High-anthropomorphic-degree Internet community comment generation system and method based on large language model

Abstract

The invention provides a high-anthropomorphic-degree Internet community comment generation system and method based on a large language model, and is applied to the technical field of data processing. The method comprises the following steps: processing a historical old post, converting picture and video contents in the historical old post into text descriptions, and establishing a historical old post content index; monitoring the new post of the user, if no comment reply is issued for a preset time, triggering an AI comment generation process, processing pictures and videos in the new post to obtain text description, and combining with the original text to form the overall text description of the new post; converting the overall text description of the new post into a text embedding vector, obtaining a preset number of similar old posts from the historical old post index through similar vector retrieval, and screening out a target number of high-quality comments; inputting the target number of high-quality comments as examples into the prompt context of the LLM, assembling the prompt and requiring the LLM to imitate the style and logic of the examples to generate the comments; and releasing the generated high anthropomorphic comments to a new comment area to generate Internet community comment information.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a highly anthropomorphic internet community comment generation system and method based on a large language model. Background Technology

[0002] In today's booming online community landscape, new posts often face the embarrassing situation of receiving zero comments, severely impacting community activity and user engagement. For example, new posts in niche interest areas, due to their limited audience, struggle to attract user comments in a short period. Furthermore, while existing AI-generated comment technology exists, the generated comments lack human-like qualities, feel mechanical, and fail to truly integrate into the community's atmosphere, thus failing to meet users' demands for high-quality, personalized interaction.

[0003] A deeper technical analysis reveals significant shortcomings in current technologies for handling multimodal content. Community posts often contain multimedia elements such as images and videos; however, traditional methods struggle to effectively extract semantic information from these elements. This results in low recall rates when searching for similar historical posts, failing to provide rich and relevant reference resources for new posts. For example, for a picture showcasing unique scenery during a mountain climb, current technology may not be able to accurately identify key elements (such as steep stone steps or magnificent sea of ​​clouds) and convert them into searchable textual semantics, leading to biases in the retrieval of similar historical posts.

[0004] Existing technologies lack effective strategies for guiding Large Language Models (LLMs) to generate comments. Because they fail to fully utilize high-quality comments from similar older posts as examples, LLM-generated comments often suffer from a monotonous style, chaotic logic, and a severe disconnect from the community's tone. For instance, comments generated for hiking-themed posts might simply list hiking precautions, lacking interactive content like those from real users sharing personal experiences and expressing emotions, thus failing to resonate with other users.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this application is to provide a highly anthropomorphic internet community comment generation system and method based on a large language model, which at least to some extent overcomes the problems existing in the prior art. By acquiring historical old posts and new posts, the system converts old post images and videos into text using a large visual model, and combines the original text with semantic vectors through an embedding model to construct a content index; it monitors new posts, and if there are no comments within a preset time, it triggers an AI process to similarly process new posts to generate an overall description and semantic vector, retrieves similar old posts and selects high-quality comments; and it uses high-quality comments as examples to input into the LLM prompt, guiding it to imitate and generate highly anthropomorphic comments and publish them.

[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to one aspect of this application, a method for generating highly anthropomorphic internet community comments based on a large language model is provided, comprising: acquiring historical posts in the community, including image and text posts, video posts, and new posts published by users; processing the historical posts by converting the image and video content into text descriptions using a large visual model, combining them with the original text of the post to form an overall text description, and then converting them into semantic embedding vectors using a text embedding model to establish a content index of historical posts; monitoring new posts by users, and if there are no comments after a preset time, triggering an AI comment generation process, processing the images and videos in the new posts using a large visual model to obtain text descriptions, and combining them with the original text to form an overall text description of the new post; converting the overall text description of the new post into a text embedding vector, retrieving a preset number of similar old posts from the historical post index using similarity vector retrieval, and filtering out a target number of high-quality comments; inputting the target number of high-quality comments as examples into the prompt context of an LLM, assembling the prompt and requiring the LLM to generate comments in the style and logic of the examples, generating highly anthropomorphic comments; and publishing the generated highly anthropomorphic comments to the comment section of new posts to generate internet community comment information.

[0009] Another aspect of this application is a highly anthropomorphic internet community comment generation device based on a large language model, characterized by comprising: an acquisition module for acquiring historical posts in the community, including image and text posts, video posts, and new posts published by users; a processing module for processing historical posts, converting the image and video content into text descriptions using a large visual model, combining them with the original text of the post to form a complete text description, and then converting it into a semantic embedding vector using a text embedding model to establish a content index for historical posts; and monitoring new posts published by users, triggering an A function if no comments are received after a preset time. The comment generation process involves processing images and videos in a new post using a large visual model to obtain text descriptions, which are then combined with existing text to form a complete text description of the new post. This complete text description is then converted into a text embedding vector. A preset number of similar old posts are retrieved from the historical post index using similarity vector retrieval, and a target number of high-quality comments are selected. These high-quality comments are then used as examples to input into the LLM's prompt context. The prompt is then assembled, and the LLM is instructed to generate comments following the style and logic of the examples, resulting in highly anthropomorphic comments. Finally, the generated highly anthropomorphic comments are published to the new post's comment section, generating internet community comment information.

[0010] According to another aspect of this application, an electronic device is characterized by comprising: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for generating highly anthropomorphic internet community comments based on a large language model by executing the executable instructions.

[0011] This application provides a highly anthropomorphic internet community comment generation system and method based on a large language model. It acquires historical and new posts, converts old post images and videos into text using a large visual model, and combines this text with semantic vectors generated through an embedding model to construct a content index. It monitors new posts; if no comments are received within a preset time, an AI process is triggered, similarly generating an overall description and semantic vector for each new post. It then retrieves similar old posts and selects high-quality comments. These high-quality comments are used as examples in the LLM's prompt, guiding it to imitate and generate highly anthropomorphic comments before posting them. The large visual model supplements multimodal semantics to improve recall, and the use of high-quality comments as examples enhances the anthropomorphism and community adaptability of the LLM comments.

[0012] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit the disclosure of the contents of this application. Attached Figure Description

[0013] Figure 1The flowchart illustrates a highly anthropomorphic internet community comment generation method based on a large language model, according to an embodiment of this application.

