extensive manual annotation. The method may further include the following steps: Step S171: Input the human evaluation feedback text and state action sequence information into the large language model.

[0061] In this embodiment, structured information is prepared to reward the training of the model. Specifically, textual evaluations of the model's generated results (such as extracted selling point keywords) by marketing experts or experienced customer service personnel, such as "This selling point is very well captured" or "The core cost-effectiveness advantage was overlooked," are used as human evaluation feedback text. Simultaneously, information on the series of states the model experiences during the generation of this result (which can be understood as intermediate representations or context within the model) and the actions taken (i.e., the generated vocabulary or decision sequence) are combined to form state-action sequence information. Both types of information are input into the LLM.

[0062] Step S172: Generate a set of states and corresponding labels based on the large language model. The labels include positive labels or negative labels.

[0063] This implementation leverages the understanding and reasoning capabilities of LLM to automatically generate sophisticated reward signals. Upon receiving input information, the Large Language Model (LLM) analyzes the association between human feedback text and the state-action sequence, automatically identifying which states in the sequence the feedback specifically refers to. Subsequently, the LLM generates a corresponding label for each identified state. This label quantifies the human feedback, typically a positive label (corresponding to positive or correct feedback) or a negative label (corresponding to negative or incorrect feedback). This process achieves the transformation from fuzzy natural language feedback to precise state-label pairs.

[0064] Step S173: Based on the reward probability distribution corresponding to the label update state, the reward probability distribution is modeled as a Beta distribution.

[0065] In this embodiment, this step establishes the core mathematical model for training the reward model. The agent in the system maintains a reward probability distribution for each tracked state, which is initialized as a beta distribution, such as β(0.5, 0.5), to represent the uncertainty regarding the reward value (0 or 1). After obtaining a state and its label from step S172, the system treats this label as an observation: If a positive label is received, the conjugate update property of the beta distribution is used to update the reward distribution corresponding to the state to β(a, b+1).

[0066] If a negative label is received, the reward distribution corresponding to that state is updated to β(a+1, b).

[0067] Through this Bayesian update-based approach, the reward model's perception of the "good" or "bad" state becomes increasingly precise as feedback accumulates.

[0068] Step S174: The reward model is trained based on the updated reward probability distribution. The reward model is used to optimize the selling point keyword recognition model through reinforcement learning algorithm.

[0069] In this embodiment, this step completes the training of the reward model and puts it into application. After iteratively updating a large number of state-label pairs in step S173, the reward probability distribution of all states tends to stabilize, at which point the training of the reward model is declared complete. This trained reward model can provide accurate reward scores for the actions taken by the selling point keyword recognition model (i.e., the actor model in RLHF) in any state. Subsequently, in reinforcement learning algorithms such as Proximal Policy Optimization (PPO), the reward values ​​generated by this reward model are used to guide and optimize the strategy of the selling point keyword recognition model, causing its behavior to continuously adjust towards obtaining high rewards (i.e., conforming to human expert preferences), ultimately becoming a high-performance selling point extraction expert.

[0070] The refined solution provided in this embodiment enables the training of a high-performance reward model in an automated and data-driven manner, providing key support for the accurate optimization of the core selling point keyword recognition model and significantly improving the performance and efficiency of the entire system.

[0071] Based on the above embodiments, in another embodiment provided in this application, the key step of constructing a high-quality training dataset for training a marketing message content selling point keyword recognition model is explained in detail. This method combines the automated extraction capabilities of a large language model with the accurate judgment of human experts, ensuring the accuracy and professionalism of the training data and laying a solid foundation for subsequent reinforcement learning training. Therefore, the method may further include the following steps: In step S181, prompt words are written based on the FABE model, and key short phrases and words for marketing selling points are extracted from high-quality text corpus using a large language model.

[0072] In this embodiment, this step aims to automatically and structurally extract marketing selling points from massive amounts of high-quality text corpora. Specifically, a set of structured cue words is first designed based on the classic FABE marketing model. These cue words assign a clear role to the large language model (such as "marketing strategy analyst") and constrain it to analyze and refine the input text according to four dimensions: F (Features), A (Advantages), B (Benefits), and E (Evidence).

[0073] For example, prompt words would require the model to: Identify and output key phrases and corresponding core words in the sales pitch that describe product features (e.g., "Package includes gigabit broadband" -> "gigabit broadband"). Identify and output key phrases and words in the sales pitch that describe the product's advantages (e.g., "fast internet speed, low latency" -> "fast internet speed, low latency"). Identify and output key phrases and words in the sales script that explain the benefits to the customer (e.g., "No more buffering when watching videos, smoother gaming experience" -> "Smooth video playback, better gaming experience"). Identify and output key phrases and short words of evidence provided in the script (e.g., "Currently, 100,000 users have signed up" -> "100,000 users").

