Live broadcast e-commerce content generation method based on hotspot fusion, computer device and storage medium

Abstract

The invention discloses a live broadcast e-commerce content generation method based on hot spot fusion, a computer device and a storage medium, and the method comprises the steps: continuously obtaining hot spot frequent events to construct a data set, and employing a dynamic convolution model to fuse text data for a multi-level text structure; fusing text sentiment features of the text data by a directional tracking method to obtain dynamic event change features; performing feature extraction on the dynamic event change features by adopting a training model to obtain a sentence-level text feature vector, and obtaining a document-level feature vector from the text feature vector through a long-short-term memory network; and constructing an encoder for the document-level feature vector through a Fsatformer model to encode the text content, and adjusting the live broadcast content in real time to realize smooth conversion between different styles and predict the degree of interest of a user during live broadcast of a product in real time, so that the flexibility and adaptability of a live broadcast generation model are improved, and the user experience is improved. The generated live content conforms to the interest of the user, and the effect of the live room is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce live streaming, and particularly to a method for generating live e-commerce content based on hotspot fusion, a computer device, and a storage medium. Background Art

[0002] Today, with the booming development of live e-commerce, content generation technology plays a crucial role in enhancing user experience and promoting sales. The content generation methods of traditional live e-commerce mostly rely on the personal expression ability of the anchor and the pre-prepared scripts. Although this method has a certain degree of flexibility, there are also problems such as single live content, difficulty in quickly responding to changes in live products, and insufficient personalization. With the development of artificial intelligence and big data technologies, a method for generating live e-commerce content based on hotspot fusion has emerged, providing a more intelligent and dynamic content generation solution for live e-commerce.

[0003] Currently, the method for generating live e-commerce content based on hotspot fusion still has the following problems:

[0004] (1) By means of web crawlers, social media monitoring, etc., various types of information are widely collected to predict hotspot events. However, in the initial stage of the event, there are situations such as lack of semantic information, diverse and sparse information, and dynamic changes, resulting in problems such as insufficient utilization of initial event information and shallow mining of dynamic information features in existing methods;

[0005] (2) Machine learning algorithms are used to extract features from live text, and then live product recommendations are made based on the extracted live features. When extracting text features, it is relatively dependent on manual work, requiring a large amount of time and effort, and the context information behind the text cannot be obtained, resulting in unsatisfactory effects in extracting user features and live features. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for generating live e-commerce content based on hotspot fusion, a computer device, and a storage medium, so as to solve the technical problems such as single live content, difficulty in quickly responding to changes in live products, and insufficient personalization in the existing technology.

[0007] To solve the above technical problems, the present invention specifically provides the following technical solutions:

[0008] In the first aspect of the present invention, a method for generating live e-commerce content based on hotspot fusion is provided, including the following steps:

[0009] Continuously obtain frequently-occurring hotspot events to construct a data set, establish a multi-level text structure for the data set with the e-commerce live product chain as the master-slave relationship, and fuse text data for the multi-level text structure using a dynamic convolution model;

[0010] Adopt a semantic orientation tracking method based on sentiment semantics time series for the text data to fuse text sentiment features and obtain dynamic event change features;

[0011] Adopt a training model for the dynamic event change features to extract sentence-level text feature vectors, and obtain document-level feature vectors for the text feature vectors through a long short-term memory network;

[0012] Encode the text content through a Fsatformer model for the document-level feature vectors, introduce a pointer vector into the Fsatformer model to inject personalized user features into the document-level feature vectors, and adjust the live content in real time.

