Method, system, medium and product for constructing social hot content semantic deduction intelligent agent

Abstract

The application discloses a kind of social hot content semantic deduction intelligent agent construction method, system, medium and product, comprising: obtaining historical hot search list data and carrying out clustering and processing, obtain the effective content set of multiple hot data categories;Respectively for each hot data category, construct intelligent agent and give preset person set collection;For each hot data category, according to the effective content set of the hot data category, obtain the comment reply strategy prompt word of the intelligent agent of the hot data category under each preset person set in its preset person set collection.The intelligent agent constructed by the application makes social hot content semantic deduction more in line with real social network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of social semantic deduction, and in particular to a construction method, system, medium and product of a social hot topic content semantic deduction agent. BACKGROUND

[0002] Hot topics are an important part of social media, usually in the form of hot search lists, attracting a large number of different users to participate in discussions, and often mixed with users with preset purposes. To reduce the negative impact of such replies from users with preset purposes, it is necessary to reason and simulate the hot topic in advance to deduce the semantic changes of the hot content and focus on simulating the reply content of users with preset purposes and the impact on the overall semantic trend of the hot content.

[0003] With the development of large language model technology, Agent (intelligent agent) is constantly maturing, and Agent can be used to simulate real social networks. The existing technology generally presets different role Agents, finds similar individuals from real social network historical data, gives the Agents corresponding personas, and then lets them comment on hot topics to deduce the semantic changes of the hot content. However, this method still has the following shortcomings: first, when giving Agents personas, the content of accounts with rich data in real social networks is usually selected, but in fact, the same real account often has multiple different personas, and directly using its full data may cause interference between personas, confusing the Agent; second, when interacting with hot content, the comments on the hot topic itself are usually simulated, but in fact, most social platforms support replies to comments, and the existing Agent lacks targeted optimization for this scenario. SUMMARY

[0004] To solve the technical problems in the background art, the present application proposes a construction method, system, medium and product of a social hot topic content semantic deduction agent.

[0005] In a first aspect, the present application proposes a construction method of a social hot topic content semantic deduction agent, comprising:

[0006] obtaining historical hot search list data of a preset social network;

[0007] clustering and processing the historical hot search list data to obtain an effective content set of multiple hot topic data categories;

[0008] constructing an intelligent agent for each hot topic data category, and giving each hot topic data category intelligent agent a preset persona set; the preset persona set includes multiple preset personas;

[0009] According to the effective content set of each hot data category, the comment reply strategy prompt word of the agent of the hot data category under each preset persona in the preset persona set of the agent of the hot data category is obtained.

[0010] Preferably, the historical hot search list data includes: hot post texts and comments, comment replies, basic information and label information of post users and comment users and comment reply users of the hot post texts.

[0011] Preferably, the historical hot search list data is clustered and processed to obtain the effective content set of the plurality of hot data categories, specifically including:

[0012] Each hot post text of the historical hot search list data is respectively input into a preset vector model to obtain vector representation of each hot post text in a semantic space;

[0013] The vector representations of all hot post texts are clustered by using a clustering algorithm to obtain a plurality of hot data categories;

[0014] The user set of each hot data category is respectively constructed according to the post users and comment users of all hot post texts in each hot data category;

[0015] The content set of each hot data category is respectively constructed according to the hot post texts and comments of all users in the user set of each hot data category;

[0016] All hot post texts of each hot data category are respectively summarized to obtain the semantics of each hot data category;

[0017] The effective content set of each hot data category is obtained by respectively eliminating the content irrelevant to the semantics of each hot data category in the content set of each hot data category.

[0018] Preferably, the agent of each hot data category is endowed with a preset persona set, specifically including:

[0019] A persona preset set is constructed; wherein the persona preset set includes a plurality of preset personas;

[0020] For each hot data category, each content in the effective content set of the hot data category is classified by using a pre-trained multi-classification model to obtain the predicted persona of each content;

[0021] According to the predicted persona of each content, the number of contents under each preset persona is counted;

[0022] The preset persona with a content number lower than a preset content number threshold is removed from the persona preset set, and the removed persona preset set is taken as the preset persona set of the hot data category.

