Commodity introduction text generation method and related device

Abstract

The invention provides a commodity introduction text generation method and a related device. The method comprises the following steps: extracting a first-level label and a second-level label from a target user question set corresponding to a commodity; wherein the target user question set comprises a plurality of user questions for the commodity; the first-level label is used for expressing a text theme; the secondary labels are used for expressing text elements required by corresponding text topics; constructing a semantic template corresponding to the first-level label; wherein the semantic template further comprises a second-level label related to the corresponding first-level label; and according to knowledge entries obtained from a data source and the semantic template, constructing a text generation prompt instruction, calling a large language model to instruct the large language model to select text elements in the knowledge entries according to the text elements expressed by the second-level tag, and forming a commodity introduction text of a text theme expressed by the first-level tag. The content comprehensiveness of the commodity introduction text can be improved to a certain extent.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and related apparatus for generating product description text. Background Technology

[0002] In the field of product knowledge construction and maintenance, product description texts for display are typically generated manually or through fixed templates, simple rules, and large language models, based on knowledge entries from various data sources such as product manuals, marketing copy submitted by merchants, historical operational scripts, and product display pages.

[0003] However, existing product description texts still sometimes lack content that users care about, resulting in insufficient comprehensiveness of knowledge in the product description texts. Summary of the Invention

[0004] In view of this, one or more embodiments of this application provide a method and related apparatus for generating product description text, which can improve the comprehensiveness of the product description text to a certain extent.

[0005] In a first aspect, one or more embodiments of this application propose a method for generating product description text, comprising: extracting primary tags and secondary tags from a set of target user questions corresponding to the product; wherein the set of target user questions includes multiple user questions about the product; secondary tags belong to primary tags, and different primary tags do not include the same secondary tags; primary tags are used to express the text theme; secondary tags are used to express the text elements required for the corresponding text theme; constructing a semantic template corresponding to the primary tags; wherein the semantic template also includes secondary tags involved in the corresponding primary tags; constructing a text generation prompt instruction based on knowledge entries obtained from a data source and the semantic template, and calling a large language model to instruct the large language model to select text elements from the knowledge entries according to the text elements expressed by the secondary tags, thereby forming product description text whose text theme is expressed by the primary tags.

[0006] Secondly, one or more embodiments of this application propose an apparatus for generating product description text, comprising: an extraction module, configured to extract primary tags and secondary tags from a set of target user questions corresponding to the product; wherein the set of target user questions includes multiple user questions related to the product; secondary tags belong to primary tags, and different primary tags do not include the same secondary tags; primary tags are used to express the text theme; secondary tags are used to express the text elements required to express the corresponding text theme; a construction module, configured to construct a semantic template corresponding to the primary tags; wherein the semantic template also includes secondary tags involved in the corresponding primary tags; and a generation module, configured to construct text generation prompt instructions based on knowledge entries obtained from a data source and the semantic template, and call a large language model to instruct the large language model to select text elements from the knowledge entries based on the text elements expressed by the secondary tags, thereby forming product description text whose text theme is expressed by the primary tags.

[0007] Thirdly, one or more embodiments of this application provide a computer device including a memory and a processor, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the method as described above.

[0008] Fourthly, one or more embodiments of this application provide a computer program product including computer instructions that, when executed by a processor, implement the method as described above.

[0009] Fifthly, one or more embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to implement the method as described above.

[0010] As can be seen from the above embodiments, multiple embodiments of this application extract primary tags for expressing text topics and secondary tags for expressing the text elements required for the corresponding text topics from the target user question set corresponding to the product. Based on the primary tags and the secondary tags involved, a semantic template containing the expected text topic and text element constraints is constructed. Then, based on the knowledge entries obtained from the data source and the semantic template, a text generation prompt instruction is constructed to call the large language model. This enables the large language model to select the corresponding text elements from the knowledge entries according to the text elements expressed by the secondary tags and generate product introduction text with the text topic expressed by the primary tags. This achieves accurate control of the text topic and text elements of the generated content in the process of generating product introduction text with user questions as the guide, making the product introduction text more in line with user concerns and covering key information more completely. Attached Figure Description

[0011] Figure 1This is a schematic diagram of a method for generating product description text according to an embodiment of this application.

[0012] Figure 2 This is a schematic diagram illustrating the relationship between primary tags, secondary tags, and questions provided in one embodiment of this application.

[0013] Figure 3 This is a flowchart illustrating a method for generating product description text according to one embodiment of this application.

[0014] Figure 4 This is a schematic diagram of a module of a product description text generation device provided in one embodiment of this application.

[0015] Figure 5 This is a schematic diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments.

[0017] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0019] In the field of product description text generation, related technologies typically rely on data sources such as product manuals, marketing copy submitted by merchants, and historical operational scripts. Content is manually organized or generated using fixed templates, simple rules, and large language models. While this approach can meet basic product description needs to some extent, the generated content often depends heavily on the original structure of the data source or the preset structure of the template, failing to adequately address the specific questions users raise in real-world inquiries. Especially in scenarios with a wide variety of products and diverse user concerns, the product description text may lack key textual elements that users focus on, rendering it less practical.

[0020] With the widespread application of large language models, their text generation methods are also being used in product description texts. However, large language models may still encounter problems such as content omissions, topic deviations, or insufficient coverage of key information when generating product description texts. Especially in fields involving medicine and healthcare, where accurate expression of key points is crucial, large language models may generate untargeted information due to insufficient focus on content of user concern, further affecting the effectiveness of product description texts.

[0021] In summary, the relevant technologies still have problems such as product description texts failing to accurately cover the content that users care about, the text themes not being focused enough, and the expression of key text elements being incomplete, which urgently need further improvement.