[0074] In step S182, the results extracted by the large language model are manually labeled to generate a training dataset for training the selling point keyword recognition model.

[0075] In this embodiment, this step is a core step in ensuring data quality, by introducing the wisdom of human experts to correct and refine the automatically extracted results of the model. Professionals with extensive marketing experience review and manually annotate the preliminary results output by the large language model in step S181. The annotation work mainly includes: Correction: Corrects inaccurate, incomplete, or biased short phrases and words extracted by the model.

[0076] Additional information: Add important selling points that may have been missed in the model.

[0077] Standardization: The extracted short words are normalized to unify the expression (for example, "fast speed" and "fast network speed" are unified into "fast network speed").

[0078] After manual verification, correction, and confirmation, a high-quality, accurately labeled training dataset was finally formed. Each sample in this dataset contains the original marketing text and its corresponding manually verified key selling points phrases and words, providing direct and reliable supervision signals for subsequent training of the selling point keyword recognition model using the RLHF framework.

[0079] The solution provided in this embodiment creatively solves the problem of obtaining high-quality training data, effectively integrates automated processes with expert knowledge, and provides crucial data support for training a high-performance selling point keyword recognition model.

[0080] Based on the above embodiments, in another embodiment provided in this application, the method further introduces a closed-loop self-optimization mechanism, enabling the marketing script generation system to continuously evolve during actual use, thereby constantly improving the quality and applicability of its generated scripts. This mechanism iteratively optimizes the system's core model by collecting feedback from the application side. Therefore, the method further includes the following steps: In step S191, the optimization results of the target marketing script are obtained.

[0081] In this embodiment, this step is the starting point for closed-loop optimization, aiming to collect effective feedback data from practical applications. After the system generates target marketing scripts according to the aforementioned embodiments and delivers them to frontline marketing personnel, the marketing personnel will evaluate, adjust, or directly optimize the scripts based on their professional experience and actual interactions with customers. This step systematically obtains these optimization results from the practical level. These results can be expressed in various forms, such as: Direct modification and polishing of dialogue texts.

[0082] A rating of the success rate or satisfaction level of the system-generated scripts.

[0083] The final version of the script marked as a success story.

[0084] These optimizations contain valuable practical experience and effectively supplement the shortcomings of the initial version of the system's scripts.

[0085] In step S192, the optimization results are added to a high-quality text corpus for training the selling point keyword recognition model.

[0086] In this embodiment, this step is crucial for completing the closed loop, realizing the transformation from practical feedback to model capabilities. All optimization results collected in step S191 (usually marketing phrases that have been manually optimized and verified to be more effective) are used as new, high-quality samples to supplement the system's existing high-quality text corpus.

[0087] This operation brings two core benefits: Expanding and updating the corpus: The continuously enriched corpus ensures that the training data can cover more diverse marketing scenarios and more cutting-edge marketing strategies, preventing the model's knowledge from becoming outdated.

[0088] Driving continuous model evolution: This continuously expanding, high-quality corpus allows for regular retraining or incremental training of the selling point keyword recognition model. By learning from these samples containing the latest practical wisdom, the model can continuously adjust and optimize its internal parameters, thereby generating more accurate selling point keywords in the future and ultimately producing more efficient and persuasive marketing messages.

[0089] Through the solution provided in this embodiment, the marketing script generation method is upgraded from a static automated tool to a dynamic optimization system with continuous learning capabilities. This method effectively transforms the later intervention of human experts into the inherent evolutionary driving force of the system, forming a virtuous cycle of "use -> feedback -> optimization -> reuse," significantly improving the system's life cycle and value in the long run.

[0090] In another embodiment provided in this application, to improve the quality of marketing script corpus, the embodiment collects audio files of marketing scripts from marketers with high historical marketing success rates to obtain high-quality audio scripts with high success rates, and then transcribes the audio files into text. Because the audio files contain background noise and some speakers' Mandarin is not standard, the transcribed text contains typos. The embodiment uses multiple algorithms, including integrated pinyin features, edit distance, and semantic similarity, to identify and match suitable words and phrases to replace potentially erroneous words in the marketing text. Since the audio files from the business hall may contain conversations of multiple people, further detection and screening of the audio content is required. The embodiment uses a fluency detection model to detect sentence fluency in the pre-trained corpus to ensure the standardization of the pre-trained corpus. To extract effective key information about marketing script selling points from the improved high-quality audio corpus, the embodiment annotates the package selling point keywords appearing in the text content and uses a feedback human reinforcement learning model to train a marketing script content selling point keyword recognition model. Finally, when conducting package marketing, the package selling point keywords are used as prompts to allow the large model to generate a complete marketing script. The specific implementation process is as follows... Figure 2 As shown, the specific steps may include: Step S210: Obtain the dialogue audio files of historical successful marketing scripts and transcribe the files into text content.