[0014] Figure 2 The diagram shows a schematic representation of a highly anthropomorphic internet community comment generation device based on a large language model, according to an embodiment of this application. Detailed Implementation

[0015] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0016] The following is combined with Figure 1 This application describes a highly anthropomorphic method for generating comments on internet communities based on a large language model, according to exemplary embodiments thereof. It should be noted that the following application scenarios are shown only to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application are applicable to any suitable scenario.

[0017] In one embodiment, this application also proposes a highly anthropomorphic internet community comment generation system and method based on a large language model. Figure 1 The diagram schematically illustrates a flowchart of a highly anthropomorphic internet community comment generation method based on a large language model, according to an embodiment of this application. Figure 1 As shown, this method is applied to a server and includes:

[0018] S101 retrieves historical posts from the community, including text and image posts, video posts, and new posts published by users.

[0019] In one implementation, acquiring historical posts and new posts from users within the community is the foundational data collection step for the entire comment generation system. This can be divided into two steps: acquiring historical posts and acquiring new posts. Specifically, acquiring historical posts requires covering all types of posts already existing within the community, including text / image posts and video posts. For example, in a food community, a post from 2024 titled "Homemade Chiffon Cake" containing multiple images and text descriptions of cake-making steps is a text / image post; a post from 2023 titled "A Century-Old Wonton Stall in an Alley" documenting the process of making street food and related text descriptions is a video post. These are all within the scope of acquiring historical posts.

[0020] When acquiring new user posts, the system primarily focuses on content that users have just published to the community. For example, a post on July 9, 2025, containing travel photos and a simple text description titled "Weekend Hiking Snapshots," or a short video sharing a pet's daily life with the caption "My furry friend's naughty moments," will be promptly captured by the system and included in the scope of new post acquisition.

[0021] S102 processes historical posts by converting images and videos into text descriptions using a large visual model. This text description is then combined with the original post text to form a complete text description. Finally, a text embedding model is used to convert the text into semantic embedding vectors to create an index of historical post content.

[0022] In one implementation, a visual big data model is used to perform text conversion on the images and videos in historical posts. This text is then combined with the original post text to form a holistic text description. A text embedding model is introduced to perform semantic vector conversion on the overall text description, achieving a vectorized representation of the historical post content. When performing text conversion on the images and videos in historical posts using the visual big data model, the key elements and dynamic processes in the images and videos are accurately analyzed. Taking a post about "homemade chiffon cake" as an example, the visual big data model can not only identify specific materials and tools such as eggs, flour, and whisks in the images, but also capture the details of the production steps presented in each image, such as the technique for separating egg yolks and whites, and the state of the batter being stirred. This is then transformed into a coherent and detailed text description, such as, "In the third image, the whisk is used to stir the egg whites at medium speed, and the egg whites gradually develop fine foam, appearing slightly white."

[0023] When combining the original text of a post to form a complete text description, the image-translated text generated by the visual model is organically integrated with the post's own text description to ensure complete information and logical flow. For the aforementioned "Homemade Chiffon Cake" post, the original text "tried making a chiffon cake at home, and surprisingly succeeded on the first try" is integrated with the image-translated text of each step, forming a complete text description such as "tried making a chiffon cake at home, and surprisingly succeeded on the first try. To make it, first prepare the eggs, flour, and other ingredients, then separate the egg yolks and whites, and then use an electric mixer to beat the egg whites at medium speed..." This comprehensively covers the post's text and image information.

[0024] When introducing a text embedding model to perform semantic vector transformation on the overall text description, the algorithm adopted, combined with relevant solution concepts, is consistent with the mainstream semantic vector generation logic to achieve deep capture and transformation of text semantic information. Taking the overall text description of "homemade chiffon cake" as an example, the text embedding model first performs word segmentation on the text, breaking down words and sentences such as "chiffon cake," "making steps," and "success" into smaller semantic units. Then, based on a pre-trained language model (such as the underlying algorithm logic used in model architectures like BERT and Word2Vec), it performs calculations through a multi-layer neural network. The model considers the contextual relationships of words in the text, such as the association between "making steps" and material words like "eggs" and "flour," and the semantic resonance between "success" and "successfully on the first try," thereby assigning a corresponding vector value to each semantic unit.

[0025] Next, the model integrates the vectors of these semantic units to form a semantic vector that represents the entire text. In this process, the model strengthens the weight of key information through an attention mechanism. For example, "chiffon cake," as the core theme, will have a higher proportion in the vector, and the logical order between each step of the process will be reflected through the correlation of the vector dimensions. The final generated numerical vector not only contains the features of the core content "chiffon cake making," but also incorporates the logical relationships between the steps and the emotional tone of the poster, providing accurate vector basis for subsequently finding related historical posts through similar vector retrieval.

[0026] By connecting the overall text description data with the semantic embedding vector results, a vector mapping relationship for old posts is constructed. The semantic relevance of content is calculated using a text embedding model, establishing the foundation for a vector index of historical old posts. The overall text description is a direct presentation of the content of old posts, including core themes, detailed information, and emotional tendencies, while the semantic embedding vector is a mathematical expression of this information; the two have a natural correspondence. Taking an old post about "homemade chiffon cake" as an example, its overall text description covers content such as "chiffon cake making steps" and "the joy of a successful attempt." The corresponding semantic vector encodes this information through numerical dimensions. Constructing a mapping relationship allows the system to clearly define the correspondence rules between "text content and vector features," providing a foundation for subsequent relevance calculations.

[0027] The semantic relevance of content is calculated using a text embedding model, based on the fundamental principle in vector space that "the closer the distance, the more similar the meaning." The model performs calculations such as cosine similarity on the semantic vectors of different old posts to quantify their semantic relevance. For example, old posts about "homemade chiffon cakes" and "beginner baking tips" both revolve around the baking theme and involve common content such as the production process. Their semantic vectors are close in the vector space, resulting in a high degree of relevance. Conversely, old posts unrelated to the theme of "outdoor camping equipment recommendations" have larger vector distances, resulting in a lower degree of relevance.