[0023] Preferably, according to the effective content set of the hotspot data category, the comment reply strategy prompt word of the agent of the hotspot data category under each preset persona in the preset persona set thereof is obtained, specifically including:

[0024] Step a, format combing is performed on the effective content set of the hotspot data category;

[0025] Step b, constructing the comment reply rule mining prompt word of the agent of the hotspot data category under each preset persona;

[0026] Step c, using a large language model to perform comment reply rule mining on the format-combined effective content set according to the comment reply rule mining prompt word under each preset persona, to obtain a comment reply strategy set under each preset persona;

[0027] Step d, for each preset persona, constructing a comment reply strategy prompt word for each comment reply strategy in the comment reply strategy set under the preset persona;

[0028] Step e, for each effective content in the effective content set, using the agent of the hotspot data category to generate a comment reply for the effective content based on each comment reply strategy prompt word under the preset persona;

[0029] Step f, using a preset judgment large language training model to judge whether there is a best reply in the comment reply generated for the effective content; if not, the effective content is no longer involved in iteration, and the effective content set is updated; if yes, the effective content is updated according to the best reply, and the effective content set is updated according to the updated effective content;

[0030] Step g, judging whether a preset convergence condition is met; wherein the preset convergence condition is that all effective contents in the effective content set are no longer involved in iteration or iteration reaches a preset iteration number; if not, step d is entered;

[0031] if yes, step h is entered;

[0032] Step h, taking the comment reply strategy prompt word of the agent of the hotspot data category under the preset persona when the preset convergence condition is met as the final comment reply strategy prompt word of the agent of the hotspot data category under the preset persona.

[0033] Preferably, the effective content includes the semantics of the hotspot data category where the hotspot post is located, the hotspot post, and the comments and comment replies thereof.

[0034] Preferably, for each comment reply strategy in the comment reply strategy set under the preset persona, a comment reply strategy prompt word is constructed, specifically including:

[0035] For each comment reply strategy in the comment reply strategy set under the preset persona, the data in the valid content set that meets the comment reply strategy forms a valid content subset;

[0036] A candidate prompt word set is constructed for each comment reply strategy, wherein each candidate prompt word in the candidate prompt word set displays a description of the comment reply strategy corresponding to the valid content subset;

[0037] Each candidate prompt word in the candidate prompt word set is used to infer the valid content subset respectively by using a large language model to obtain a predicted comment reply set; the semantic similarity between each predicted comment reply in the predicted comment reply set and the corresponding real comment reply is calculated, and the sum of the semantic similarities is taken as the score of each candidate prompt word;

[0038] The candidate prompt word with the highest score is taken as the comment reply strategy prompt word of the comment reply strategy.

[0039] In a second aspect, the present application further provides a construction system of a social hot content semantic deduction agent, comprising:

[0040] An acquisition module acquires historical hot search list data of a preset social network;

[0041] A processing module is used for clustering and processing the historical hot search list data to obtain valid content sets of multiple hot data categories;

[0042] A construction module is used for constructing an agent for each hot data category respectively; and each agent of each hot data category is endowed with a preset persona set; wherein the preset persona set comprises multiple preset personas;

[0043] A comment reply strategy construction module is used for obtaining, for each hot data category, a comment reply strategy prompt word of the agent under each preset persona in the preset persona set of the agent according to the valid content set of the hot data category.

[0044] In a third aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the construction method of the social hot content semantic deduction agent according to any one of the first aspect.

[0045] In a fourth aspect, the present application further provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the construction method of the social hot content semantic deduction agent according to any one of the first aspect.