[0013] As a preferred solution of the present invention, establish a multi-level text structure for the data set based on the master-slave relationship of the e-commerce live product chain, including:

[0014] Take each category of products in the e-commerce live broadcast as a node, search for product relevance centered on the node, and sort the corresponding products in the order of positive correlation;

[0015] Search for descriptive words in the data set for the sorted product nodes, generate a document from the descriptive words with a word relationship graph, and calculate the weight w of the descriptive words o , and the expression is:

[0016]

[0017] Among them, i represents the number of descriptive words, j represents the number of documents, m represents the total number of times the descriptive words appear in the document, k represents the number of times each descriptive word appears in document j, and n i,j represents the number of times the descriptive word i appears in document j, and n k,j represents the intensity value of a descriptive word appearing in document j, and D j represents the number of documents;

[0018] Arrange the relationship graph between descriptive words according to the descriptive word weights, construct a master-slave relationship for the word relationship graph, and obtain a master word relationship graph and a slave word connection graph;

[0019] Merge the descriptive word content with the slave word connection graph corresponding to the product node of the master word relationship graph to obtain a text structure;

[0020] Inject text keywords into the text structure for the corresponding product nodes, calculate the importance of the text content by calculating the similarity between words, and divide the multi-level text structure according to the importance of the text content.

[0021] As a preferred solution of the present invention, fuse text data for the multi-level text structure by using a dynamic convolution model, including:

[0022] The data set is used to edit product features according to live products to obtain product styles, the product styles are compiled to obtain style coding features, the style coding features are corresponded to the multi-level text structure, and the multi-level text structure is segmented into multiple overlapping small blocks of different N*N. True or false judgment is performed on each overlapping small block through the first layer of convolution to obtain the true score of each overlapping small block;

[0023] According to the true scores, the style coding features are sequentially input into the coding network to obtain key coding features. The coding network is composed of two layers of convolution and an adaptive pooling layer. The key coding features perform dynamic convolution operations within the corresponding coding network to obtain hot data;

[0024] The hot data is incorporated into the multi-level text structure to obtain text data.

[0025] As a preferred solution of the present invention, a semantic orientation tracking method based on emotional semantic time series is used to fuse text emotional features for the text data, including:

[0026] The text data is divided into multiple units according to the real-time data stream, and the burst words in each unit are detected. The real-time data stream divides the text data according to a fixed time length;

[0027] Emotional scores are generated for the burst words through an emotional dictionary, the emotional changes of the text data are detected using a Gaussian distribution, and the emotional scores corresponding to the burst words within a unit time are calculated;

[0028] The same emotional scores within the same time period are attributed to the same emotional time series, and a word relationship graph is constructed in each unit using emotional semantic time series to obtain text emotional features.

[0029] As a preferred solution of the present invention, dynamic event change features are obtained based on the text emotional features, including:

[0030] The text emotional features are used as the original live information, and a Chinese text processing library is used to extract the text emotional features, and the emotional coefficient of the original live information is calculated;

[0031] According to the emotional coefficient, the number of forwards of the current live content is judged, the forward frequency feature of the original live information in the current state is calculated, and the popularity V of the current topic is calculated based on the forward frequency feature p , and its expression is:

[0032]

[0033] where s represents the current topic time, p represents the time window, denote the forwarding frequency of the s-th event within the p-th time window, and T denote the continuous observation time;

[0034] According to the popularity V of the topic p Continuously extract the total number of participating users within T time periods as a fan attribute feature, and use a hierarchical clustering algorithm to monitor the fan attribute feature in real time to obtain dynamic event change features.

[0035] As a preferred solution of the present invention, a training model is used to extract text feature vectors at the sentence level from the dynamic event change features, including:

[0036] The training model uses the RoBERTa algorithm to vectorize the dynamic event change features to obtain text relationship features, and encodes the text relationship features to generate sentence vectors;

[0037] The sentence vectors are incorporated with descriptive characters in the mode of position embedding vectors, and text feature vectors at the sentence level are obtained according to prior knowledge.

[0038] As a preferred solution of the present invention, document-level feature vectors are obtained from the text feature vectors through a long short-term memory network, including:

[0039] The text feature vectors are input into two long short-term memory networks with opposite time sequences for bidirectional combination, and the text feature vectors are bidirectionally encoded to extract comprehensive text features;

[0040] The comprehensive text features are used to calculate weights through a self-attention mechanism to obtain a weight matrix of the comprehensive text features;

[0041] The text features are sorted according to the weight matrix to construct document-level feature vectors.