[0028] This vector indexing foundation based on relevance allows the system to quickly locate older posts with similar semantics. Because the index records the vector features and relevance information of each older post, when it is necessary to retrieve historical posts similar to new posts, it is only necessary to calculate the relevance between the semantic vector of the new post and the vector of the older posts in the index. This allows for efficient filtering of content with high matching degrees. This process conforms to the conventional logic of vector retrieval and lays a reliable foundation for building a more comprehensive multidimensional vector matrix.

[0029] Using historical posts as the core dimension, this approach integrates translated text data from images and videos, original post text data, and semantic embedding vector data to construct a multi-dimensional vector matrix of historical post content, achieving comprehensive vectorization of historical post content. Historical posts often contain multimodal information such as text, images, and videos; a single type of data cannot fully reflect the core content of the post. Taking the video post "A Century-Old Wonton Stall in an Alley" as an example, the visual big data model translates the text "The video shows the process of making wonton wrappers and the method of preparing the filling," supplementing the actions and details in the video footage; the original text "This wonton stall tastes authentic, and there are many people queuing up every day" directly expresses the poster's evaluation and scene description; and the semantic embedding vector data is a mathematical encoding of this textual information, containing core semantic features such as "wonton making" and "authentic taste." By integrating these three types of data, the content details of posts can be covered from different levels, avoiding incomplete semantic representation caused by the lack of a single data type.

[0030] Each dimension of the multidimensional vector matrix corresponds to a feature of a data type, and the correlation between dimensions enables a comprehensive depiction of the post content. For example, the translated text of "wonton wrapper making process" and the original text of "authentic taste" have a logical semantic connection of "the production process affects the taste," which is reflected in the numerical correlation of the corresponding dimensions in the vector matrix. The semantic embedding vector acts as a bridge, transforming textual information into calculable mathematical features, allowing different types of data to form a unified representation system in the same vector space. This comprehensive vectorization provides a more accurate basis for subsequent similar post retrieval. Because the multidimensional vector matrix contains both the explicit content of the post (such as text descriptions and image translations) and implicit semantic connections (such as the logic of production and evaluation, and the atmosphere of the scene), when a new post is compared with a vector, similar features with historical posts can be found from more dimensions, which meets the goal of "improving the recall probability of similar posts" in the solution and lays a solid content foundation for subsequent screening of high-quality comments and generation of anthropomorphic comments.

[0031] This paper addresses vector bias issues through a semantic vector optimization model, enhances vector representation accuracy by combining it with text embedding reinforcement mechanisms, and integrates it with the vector mapping relationship features of old posts to generate a historical post content index. During the conversion of images and videos to text, misjudgments by the visual model may lead to biases in the translated text, causing the generated semantic vectors to deviate from the true meaning. For example, a historical post titled "Handmade Cotton Bags" might be mistranslated by the visual model as "linen," resulting in a semantic vector biased towards "linen products," which deviates from the post's true content. The semantic vector optimization model identifies semantic conflicts between "linen" and "cotton" by comparing the logical consistency between the translated text and the original text (e.g., the original text mentions "the soft texture of cotton"). It then corrects the values ​​of the corresponding dimensions in the vectors, making them more aligned with the core semantics of "cotton bag making." This process relies on the principle of self-consistency in the internal semantic logic of the text, ensuring that vector bias is effectively corrected.

[0032] The text embedding enhancement mechanism improves the accuracy of vector representation by strengthening the weights of key semantics. For the corrected vector, the mechanism increases the weights of core information in the post (such as "handmade" and "cotton bag") while weakening the influence of irrelevant information. For example, in the aforementioned old post about "handmade cotton bags," after enhancement, the vector dimensions corresponding to core words such as "handmade," "cotton," and "bag" are more prominent, allowing the vector to more accurately reflect the core features of the post. This aligns with the approach in this application of "improving the comprehensiveness of understanding by supplementing textual semantic information," providing a more reliable vector foundation for subsequent correlation calculations.

[0033] The index is generated by fusing optimized vectors with the mapping features of old post content vectors, based on the systematic nature of the associations between vectors. Each old post's vector not only needs to accurately represent its own content but also clearly define its semantic association with other old posts (e.g., the high correlation between "handmade cotton bags" and "introduction to fabric crafts"). The fusion process incorporates the optimized vectors of individual old posts into the overall vector relationship network, ensuring that the index contains both the precise vector features of each old post and records the strength of associations between vectors, thus forming a structured index system. This index ensures that subsequent similarity searches can find precise matches using individual post vectors while also expanding the search scope based on the relationship network, meeting the requirement of "improving the recall probability of similar posts" and providing efficient index support for selecting high-quality comments from historical old posts.

[0034] S103 monitors new user posts. If no comments are received after a preset time, an AI-generated comment process is triggered. The visual big data model processes the images and videos in the new post to obtain a text description, which is then combined with the original text to form a complete text description of the new post.

[0035] In one implementation, a new post publication time monitoring model and a comment reply status tracking mechanism are combined to monitor and judge the publication duration and comment interaction of a user's new post in real time, generating new post interaction status data. For example, if a user publishes a new post titled "Weekend Hiking Snapshots" on July 9, 2025, the system records the publication time using the time monitoring model and simultaneously monitors the comment section activity in real time using the comment status tracking mechanism. If there are no comments by 3 PM on July 9 (one hour after publication), new post interaction status data containing "publication duration 1 hour, number of comments 0" is generated.

[0036] The system analyzes the interaction data of new posts by combining preset time threshold rules and no-comment trigger conditions, and then activates the AI ​​comment generation trigger mechanism to generate an AI process start command. Assuming the community's preset time threshold is 1 hour and the no-comment trigger condition is 0 comments, the system analyzes the interaction data of the aforementioned "Weekend Hiking Snapshots" post and finds that it meets the condition of "posting time reaches 1 hour and no comments." Therefore, the AI ​​comment generation trigger mechanism is activated, generating a start command to "start executing AI comment generation process."

[0037] The AI-driven workflow initiates a command to invoke a large-scale visual model, which then performs text translation on the images and videos in the new post, generating a multimedia content description. For example, a new post titled "Weekend Hiking Snapshots" contains multiple photos of hiking scenery. After analyzing the images, the large-scale visual model generates a text description such as, "The first image shows steep mountain steps surrounded by lush green trees; the second image is a view of the sea of ​​clouds from the mountaintop, with distant peaks appearing and disappearing in the mist."