[0091] This example involves collecting and filtering valid historical successful marketing phrases and converting the audio files into text data. The specific process is as follows: The audio files of marketing scripts are large in volume and contain issues such as audio interruptions and non-standard pronunciation. After transcription, the marketing script texts often contain typos and filler words, and the marketing content within the texts spans different scenarios (or different types of marketing content). Therefore, rigorous screening and data mining are necessary when using this data. Specific methods are as follows: Obtain historical dialogue data of call center agents with high marketing success rates. The marketing success rate is calculated as the ratio of the number of user services successfully handled by the call center agent to the number of users recruited through marketing.

[0092] After assessing the marketing success rate, it's also necessary to consider the actual number of successful transactions handled by each agent. Agents with high marketing success rates but low actual transaction volume should be excluded. For example, the implementation example uses statistical data on successful marketing scripts from all XXX marketing agents in a given quarter of a given year. .

[0093] in, x i Indicates the first i The number of successful transactions handled by each marketing agent nThis represents the number of marketing agents. The following uses a box plot to calculate successful marketing script data. X The marketing script data is filtered using the quartile Q1 value, calculated as follows: (1) First, all the data need to be sorted in ascending order.

[0094] (2) Then, calculate X The quartile position is calculated as (n+1)×0.25, where n is the number of elements in the dataset. If the result is an integer, the value corresponding to that position is taken directly; if it is a decimal, the value of Q1 is calculated using interpolation.

[0095] (3) Screen marketing agents whose number of successful business transactions exceeded that of Q1, and transcribe the audio files of their successful marketing scripts from year x quarter into text data using the Whisper (Large-v2) model. After screening, approximately 24,000 successful marketing script records were collected.

[0096] Step S220 involves fine-tuning and training an open-source large-scale model (such as Qianwen 2.0) with the experience of marketing script experts. This trained large language model then identifies and marks potential erroneous words in the marketing script text. Next, multiple methods, including pinyin features, edit distance, and semantic similarity, are used to replace suspected erroneous words. In essence, this process mainly includes two steps: locating erroneous words and correcting potential erroneous words.

[0097] (1) Locating incorrect words.

[0098] First, the experience of marketing script experts is fed into an open-source large model (such as Qianwen 2.0) for fine-tuning and training. Then, the model is used to label the probability of each word in the marketing script, and suspected erroneous words are located based on the probability. The specific method is as follows: Step 1: Generate word probability annotations in the marketing text.

[0099] The fine-tuned large language model generates a probability score for each word in the marketing text. This probability reflects the likelihood of each word appearing in a given context.

[0100] Step 2: Screening for suspected erroneous words.

[0101] First, a probability threshold is defined to identify potential erroneous words. The example first calculates the mean and standard deviation of all word probability values ​​based on the probability value generated for each word in each batch of marketing terminology audio text by the large model. The mean minus three times the labeled difference is used as the threshold for filtering erroneous word probabilities (this method can also be implemented using box plots, clustering, etc.). This threshold can be adjusted according to actual needs and model performance.

[0102] Then, words below a certain threshold are marked as potential errors. These words are considered unlikely to be correct in the current context, possibly due to spelling errors, grammatical errors, or inappropriate word use.

[0103] Step 3: Create a list of potential error words.

[0104] All low-probability words that have been tagged are collected into a potential error vocabulary list. This list will be used in subsequent error correction processes, including word replacement and user feedback loops.

[0105] Through this series of steps, the system can effectively identify and label potentially erroneous words, providing data support for further error correction. This probabilistic error word localization method, combined with the powerful language understanding capabilities of large-scale language models, makes error detection more accurate and efficient.

[0106] (2) Correct potential errors in words.

[0107] By constructing a high-frequency word list (i.e., the fixed phrases used to describe each package in the marketing materials), and matching this list with potential error words identified in the text, the most suitable replacement option is found. This process involves multiple matching dimensions, including syllable structure and phonological features, edit distance, and semantic similarity. The specific implementation method is as follows: Step 1: Set the matching dimensions.

[0108] The marketing scripts in the examples were all transcribed from audio files, which resulted in translation errors due to inaccurate pinyin pronunciation and multiple expressions for the same word in colloquial marketing scripts. To solve the matching of words with pinyin translation errors or synonyms, multiple methods based on pinyin similarity, semantic similarity, and edit distance were used for high-frequency word matching.