[0042] As a preferred solution of the present invention, an encoder is constructed through the Fsatformer model to encode the text content of the document-level feature vectors, and a pointer vector is introduced into the Fsatformer model to inject personalized user features into the document-level feature vectors, including:

[0043] The Fsatformer model constructs an encoder through an independent linear Transformation layer, inputs the document-level feature vectors into an attention matrix of length N, and obtains global document vectors;

[0044] Calculate the key-value vectors of the weight matrix, combine the global document vectors with each key-value vector through an element-wise product operation, and use an attention mechanism to transform the key-value vectors into global key-value vectors;

[0045] Obtain the attention sequence for global information through linear transformation, and use the attention sequence as a pointer vector to summarize personalized user features containing global context information;

[0046] Integrate the personalized user features into the document-level feature vector and adjust the live content in real time.

[0047] The second aspect of the present invention provides a system for a live e-commerce content generation method based on hotspot fusion, including:

[0048] A data collection module that collects the latest hot topics from channels such as social media platforms and news websites, as well as user interaction data within the e-commerce platform;

[0049] A hotspot identification and analysis module that performs semantic analysis on the collected data to identify the current most popular product categories, brands, or events, and evaluates their popularity levels;

[0050] A personalized evaluation module that predicts the products and services that each user may be interested in based on the user's interests, and implements personalized recommendations according to the user's interests;

[0051] A content generation module that generates a live script, selects products to display, and designs interactive sessions according to user interests and hot events;

[0052] A feedback adjustment module that monitors the live effect in real time and dynamically adjusts the live content to optimize the user experience.

[0053] The second aspect of the present invention provides a storage medium in which computer execution instructions are stored. When the processor executes the computer execution instructions, the method described in any one of claims 1-8 is implemented.

[0054] The present invention has the following beneficial effects compared with the prior art:

[0055] The present invention adopts a dynamic convolution model to dynamically adjust the normalization parameters according to different live product styles to achieve smooth conversion between different styles, and real-time predicts the degree of user interest when live products, thereby improving the flexibility and adaptability of the live generation model;

[0056] The RoBERTa algorithm is used to extract text features to obtain sentence-level text feature vectors, which can obtain the context information contained behind the text in real time. Through the combination of long short-term memory network and attention mechanism, document-level feature vectors are obtained. An encoder is constructed based on the Fastformer model to encode the text vectors, and the user features extracted from the personalized live recommendation model are injected into the pointer generation network decoder, thereby affecting the generated live content and making the generated live content meet the user's interests and improve the live broadcast room effect. Brief Description of the Drawings

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only exemplary, and for those of ordinary skill in the art, without creative efforts, other implementation drawings can also be obtained according to the provided drawings by extension.

[0058] Figure 1 It is a flowchart of the live e-commerce content generation method provided by the embodiment of the present invention;

[0059] Figure 2 It is a system structure block diagram of the live e-commerce content generation method provided by the embodiment of the present invention. Detailed Embodiments

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0061] As Figure 1 shown, the present invention provides a live e-commerce content generation method based on hot spot fusion, including the following steps:

[0062] Continuously obtain frequently-occurring hot spot events to construct a data set, establish a multi-level text structure for the data set with the e-commerce live product chain as the master-slave relationship, and fuse the text data of the multi-level text structure using a dynamic convolution model;

[0063] In this embodiment, real-time hot spot information is obtained from multiple online resources and organized into a data set. The hot spot information is semantically parsed to determine highly relevant products or services, understand user attributes and interest points, and establish a multi-level text structure with the e-commerce live product chain as the master-slave relationship, which is used as the primary data for predicting user behavior preferences.

[0064] In this embodiment, the multi-level text structure is integrated into data such as user preferences and emotional characteristics during the live broadcast through a dynamic convolution model to obtain text data, which can effectively improve the accuracy of live broadcast effect monitoring.