[0038] Based on the original text information of the new post and the multimedia content text description, information is fused through a content integration mechanism to generate comprehensive text material for the new post. The original text of the new post was "Went hiking this weekend, the scenery was amazing, highly recommended!" After merging it with the translated text of the images, the comprehensive text material is formed as follows: "Went hiking this weekend, the scenery was amazing, highly recommended! The first picture shows steep mountain steps surrounded by lush green trees; the second picture is a view of the sea of ​​clouds from the mountaintop, with distant peaks appearing and disappearing in the mist."

[0039] Based on the overall text construction rules, the comprehensive text materials of the new post are structurally integrated and semantically optimized to generate a comprehensive text description of the new post that includes the original text and multimedia translations. Following the structural rule of "original text first, then multimedia translations," the comprehensive text materials are optimized to make the sentences more coherent, ultimately generating a comprehensive text description of the new post: "I went hiking this weekend, the scenery was amazing, highly recommended! Along the way, you can see steep mountain steps surrounded by lush green trees; after reaching the summit, the sea of ​​clouds was particularly spectacular, with distant mountain peaks appearing and disappearing in the mist."

[0040] S104: Convert the overall text description of the new post into a text embedding vector, retrieve a preset number of similar old posts from the historical old post index through similarity vector retrieval, and filter out the target number of high-quality comments.

[0041] In one implementation, the overall text description of the new post is correlated with a text embedding model. This process converts the overall text description into a corresponding semantic embedding vector, generating semantic vector data for the new post. The overall text description includes the original text information and the text descriptions translated from images and videos. The semantic embedding vector represents the semantic features of the new post's content. The overall text description encompasses the original text and the translated text from images and videos, containing rich semantic details (such as "mountain climbing," "stone steps," and "sea of ​​clouds" in a post titled "Weekend Mountain Climbing Snapshots"). The core function of the text embedding model is to convert the semantic information in the text into a computable vector form. This correlation process allows the model to fully capture the core themes, scene elements, and emotional tendencies in the text. For example, the positive evaluation conveyed by "the scenery is incredibly beautiful" and the specific scene described by "steep mountain stone steps" are both converted into corresponding dimensional features in the vector, ensuring that the vector comprehensively reflects the content of the new post.

[0042] Text embedding models utilize pre-trained neural network structures (such as the underlying logic of models like BERT and Word2Vec) to process text through word segmentation and contextual semantic analysis. Taking the entire text of a new post titled "Weekend Hiking Photos" as an example, the model first breaks the text down into semantic units such as "weekend," "hiking," and "scenery." Then, it combines the context to determine the association between "spectacular" and "sea of ​​clouds" in "the sea of ​​clouds is particularly spectacular," assigning a corresponding vector value to each unit. Simultaneously, an attention mechanism strengthens the weight of core words. For example, "hiking" and "sea of ​​clouds," as the core theme of the new post, will have a higher proportion in the vector than auxiliary information such as "weekend," ensuring that the generated semantic vector accurately points to the core content of the new post. This provides a reliable basis for subsequent similarity retrieval with historical post vectors. By transforming multimodal information into a unified semantic vector, the model ensures that new posts and historical posts are compared in the same vector space, improving the accuracy of similar post retrieval and laying the foundation for selecting high-quality comments.

[0043] The semantic vector data of new posts is compared with the semantic embedding vectors in the index of historical posts to generate a candidate list of similar historical posts. This candidate list includes a predetermined number of historical posts with high semantic relevance to the new posts. The essence of semantic embedding vectors is to transform textual semantics into numerical vectors in a high-dimensional space. The distance between vectors (such as cosine similarity) directly reflects the semantic relevance—the closer the distance, the higher the similarity between the two posts in terms of core themes, scene descriptions, and emotional tendencies. For example, the semantic vector of the new post "Weekend Hiking Snapshots" contains core features such as "hiking," "stone steps," and "sea of ​​clouds." Similarly, the semantic vectors of posts in the historical post index such as "One-Day Hiking Trip to Huangshan" and "Taishan Summit Viewing Guide" also contain similar features such as "hiking route" and "summit view." The vector distances are relatively close, and through calculation, these posts can be accurately identified as highly relevant content.

[0044] Setting the preset number to 20 avoids both insufficient sources of high-quality comments (e.g., searching only 5 old posts might miss some highly relevant content) and excessive sources that would increase redundancy in subsequent filtering (e.g., searching 100 old posts might mix in low-relevance content), thus balancing retrieval efficiency and recall quality. For example, for a new post titled "Weekend Hiking Snapshots," the 20 candidate old posts include content that directly describes the hiking experience, as well as relevant information such as recommendations for nearby equipment and weather precautions, providing a rich reference sample for subsequent selection of high-quality comments.

[0045] By using vector similarity retrieval, we have overcome the dependence of traditional keyword retrieval on the number of words (e.g., it is difficult to match multimodal content with just the word "mountain climbing"). With the help of the textual semantic information supplemented by the visual big model, new posts containing pictures and videos can be semantically compared with historical old posts across modalities, which greatly improves the recall accuracy of similar posts and lays a reliable foundation for extracting high-quality comments from the candidate list.

[0046] Based on a candidate list of similar old posts, comment information is extracted from each similar old post and its quality is assessed to generate a high-quality comment candidate set. This high-quality comment candidate set is used to select comment content that meets the quality standards. Posts in the candidate list of similar old posts are highly semantically related to new posts (e.g., both "A Day Trip to Huangshan" and "Weekend Hiking Snapshots" revolve around the theme of hiking and sightseeing), and their comment sections naturally align with the potential discussion direction of the new post. Extracting these comments as candidates ensures the relevance of the reference sample to the content of the new post, avoiding the introduction of irrelevant comments that could cause LLM to generate content off-topic. For example, a comment in the old post "A Day Trip to Huangshan" stating "The stone steps are indeed very steep... the sea of ​​clouds at the summit is definitely worth the ticket price" is directly related to the descriptions of "steep mountain stone steps" and "the sea of ​​clouds is particularly spectacular" in the new post, providing a precise reference for LLM to generate comments that fit the context of the new post.