[0046] In the present application, the proposed construction method of the social hot content semantic deduction intelligent agent constructs an intelligent agent for each hot data category; each hot data category is assigned a preset person set, and according to the effective content set of each hot data category, the comment reply strategy prompt word of the intelligent agent under each preset person set in each hot data category is obtained. The intelligent agent constructed in the present application makes the social hot content semantic deduction more in line with the function of the real social network, facilitates timely intervention, and avoids negative effects. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The flowchart of the construction method of the social hot content semantic deduction intelligent agent in an embodiment of the present application.

[0048] Figure 2 The acquisition diagram of the comment reply strategy prompt word in an embodiment of the present application. DETAILED DESCRIPTION

[0049] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] In a first aspect, with reference to Figure 1 The construction method of the social hot content semantic deduction intelligent agent proposed in the present application comprises:

[0051] Obtain the historical hot search list data of a preset social network;

[0052] Cluster and process the historical hot search list data to obtain an effective content set of multiple hot data categories;

[0053] Construct an intelligent agent for each hot data category, and assign a preset person set to the intelligent agent of each hot data category; the preset person set comprises multiple preset persons;

[0054] For each hot data category, according to the effective content set of the hot data category, obtain the comment reply strategy prompt word of the intelligent agent of the hot data category under each preset person set in the preset person set of the hot data category.

[0055] The present application constructs an intelligent agent for each hot data category; each hot data category is assigned a preset person set, and according to the effective content set of each hot data category, the comment reply strategy prompt word of the intelligent agent under each preset person set in each hot data category is obtained; the social hot content deduced by the constructed intelligent agent is more in line with the function of the real social network, facilitates timely intervention, and avoids negative effects.

[0056] Of course, the basic information and the label information of each type of user in the embodiment have been desensitized.

[0057] In the embodiment, the historical hot search list data is clustered and processed to obtain an effective content set of multiple hot data categories, specifically including:

[0058] Each hot post in the historical hot search list data is input into a preset vector model to obtain a vector representation of each hot post in a semantic space;

[0059] The vector representations of all hot posts are clustered by using a clustering algorithm to obtain a plurality of hot data categories;

[0060] A user set of each hot data category is constructed according to the posting users and the commenting users of all hot posts in each hot data category;

[0061] A content set of each hot data category is constructed according to the hot posts and the comments of all users in the user set of each hot data category;

[0062] The semantic of each hot data category is obtained by summarizing all hot posts in each hot data category;

[0063] The effective content set of each hot data category is obtained by removing the content irrelevant to the semantic of each hot data category from the content set of each hot data category.

[0064] The vector model and the clustering algorithm in the embodiment can use existing technologies.

[0065] The semantic of each hot data category is obtained by summarizing all hot posts in each hot data category, specifically including: all hot posts in each hot data category are summarized by using a preset first large language model to obtain the semantic of each hot data category.

[0066] The effective content set of each hot data category is obtained by removing the content irrelevant to the semantic of each hot data category from the content set of each hot data category, specifically including: whether each content in the content set of each hot data category is related to the semantic of the hot data category is judged by using a preset second large language model; if yes, it is retained; if no, it is deleted.

[0067] Specifically, first, the historical hot search list data, i.e., the social post on the hot search list, is denoted as ;

[0068] Wherein, ; in the formula, represents the historical hot search list data, represents the ith hot post in the hot search list, i = 1, 2, …, m, m represents the total number of hot contents in the hot search list;

[0069] The hot search list data Each hot post Input vector model to get the vector representation of each hot post in the semantic space, and further use clustering algorithm to cluster the vector representation of all hot posts. All posts can be divided into several clusters, and each cluster is a hot data category;

[0070] Secondly, on the basis of clustering to obtain various hot data categories, the embodiment is not limited to real social account individuals, but reorganizes all contents in the same hot data category; for example, all post authors and all commenters in the current hot data category are recorded as a user set The post and comment set of the user in the user set is First, the contents of all users in the user set are merged into a content set , that is, the original real user is no longer distinguished; in the formula, represents the content set, represents the jth content in the content set, and the content is a hot post or a comment;

[0071] Finally, the records irrelevant to the semantic of the current hot data category are removed from the content set of the current hot data category.