[0109] Pinyin similarity matching: Each word is converted into its corresponding pinyin representation using a pinyin conversion tool. The pinyin of each potential error word and high-frequency word is compared. By calculating the edit distance between pinyin or using character-level similarity analysis, it is determined which high-frequency words are similar in pronunciation to the error words. In this example, since inaccurate pinyin pronunciation leads to many translation errors, edit distance is used to calculate pinyin similarity. The calculation method is as follows: Suppose we have two strings A and B, with lengths respectively. and , Indicates the beginning of string A The first character is converted into string B. The minimum number of editing operations required to edit a character.

[0110]

[0111] in, If the pinyin characters are different, it is set to 1; if they are the same, it is set to 0.

[0112] Recursive calculation D [ m ][ n The minimum number of edit operations required to completely convert string A into string B.

[0113] Semantic similarity: Generate vector representations for each word using semantic embedding models (such as Word2Vec or BERT) and calculate the cosine similarity between these vectors.

[0114] (2) in, v 1 and v 2 is the vector representation of two words.

[0115] Character edit distance: Calculates the edit distance between each potential error word and a high-frequency word, which is the minimum number of single-character edit operations (insertion, deletion, replacement) required to transform one word into another.

[0116] Step 2: Generate a list of candidate words for replacement.

[0117] Integrated Matching Results: For each potential erroneous word, a comprehensive score is calculated for each high-frequency word by combining the matching results of pinyin features, edit distance, and semantic similarity. A weighted average method is used to merge the scores from different dimensions, where the weights can be adjusted according to the importance of each dimension in practical applications. In this embodiment, the weights for pinyin features, semantic similarity, and edit distance are calculated as follows: First, for pinyin feature and edit distance matching, the minimum number of edit operations between the potential error word and each high-frequency word is calculated, and the minimum value is taken as the minimum number of edit operations X for that potential error word (for semantic similarity, the maximum similarity with each high-frequency word is taken). Then, the N potential error words are normalized. The calculation formula is as follows: (3) in, For X's i One value, Let X be the minimum value. Let X be the maximum value of X. The method for calculating semantic similarity is as follows: .

[0118] Calculate each similarity matching index separately. The proportion of all values ​​under this indicator: (4) The above formula essentially defines a "probability" and links it to its possible values. This is done because the entropy calculated below is only related to the probability. If the possible values ​​of are highly diverse, the probability calculated using this value will also be highly diverse, varying greatly, thus reducing the entropy. The entropy calculations for each similarity matching feature are as follows: (5) The difference coefficients for each similarity matching feature are then calculated: g = 1 - e (6) In this embodiment, to further optimize the weight calculation of each similarity matching feature, the accuracy of each similarity matching is manually verified by randomly selecting 1000 latent word matches. The difference coefficient g corresponding to the accuracy of each similarity matching method is then normalized by weighting. The calculation method is as follows: (7) in, r j Indicates the first j The accuracy of several similarity matching methods, j=1,2,3. g j Indicates the first j The difference coefficient of similarity matching methods.

[0119] (8) Selecting the best replacement: For each potential incorrect word, select the highest-frequency word with the highest overall score as a replacement candidate. Create a word replacement candidate table, listing each incorrect word and its recommended high-frequency replacement word. i The first potential misspelling and the first m The comprehensive score calculation method for each high-frequency word is as follows: in, Indicates the first i The first potential misspelling word utilizes the first j The first similarity matching method calculates the second similarity matching method. m The normalized value of the similarity of the nth high-frequency words. Take the nth... i Each potential error word is matched with the word that corresponds to the highest matching score of all high-frequency words as a candidate matching word.

[0120] By comprehensively utilizing pinyin features, edit distance, and semantic similarity, this multi-dimensional word matching method can more accurately identify and replace incorrect words.

[0121] Step 230: Use a fluency detection model to detect sentence fluency in the pre-training corpus to ensure the standardization of the pre-training corpus.

[0122] In this embodiment, the large language model is a statistical data system built from a large corpus through training, word segmentation, word frequency calculation, and other processes. Among them, N-Gram is a common language algorithm model based on statistical rules. Its input is a sentence (a sequential sequence of words), and its output is the probability of the sentence, that is, the joint probability of these words. In this paper, N=2 is selected, as shown in Equations (9) and (10).

[0123] (9) (10) In the formula: count ( a )express a The number of times it appears in the training set.

[0124] In this embodiment, due to the different actual dialogue scenarios and varying lengths of the dialogue, the calculated probabilities of sentences differ significantly, greatly affecting the final score calculation. Therefore, this paper aggregates sentences of the same length and normalizes the calculated sentence probabilities, thus controlling the sentence probabilities within a certain range. Finally, different types of sentences are distinguished based on a set threshold.