[0047] The evaluation metrics (relevance, fluency, and emotional sincerity) all serve the goal of "generating highly human-like comments": relevance ensures that comments revolve around the theme, such as "recommending comfortable sneakers" being closely related to the hiking scenario; fluency avoids fragmented and grammatically incorrect expressions, ensuring the language standardization of the reference samples; emotional sincerity filters out comments with personal experiences and emotions (such as "my legs are sore but it was worth the ticket price"), conveying the tone of genuine human interaction. These characteristics can be learned and imitated by the LLM, reducing the "AI mechanical feel" of the comments. By extracting relevant comments and rigorously evaluating their quality, the generated high-quality comment candidate set ensures both content relevance and human-like expression, providing high-quality material for subsequent selection of target high-quality comments and driving the LLM to generate comments that conform to the community's tone, ultimately achieving the effect of "making the generated comments truly evoke empathy and resonance from users."

[0048] The system selects a target number of comments from the candidate set of high-quality comments as the final high-quality comments, generating a target number of high-quality comment results. These results serve as reference examples for LLM-generated comments. Setting the target number to 10 ensures sufficient sample size while avoiding an excessive number that could distract the LLM learning focus. For example, for a new post titled "Weekend Hiking Snapshots," 10 comments can cover various comment types, such as "description of hiking experience," "sharing of practical advice," and "interactive feedback on the post content," while allowing LLM to clearly capture the expressive characteristics of different styles. This solves the problem of existing technologies where "comments generated directly by LLM have a monotonous style and feel mechanically generated by AI." From the perspective of the targeted selection criteria, "most representative" ensures that the selected comments reflect the typical interaction methods of community users on similar topics. For example, "Climbing is tiring, but seeing such a sea of ​​clouds is really worth it" accurately reflects the common experience of "effort and reward" in the climbing scenario, which meets the requirement of "making the generated comments conform to the community tone" in this application. "Diverse styles" covers different types such as colloquial expressions, rational suggestions, and emotional resonance, avoiding mechanical repetition of LLM-generated comments. For example, interactive comments such as "The photos taken by the poster are amazing" and suggestive comments such as "Beginners should climb slowly" complement each other in style. "Closely related to the content of the new post" focuses on the core elements such as "stone steps" and "sea of ​​clouds" in the new post, ensuring that the reference examples can directly provide a basis for LLM to generate comments that fit the scenario of the new post, avoiding content that deviates from the topic.

[0049] This selection method directly serves the key innovation of this application, namely, "enhancing the human-like quality of LLM-generated comments through high-quality comment examples." The 10 final high-quality comments serve as prompt input for the LLM, conveying the authentic expression habits of community users (such as colloquial tone and personalized experience sharing) while providing diverse logical frameworks (such as experience descriptions, practical suggestions, and emotional interactions). This allows the LLM to have a richer pool of samples to emulate when generating comments, resulting in comments that are more likely to evoke empathy and resonance from users. This effectively addresses the shortcomings of existing technologies where "AI-generated comments lack depth of thought and are not human-like."

[0050] S105: Input the target number of high-quality comments as examples into the LLM's prompt context, assemble the prompt, and ask the LLM to generate comments in the style and logic of the examples, generating highly anthropomorphic comments.

[0051] In one implementation, a target number of high-quality comments are associated with the LLM's prompt context. These high-quality comments are input as examples into the prompt context to generate basic prompt content with examples. The target number of high-quality comments includes those selected from similar older posts that meet quality standards. The prompt context provides a generation reference for the LLM. The quality of the generated LLM highly depends on the guidance of the prompt context. High-quality comments, as products of genuine community user interaction, contain expressive styles, logical structures, and emotional inclinations that align with the community's tone. Inputting them as examples into the prompt context provides the LLM with concrete and referable generation templates. For example, in a new post titled "Weekend Hiking Snapshots," among the 10 high-quality comments, "The stone steps on this mountain path are so steep, but seeing the sea of ​​clouds at the top made all the fatigue worthwhile!" embodies the expressive logic of "description of experience + emotional resonance," while "I suggest everyone wear non-slip shoes, otherwise it's easy to slip on the way down" demonstrates the content type of "practical advice." These examples help the LLM clarify "what to say" and "how to say it."

[0052] The prompt context needs to be highly relevant to the target task to be effective in guiding the discussion. Associating high-quality comments with the prompt context essentially binds "high-quality expressions in similar scenarios" to the "current generation task," ensuring that the LLM focuses on content relevant to the new post. For example, for a new post titled "Weekend Hiking Snapshots," the selected high-quality comments all revolve around core elements such as "hiking," "stone steps," and "sea of ​​clouds," demonstrating a high degree of relevance to the overall text description of the new post. This association prevents the LLM from generating content that deviates from the topic.

[0053] This approach directly serves the key innovation of this application: "enhancing the human-like quality of LLM-generated comments through high-quality comment examples." The prompt content with examples provides LLM with "human expression samples" to emulate, enabling it to learn colloquial tone (such as "so steep," "otherwise"), personalized experiences (such as "all the fatigue was worth it"), and interactive logic (such as directly giving advice) from comments. This, in turn, generates comments that more closely resemble real human communication, effectively addressing the shortcomings of existing technologies where "AI-generated comments lack depth of thought and are not human-like," thus laying the foundation for generating highly human-like comments in the future.

[0054] This application integrates the basic content of the example prompts with the requirements for comment generation, explicitly requiring the LLM to imitate the style and logic of high-quality comments and generate structured prompt instructions. It points out that "the quality of the prompt directly affects the quality of the LLM's generated results," and the completeness, style, and clarity of the example comments are the foundation of a high-quality prompt. For the structured prompt instructions in the new post "Weekend Hiking Snapshots," the verification process must ensure that the 10 example comments fully retain core information (such as key expressions like "the stone steps are so steep" and "the sea of ​​clouds is exhausting") to avoid misinterpretation by the LLM due to missing information. Simultaneously, it must confirm that requirements such as "imitating the style of colloquial expression and personal experience sharing" and "describing the hiking experience first and then adding practical suggestions" are clearly stated to prevent the LLM from deviating from the generation goal due to ambiguous instructions.