[0022] In several embodiments provided in this application, the method for generating product description text can be applied to electronic devices with certain computing power and network access capabilities. This electronic device can be a desktop computer, laptop computer, tablet computer, smartphone, or a server. Specifically, the electronic device includes a processor, memory, and a network access module for network communication. The server can be an electronic device with strong data processing capabilities; of course, a server can also refer to a server cluster formed by multiple electronic devices, or a quantum server built using a quantum computer.

[0023] Please see Figure 1 and Figure 2 One embodiment of this application provides an application scenario example of a method for generating product description text. This method can be applied to a generation device. The device can generate product description text by acquiring a set of user questions about the product, extracting a tag system, constructing a semantic template, and calling a large language model. It also combines multi-dimensional quality assessment and conflict correction to achieve structured generation and iterative optimization of the product description text. In this scenario example, the processing procedure of the generation device is described using healthcare products as an example.

[0024] When a merchant prepares to update the product description of a vitamin supplement, the generation device can first receive a set of user questions related to the product. For example, the user question set may include data from historical searches, online inquiries, and user reviews, such as "How many tablets should I take daily?", "Who is this product suitable for?", "Are there any adverse reactions?", and "Can it be taken with other health supplements?". When generating product description text for the first time using the method provided in this application, the device can count the number of user questions and select a specified number of questions with the highest frequency to form a target user question set. Of course, in some cases, the generation device can also analyze this user question set to identify which user questions are not yet covered in the existing product description text, thus obtaining the target user question set. For example, the generation device can identify that the current product description text does not include content such as "adverse reaction instructions" and "applicable population restrictions," and the corresponding user questions become the target user question set.

[0025] After receiving a set of questions from the target users, the generation device can cluster the questions to obtain multiple clusters expressing different semantic topics. In this example scenario, the generation device can categorize user questions such as "how many tablets to take daily," "how to take," and "when to take" into the "usage method" cluster; questions such as "is it suitable for children," "can pregnant women take it," and "who is not suitable" into the "suitable population" cluster; and questions such as "what are the adverse reactions" and "will it cause allergies" into the "precautions" cluster. The generation device can automatically generate corresponding first-level tags based on the semantic content of each cluster, such as "usage method," "suitable population," and "precautions."

[0026] After the primary tags are generated, the generation device can further divide the user questions within each primary tag's corresponding cluster into sub-clusters to extract more granular expressive elements for constructing secondary tags. For example, under the primary tag "How to use", the generation device can categorize "How many tablets to take per day" as "Dosage instructions", "Before or after meals" as "Timing of administration", and "Chewing or swallowing" as "Method of administration". Under the primary tag "Suitable population", the generation device can categorize "Can pregnant women use it?" as "Suitability during pregnancy" and "Can children use it?" as "Suitability for children". Under the primary tag "Precautions", "What are the adverse reactions?" as "Contraindications", and "Will there be allergies?" as "Allergy history". The generation device can automatically generate corresponding secondary tags based on these sub-clusters to express the text elements to be covered under each text theme.

[0027] After obtaining the primary and secondary tag system, the generation device can construct a corresponding semantic template for each primary tag. In this scenario example, the generation device can construct a statement template for "Usage Method" that includes reserved spaces for "Dosage Instructions," "Directions for Use," and "Timing of Use," such as "The daily dosage is {Dosage Instructions}, and it is recommended to take it at {Timing of Use} using {Directions for Use}." The generation device can construct a statement template for "Suitable Population" that includes reserved spaces for "Suitability During Pregnancy" and "Suitability for Children," such as "This product is suitable for {Suitable for Children}, but caution should be exercised when using it for {Suitable for Pregnant Women}." These statement templates constitute semantic templates, used to guide the large language model in filling in the correct text elements when generating text.

[0028] Before the generation stage, the generation device can pre-optimize the large language model through supervised fine-tuning (SFT), enabling it to summarize and annotate literature sources based on knowledge source content. To this end, the generation device can construct training datasets based on sources such as product manuals, marketing copy, internal customer service knowledge bases, and user reviews, and perform structured annotations on the content, such as labeling "Dosage instructions are from paragraph X of the manual" or "Applicable audience information is from entry Y of the customer service knowledge base." The generation device can use these training samples with clearly defined citations to perform supervised fine-tuning of the large language model, allowing the model to better retain and accurately cite knowledge sources when generating subsequent product description text, thus obtaining a preliminarily usable large language model.

[0029] During the generation phase, the device can extract product-related knowledge items from multiple data sources, including product manuals, marketing copy, historical customer service Q&As, and operational data. For image and video content, the device can use OCR (Optical Character Recognition) technology to extract text from images and ASR (Automatic Speech Recognition) technology to extract video subtitles. Furthermore, it uses NLP (Natural Language Processing) technology to clean and normalize the multimodal text, removing redundant descriptions and standardizing sentence formatting to obtain the knowledge items. The manual might contain information such as "adults take one tablet daily" or "use with caution during pregnancy," while the marketing copy might contain selling points such as "boosts immunity." The device can construct text generation prompts based on semantic templates and call a large language model fine-tuned using SFT to execute the text generation task. The large language model can select text elements corresponding to secondary tags based on prompts. For example, it can fill in the dosage instructions for "1 tablet daily for adults" from the instruction manual, and fill in the pregnancy applicability for "use with caution in pregnant women" from the medical guidelines, thus forming a product description text segment that expresses a complete text theme. The generation device can generate multiple text paragraphs for different primary tags and combine them to form the product description text.