[0065] Fuse the text emotion features of the text data using a semantic orientation tracking method based on emotional semantic time series to obtain dynamic event change features;

[0066] In this embodiment, text sentiment features are integrated into the text data, and an event tracking method based on sentiment time series is adopted to enhance the sensitivity of accurately detecting event changes.

[0067] The training model is used to extract features from the dynamic event change features to obtain text feature vectors at the sentence level, and the long short-term memory network is used to obtain document-level feature vectors from the text feature vectors;

[0068] In this embodiment, the dynamic event change features are trained through a training model, and an event heat calculation method integrating text popularity, content sensitivity, sentiment fluctuation value, and user participation is used to predict the event heat, obtain document-level feature vectors with directivity, and obtain accurate summaries of each stage of the event.

[0069] The Fsatformer model is used to construct an encoder to encode the text content of the document-level feature vectors. A pointer vector is introduced into the Fsatformer model to inject personalized user features into the document-level feature vectors and adjust the live content in real time.

[0070] In this embodiment, an encoder is constructed based on the Fastformer model to encode the document-level feature vectors, and the user features extracted from the personalized recommendation model are injected into the pointer generation network decoder, thereby generating high-heat topics and making the generated live content conform to the interests of users.

[0071] A multi-level text structure is established for the data set with the main and subordinate relationships of the e-commerce live product chain, including:

[0072] Each category of products in the e-commerce live is used as a node, and the product relevance is searched centered on the node, and the corresponding products are sorted in the order of positive correlation;

[0073] In this embodiment, each category of products in the live is used as a node, and the live content is written in the order of the live products, and the products in the live are sorted in positive correlation according to the popularity of the live products among users and the user emotion characteristics, so as to enhance the user's interest in the live content.

[0074] The sorted product nodes are searched in the data set for descriptive words, and the descriptive words are used to generate a document in a word relationship graph, and the weight w of the descriptive words is calculated o , and the expression is:

[0075]

[0076] where i represents the number of descriptive words, j represents the number of documents, m represents the total number of times the descriptive words appear in the document, k represents the number of times each descriptive word appears in document j, and n i,jDenotes the number of times the corresponding descriptor i appears in document j, n k,j Denotes the intensity value of a descriptor appearing in document j, D j Denotes the number of documents;

[0077] Arrange the relationship graph between descriptors according to the descriptor weights, construct the master-slave relationship for the word relationship graph, and obtain the master-word relationship graph and the slave-word connection graph;

[0078] In this embodiment, the TextRank algorithm is used to extract keywords from all descriptors. A network graph model is constructed by splitting words. The importance of words is calculated according to the similarity between words. Finally, the keywords are obtained by sorting according to the word weights, and the master-word relationship graph and the slave-word connection graph are obtained according to the word order between the keywords.

[0079] Merge the descriptor content with the slave-word connection graph corresponding to the product node of the master-word relationship graph to obtain the text structure;

[0080] Inject text keywords into the text structure corresponding to the product node, calculate the importance degree of the text content by calculating the similarity between words, and divide the multi-level text structure according to the importance degree of the text content.

[0081] In this embodiment, the weights are calculated according to the descriptors of the users appearing in the live broadcast room, the popularity of the live broadcast content by the users is analyzed, and the word relationship graph is constructed according to the popularity to generate the text structure, which is convenient to grasp the emotional fluctuations of the users.

[0082] Fuse the text data for the multi-level text structure using a dynamic convolution model, including:

[0083] Edit the product features of the dataset according to the live broadcast products to obtain the product style, compile the product style to obtain the style encoding features, correspond the style encoding features to the multi-level text structure, and divide the multi-level text structure into multiple overlapping small blocks of different N*N. Perform true / false judgment on each overlapping small block through the first layer of convolution to obtain the true score of each overlapping small block;

[0084] Input the style encoding features into the encoding network in turn according to the true scores to obtain the key encoding features. The encoding network consists of two layers of convolution and an adaptive pooling layer. The key encoding features perform dynamic convolution operations within the corresponding encoding network to obtain the hot data;

[0085] In this embodiment, in the dynamic convolution transformation, the weights and biases of the encoding network are dynamically generated by encoding the input dataset, which can process any data and improves the flexibility and adaptability of the model.