[0072] To this end, first, the semantics of all hot posts

[0073] in the current hot data category need to be summarized to obtain the semantic of the current hot data category, that is: ; in the formula, represents the semantic of the hot data category, represents the first large language model, represents the jth hot post in the current hot data category, j = 1, 2, …, k, k represents the total number of hot posts in the current hot data category;

[0074] Further, it is judged whether each content in the content set of the current hot data category is related to the semantic of the hot data category , and only the related ones are retained to obtain an effective content set;

[0075] wherein, ; in the formula, represents the effective content set, represents the second large language model, represents the jth content in the valid content set.

[0076] The first large language model and the second large language model in the embodiment can be selected from mainstream models, and will not be described again.

[0077] There are a large number of ordinary users in the social network, some of which will participate in the interaction of hot content; there are also a small number of special users, which may interact with the preset persona of the hot content through interaction. In order to interact with the preset persona of the hot content through interaction, in the embodiment, a preset persona set is assigned to the agent of each hot data category, which specifically includes:

[0078] The preset persona set is constructed; wherein the preset persona set includes a plurality of preset personas;

[0079] For each hot data category, a pre-trained multi-classification model is used to classify each content in the valid content set of the hot data category, to obtain the predicted persona of each content.

[0080] According to the predicted persona of each content, the number of contents under each preset persona is counted.

[0081] The preset persona with a content number lower than the preset content number threshold is removed from the preset persona set, and the preset persona set after removal is taken as the preset persona set of the hot data category.

[0082] The embodiment can assign a preset persona set to each hot data category, and remove the preset personas that are less faced by each hot data category, so that the subsequent social hot content semantic deduction is more in line with the function of the real social network.

[0083] Specifically, for the valid content set of each hot data category , a multi-classification model is constructed by using the BERT connection classification head; the multi-classification model is used to calculate whether the jth comment under the ith hot post is a preset persona of a certain type.

[0084] wherein, ; in the formula, is one of n+1 classifications, n is the total number of preset personas, represents the predicted persona of the jth comment under the ith hot post, represents the ith hot post, represents the jth comment under the ith hot post. represents the multi-classification model.

[0085] ​Thirdly, the number of contents of each predicted persona under each hot data category is counted, and the preset persona whose number of contents is less than the preset content number threshold is removed from the preset persona set, and it is considered that the hot content of this semantic is less, and the interaction with those removed preset personas is faced;

[0086] Finally, for the remaining preset personas, the corresponding agent is given a preset persona, that is, the content corresponding to the classification label is given to the agent as long-term memory.

[0087] It should be understood that 1 in n+1 categories is a reserved category, which is used to represent other personas that are not identified or not in the preset persona set.

[0088] As shown in Figure 2 In this embodiment, according to the valid content set of the hot data category, the comment reply strategy prompt word of the agent of the hot data category under each preset persona in its preset persona set is obtained, which specifically includes:

[0089] Step a, format combing the valid content set of the hot data category;

[0090] Step b, constructing the comment reply rule mining prompt word of the agent of the hot data category under each preset persona;

[0091] Step c, using a large language model to mine the comment reply rule of the valid content set after format combing according to the comment reply rule mining prompt word under each preset persona, to obtain a comment reply strategy set under each preset persona;

[0092] Step d, for each preset persona, constructing a comment reply strategy prompt word for each comment reply strategy in the comment reply strategy set under the preset persona;

[0093] Step e, for each valid content in the valid content set, using the agent of the hot data category to generate a comment reply for the valid content based on each comment reply strategy prompt word under the preset persona;

[0094] Step f, using a preset judgment large language training model to judge whether there is a best reply in the comment reply generated for the valid content; if not, the valid content is no longer involved in iteration, and the valid content is marked as "no longer involved in iteration", and the valid content set is updated; if yes, the valid content is updated according to the best reply, and the valid content set is updated according to the updated valid content;

[0095] Step g, judging whether the preset convergence condition is met; wherein the preset convergence condition is that all valid contents in the valid content set are no longer involved in iteration or iteration to a preset iteration number; if not, go to step d;

[0096] If yes, go to step h;

[0097] Step h, if the agent of the hot data category under the preset persona meets the preset convergence condition, the comment reply strategy prompt word of the agent of the hot data category under the preset persona is taken as the final comment reply strategy prompt word of the agent of the hot data category under the preset persona.