[0030] After generating the product description text, the generation device can further verify the quality of the text content. In this example scenario, the generation device can perform semantic conflict matching between the text content and knowledge entries from various data sources to identify content that may contradict correct knowledge entries. For example, if the statement "Safe for children to take" generated by the large language model conflicts with "Not recommended for children" in the instruction manual, the generation device can identify the conflict and, if the instruction manual content is deemed to have higher semantic credibility, correct the product description text based on the instruction manual content.

[0031] After correcting conflicting content, the generation device can score the quality of product description text based on viewpoint richness, tag fill rate, source reliability, external knowledge votes, and timeliness. For example, in this scenario, the generation device can calculate the number of unique viewpoints in the text to obtain viewpoint richness; determine whether each secondary tag is fully filled using a semantic template to obtain the tag fill rate; calculate source reliability based on the historical accuracy weights of different data sources; determine the semantic consistency between the text's conclusions and authoritative knowledge by comparing them with external medical literature to obtain external knowledge votes; and calculate timeliness by combining the update time of knowledge entries. The generation device can input these dimensional scores into a reward model to obtain a quality score, including multiple dimensions such as VRS, LCS, SRS, EVS, and TS.

[0032] Building upon quality scoring, the generation device can further enhance the large language model using reinforcement learning (RLHF) and policy optimization algorithms, such as Proximal Policy Optimization (PPO) or Direct Preference Optimization (DPO). The generation device can use the scores provided by the reward model as feedback signals to update the parameters of the large language model through policy gradients. This allows the large language model to automatically strive for higher viewpoint richness, more complete label filling rate, higher source reliability, higher external knowledge consistency, and higher knowledge timeliness in subsequent text generation. Through this reinforcement learning process, the generation device obtains an optimized large language model, resulting in continuous improvement in the quality of the generated text.

[0033] If the quality score does not meet the preset threshold, the generation device can continue to call the large language model to update the product description text, such as supplementing missing text elements, correcting content with low credibility, or replacing outdated knowledge fragments, until the product description text reaches a high level of credibility and content completeness.

[0034] In this scenario example, after completing the quality assessment, the generation device can further perform automated knowledge completion and T+1 updates. When the knowledge entries corresponding to certain user questions are not present in the existing knowledge base, the generation device can automatically retrieve and generate candidate knowledge from the original document based on the uncovered knowledge list. For example, it can extract "not recommended for children" from the full text of the instruction manual or generate new adverse reaction prompts from the latest announcements from the drug regulatory authority. Subsequently, the generation device can automatically push these candidate knowledge to the operations platform for manual review and supplementation. Every day at midnight, the generation device can automatically execute a T+1 batch processing task, incrementally writing the product description text approved the previous day into the knowledge base, ensuring that the knowledge base content remains continuously updated.

[0035] Once the product description text has undergone final quality verification after being updated, the generation device can widely apply it to various business scenarios. In this example scenario, the generation device can directly call the generated product description text in an AI customer service scenario to provide accurate and traceable answers to user questions; in product search, the tag system based on the product description text enhances semantic understanding capabilities, thereby improving recall and ranking relevance; in the recommendation system, the product description text serves as a reasoning link, providing structured reasons such as "suitable audience" and "core selling points" to support "why recommend this product"; in operational scenarios, it automatically generates a product launch checklist and intuitively displays the knowledge gaps and review progress of the product description text in the knowledge maintenance dashboard, enabling the operations team to efficiently carry out batch product maintenance tasks.

[0036] Please see Figure 3 One embodiment of this application provides a method for generating product description text. The method for generating product description text can be applied to a generating device, which can be applied to the aforementioned electronic device possessing certain computing power and network access capabilities. Of course, in some embodiments, the generating device can also be software running on the electronic device. The method for generating product description text may include the following steps.

[0037] Step S110: Extract primary tags and secondary tags from the target user question set corresponding to the product; wherein, the target user question set includes multiple user questions for the product; secondary tags belong to primary tags, and different primary tags do not include the same secondary tags; primary tags are used to express the text theme; secondary tags are used to express the text elements required for the corresponding text theme.

[0038] Step S120: Construct a semantic template corresponding to the first-level tag; wherein, the semantic template also includes the second-level tags involved in the corresponding first-level tag.

[0039] Step S130: Construct text generation prompt instructions based on the knowledge entries obtained from the data source and the semantic template, call the large language model to instruct the large language model to select text elements in the knowledge entries according to the text elements expressed by the secondary tags, and form product introduction text expressing the text theme by the primary tags.

[0040] In this embodiment, product description text can be used to present product information to users. Product description text can be displayed on the product page or used to generate answers to user questions, allowing users to quickly obtain the content they are interested in when browsing the product page or inquiring about related information. In some embodiments, the same product can have multiple product description texts, each of which can include at least one text paragraph and expresses a text theme for the product. Different product description texts for the same product correspond to different text themes. In some embodiments, product description text can be generated from text content used for product display in the merchant's backend, or it can be used to support various application scenarios such as customer service assistance, mobile display, and search scenario display.

[0041] In this embodiment, the generation device can extract primary and secondary tags based on the target user question set corresponding to the product. The target user question set can include multiple user questions about the product, representing the concerns raised by users regarding the product during actual consultations or searches. Examples include "Who is this product suitable for?", "How many times a day should it be used?", and "Are there any adverse reactions?". In some embodiments, the target user question set can be generated based on various data sources such as historical search records, online consultation dialogues, user reviews, and a question database compiled by operations. Primary tags can be used to express text topics and categorize user questions covered in the target user question set. For example, user questions related to "usage method", "applicable population", and "precautions" can be categorized into different text topics. Secondary tags belong to the primary tags and are used to express the text elements required for the corresponding text topic. For example, under the text topic of "usage method", secondary tags can be used to distinguish different text elements such as "dosage frequency", "duration of use", and "method of use". To ensure a clear hierarchical structure in the tagging system, different first-level tags do not include the same second-level tags. This ensures that each second-level tag has a unique text theme within the entire tagging system, which is beneficial for generating product description text based on first-level and second-level tags.