[0086] Integrate the hot data into the multi-level text structure to obtain the text data.

[0087] In this embodiment, the input product style is segmented into multiple overlapping small blocks of N*N, and each overlapping small block is judged as true or false by means of convolution. In this process, each small block is regarded as independent, and the authenticity score of each image can be regarded as the average value of the authenticity scores of all small blocks. In this way, the output of the generator can be better constrained, so that it can more accurately learn to generate descriptive statements with user attributes, thereby helping the generator generate more accurate text data.

[0088] The semantic orientation tracking method based on emotional semantic time series is used to fuse the text emotional features of the text data, including:

[0089] The text data is divided into multiple units according to the real-time data stream, and the burst words in each unit are detected. The real-time data stream divides the text data according to a fixed time length;

[0090] The emotional score of the burst word is generated through an emotional dictionary, the emotional change of the text data is detected by using a Gaussian distribution, and the emotional score corresponding to the burst word within a unit time is calculated;

[0091] The same emotional score within the same time period is attributed to the same emotional time series, and a word relationship graph is constructed in each unit by using emotional semantic time series to obtain text emotional features.

[0092] In this embodiment, task events are configured for the text data, the occurrence frequency of the events is tracked regularly, and it is set to run once every two hours. The burst words in each unit are detected. First, the classifyWords function is written to calculate the emotional score, and it is judged which emotional words, negative words, and adverbs of degree are contained in the text. After using the open function to obtain all text words in the way of reading line by line, a dictionary is constructed by using the defaultdict function of the collections module. The Keys function is used to judge whether the word exists in the corresponding dictionary. All emotional words and corresponding emotional weights are traversed, and the emotional score is calculated in combination with negative words and adverbs of degree. After adding up all the text emotional scores, the average value is taken as the time series emotional score. A word relationship graph is constructed in each unit by using emotional semantic time series to obtain text emotional features.

[0093] The dynamic event change features are obtained according to the text emotional features, including:

[0094] The text emotional features are used as the original live broadcast information, and a Chinese text processing library is used to extract the text emotional features and calculate the emotional coefficient of the original live broadcast information;

[0095] Judge the forwarding quantity of the current live content according to the sentiment coefficient, calculate the forwarding frequency feature of the original live information in the current state, and calculate the popularity V of the current topic according to the forwarding frequency feature p , and its expression is:

[0096]

[0097] where s represents the current topic time, p represents the time window, represents the forwarding frequency of the s-th event within the p-th time window, and T represents the continuous observation time;

[0098] According to the popularity V of the topic p continuously extract the total number of participating users as the fan attribute feature within T time periods, and use the hierarchical clustering algorithm to monitor the fan attribute feature in real time to obtain the dynamic event change feature.

[0099] Use the training model to extract feature vectors at the sentence level from the dynamic event change feature, including:

[0100] The training model uses the RoBERTa algorithm to vectorize the dynamic event change feature to obtain the text relationship feature, and encodes the text relationship feature to generate a sentence vector;

[0101] In this embodiment, the word embedding, segment embedding, and position embedding information are fused as the input of the RoBERTa algorithm to obtain the text context relationship feature and encode it to generate a sentence vector containing rich information.

[0102] Integrate the description characters into the sentence vector in the mode of the position embedding vector, and obtain the text feature vector at the sentence level according to the prior knowledge.

[0103] In this embodiment, the word embedding vector is obtained by the text encoder of the RoBERTa model, the segment embedding vector is obtained by querying the word vector table, the position embedding vector is obtained by modeling the position of the characters in the sentence using relative position encoding, and then the three vectors are used as the input vector of the RoBERTa pre-training model in a summing way, and the text feature vector is obtained through the prior knowledge of the RoBERTa pre-training model.