[0098] In this way, the embodiment can iteratively optimize the comment reply strategy prompt word of the agent of each hot data category under each preset persona and expand the effective content set, so that the agent finally obtained has a better comment reply ability to hot content.

[0099] The comment reply rule mining prompt word in the present example is, for example, “… summarize the angle and method of this comment reply … describe it in no more than 10 words…”.

[0100] In the comment reply rule mining, each data in the effective content set after format sorting is reasoned by using a preset third large language model according to the comment reply rule mining prompt word under each preset persona, to obtain the angle and method of the real comment reply of each data; the angle and method of the real comment reply of each data are summarized by using a preset fourth large language model according to a preset induction prompt word, to obtain a comment reply strategy set under the hot data category.

[0101] It should be understood that each large language model in the embodiment can be the same or different.

[0102] In the present embodiment, a comment reply strategy prompt word is constructed for each comment reply strategy in the comment reply strategy set under the preset persona, specifically including:

[0103] For each comment reply strategy in the comment reply strategy set under the preset persona, the data in the effective content set that meets the comment reply strategy is formed into an effective content subset;

[0104] A candidate prompt word set is constructed for each comment reply strategy; each candidate prompt word in the candidate prompt word set displays a description of the comment reply strategy corresponding to the effective content subset;

[0105] Each candidate prompt word in the candidate prompt word set is used to reason the effective content subset respectively by using a fifth large language model, to obtain a predicted comment reply set; the semantic similarity between each predicted comment reply in the predicted comment reply set and the corresponding real comment reply is calculated, and the sum of the semantic similarities is taken as the score of each candidate prompt word;

[0106] The candidate prompt word with the highest rating will be used as the prompt word for this comment reply strategy.

[0107] This embodiment can construct targeted comment reply strategy prompts for each comment reply strategy in the comment reply strategy set, so that the semantic inference of subsequent social hot topics is more in line with the functions of real social networks.

[0108] Specifically, firstly, we sort out the various hot data categories. The data format below is: That is, each comment reply is marked with its corresponding semantics, original post, and comment.

[0109] Secondly, the prompt project is used to extract patterns from comment replies. The prompt format is something like, "...please summarize the angle and techniques used in replying to this comment...describe in no more than 10 words...". For each data point, a large language model is used for inference to obtain the perspectives and techniques of real comment responses within the semantic hot content of that hot data category. Then... The perspectives and techniques used to reply to all individual comments are summarized into several general perspectives and techniques using the prompt project and LLM, and recorded as a strategy set. .

[0110] Again, for each strategy To obtain a better prompt, follow these steps. Next matching strategy valid content data subset First, construct a set of candidate prompt words. Note that the strategy should be explicitly included within the candidate suggestions. The description.

[0111] Using the fifth language model to analyze each candidate prompt word Attempt to reason about the aforementioned subset of data, namely: In the formula, and yes The post and comment in the i-th record, Indicates passage The prompt indicates the predicted comment response obtained from the fifth language model.

[0112] The expectation is to predict comment replies. Comments and replies from real users Strive for semantic consistency. Therefore, measure the degree of semantic consistency and use semantic similarity as a candidate prompt word. rating .

[0113] in, ; wherein, represents the score of the jth candidate prompt word, represents the semantic similarity, represents the predicted comment reply obtained by the fifth large language model, represents the comment reply of the real user.