[0042] In this embodiment, the generation device can construct a semantic template corresponding to the primary tag based on the primary tag. The semantic template can be used to describe the expected expression structure and content of the product description text under a certain text theme. Specifically, for each primary tag, a semantic template corresponding to that primary tag can be constructed. The semantic template includes secondary tags related to the corresponding primary tag, used to indicate the various text elements that need to be covered under the text theme. For example, for a primary tag expressing the text theme of "usage method", its corresponding semantic template can include descriptions of text elements related to secondary tags such as "usage frequency", "usage method", and "usage period". In some embodiments, the semantic template can adopt a structured data form or a predefined template description form to clarify the text elements and expression order that need to be covered under different text themes, so that the subsequently generated product description text can revolve around the text theme expressed by the primary tag and achieve relatively complete content coverage at the level of text elements corresponding to the secondary tags.

[0043] In this embodiment, the generation device can also construct text generation prompts based on knowledge entries obtained from the data source and the semantic template, and call a large language model to generate product description text. Knowledge entries can be used to store structured or unstructured information related to the product, such as functional descriptions, usage and dosage information, and contraindications from product manuals; selling point descriptions from marketing copy submitted by merchants; frequently asked and answered questions from historical operational scripts; and graphic descriptions from product display pages.

[0044] In some implementations, the generation device can organize multiple knowledge entries related to the product according to a preset format, facilitating subsequent selection and combination by a large language model. Text generation prompts can be constructed based on semantic templates and knowledge entries, explicitly instructing the large language model of the text theme and text elements to be filled in for the product description text to be generated. Specifically, the text generation prompts can carry basic product information, a set of knowledge entries related to the product, and text theme and text element constraint information represented by the primary and secondary tags, instructing the large language model to select and organize text content matching the corresponding text elements from the knowledge entries based on the text elements expressed by the secondary tags. After receiving the text generation prompts, the large language model can combine the constraints on text theme and text elements in the semantic template to filter and reorganize the knowledge entries, forming a product description text with a text theme expressed by primary tags and relatively complete information coverage at the text element level corresponding to the secondary tags.

[0045] In some implementations, the generation device can construct corresponding text generation prompts for different primary tags, enabling the large language model to generate multiple paragraphs or modular product description texts by topic, thereby improving the overall structure of the product description text.

[0046] Multiple embodiments of this application extract primary tags for expressing text topics and secondary tags for expressing the necessary text elements from the target user question set corresponding to the product. A semantic template containing the expected text topic and text element constraints is constructed based on the primary tags and their associated secondary tags. Then, a text generation prompt instruction is constructed based on the knowledge entries obtained from the data source and the semantic template to invoke a large language model. This enables the large language model to select corresponding text elements from the knowledge entries according to the text elements expressed by the secondary tags and generate product description text with the primary tags expressing the text topic. This achieves accurate control over the text topic and text elements of the generated content in the product description text generation process, guided by user questions, resulting in product description text that better matches user concerns and provides more complete coverage of key information.

[0047] In some implementations, the generating device may perform clustering processing on the target user question set to obtain multiple clusters; wherein each cluster includes at least one user question; and generate a first-level label expressing the semantics of the cluster for each cluster.

[0048] In this embodiment, the generation device can cluster the user questions in the aforementioned target user question set to obtain multiple clusters. By clustering user questions, multiple user questions that are semantically similar and have similar concerns can be grouped into the same cluster, so that each cluster contains one or more user questions that are related in expression, thereby facilitating the subsequent generation of corresponding first-level tags for different clusters. In some embodiments, constraints can be set on the number of clusters and the number of user questions within each cluster according to transaction requirements to avoid the clusters being too scattered or too concentrated.

[0049] In some implementations, the generation device can utilize a text vectorization model or semantic embeddings obtained based on a large language model to vectorize each user question in the target user question set, and perform a clustering algorithm based on the semantic similarity calculation results to obtain multiple clusters. For example, density-based clustering algorithms, hierarchical clustering algorithms, or other algorithms suitable for short text semantic clustering can be used to group user questions with close semantic distances into the same cluster. This approach ensures that each cluster has good concentration in semantic topics, facilitating the subsequent generation of first-level tags that summarize the semantic content of the cluster.

[0050] In this embodiment, the generation device can generate a first-level tag expressing the semantics of each cluster. Specifically, the generation device can summarize the common concerns reflected by the cluster based on multiple user questions contained in the cluster, and determine the words or phrases used to summarize the concerns as first-level tags. For example, when the user questions in a cluster are all about "how to use" and "how many times a day to use", the generation device can determine the first-level tag for that cluster as "usage method"; when the user questions in a cluster are concentrated on "who can use it" and "is it suitable for the elderly or children", the first-level tag for that cluster can be determined as "suitable population". In some embodiments, the generation device can call a large language model to summarize the user questions in the cluster and automatically generate first-level tags to express the semantics of the cluster, thereby improving the automation and semantic accuracy of tag generation. The generated first-level tags can serve as the basis for subsequent construction of semantic templates and generation of product description text based on semantic templates.

[0051] In some implementations, the generating device can be used to perform clustering processing on the user questions included in the clusters corresponding to the primary tags to obtain multiple sub-clusters; and generate secondary tags expressing the semantics of each sub-cluster for each sub-cluster.