[0104] Use the long short-term memory network to obtain the document-level feature vector from the text feature vector, including:

[0105] Input the text feature vector into two long short-term memory networks with opposite time sequences for bidirectional combination, perform bidirectional encoding on the text feature vector, and extract the comprehensive text feature;

[0106] Calculate the weights of the comprehensive text features through the self-attention mechanism to obtain the weight matrix of the comprehensive text features;

[0107] Sort the text features according to the weight matrix to construct the document-level feature vector.

[0108] In this embodiment, a long short-term memory network is introduced after the RoBERTa pre-trained model layer. By combining two long short-term memory network models with opposite time sequences, the text sequence can be encoded in both forward and backward directions. Text feature extraction is performed in the bidirectional long short-term memory network model in both forward and backward directions, and finally the forward and backward features are fused to obtain the document-level text feature vector.

[0109] Encode the document-level feature vector through the Fsatformer model to construct an encoder for the text content. Inject personalized user features into the document-level feature vector in the Fsatformer model, including:

[0110] The Fsatformer model constructs an encoder through an independent linear Transformation layer, inputs the document-level feature vector into an attention matrix of length N, and obtains the global document vector;

[0111] Calculate the key-value vectors of the weight matrix, combine the global document vector with each key-value vector through an element-wise product operation, and transform the key-value vectors into global key-value vectors using the attention mechanism;

[0112] Obtain the attention sequence of the global information through a linear transformation, and use the attention sequence as the pointer vector to summarize the personalized user features containing the global context information;

[0113] In this embodiment, encoding the text data using the attention sequence can combine multiple self-attention mechanisms. The focus of the features learned by each self-attention mechanism may be slightly different, thereby reducing the overfitting phenomenon and being able to learn more comprehensive text feature information.

[0114] Integrate the personalized user features into the document-level feature vector and adjust the live content in real time.

[0115] In this embodiment, using the Fsatformer model to model and analyze the document-level feature vector and output the feature vector containing context information can combine the context content to obtain more information, greatly improving the ability of the language model to obtain text vector features. At the same time, injecting the user representation information extracted by the personalized recommendation model during the decoding process enables the generated live content to conform to the behavior preferences of different users.

[0116] AsFigure 2 As shown in the figure, Second Embodiment: A system for a live e-commerce content generation method based on hotspot fusion, comprising:

[0117] A data collection module that collects the latest hot topics from channels such as social media platforms and news websites, as well as user interaction data within the e-commerce platform;

[0118] A hotspot identification and analysis module that performs semantic analysis on the collected data to identify the currently most popular product categories, brands, or events, and evaluates their popularity levels;

[0119] A personalized evaluation module that predicts the products and services that a user may be interested in based on each user's interests, and implements personalized recommendations according to user interests;

[0120] A content generation module that generates a live script, selects displayed products, and designs interactive sessions according to user interests and hot events;

[0121] A feedback adjustment module that monitors the live broadcast effect in real time and dynamically adjusts the live broadcast content to optimize the user experience.

[0122] In this embodiment, the RoBERTa pre-trained model is used to obtain real-time hotspot information from multiple online resources, perform semantic parsing on the hotspot information, determine the popularity of live products for users and the semantic feature representations of related products, establish a personalized model based on the user's behavior pattern, formulate a live broadcast plan, and adjust the index content in real time according to the audience interaction during the live broadcast.

[0123] Third Embodiment: A storage medium in which computer-executable instructions are stored, and when a processor executes the computer-executable instructions, the method described in any one of claims 1-8 is implemented.