[0125] The quality of a sentence is judged by the following steps: First, the average of multiple statistical metrics of the 2-Gram probability in the training set is calculated as a threshold. Second, the probability of each sentence in the training set is normalized. Finally, the average probability of sentences in the training set is compared with the normalized probability values ​​of sentences in the test set. If a sentence of the same length as a sentence in the test set cannot be found in the training set, the probability value of that sentence in the test set is directly compared with the probability value calculated from the overall training samples.

[0126] Step 240: Train the marketing message content selling point keyword recognition model by pre-training the model and using the RLHF framework.

[0127] (1) Construct the dataset.

[0128] First, by creating prompts, key short phrases and words highlighting marketing selling points are extracted from the audio content of marketing terminology. The prompts are shown below: Role: Marketing strategy analyst and communication skills expert.

[0129] Background: Users need to extract key information from a marketing statement and present it in a structured manner using the FABE (Features, Advantages, Benefits, Evidence) model to improve marketing efficiency and persuasiveness.

[0130] Profile: You are an experienced marketing strategy analyst who excels at transforming complex marketing messages into concise and powerful language. You are also a skilled communicator who can guide users on how to deliver information more effectively.

[0131] Skills: You possess excellent information extraction, language organization, and communication skills. You can quickly identify key elements in marketing rhetoric and organize them effectively according to the FABE model.

[0132] Goals: Help users extract key phrases and short words from marketing messages to improve the efficiency and persuasiveness of marketing messages.

[0133] Constraints: The extracted information should be concise, clear, easy to understand and remember, while maintaining the original intent and information integrity.

[0134] OutputFormat: Provides key phrases and corresponding words for the FABE model.

[0135] Workflow: 1. Read and understand the full text of the marketing script.

[0136] 2. Based on the FABE model, identify and extract key information from the speech.

[0137] 3. Organize the extracted information into concise short sentences, and provide one or more corresponding short words for each key point.

[0138] Examples: Example 1: F (Features).

[0139] Short sentence: The package includes home broadband, WiFi, home monitoring, and overseas game acceleration service.

[0140] Short terms: home broadband, WiFi, home monitoring, game acceleration.

[0141] Example 2: A (Advantages).

[0142] Short sentence: High cost performance, offers a variety of usage rights, and has a fast and stable network.

[0143] Short phrases: high cost-performance ratio, multiple benefits, stable network.

[0144] Example 3: B (Benefits).

[0145] Short sentence: Meets the needs of home networking, security monitoring, and gaming experience, making it more reassuring to use.

[0146] Short phrases: meet needs, feel at ease.

[0147] Example 4: E (Evidence).

[0148] Short sentence: I had a good experience after installing it in my own home.

[0149] Short word: for one's own use.

[0150] Then, the large model, combined with cue words, extracts key phrases and corresponding short words from the processed marketing rhetoric text. Simultaneously, short phrases and short words in the dialogue content are manually annotated based on the marketing rhetoric.

[0151] (2) Model Training: This example introduces the RLHF framework and PPO algorithm, utilizing KL divergence and updated ActorModel policies. Through iterative updates, the policy learning enhances the expected reward results, striving to make the ActorModel's behavior more consistent with high-quality dialogue standards. The model is defined as follows: Actor Model: defined as The function of the model is to represent the state. s Its corresponding action a Association is used to generate strategies for Actors, quantifying the probability of each possible action. The example uses the Qianwen 2.0 model to initialize the Actor Model weights.

[0152] Critic Model: Defined as The Critic model evaluates the value of a given state s, providing key values ​​to guide the training trajectory.

[0153] Reward Model: Defined as R ( s , a This model, through action... a and state s The evaluation is used to assign reward values ​​that reflect the quality of the generated sequence. In this example, to address the issue of the reward model relying on preference feedback, an LLM (Limited Least Metric) is used to train the reward model. Figure 3 As shown, Figure 3 The provided example provides a framework diagram of RLHF learning a reward model from human text input.

[0154] The example uses LLM to transform human evaluation feedback text in natural language form into state-level rewards. These labeled states are used to train a reward model, which is then used in conjunction with a traditional reinforcement learning algorithm to train an agent.

[0155] Human evaluations, along with information about the environment and state-action sequences, are input into the LLM. The LLM then processes this information and determines which states the human feedback refers to. That is, the LLM generates a set of states and their corresponding labels (positive or negative). This set is used as a training dataset to train the reward model.