[0052] In this embodiment, the generation device can further cluster user questions within the clusters corresponding to the aforementioned primary labels to obtain multiple sub-clusters. Primary labels summarize the semantic themes of user questions at a higher level, while multiple user questions within the same cluster may still contain different focuses or subtle differences. Therefore, by performing clustering again within the clusters corresponding to the primary labels, the semantic differences of user questions can be distinguished at a finer granular level, enabling each sub-cluster to represent a more concentrated and segmented set of user concerns under the text theme corresponding to the primary label. In some embodiments, the generation device can, based on the semantic vectors of user questions within a cluster, utilize short text clustering algorithms, semantic similarity calculation methods, or semantic embedding methods based on large language models to group user questions with closer semantic distances into the same sub-cluster, thereby obtaining multiple sub-clusters.

[0053] In this embodiment, sub-clusters can be used to represent more specific and refined semantic groups under the same text topic. For example, if the primary tag is "usage method," and the user questions in the cluster involve different focuses such as "how many times a day," "how to control the dosage each time," and "how long to use continuously for it to be effective," then after dividing the sub-clusters, the generating device can form sub-clusters reflecting detailed semantics such as "usage frequency," "dosage," and "usage duration." By dividing the sub-clusters, the system can further distinguish different text elements when generating product description text, thereby improving the completeness of content coverage.

[0054] In this embodiment, the generation device can generate secondary tags for each sub-cluster, expressing the semantics of that sub-cluster. The secondary tags are used to express the text elements that need to be covered under the text topic corresponding to the primary tag. Each secondary tag can be used to summarize the common concerns reflected in user questions within the sub-cluster. For example, for a sub-cluster containing content such as "how many times a day" or "whether it should be used morning and evening," the generation device can determine the secondary tag corresponding to that sub-cluster as "usage frequency"; for a sub-cluster containing content such as "how much to use each time" or "whether strict measurement is required," the secondary tag can be determined as "dosage."

[0055] In some implementations, the generation device can invoke a large language model to automatically summarize the semantic content of each sub-cluster to generate secondary tags that express the semantics of the sub-clusters, thereby improving the automation of tag extraction. The generated secondary tags can serve as the basis for text elements in the semantic template, guiding the generation of subsequent product description text, enabling the product description text to cover more details that users are particularly interested in while maintaining a structured expression.

[0056] In some embodiments, the generating device can be used to generate corresponding statement templates for each of the secondary tags included in the primary tag; wherein, the statement template has a reserved space for filling in text elements that conform to the secondary tag; the statement template constitutes the semantic template.

[0057] In this embodiment, the generation device can construct a semantic template corresponding to the primary tag based on the aforementioned primary tag and the secondary tags corresponding to the primary tag. Specifically, the generation device can generate a corresponding statement template for each secondary tag included in the primary tag, so that the text elements corresponding to different secondary tags are constrained by their respective statement templates when generating product description text in the subsequent process.

[0058] In this embodiment, the statement template can be a preset sentence structure used to carry the text elements corresponding to the aforementioned secondary tags. The statement template can have at least one reserved position for filling in text elements matching the secondary tag when generating product description text, thus allowing the same statement template to adapt to the specific content of different products. For example, when a secondary tag is used to represent "usage frequency," its corresponding statement template can adopt the sentence structure "This product is recommended to be used at {usage frequency}," where "{usage frequency}" is a reserved position for filling in the specific text elements corresponding to the "usage frequency" secondary tag; when a secondary tag is used to represent "applicable population," its corresponding statement template can adopt the sentence structure "This product is suitable for {applicable population}," where "{applicable population}" is a reserved position for filling in the specific text elements corresponding to the "applicable population" secondary tag.

[0059] In some implementations, reserved positions can be represented by placeholder symbols, variable markers, or other identifiable markers, and explicitly associated with corresponding secondary tags. When constructing text generation prompts, the generation device can include each reserved position and its corresponding secondary tag identifier in the prompt information to indicate to the large language model "which secondary tag the reserved position corresponds to". For example, the text generation prompts can simultaneously include constraint information such as "secondary tag: usage frequency, corresponding placeholder: {usage frequency}", as well as candidate text elements related to "usage frequency". This allows the large language model to identify which reserved position needs to be filled with text content corresponding to the "usage frequency" secondary tag when generating product description text, thereby avoiding confusion between different text elements.

[0060] In this embodiment, the generation device can combine multiple statement templates according to a preset arrangement order or a logical order determined by transaction requirements. The overall structure formed by these statement templates serves as a semantic template, constraining the text organization method of the large language model when generating product description text. When constructing text generation prompts, the generation device can incorporate the statement templates, reserved positions, and corresponding secondary tag identifiers from the semantic template into the prompt information. This allows the large language model to fill in matching text elements in the corresponding reserved positions based on the constraints on the text structure in the semantic template and the correspondence between secondary tags and reserved positions when generating product description text. This improves the structural consistency of the generated text and the degree of matching between the content and secondary tags.

[0061] In some implementations, the generating device can acquire a set of user questions corresponding to a product; wherein the set of user questions includes at least one user question; and from the set of user questions for a product, identify a target set of user questions for which no answers to questions exist in the product description text of the product.

[0062] In this embodiment, the generating device can further acquire a set of user questions corresponding to the product. User questions can include questions raised by users when actually browsing, inquiring about, or searching for the product. The set of user questions can include at least one user question and can originate from various data sources such as historical search records, online inquiry logs, user reviews, and a question database compiled by the merchant. The set of user questions typically covers the points of interest users have regarding the product in real-world scenarios.