[0124] The present invention uses a dynamic convolution model to dynamically adjust the normalization parameters according to different live product styles to achieve smooth conversion between different styles, and real-time predicts the degree of user interest when live products are being broadcast, thereby improving the flexibility and adaptability of the live broadcast generation model;

[0125] The RoBERTa algorithm is used to extract text features to obtain sentence-level text feature vectors, which can obtain the context information contained behind the text in real time. Through a long short-term memory network combined with an attention mechanism, document-level feature vectors are obtained. An encoder is constructed based on the Fastformer model to encode the text vectors, and the user features extracted from the personalized live broadcast recommendation model are injected into the pointer generator network decoder, thereby affecting the generated live broadcast content and making the generated live broadcast content conform to the user's interests and improve the live broadcast room effect.

[0126] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present application. The protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions within the essence and protection scope of the present application, and such modifications or equivalent substitutions should also be regarded as falling within the protection scope of the present application.

Claims

1. A method for generating live e-commerce content based on hotspot fusion, characterized in that: The following steps are involved: Continuously obtain hot frequent events to construct a data set, establish a multi-level text structure for the data set based on the master-slave relationship of the e-commerce live broadcast product chain, and use a dynamic convolution model to fuse text data for the multi-level text structure; The text data is subjected to a semantic directional tracking method based on the time sequence of sentiment semantics to fuse the text sentiment features and obtain the dynamic event change features; The training model is used to extract the dynamic event change features to obtain a sentence-level text feature vector, and the text feature vector is used to obtain a document-level feature vector through a long short-term memory network; The document-level feature vector is encoded by constructing an encoder through the Fsatformer model to encode the text content, and a pointer vector is introduced into the Fsatformer model to inject personalized user features into the document-level feature vector, so as to adjust the live broadcast content in real time.

2. According to the method for generating live e-commerce content based on hotspot fusion according to claim 1, it is characterized in that: A multi-level text structure is established for the data set based on the e-commerce live broadcast product chain as the master-slave relationship, including: Take each category of products in the e-commerce live broadcast as a node, search for product relevance with the node as the center, and sort the corresponding products in order of positive relevance; Search the sorted product nodes for descriptive words in the data set, generate documents with the descriptive words as word relationship graphs, and calculate the weights w of the descriptive words. o , the expression is: Where i represents the number of descriptive words, j represents the number of documents, m represents the total number of occurrences of descriptive words in the documents, k represents the number of occurrences of each descriptive word in document j, and n represents the total number of occurrences of each descriptive word in document j. i,j Indicates the number of times the corresponding description word i appears in document j, n k,j represents the intensity value of a description word appearing in document j, D j Indicates the number of documents; Arrange the relationship graph between description words according to the description word weights, construct a master-slave relationship for the word relationship graph, and obtain a master word relationship graph and a slave word connection graph; The main word relationship graph corresponds to the product node, and the description word content is merged with the sub-word connection graph to obtain a text structure; Text keywords are injected into the text structure with corresponding product nodes, the importance of the text content is calculated by calculating the similarity between words, and the text structure is divided into multiple levels according to the importance of the text content.

3. According to the method for generating live e-commerce content based on hotspot fusion according to claim 2, it is characterized in that: The multi-level text structure is subjected to text data fusion using a dynamic convolution model, including: The data set is edited according to the product features of the live broadcast product to obtain the product style, the product style is compiled to obtain the style coding feature, the style coding feature is corresponded to the multi-level text structure, and the multi-level text structure is divided into different N*N multiple overlapping small blocks, and each overlapping small block is judged to be true or false through the first layer of convolution to obtain the real score of each overlapping small block; According to the true score, the style coding features are sequentially input into the coding network to obtain key coding features, wherein the coding network is composed of two layers of convolution and an adaptive pooling layer, and the key coding features are dynamically convolved in the corresponding coding network to obtain hotspot data; The hotspot data is integrated into the multi-level text structure to obtain text data.