[0055] From the perspective of parameter optimization adaptability, LLMs exhibit specific patterns in their understanding of natural language. Optimizing the expression to better align with LLMs' cognitive habits can improve the efficiency of instruction execution. For example, adjusting the style requirement from "use conversational expressions that include personal experience" to "imitate conversational expressions and share your mountaineering experience like the example" strengthens the connection with the example comment by using the analogy of "like the example," making it easier for LLMs to understand the specific direction of style imitation. Similarly, optimizing the logical requirement from "follow the order of experience first, then suggestions" to "first describe your feelings about climbing this mountain, then offer some practical tips to other climbers, just like the example comment's approach" lowers the understanding threshold for LLMs with more concrete expressions like "describe your feelings" and "offer some tips." This optimization aligns with the requirement in this application to "enable LLMs to generate comments as closely as possible to the language style and logic of high-quality comments," ensuring smoother interaction between instructions and LLMs. The final generated prompt text, which includes example comments, style imitation requirements, and logical rules, directly serves the innovation point of this application: "guiding LLM to imitate high-quality comments through structured prompts." The complete and optimized prompt text not only provides LLM with specific imitation samples (example comments) but also clarifies the dimensions of imitation (style and logic), enabling LLM to accurately capture the expressive characteristics of community users and generate highly anthropomorphic comments that conform to the community's tone.

[0056] In another implementation, the basic content of the prompt with examples and the core elements of comment generation are extracted and processed to generate style imitation points, logical compliance rules, content relevance requirements, and example feature annotation data. For example, for 10 high-quality example comments on a new post titled "Weekend Hiking Snapshots," the style imitation points are extracted as "colloquial expression, using everyday words such as 'amazing' and 'almost gave up,' with direct expression of personal feelings"; the logical compliance rules are "first describe the specific experience during the hike, and then naturally transition to practical advice or interactive feedback on the post content"; the content relevance requirements are "must include content related to the core elements of the new post, such as 'stone steps,' 'sea of ​​clouds,' and 'hiking'"; and the example feature annotation data are "comments such as 'The poster's photos are so realistic' reflect interactivity with the original post, and 'be sure to bring enough water' reflects practical advice."

[0057] The key points of style imitation, logical rules, and content relevance requirements are integrated and processed to generate a standardized framework for comment generation. After integrating the extracted key points, rules, and requirements, the standardized framework is: "Share your real feelings during the mountain climbing process in a conversational style, first describing specific experiences such as steep stone steps and spectacular sea of ​​clouds, and then providing practical suggestions such as bringing water and wearing non-slip shoes, or interactively commenting on the photos taken by the poster, ensuring that the content revolves around the mountain climbing scene and the core elements of the new post."

[0058] The framework for comment generation is defined in a structured manner, generating an initial prompt template. This initial prompt template includes instruction dimensions, instruction expression methods, core requirements, and example citation rules. Instruction dimensions are divided into "style imitation," "logical structure," and "content relevance." Instructions are expressed in an imperative sentence format, such as "Please generate a comment in the manner of...". Core requirements include "using everyday vocabulary, reflecting the logic of experience + suggestions / interactions, and including the core elements of the new post." Example citation rules are "Refer to the descriptions and interactive tone of 'stone steps' and 'sea of ​​clouds' in the example comments."

[0059] Based on the example feature annotation data, the initial prompt instruction template was adapted and adjusted to generate structured prompt instructions that conform to the understanding logic of LLM. Combining the features such as "interactivity" and "practicality" in the examples, the initial template was adjusted to "Please refer to the following example comments and generate comments in a conversational manner: 1. First, describe the real feelings brought by the stone steps and sea of ​​clouds when climbing, using everyday words such as 'amazing' and 'almost gave up'; 2. Then, naturally give practical suggestions such as bringing water and wearing non-slip shoes, or interact with the original post like 'The photos taken by the poster are so realistic'; 3. The content should revolve around 'climbing,' 'stone steps,' and 'sea of ​​clouds.' Example comments: [List 10 example comments]", making the instructions more in line with the understanding habits of LLM and ensuring that the generated comments conform to the specifications.

[0060] The structured prompt instructions undergo integrity verification and parameter optimization to generate a final prompt text adapted to LLM. This final prompt text includes example comments, style imitation requirements, and logical adherence rules. For the structured prompt instructions in the new "Weekend Hiking Snapshots" post, verification must ensure that the 10 example comments fully retain core information (such as key expressions like "the stone steps are so steep" and "the sea of ​​clouds made me tired") to avoid misinterpretation by the LLM due to missing information. Simultaneously, requirements such as "imitating colloquial expressions and a personal experience sharing style" and "describing the hiking experience first, then adding practical advice" must be clearly stated to prevent the LLM from deviating from its generation goal due to ambiguous instructions.

[0061] LLMs (Layered Language Managers) exhibit specific patterns in their understanding of natural language. Optimizing the expression to better align with LLMs' cognitive habits can improve the efficiency of instruction execution. For example, adjusting the style requirement from "use conversational language and expressions that include personal experience" to "imitate conversational expressions and share your mountaineering experience like the example" strengthens the connection with the example commentary through the analogy of "like the example," making it easier for LLMs to understand the specific direction of style imitation. Similarly, optimizing the logical requirement from "follow the order of experience first, then advice" to "first describe your feelings about climbing this mountain, then offer some practical tips to other climbers, just like the example commentary's approach" lowers the comprehension threshold for LLMs with more concrete expressions like "describe your feelings" and "offer some tips." This optimization aligns with the requirement in this application to "enable LLMs to generate comments that, as far as possible, follow the language style and logic of high-quality comments," ensuring smoother interaction between instructions and LLMs.

[0062] The final generated prompt text, which includes example comments, style imitation requirements, and logical rules, directly serves the innovation point of this application: "guiding LLM to imitate high-quality comments through structured prompts." The complete and optimized prompt text not only provides LLM with specific imitation samples (example comments) but also clarifies the dimensions of imitation (style and logic), enabling LLM to accurately capture the expressive characteristics of community users and generate highly anthropomorphic comments that conform to the community's tone.