[0063] In this embodiment, the generation device can analyze the set of user questions item by item based on the generated product description text to identify user questions that exist in the set but have not yet been answered in the product description text, thereby forming a target set of user questions. Specifically, the generation device can match the semantics of each user question in the set with the text content organized by primary and secondary tags in the product description text to determine whether the current product description text contains text elements corresponding to the semantics of the user question. For example, if the set of user questions includes the question "Is it suitable for pregnant women?", but the product description text does not contain explicit text elements related to the secondary tag "Suitable population", the generation device can identify this question as a user question with a missing answer.

[0064] In some implementations, the generating device can identify whether a user question has been answered by the product description text by calculating the semantic similarity, tag matching degree, or whether key text elements are covered, thus enabling the accurate identification of questions with missing answers. The target user question set obtained in this way can be used to represent questions that are not yet covered in the product description text but are of genuine concern to the user, providing a basis for subsequent supplementation of the product description text.

[0065] In this embodiment, the generation device can further generate new primary and secondary tags based on the target user question set, or construct corresponding semantic templates and text generation prompts to call a large language model to supplement and improve the product introduction text. This allows the final product introduction text to not only express the established text theme and text elements around the existing tag system, but also achieve more comprehensive content coverage in the dimension of real user concerns, thereby improving the product introduction text's responsiveness to user questions.

[0066] In some implementations, the generating device can perform semantic conflict matching between the product description text and knowledge entries from multiple data sources to obtain knowledge entries that semantically conflict with the product description text; if the knowledge entries with semantic conflicts are determined to be semantically correct, the product description text can be corrected based on the knowledge entries.

[0067] In this embodiment, after completing the initial generation of product description text, the generation device can also perform semantic consistency verification on the product description text to improve the knowledge accuracy of the text content. To this end, the generation device can perform semantic conflict matching between the product description text and knowledge entries from multiple data sources to identify whether the content in the product description text is semantically inconsistent, factually contradictory, or informationally conflicting with knowledge entries from different sources.

[0068] In this embodiment, knowledge entries from multiple data sources may include functional descriptions and usage precautions from product manuals, selling point information from marketing copy submitted by merchants, Q&A content from historical operational scripts, graphic descriptions on product display pages, and other structured or unstructured data that can reflect the true information of the product. The generation device can compare each text segment in the product description text with the knowledge entries one by one based on the comparison capabilities of semantic matching algorithms, similarity calculation models, or large language models, thereby identifying whether the product description text contains content that semantically conflicts with the knowledge entries. For example, if the product description text, generated based on a secondary tag, describes "use three times a day," but the relevant knowledge entry in the product manual clearly states "not more than twice a day," the generation device can identify a semantic conflict between this text element and the data source information. After identifying knowledge entries with semantic conflicts, the generation device can further invoke large language models or employ preset semantic verification rules to determine the semantic correctness of the conflicting knowledge entries. In some embodiments, the generation device can comprehensively evaluate whether the semantics expressed by the knowledge entries are accurate based on factors such as the source credibility of the knowledge entries, the timeliness of the knowledge entries, and the historical accuracy level of the data source. When the semantics expressed by the knowledge entry are determined to be correct, the generation device can modify the product description text based on the knowledge entry to ensure that the product description text accurately reflects the true information of the product.

[0069] In some implementations, the correction operation may include directly replacing text content in the product description text that contradicts the knowledge entry, or it may include reorganizing the text elements related to the secondary tag according to a semantic template, so that the generated product description text is consistent with the correct knowledge entry. For example, when the "usage frequency" text element under the "how to use" text topic in the semantic template is affected by a conflicting knowledge entry, the generation device can regenerate the text element based on the latest recognized correct semantics, so that the final output product description text is clear, accurate, and free of semantic conflicts.

[0070] In some implementations, the generating device can generate a quality score for the product description text based on scoring across multiple evaluation dimensions. The quality score represents the credibility of the product description text. These multiple evaluation dimensions include viewpoint richness, tag fill rate, source reliability, external knowledge voting, and timeliness. Viewpoint richness evaluates the number of unique viewpoints in the product description text. The tag fill rate represents the fill rate of text elements in the product description text that have been accurately labeled with secondary tags. Source reliability represents the stability of the data source in providing accurate information in past use. External knowledge voting represents the degree of consistency and semantic similarity between the product description text and multiple authoritative external knowledge sources. Timeliness represents the timeliness of the knowledge items used in the product description text relative to the current time. The product description text is updated with the goal of improving the quality score.

[0071] In this embodiment, the generation device can conduct a comprehensive quality assessment of the product description text based on multiple evaluation dimensions to further improve the completeness of the product description text. To this end, the generation device can generate a quality score that characterizes the overall credibility of the product description text based on the dimensional scores of the product description text across multiple evaluation dimensions.

[0072] In some implementations, multiple evaluation dimensions may include, but are not limited to, richness of viewpoints, tag fill rate, source reliability, external knowledge voting, and timeliness. Richness of viewpoints reflects the diversity of product description text in expressing product information and is used to evaluate the number of unique viewpoints in the product description text. Tag fill rate measures whether the product description text fully covers the text elements corresponding to each secondary tag, thus reflecting whether the text content comprehensively meets the user's concerns. Source reliability characterizes the stability of the data source in providing accurate information in past use, assigning credibility weights to different knowledge items by assessing the historical accuracy of the knowledge items. External knowledge voting semantically compares the product description text with multiple external authoritative knowledge sources (such as materials published by authoritative institutions, professional databases, industry guidelines, etc.) to characterize the consistency of viewpoints and semantic similarity between the product description text and external authoritative knowledge. Timeliness evaluates the timeliness of the knowledge items used in the product description text relative to the current time, reflecting whether the generated content is based on relatively new data sources.