[0156] The agent tracks the reward probability distribution for each state, which is modeled as a beta distribution. Specifically, we initialize the reward distribution to β(0.5, 0.5) to introduce a bias towards binary rewards: 0 (negative) or 1 (positive). For ease of analysis, the embodiment models the observed labels using conjugate priors, i.e., a Bernoulli distribution. Thus, the posterior of a given state probability distribution β(a, b) becomes β(a, b + 1) if a positive label is received, and β(a + 1, b) if the label is negative.

[0157] Ref Model: defined as This model serves as a pre-trained reference point, providing a standard for action evaluation. The Actor model and Ref model are two copies of the supervised fine-tuning model from the first stage of RLHF, while the Reward model and Critic model are two copies of the reward model. The Actor model's learning process is guided by the PPO algorithm, updating the policy in a way that maximizes performance while avoiding large deviations from previous policies. The objective function for updating the Actor model is shown below: (11) in, This represents the ratio of the probability of state a occurring under the current policy to the probability of it occurring under the old policy. This indicates the advantage function estimated using generalized advantage estimation (GAE). These are hyperparameters that the control strategy allows to change to a certain extent. The GAE calculation uses a weighted sum of n-step time difference (TD) residuals, formally defined as: (12) In the formula, This is the TD residual at time t. The advantage estimate of GAE is: (13) It is the parameter of the bias-variance tradeoff in the balance advantage estimation.

[0158] This example uses the distributed computing training framework DeepSpeed, employing eight Ascend 910B GPUs to train the Qianwen 2.0 model. The learning rate for the Actor Model is set to 5e-6, the learning rate for the Critic Model is set to 5e-7, and the clipping parameter is set to 0.2. To ensure a balanced training process, a discount factor is set. gamma Keep it at 0.95.

[0159] Step 250: Generate package marketing script. After completing the keyword identification model for the selling points of the marketing script, key short phrases for package selling points are extracted from historical successful marketing scripts. When conducting package marketing, these key short phrases for package selling points and basic package information are used as prompts to allow the large model to generate a complete marketing script.

[0160] After marketing agents complete customer transactions based on the generated scripts, they can score the scripts, improve and optimize the content, extract key marketing points from the revised scripts, and store them in the script corpus as high-quality samples for continuous optimization of the marketing script content and keyword recognition model.

[0161] The embodiment proposes a method for recommending package marketing scripts. First, high-quality marketing script pronunciations are collected to increase the frequency of marketing selling points. Second, the marketing script text content is revised to improve the quality of the marketing script data. At the same time, the fluency of the marketing script text is further detected to filter out high-quality marketing script text data. Then, a marketing script content selling point keyword recognition model is trained to accurately identify the key marketing selling points of different packages. Finally, the package marketing selling points and basic package information are used as prompt words, and a large model is used to generate marketing scripts for different packages to improve the success rate of package marketing.

[0162] This embodiment employs a comprehensive approach, utilizing pinyin features, edit distance, and semantic similarity to calculate the similarity between potential error words and high-frequency words. To improve the accuracy of matching with high-frequency words, it proposes normalizing the accuracy rates of different similarity matching methods and using this as the initial weight. Furthermore, it uses a weighted difference coefficient normalization method to calculate the weights of different similarity methods, ultimately calculating the comprehensive matching score between potential error words and high-frequency words. This multi-dimensional word matching method can more accurately identify and replace error words.

[0163] The human-feedback RLHF provided in the embodiment uses LLM to efficiently train the reward model by leveraging information from human-input text. Human evaluations, along with information about the environment and state-action sequences, are input into the LLM. The LLM generates a set of states and their corresponding positive or negative labels. The agent tracks the reward probability distribution for each state, which is a beta distribution. By combining text feedback with LLM with minimal human-computer interaction, the proposed method accurately approximates the reward function, thereby significantly improving model performance.

[0164] By dividing each functional module according to its corresponding function, this application provides a marketing script generation device, which can be a server, a terminal, or a chip applied to a server. Figure 4 A schematic block diagram of the functional modules of a marketing script generation device provided for an exemplary embodiment of this application. Figure 4 As shown, the marketing script generation device includes: Data acquisition module 41 is used to acquire marketing script recordings with a historical marketing success rate higher than a first threshold, and convert the marketing script recordings into initial text data; Data correction module 42 is used to identify and correct erroneous words in the initial text data to obtain corrected text data; Data filtering module 43 is used to perform sentence fluency detection on the corrected text data and filter out high-quality text corpus; Training module 44 is used to train a marketing message content selling point keyword recognition model based on the high-quality text corpus using the feedback human reinforcement learning (RLHF) framework. The key phrase output module 45 is used to input the marketing text to be analyzed into the trained selling point keyword recognition model and output the key phrases of the package selling points. The marketing script generation module 46 is used to combine the key short phrases of the package's selling points and the basic information of the package into prompt words, and generate target marketing scripts based on the prompt words and the large language model.