[0114] Among them, the large language model and the algorithm for calculating the semantic similarity can use common models, which will not be repeated.

[0115] Finally, under each hot data category, the strategy The optimal candidate prompt word of the reply is assigned to the agent according to certain rules, so that the agent has a better comment reply ability to hot content, and the subsequent social hot content semantic deduction is more in line with the function of the real social network.

[0116] Among them, the format of each valid content in the valid content set after format arrangement is That is, each comment reply indicates its corresponding cluster semantics (the semantics of the hot data category where the hot post is located), the original hot post and the comment.

[0117] In the optimization process, the valid content includes cluster semantics, hot post and its comment and comment reply.

[0118] For example, the valid content is .

[0119] For example, the valid content is .

[0120] Therefore, in the optimization process, the valid content and the valid content set in the embodiment are constantly iterated.

[0121] Considering that the comment reply in the real social network is usually only a few, in one specific embodiment, the preset convergence condition is set as: if all valid contents in step g are marked as "not participating in the next iteration", or the iteration has been full for 10 rounds, the iteration is stopped.

[0122] Secondly, the application also proposes a construction system of a social hot content semantic deduction agent, comprising:

[0123] The acquisition module acquires the historical hot search list data of the preset social network;

[0124] The processing module is used for clustering and processing the historical hot search list data to obtain a valid content set of multiple hot data categories;

[0125] The constructing module is configured to construct an agent for each hot data category respectively; the agent of each hot data category is endowed with a preset persona set; the preset persona set includes a plurality of preset personas;

[0126] The comment reply strategy constructing module is configured to, for each hot data category, obtain a comment reply strategy prompt word of the agent of the hot data category under each preset persona in the preset persona set of the agent according to the valid content set of the hot data category.

[0127] The historical hot search list data in the embodiment includes: hot post texts and comments, comment replies, and basic information and label information of post users and comment users and comment reply users of the hot post texts.

[0128] In the embodiment, the historical hot search list data is clustered and processed to obtain valid content sets of a plurality of hot data categories, specifically including:

[0129] Each hot post text of the historical hot search list data is respectively input into a preset vector model to obtain vector representations of the hot post texts in a semantic space;

[0130] The vector representations of all hot post texts are clustered by using a clustering algorithm to obtain a plurality of hot data categories;

[0131] A user set of each hot data category is constructed according to post users and comment users of all hot post texts in the hot data category respectively;

[0132] A content set of each hot data category is constructed according to hot post texts and comments of all users in the user set of the hot data category respectively;

[0133] All hot post texts of each hot data category are summarized to obtain a semantic of the hot data category;

[0134] Contents irrelevant to the semantic of each hot data category in the content set of the hot data category are respectively removed to obtain a valid content set of the hot data category.

[0135] In the embodiment, the agent of each hot data category is endowed with a preset persona set, specifically including:

[0136] The preset persona set is constructed; the preset persona set includes a plurality of preset personas;

[0137] For each hot data category, each content in the valid content set of the hot data category is classified by using a pre-trained multi-classification model to obtain a predicted persona of the content;

[0138] According to the predicted persona of each content, the number of contents under each preset persona is counted.

[0139] Pruning the preset person setting whose content quantity is less than the preset content quantity threshold from the preset person setting preset set, and taking the pruned person setting preset set as the preset person setting set of the hot data category.

[0140] In the embodiment, according to the effective content set of the hot data category, the comment reply strategy prompt word of the intelligent agent under each preset person setting in its preset person setting set is obtained, specifically including:

[0141] Step a, format combing the effective content set of the hot data category;

[0142] Step b, constructing the comment reply rule mining prompt word of the intelligent agent under each preset person setting of the hot data category;

[0143] Step c, using a large language model to mine the comment reply rule of the effective content set after format combing according to the comment reply rule mining prompt word under each preset person setting, to obtain a comment reply strategy set under each preset person setting;

[0144] Step d, for each preset person setting, constructing a comment reply strategy prompt word for each comment reply strategy in the comment reply strategy set under the preset person setting;