[0073] In this embodiment, viewpoint richness can be used to characterize the number of unique viewpoints in the product description text, thus reflecting the information richness of the text content. In some embodiments, the viewpoint richness dimension score can be calculated as follows: VRS = α·log(1 + unique_viewpoints) + β·diversity_score, In this framework, VRS represents viewpoint richness, unique_viewpoints represents the number of unique viewpoints in the product description text, and diversity_score represents the viewpoint diversity score (e.g., quantifying the degree of difference between different viewpoints based on text similarity distribution). α and β are weighting parameters used to balance the impact of viewpoint quantity and viewpoint diversity on viewpoint richness in the overall evaluation. For example, when product description text under a certain "usage method" topic covers multiple different viewpoints such as dosage instructions, usage steps, and precautions, and these viewpoints are semantically significantly different, both unique_viewpoints and diversity_score will be high, and the corresponding VRS score can also be high.

[0074] In some implementations, the tag fill rate can be used to reflect the degree to which the product description text covers the text elements corresponding to the secondary tags. The generation device can determine whether the reserved positions in the semantic template have been filled with the required text elements. In some implementations, the dimensional score of the tag fill rate can be calculated based on a weighted average of the fill status of each tag field, for example: LCS = ∑(li × wi) / T, Here, LCS represents the tag fill rate. `li` represents the fill status of the i-th tag field; `li` is 1 when the corresponding secondary tag text element has been accurately filled in the product description text, and 0 otherwise. `wi` represents the weight of the i-th tag field, which can be a combination of the tag's importance (importance_factor_i, e.g., 1-5 points) and fill quality score (quality_score_i, e.g., a normalized score of 0-1). This reflects the importance and actual fill quality of a tag in the overall text. Specifically, `wi = li × importance_factor_i × quality_score_i`. `T` represents the total number of tag fields to be considered under the current text theme. For example, under the text theme "Target Audience," if the "Core Target Audience" tag has a high importance and the corresponding text description is clear and complete, then the contribution of `li·wi` to the LCS of this tag will be greater.

[0075] In some implementations, source reliability can be used to characterize the credibility of knowledge entries used in product description text. The generation device can assign different historical reliability weights to different data sources based on the data source to which the knowledge entries belong; for example, data sources such as instruction manuals and authoritative medical literature have higher weights, while marketing copy has lower weights. The source reliability score can be calculated by weighting the source weights of the knowledge entries cited by each text element. In some implementations, the dimensional score of source reliability can be determined as follows: SRS = ∑(wi × quality_score_i) / ∑wi, Here, SRS represents the reliability of the source, quality_score_i represents the reliability score of the i-th source in providing correct information in past use (e.g., a value between 0 and 1 obtained based on historical error correction records or manual review records), and wi represents the weight of the i-th source (e.g., it can be set based on call frequency, coverage, or source level). When the key content of the product description text mainly relies on highly reliable sources (e.g., instruction manuals, authoritative medical information, etc.), the corresponding Score_src score can be higher.

[0076] In some implementations, external knowledge voting can be used to characterize the consistency of viewpoints between product description text and multiple external authoritative knowledge sources. The generation device can construct a comparison query for key conclusions in the text, obtain corresponding conclusions from multiple external knowledge sources, and count the number of conclusions consistent with the product description text. In some implementations, the dimensional score of external knowledge voting can be calculated as follows: EVS = γ × ∑(vi × similarity(K, si)) / m, Here, EVS represents external knowledge voting. K represents a text segment in the product description text addressing a key conclusion, si represents the relevant reference content from the i-th external knowledge source, similarity(K,si) represents the semantic similarity between K and si (e.g., measured by cosine similarity of text embedding vectors), vi represents the weight of the i-th external knowledge source (e.g., authoritative medical journals, guidelines, etc., can be given higher weights), m represents the number of external knowledge sources participating in the voting, and γ is a normalization coefficient or weighting coefficient used to control the influence of external knowledge voting in the overall evaluation. For example, when the description of "maximum daily usage" in the product description text is highly consistent with most authoritative external information, the EVS score can approach the highest level.

[0077] In some implementations, timeliness can be used to reflect the update timeliness of the knowledge entries on which the product description text depends. The generation device can generate a timeliness score based on the time difference between the update time of the knowledge entry and the current time, using a time decay function. In some implementations, an exponential decay scoring method can be used, for example: TS = δ × exp(-λ · time_diff), Here, TS represents timeliness. time_diff represents the difference between the creation time or most recent update time of a knowledge item and the current time, λ is the time decay factor, used to control the rate at which the score decays over time, and δ is the timeliness weight, used to control the degree of influence of the timeliness score on the overall quality score. For example, when a knowledge item related to "latest adverse reaction warning" is updated more recently, time_diff is smaller, and the TimeScore is higher.

[0078] In this embodiment, the generation device can input the scores of each of the above evaluation dimensions into the reward model and obtain the final quality score through a weighted method. In some embodiments, the generation device can weight the scores of each dimension based on preset weights or based on a trained weight model, for example: in, to Each dimension has a weight, and the sum is 1. In some implementations, the generation device can use the quality score as a feedback signal to guide the updating of the product description text. For example, when the tag fill rate score is low, the generation device can call a large language model to supplement missing text elements; when the reliability of external knowledge voting or sources is low, the corresponding fragments can be regenerated based on more credible knowledge sources; when the timeliness score is low, outdated content can be replaced based on the latest knowledge entries. Through the above mechanism, the product description text can be gradually optimized under the feedback of the quality score, improving the credibility, structural integrity, and information accuracy of the final generated content.