4. According to the method for generating live e-commerce content based on hotspot fusion according to claim 3, it is characterized in that: The text data is subjected to a semantic directional tracking method based on the time sequence of sentiment semantics to fuse the text sentiment features, including: Dividing the text data into a plurality of units according to a real-time data stream, detecting burst words in each unit, wherein the real-time data stream divides the text data according to a fixed time length; Generate a sentiment score for the burst word through a sentiment dictionary, detect sentiment changes of the text data using Gaussian distribution, and calculate the sentiment score corresponding to the burst word in unit time; The same emotion score in the same time period is attributed to the same emotion time series, and the emotion semantic time series is used to construct a word relationship graph in each unit to obtain the text emotion features.

5. According to the method for generating live e-commerce content based on hotspot fusion according to claim 3, it is characterized in that: The dynamic event change features are obtained based on the text sentiment features, including: Taking the text sentiment feature as the original live broadcast information, extracting the text sentiment feature using a Chinese text processing library, and calculating the sentiment coefficient of the original live broadcast information; The number of reposts of the current live broadcast content is determined based on the sentiment coefficient, and the repost frequency characteristics of the original live broadcast information in the current state are calculated. The popularity V of the current topic is calculated based on the repost frequency characteristics. p , whose expression is: Among them, s represents the current topic time, p represents the time window, represents the forwarding frequency of the s-th event in the p-th time window, and T represents the continuous observation time; According to the popularity of the topic V p The total number of participating users is continuously extracted in T time periods as fan attribute features, and a hierarchical clustering algorithm is used to monitor the fan attribute features in real time to obtain dynamic event change features.

6. A method for generating live e-commerce content based on hotspot fusion according to claim 5, characterized in that: The training model is used to extract the dynamic event change features to obtain a sentence-level text feature vector, including: The training model uses the RoBERTa algorithm to perform text vectorization on the dynamic event change features, obtain text relationship features, and encode the text relationship features to generate sentence vectors; The sentence vector is embedded with description characters using a position embedding vector model, and a text feature vector at the sentence level is obtained based on prior knowledge.

7. A method for generating live e-commerce content based on hotspot fusion according to claim 6, characterized in that: Obtaining a document-level feature vector from the text feature vector through a long short-term memory network includes: Inputting the text feature vector into two long short-term memory networks with opposite time sequences for bidirectional combination, bidirectionally encoding the text feature vector, and extracting comprehensive text features; Calculate the weight of the comprehensive text features through a self-attention mechanism to obtain a weight matrix of the comprehensive text features; The text features are sorted according to the weight matrix to construct a document-level feature vector.

8. A method for generating live e-commerce content based on hotspot fusion according to claim 7, characterized in that: The document-level feature vector is encoded by constructing an encoder through the Fsatformer model to encode the text content, and a pointer vector is introduced into the Fsatformer model to inject personalized user features into the document-level feature vector, including: The Fsatformer model constructs an encoder through an independent linear Transformation layer, inputs the document-level feature vector into an attention matrix of length N, and obtains a global document vector; Calculating the key value vector of the weight matrix, combining the global document vector with each key value vector through an element-wise product operation, and converting the key value vector into a global key value vector using an attention mechanism; Obtaining an attention sequence of global information through linear transformation, and using the attention sequence as a pointer vector to summarize personalized user features containing global context information; The personalized user features are integrated into the document-level feature vector, and the live broadcast content is adjusted in real time.

9. A system for generating live e-commerce content based on hotspot fusion, characterized in that: include: Data collection module, which collects the latest hot topics from social media platforms, news websites and other channels, as well as user interaction data within e-commerce platforms; The hotspot identification and analysis module performs semantic analysis on the collected data, identifies the most popular product categories, brands or events, and evaluates their popularity levels; The personalized evaluation module predicts the products and services that each user may be interested in based on their interests, and implements personalized recommendations based on the user's interests; Content generation module, which generates live broadcast scripts, selects display products and designs interactive links based on user interests and hot events; Feedback adjustment module monitors the live broadcast effect in real time and dynamically adjusts the live broadcast content to optimize the user experience.

10. A storage medium, characterized in that: The storage medium stores computer-executable instructions, and when the processor executes the computer-executable instructions, the method according to any one of claims 1 to 8 is implemented.

Images