[0145] Step e, for each effective content in the effective content set, using the intelligent agent of the hot data category to generate a comment reply for the effective content based on each comment reply strategy prompt word under the preset person setting;

[0146] Step f, using a preset judgment large language training model to judge whether there is a best reply in the comment reply generated for the effective content; if not, the effective content is no longer involved in iteration, and the effective content is marked as "no longer involved in iteration", and the effective content set is updated; if yes, the effective content is updated according to the best reply, and the effective content set is updated according to the updated effective content;

[0147] Step g, judging whether the preset convergence condition is met; wherein the preset convergence condition is that all effective contents in the effective content set are no longer involved in iteration or iteration to a preset iteration number; if not, go to step d;

[0148] If yes, go to step h;

[0149] Step h, taking the comment reply strategy prompt word of the intelligent agent of the hot data category under the preset person setting when the preset convergence condition is met as the final comment reply strategy prompt word of the intelligent agent of the hot data category under the preset person setting.

[0150] Each data of the format-arranged effective content set comprises a comment reply, a comment corresponding to the comment reply, a hot post corresponding to the comment, and a semantic of a hot data category where the hot post is located.

[0151] A comment reply strategy prompt word is constructed for each comment reply strategy in the comment reply strategy set under the preset persona, and the comment reply strategy prompt word specifically comprises:

[0152] For each comment reply strategy in the comment reply strategy set under the preset persona, data in the effective content set that meets the comment reply strategy is formed into an effective content subset;

[0153] A candidate prompt word set is constructed for each comment reply strategy, and each candidate prompt word in the candidate prompt word set displays a description of the comment reply strategy corresponding to the effective content subset;

[0154] Each candidate prompt word in the candidate prompt word set is used to infer the effective content subset by using a large language model, and a predicted comment reply set is obtained; the semantic similarity between each predicted comment reply in the predicted comment reply set and a corresponding real comment reply is calculated, and the sum of the semantic similarities is taken as the score of each candidate prompt word;

[0155] The candidate prompt word with the maximum score is taken as the comment reply strategy prompt word of the comment reply strategy.

[0156] In a third aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the method for constructing a social hot content semantic deduction agent according to any one of the first aspect.

[0157] In a fourth aspect, the present application further provides a computer program product comprising a computer program, the computer program being executed by a processor to implement the steps of the method for constructing a social hot content semantic deduction agent according to any one of the first aspect.

[0158] The above description is only the preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for constructing a semantic inference agent for social trending content, characterized in that, include: Retrieve historical trending topics data from preset social networks; Clustering and processing of historical trending search data yields effective content sets for multiple trending data categories; Build an intelligent agent for each hot data category and assign a set of preset personas to the intelligent agent for each hot data category; For each hot data category, based on the effective content set of that hot data category, obtain the comment reply strategy prompts for the agent in that hot data category under each preset persona in its preset persona set.

2. The method for constructing a semantic inference agent for social hotspot content according to claim 1, characterized in that, Historical trending search data includes: trending posts and their comments, comment replies, basic information and tags of the users who posted and commented on the trending posts and the users who replied to the comments.

3. The method for constructing a semantic inference agent for social hotspot content according to claim 1, characterized in that, Clustering and processing of historical trending search data yields effective content sets across multiple trending data categories, specifically including: Each trending post from the historical hot search list is input into a preset vector model to obtain the vector representation of each trending post in the semantic space. Clustering algorithms are used to cluster the vector representations of all trending posts to obtain several trending data categories; Construct a user set for each hot data category based on the users who posted and commented on all hot posts within each hot data category; Construct a content set for each hot data category based on the hot posts and comments of all users within the user set of each hot data category; The semantics of each hot data category are obtained by summarizing all hot posts for each hot data category. By removing content from the content set of each hot data category that is semantically irrelevant to that hot data category, the effective content set for each hot data category is obtained.