[0079] Please see Figure 4 The embodiments of this application also provide an apparatus for generating product description text. The generating apparatus includes: an extraction module, a construction module, and a generation module.

[0080] The extraction module is used to extract primary tags and secondary tags from the target user question set corresponding to the product; wherein, the target user question set includes multiple user questions for the product; secondary tags belong to primary tags, and different primary tags do not have the same secondary tags; primary tags are used to express the text theme; secondary tags are used to express the text elements required for the corresponding text theme.

[0081] The construction module is used to construct the semantic template corresponding to the first-level tag; wherein, the semantic template also includes the second-level tags involved in the corresponding first-level tag.

[0082] The generation module is used to construct text generation prompts based on knowledge entries obtained from the data source and the semantic template, and to call the large language model to instruct the large language model to select text elements from the knowledge entries based on the text elements expressed by the secondary tags, so as to form product introduction text expressing the text theme by the primary tags.

[0083] In this embodiment, the functions and effects of the product description text generation device can be explained in comparison with the aforementioned embodiments, and will not be repeated here.

[0084] Please see Figure 5 This application also provides a computer device comprising: a memory and a processor, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the method described above.

[0085] The memory, processor, and communication interface in the computer device can communicate with each other via the system bus and network communication.

[0086] In this embodiment, the functions and effects implemented by the computer device can be explained by referring to the foregoing embodiments, and will not be repeated here.

[0087] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to implement the method as described above.

[0088] The functions and effects achieved in this embodiment can be explained by referring to other embodiments, and will not be repeated here.

[0089] This application also provides a computer program product containing instructions, including a computer program / instructions that, when executed by a processor, implement the method as described above.

[0090] The functions and effects achieved in this embodiment can be explained by referring to other embodiments, and will not be repeated here.

[0091] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments of this application, and are not intended to limit the scope of the invention.

[0092] It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0093] It is understood that the various implementation methods described in this application can be implemented individually or in combination, and the implementation methods in this application are not limited in this respect.

[0094] Unless otherwise stated, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0095] It is understood that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in 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. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0096] It is understood that the memory in the embodiments of this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Specifically, non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may be random access memory (RAM). It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0097] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0098] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0099] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0101] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0102] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for generating a product introduction text, characterized by, The method comprises the following steps: extracting a primary label and a secondary label from a target user question set corresponding to a commodity; wherein the target user question set comprises a plurality of user questions about the commodity; the secondary label belongs to the primary label, and different primary labels do not include the same secondary label; the primary label is used to express a text theme; and the secondary label is used to express a text element required by the corresponding text theme; constructing a semantic template corresponding to the primary label; wherein the semantic template also includes the secondary label related to the corresponding primary label; constructing a text generation prompt instruction according to a knowledge item obtained from a data source and the semantic template, calling a large language model, instructing the large language model to select a text element of the knowledge item according to the text element expressed by the secondary label, and forming a commodity introduction text in which the primary label expresses a text theme.

2. The method of claim 1, wherein, extracting a primary label and a secondary label from a target user question set corresponding to a commodity, comprising: performing clustering processing on the target user question set to obtain a plurality of clusters; wherein each cluster includes at least one user question; generating a primary label expressing the semantics of each cluster.

3. The method of claim 2, wherein, The method further comprises: performing clustering processing on the user questions included in the primary label corresponding cluster to obtain a plurality of sub-clusters; generating a secondary label expressing the semantics of each sub-cluster.

4. The method of claim 1, wherein, constructing a semantic template corresponding to the primary label, comprising: generating a corresponding sentence template for each secondary label included in the primary label; wherein the sentence template has a reserved position for filling in a text element conforming to the secondary label; and the sentence templates constitute the semantic template.

5. The method of claim 1, wherein, The method further comprises: obtaining a user question set corresponding to a commodity; wherein the user question set includes at least one user question; identifying, from the user question set of the commodity, a target user question set for which there is no question answer in the commodity introduction text of the commodity.

6. The method of claim 1, wherein, The method further comprises: performing semantic conflict matching on the commodity introduction text and the knowledge items of a plurality of data sources to obtain knowledge items that have semantic conflicts with the commodity introduction text; in the case of identifying that the knowledge item with semantic conflict is correct semantic, correcting the commodity introduction text according to the knowledge item.

7. The method of claim 1, wherein, The method further comprises: According to the dimension score of the product introduction text in multiple evaluation dimensions, a quality score of the product introduction text is generated; wherein the quality score is used to represent the credibility of the product introduction text; wherein the multiple evaluation dimensions include opinion richness, label filling rate, source reliability, external knowledge voting and timeliness; the opinion richness is used to evaluate the number of non-repeated opinions of the product introduction text; the label filling rate is used to represent the filling rate of the text elements accurately added with the secondary label in the product introduction text; the source reliability is used to represent the stability of the data source in providing correct information in the past use; the external knowledge voting is used to represent the matching degree of the opinion consistency and semantic similarity between the product introduction text and multiple external authoritative knowledge sources; the timeliness is used to represent the time effectiveness proximity of the knowledge items used in the product introduction text relative to the current time; A product introduction text is updated with the goal of improving the quality score.

8. A computer-readable storage medium, characterized in that, A computer program is stored thereon, which, when executed by a processor, causes the processor to implement the method of any one of claims 1 to 7.

9. A computer device, comprising: The computer device includes a memory and a processor, and at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the method of any one of claims 1 to 7.

10. A computer program product, characterised in that, Computer instructions are included, which, when executed by a processor, implement the method of any one of claims 1 to 7.

Images