[0165] In another embodiment provided in this application, the data correction module 42 is specifically used for: The initial text data is probabilistically labeled using a finely tuned large language model, and words with probabilities below a second threshold are marked as potential error words. Calculate the comprehensive score of the potential erroneous words and the words in the preset high-frequency word list in terms of pinyin features, edit distance, and semantic similarity. The most frequently used words with the highest overall scores were selected to replace the potentially incorrect words.

[0166] In another embodiment provided in this application, the data filtering module 43 is specifically used for: The probability of sentences in the corrected text data is calculated using the N-Gram language model; The sentences are grouped according to their length, and the probabilities of the sentences in each group are normalized. The normalized sentence probability is compared with a preset probability threshold to determine whether the fluency of the sentences in the corrected text data meets the standard.

[0167] In another embodiment provided in this application, during the training of the marketing script content selling point keyword recognition model using the Feedback Human Reinforcement Learning (RLHF) framework, the training of the marketing script content selling point keyword recognition model is guided by a reward model. The device further includes a training processing module, specifically used for: Input human evaluation feedback text and state action sequence information into a large language model; Based on the large language model, a set of states and corresponding labels are generated, wherein the labels include positive labels or negative labels; The reward probability distribution corresponding to the state is updated based on the label, and the reward probability distribution is modeled as a Beta distribution. The reward model is trained based on the updated reward probability distribution, and the reward model is used to optimize the selling point keyword recognition model through a reinforcement learning algorithm.

[0168] In another embodiment provided in this application, the apparatus further includes a training dataset generation module, specifically used for: Prompt words were written based on the FABE model, and key short phrases and words for marketing selling points were extracted from the high-quality text corpus using a large language model. The results extracted by the large language model are manually labeled to generate a training dataset for training the selling point keyword recognition model.

[0169] In another embodiment provided in this application, the apparatus further includes a model optimization module, used for: Obtain the optimization results for the target marketing script; The optimization results are added to the high-quality text corpus for training the selling point keyword recognition model.

[0170] This application also provides an electronic device, including: at least one processor; a memory for storing executable instructions of the at least one processor; wherein the at least one processor is configured to execute the instructions to implement the method disclosed in the embodiments of this application.

[0171] Figure 5This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application. For example... Figure 5 As shown, the electronic device 1800 includes at least one processor 1801 and a memory 1802 coupled to the processor 1801. The processor 1801 can perform the corresponding steps in the methods disclosed in the embodiments of this application.

[0172] The processor 1801 described above can also be called a central processing unit (CPU), which can be an integrated circuit chip with signal processing capabilities. Each step in the method disclosed in this application can be implemented by the integrated logic circuitry in the hardware of the processor 1801 or by instructions in software form. The processor 1801 can be a general-purpose processor, a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this application can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can be located in the memory 1802, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The processor 1801 reads information from the memory 1802 and, in conjunction with its hardware, completes the steps of the above method.

[0173] Furthermore, the various operations / processes according to this application, when implemented via software and / or firmware, can be transmitted from a storage medium or network to a computer system with a dedicated hardware architecture, such as... Figure 6 The computer system 1900 shown is equipped with the programs that constitute the software. When various programs are installed, the computer system is able to perform various functions, including those described above. Figure 6 A structural block diagram of a computer system provided for an exemplary embodiment of this application.

[0174] Computer System 1900 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present application described and / or claimed herein.

[0175] like Figure 6 As shown, the computer system 1900 includes a computing unit 1901, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 1902 or a computer program loaded from a storage unit 1908 into a random access memory (RAM) 1903. The RAM 1903 may also store various programs and data required for the operation of the computer system 1900. The computing unit 1901, ROM 1902, and RAM 1903 are interconnected via a bus 1904. An input / output (I / O) interface 1905 is also connected to the bus 1904.

[0176] Multiple components in computer system 1900 are connected to I / O interface 1905, including: input unit 1906, output unit 1907, storage unit 1908, and communication unit 1909. Input unit 1906 can be any type of device capable of inputting information into computer system 1900. Input unit 1906 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 1907 can be any type of device capable of presenting information and may include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 1908 may include, but is not limited to, hard disks and optical disks. Communication unit 1909 allows computer system 1900 to exchange information / data with other devices via a network such as the Internet, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0177] The computing unit 1901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1901 performs the various methods and processes described above. For example, in some embodiments, the methods disclosed in the embodiments of this application can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 1908. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 1902 and / or communication unit 1909. In some embodiments, the computing unit 1901 can be configured to perform the methods disclosed in the embodiments of this application by any other suitable means (e.g., by means of firmware).