4. The method for constructing a semantic inference agent for social hotspot content according to claim 1, characterized in that, Assign a pre-defined set of personas to the agent for each hot data category, specifically including: Construct a set of character presets; whereby the set of character presets includes multiple preset character settings; For each hot data category, a pre-trained multi-classification model is used to classify each piece of content in the effective content set of that hot data category, and the predicted persona for each piece of content is obtained. Based on the predicted character settings for each content, count the number of content items under each preset character setting; Preset personas with content quantity below a preset content quantity threshold are removed from the persona preset set, and the set of preset personas after removal is used as the preset persona set for that hot data category.

5. The method for constructing a semantic inference agent for social hotspot content according to claim 1, characterized in that, Based on the valid content set of this hot data category, the comment reply strategy prompts for the intelligent agent in this hot data category under each preset persona in its preset persona set are obtained, specifically including: Step a: Organize the format of the valid content set for this hot data category; Step b: Construct an intelligent agent for this hot data category to mine prompt words based on the pattern of comment replies under various preset personas; Step c: Using a large language model, extract prompt words based on the comment reply patterns under each preset persona. Perform comment reply pattern mining on the effective content set after formatting to obtain a set of comment reply strategies under each preset persona. Step d: For each preset persona, construct a comment reply strategy prompt for each comment reply strategy in the comment reply strategy set under that preset persona; Step e: For each valid piece of content in the set of valid content, the agent of the hot data category generates a comment reply for each valid piece of content based on the comment reply strategy prompts under the preset persona; Step f: Use a pre-set large language training model to determine whether there is a best reply among the comment replies generated for the valid content; if not, then stop the valid content from participating in the iteration and update the set of valid content; if yes, then update the valid content according to the best reply and update the set of valid content according to the updated valid content. Step g: Determine whether the preset convergence condition is met; wherein, the preset convergence condition is that all valid content in the set of valid content no longer participates in the iteration or iterates to the preset number of iterations; if not, proceed to step d; If so, proceed to step h; Step h: The comment reply strategy prompt words of the agent of the hot data category under the preset persona when the preset convergence condition is met are taken as the final comment reply strategy prompt words of the agent of the hot data category under the preset persona.

6. The method for constructing a semantic inference agent for social hotspot content according to claim 5, characterized in that, Each valid piece of content includes the semantics of the trending data category to which the trending post belongs, the trending post itself, its comments, and comment replies.

7. The method for constructing a semantic inference agent for social hot topic content according to claim 5, characterized in that, For each comment reply strategy in the set of comment reply strategies under the preset persona, construct a comment reply strategy prompt word, specifically including: For each comment reply strategy in the set of comment reply strategies under the preset persona, the data in the set of valid content that conforms to the comment reply strategy constitutes a subset of valid content; For each comment reply strategy, construct a set of candidate suggestion words; wherein, each candidate suggestion word in the set of candidate suggestion words displays a description of the comment reply strategy corresponding to the effective subset of content; The large language model is used to reason about the effective content subset for each candidate prompt word in the candidate prompt word set to obtain the predicted comment response set; the semantic similarity between each predicted comment response and the corresponding real comment response in the predicted comment response set is calculated, and the sum of the semantic similarity is taken as the score of each candidate prompt word; The candidate prompt word with the highest rating will be used as the prompt word for this comment reply strategy.

8. A system for constructing a semantic inference intelligent agent for social hot topic content, characterized in that, include: The acquisition module retrieves historical trending search data from preset social networks; The processing module is used to cluster and process historical hot search data to obtain a set of effective content for multiple hot data categories; The module is used to build agents for each hot data category and assign a set of preset personas to the agents for each hot data category. The comment reply strategy construction module is used to obtain comment reply strategy prompts for each hot data category based on the effective content set of that hot data category and the intelligent agent in that hot data category under each preset persona set.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for constructing a social hotspot content semantic inference agent as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method for constructing a social hotspot content semantic inference agent as described in any one of claims 1-7.

Images