[0063] Based on the final prompt text, LLM is used to generate comments, producing highly human-like comments that conform to the style and logic of the examples. Taking the new post "Weekend Hiking Snapshots" as an example, the example comments "The stone steps on this mountain path are so steep, but the moment I reached the top and saw the sea of ​​clouds, all the fatigue was worth it!" demonstrate a conversational expression of "description of experience + emotional resonance," while "I suggest everyone wear non-slip shoes, otherwise it's easy to slip on the way down" reflects the logical structure of "practical advice." By learning from these examples, LLM can capture the typical expression habits of community users in hiking scenarios, such as using conversational words like "almost made me give up" and "it's so beautiful" to convey genuine feelings, and using transition words like "that's right" to naturally transition to advice, making the generated comments closer to the tone of everyday human communication.

[0064] In terms of the relevance of the generated comments to the new post content, the final prompt text explicitly requires comments to be related to the core elements of the new post, such as "stone steps" and "sea of ​​clouds." LLM prioritizes these elements during the generation process. For example, the generated comment, "The stone steps almost made me give up while climbing this mountain, but thankfully the sea of ​​clouds I saw after reaching the top was so beautiful, and the photos you took are so realistic!" directly echoes the descriptions of "steep mountain stone steps" and "the sea of ​​clouds is particularly spectacular" in the new post. The comment, "Be sure to bring enough water, there are no supply points halfway up the mountain," adds practical information on top of being relevant to the scenario, meeting the requirement in this application that "the generated comments should be closely related to the post content," and avoiding generalized expressions that deviate from the topic. The comments generated by LLM not only mimic the logical structure of "experience + suggestion" in the example but also absorb colloquial and personalized expressive features (such as "The photos you took are so realistic" reflecting interactivity), while closely adhering to the content of the new post. This effectively solves the shortcomings of existing technologies where "AI-generated comments lack depth of thought and are not human-like."

[0065] S106: Publish the generated highly anthropomorphic comments to the new post comment section, generating internet community comment information.

[0066] In one implementation, for a new post titled "Weekend Hiking Snapshots," the preceding process generates highly anthropomorphic comments such as, "The stone steps almost made me give up while climbing this mountain, but thankfully the sea of ​​clouds I saw after reaching the top was so beautiful! The photos you took are so realistic! By the way, be sure to bring enough water, as there are no supply points halfway up the mountain."

[0067] The system will automatically post this comment to the comment section of the new post. At this point, the comment becomes part of the internet community's comment information. Other users will see this AI-generated comment when browsing the new post, which may trigger further interaction and discussion, enriching the community's communication atmosphere.

[0068] The core of this application lies in addressing the issues of zero comments on new posts and low anthropomorphism in AI comments. The process involves first acquiring historical posts and new user posts. For historical posts, images and videos are converted to text using a large visual model, and combined with the original text to form a comprehensive description. This description is then converted to a semantic vector using a text embedding model to construct a content index. New posts are monitored; if no comments are received within a preset time, the AI ​​process is triggered. Similarly, the multimedia content of new posts is processed to form a comprehensive description, converted to a semantic vector, and then similar historical posts are searched to select high-quality comments.

[0069] Then, high-quality comments are used as examples to input into the LLM prompt, and instructions are assembled to allow it to mimic and generate highly human-like comments before publishing them. The key innovation lies in using a large visual model to supplement the semantics of multimodal content, thereby improving the recall rate of similar posts; using high-quality comments from similar posts as examples, the LLM is guided to generate comments that conform to the community's tone and are close to human expression, thus improving the user experience.

[0070] In one implementation, such as Figure 2 As shown, this application also provides a highly anthropomorphic internet community comment generation device based on a large language model, comprising:

[0071] The acquisition module 201 is used to acquire historical posts in the community, including text and image posts, video posts, and new posts published by users;

[0072] Processing module 202 is used to process historical posts. It uses a large visual model to convert images and videos into text descriptions, combines this with the original post text to form a complete text description, and then uses a text embedding model to convert it into a semantic embedding vector, establishing a historical post content index. It monitors new user posts; if no comments are received after a preset time, it triggers an AI-generated comment process. It uses a large visual model to process images and videos in new posts to obtain text descriptions, combining this with the original text to form a complete text description of the new post. It converts the complete text description of the new post into a text embedding vector, retrieves a preset number of similar posts from the historical post index using similarity vector retrieval, and selects a target number of high-quality comments. It uses these target number of high-quality comments as examples to input into the LLM's prompt context, assembles the prompt, and requires the LLM to generate comments following the example style and logic, generating highly anthropomorphic comments. Finally, it publishes the generated highly anthropomorphic comments to the new post comment section, generating internet community comment information.

[0073] The computer-readable storage medium provided in the above embodiments of this application and the highly anthropomorphic Internet community comment generation method based on a large language model provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0074] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the highly anthropomorphic internet community comment generation method, electronic device, electronic device, and readable storage medium based on a large language model are basically similar to the embodiments of the highly anthropomorphic internet community comment generation method based on a large language model described above, and are therefore described relatively simply. Relevant parts can be referred to in the descriptions of the embodiments of the highly anthropomorphic internet community comment generation method based on a large language model described above.

Claims

1. A highly anthropomorphic method for generating comments in internet communities based on a large language model, characterized in that, include: Retrieve historical posts from the community, including text and image posts, video posts, and new posts published by users; Historical posts are processed by converting images and videos into text descriptions using a large visual model. These descriptions are then combined with the original text of the posts to form a complete text description. Finally, a text embedding model is used to convert the text into semantic embedding vectors to create an index of historical post content. The system monitors new user posts. If no comments are received after a preset time, it triggers an AI-generated comment process. The system uses a large visual model to process images and videos in the new post to obtain a text description, which is then combined with the original text to form a complete text description of the new post. The entire text description of the new post is converted into a text embedding vector. A preset number of similar old posts are retrieved from the historical old post index through similarity vector retrieval, and a target number of high-quality comments are selected. The target number of high-quality comments are used as examples to input into the LLM's prompt context. The prompt is assembled and the LLM is asked to generate comments in the style and logic of the examples, resulting in highly human-like comments. The generated highly anthropomorphic comments are published in the new post's comment section, generating internet community comment information.