[0178] This application also provides a computer-readable storage medium, wherein when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is able to perform the methods disclosed in this application.

[0179] The computer-readable storage medium in this application embodiment may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. The aforementioned computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specifically, the aforementioned computer-readable storage medium may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.

[0180] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0181] This application also provides a computer program product, including a computer program, wherein the computer program, when executed by a processor, implements the methods disclosed in the embodiments of this application.

[0182] In embodiments of this application, computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network (including a local area network (LAN) or a wide area network (WAN)), or it can be connected to an external computer.

[0183] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0184] The modules, components, or units described in the embodiments of this application can be implemented in software or hardware. The names of the modules, components, or units do not necessarily constitute a limitation on the module, component, or unit itself.

[0185] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary hardware logic components that can be used include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0186] The above description is merely an embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

[0187] While specific embodiments of this application have been described in detail by way of examples, those skilled in the art should understand that the above examples are for illustrative purposes only and are not intended to limit the scope of this application. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this application. The scope of this application is defined by the appended claims.

Claims

1. A method for generating marketing scripts, characterized in that, The method includes: Acquire marketing script recordings with a historical marketing success rate higher than a first threshold, and convert the marketing script recordings into initial text data; Error words are identified and corrected in the initial text data to obtain corrected text data; Sentence fluency was tested on the corrected text data to select high-quality text corpus. Based on the high-quality text corpus, a marketing message content selling point keyword recognition model was trained using the Feedback Human Reinforcement Learning (RLHF) framework. Input the marketing text to be analyzed into the trained keyword recognition model for selling points, and output key short phrases for the package's selling points; The key selling points of the package and basic package information are combined into prompt words, and target marketing scripts are generated based on the prompt words and a large language model.

2. The method according to claim 1, characterized in that, The process of identifying and correcting erroneous words in the initial text data includes: The initial text data is probabilistically labeled using a finely tuned large language model, and words with probabilities below a second threshold are marked as potential error words. Calculate the comprehensive score of the potential erroneous words and the words in the preset high-frequency word list in terms of pinyin features, edit distance, and semantic similarity. The most frequently used words with the highest overall scores were selected to replace the potentially incorrect words.

3. The method according to claim 1, characterized in that, The sentence fluency detection of the corrected text data includes: The probability of sentences in the corrected text data is calculated using the N-Gram language model; The sentences are grouped according to their length, and the probabilities of the sentences in each group are normalized. The normalized sentence probability is compared with a preset probability threshold to determine whether the fluency of the sentences in the corrected text data meets the standard.

4. The method according to claim 1, characterized in that, In the process of training the marketing script content selling point keyword recognition model using the Feedback Human Reinforcement Learning (RLHF) framework, the training of the marketing script content selling point keyword recognition model is guided by a reward model. The method further includes: Input human evaluation feedback text and state action sequence information into a large language model; Based on the large language model, a set of states and corresponding labels are generated, wherein the labels include positive labels or negative labels; The reward probability distribution corresponding to the state is updated based on the label, and the reward probability distribution is modeled as a Beta distribution. The reward model is trained based on the updated reward probability distribution, and the reward model is used to optimize the selling point keyword recognition model through a reinforcement learning algorithm.

5. The method according to claim 1, characterized in that, The method further includes: Based on the FABE model, prompt words were written, and a large language model was used to extract key short phrases and words of marketing selling points from the high-quality text corpus. The results extracted by the large language model are manually labeled to generate a training dataset for training the selling point keyword recognition model.

6. The method according to claim 1, characterized in that, The method further includes: Obtain the optimization results for the target marketing script; The optimization results are added to the high-quality text corpus for training the selling point keyword recognition model.

7. A marketing script generation device, characterized in that, The device includes: The data acquisition module is used to acquire marketing script recordings with a historical marketing success rate higher than a first threshold, and convert the marketing script recordings into initial text data; The data correction module is used to identify and correct erroneous words in the initial text data to obtain corrected text data. The data filtering module is used to perform sentence fluency detection on the corrected text data and filter out high-quality text corpora. The training module is used to train a marketing message content selling point keyword recognition model based on the high-quality text corpus using the Feedback Human Reinforcement Learning (RLHF) framework. The key phrase output module is used to input the marketing text to be analyzed into the trained selling point keyword recognition model and output the key phrases of the package selling points. The marketing script generation module is used to combine key short phrases of the package's selling points and basic package information into prompt words, and generate target marketing scripts based on the prompt words and a large language model.

8. An electronic device, characterized in that, include: At least one processor; Memory for storing the at least one processor-executable instruction; The at least one processor is configured to execute the instructions to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-6.

Images