2. The method as described in claim 1, characterized in that, Historical posts are processed by converting their images and videos into text descriptions using a large visual model. This text description is then combined with the original post text to form a complete description. Finally, a text embedding model is used to convert this description into semantic embedding vectors, creating an index of historical post content, including: By using a large visual model to convert images and videos in historical posts into text, and combining them with the original text of the posts to form a complete text description, a text embedding model is introduced to perform semantic vector conversion on the complete text description, thereby realizing the vectorized representation of the content of historical posts. By connecting the overall text description data with the semantic embedding vector results, a vector mapping relationship of old posts is constructed. The semantic relevance of the content is calculated through the text embedding model to establish the basis for the vector index of historical old posts. Using historical posts as the core dimension, we integrate image and video translated text data, original post text data, and semantic embedding vector data to construct a multi-dimensional vector matrix of historical posts, thereby achieving comprehensive vectorization of historical posts. The semantic vector optimization model addresses the vector bias problem, and the text embedding enhancement mechanism improves the accuracy of vector representation. This is then integrated with the vector mapping relationship features of old posts to generate an index of historical old posts.

3. The method as described in claim 1, characterized in that, The system monitors new user posts. If no comments are received within a preset time after posting, an AI-generated comment process is triggered. This process uses a large visual model to process images and videos in the new post to obtain a text description. This description is then combined with the existing text to form a complete text description of the new post, including: By combining a new post publication time monitoring model with a comment reply status tracking mechanism, the system monitors and judges the publication time of new posts and comment interaction in real time, and generates new post interaction status data. By combining preset time threshold rules and no-comment trigger conditions, the interaction status data of new posts is analyzed, and the AI ​​comment generation trigger mechanism is activated to generate AI process start instructions; Based on the AI ​​process startup command, the visual big model is invoked to perform text translation processing on the images and videos in the new post, and generate multimedia content text descriptions. Based on the original text information of the new post and the multimedia content text description, information is integrated through a content integration mechanism to generate comprehensive text material for the new post. Based on the overall text construction rules, the comprehensive text materials of the new post are structurally integrated and semantically optimized to generate an overall text description of the new post that includes the original text and multimedia translation content.

4. The method as described in claim 3, characterized in that, The entire text description of the new post is converted into a text embedding vector. A preset number of similar old posts are retrieved from the historical post index using similarity vector retrieval. A target number of high-quality comments are then selected, including: The overall text description of the new post is associated with the text embedding model. The overall text description of the new post is converted into the corresponding semantic embedding vector to generate semantic vector data of the new post. The overall text description of the new post includes the original text information of the new post and the text description translated from the images and videos. The semantic embedding vector is used to represent the semantic features of the content of the new post. The semantic vector data of new posts and the semantic embedding vectors in the content index of historical posts are subjected to similarity retrieval processing to generate a candidate list of similar old posts. The candidate list of similar old posts includes a preset number of historical old posts that are highly semantically related to new posts. Based on the candidate list of similar old posts, the comment information in each similar old post is extracted and the quality is evaluated to generate a candidate set of high-quality comments. The candidate set of high-quality comments is used to select comment content that meets the quality standards. The target number of comments are selected from the candidate set of high-quality comments as the final high-quality comments, and the target high-quality comment results are generated. The target high-quality comment results are used as a reference example for LLM to generate comments.

5. The method as described in claim 4, characterized in that, The target number of high-quality comments are used as examples to input into the LLM's prompt context. The prompt is then assembled, and the LLM is instructed to generate comments that mimic the style and logic of the examples, producing highly human-like comments, including: The target number of high-quality comments are associated with the LLM's prompt context. The high-quality comments are used as examples to input into the prompt context to generate basic prompt content with examples. The target number of high-quality comments includes comments that meet the quality standards and are selected from similar old posts. The prompt context is used to provide a generation reference for the LLM. The basic content of the prompt with examples and the requirements for comment generation are integrated and processed, and the LLM is explicitly required to generate structured prompt instructions by imitating the style and logic of high-quality comments. The structured prompt instructions are validated for integrity and optimized for parameters to generate the final prompt text adapted to LLM. The final prompt text includes example comments, style imitation requirements, and logical compliance rules. Based on the final prompt text, LLM is called to generate comments, producing highly human-like comments that conform to the style and logic of the example.

6. The method as described in claim 5, characterized in that, The basic content of the prompt with examples is integrated with the requirements for comment generation. The LLM is explicitly required to generate structured prompt instructions that mimic the style and logic of high-quality comments, including: Extract and process the basic content of the prompt with examples and the core elements of comment generation to generate style imitation points, logical rules, content association requirements and example feature annotation data; The key points of style imitation, logical rules, and content relevance requirements are integrated and processed to generate a standardized framework for comment generation; The framework for comment generation is defined in a structured manner, and an initial prompt template is generated. The initial prompt template includes the division of prompt dimensions, the way the prompts are expressed, the core requirements, and the rules for referencing examples. The initial prompt instruction template is adapted and adjusted based on example feature annotation data to generate a structured prompt instruction that conforms to the understanding logic of LLM.

7. A highly anthropomorphic internet community comment generation device based on a large language model, characterized in that, The device includes: The acquisition module is used to retrieve historical posts from the community, including text and image posts, video posts, and new posts published by users. The processing module handles historical posts by converting images and videos into text descriptions using a large visual model. This text description is then combined with the original post text to form a complete description, which is further converted into semantic embedding vectors using a text embedding model to create an index of historical post content. The module monitors new user posts and, if no comments are received within a preset time, triggers an AI-generated comment process. This process uses a large visual model to process images and videos in the new post to obtain text descriptions, which are then combined with the original text to form a complete description of the new post. The complete description of the new post is converted into a text embedding vector, and a preset number of similar posts are retrieved from the historical post index using similarity vector retrieval. A target number of high-quality comments are then selected. These high-quality comments are used as examples to input into the LLM's prompt context. The prompt is then assembled, and the LLM is instructed to generate comments following the example style and logic, resulting in highly anthropomorphic comments. Finally, the generated highly anthropomorphic comments are published to the new post comment section, generating internet community comment information.

8. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the highly anthropomorphic internet community comment generation method based on a large language model as described in any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the highly anthropomorphic Internet community comment generation method based on a large language model as described in any one of claims 1 to 6.

Images