A method, computer device and storage medium for generating live e-commerce content based on hotspot fusion

Abstract

The present invention discloses a live e-commerce content generation method based on hotspot fusion, a computer device and a storage medium, comprising: continuously acquiring a data set of frequently occurring hotspot events, fusing text data with a dynamic convolution model for the multi-level text structure; fusing text sentiment features with a directional tracking method for the text data to obtain dynamic event change features; extracting features from the dynamic event change features using a training model to obtain sentence-level text feature vectors, and obtaining document-level feature vectors from the text feature vectors through a long short-term memory network; encoding the text content with an Fsatformer model for the document-level feature vectors, adjusting the live content in real time to achieve smooth conversion between different styles, and predicting the user's interest level in live products in real time, thereby improving the flexibility and adaptability of the live generation model, making the generated live content meet the user's interests, and improving the live broadcast effect.

Description

Technical Field

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

[0002] With the booming development of live e-commerce, content generation technology plays a vital role in improving user experience and promoting sales. Traditional live e-commerce content generation methods rely heavily on the host's personal expressive ability and pre-prepared scripts. Although this method has a certain degree of flexibility, it also has problems such as single live content, difficulty in quickly responding to changes in live products, and lack of personalization. With the development of artificial intelligence and big data technology, live e-commerce content generation methods based on hotspot fusion have emerged, providing live e-commerce with a smarter and more dynamic content generation solution.

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

[0004] (1) Various types of information are widely collected through web crawlers and social media monitoring to predict hot events. However, in the early stages of an event, there is a lack of semantic information, and the information is diverse, sparse, and dynamically changing. As a result, existing methods have problems such as insufficient utilization of initial event information and shallow mining of dynamic information features.

[0005] (2) Features are extracted from live broadcast texts through machine learning algorithms, and then live broadcast product recommendations are made based on the extracted live broadcast features. Extracting text features is relatively dependent on manual labor, which requires a lot of time and effort, and is unable to obtain the contextual information behind the text, resulting in unsatisfactory results in extracting user features and live broadcast features. Summary of the Invention

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

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

[0008] A first aspect of the present invention provides a method for generating live e-commerce content based on hotspot fusion, comprising the following steps:

[0009] Continuously acquire hot and 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;

[0010] 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 dynamic event change features;

[0011] Extracting the dynamic event change features using a training model to obtain sentence-level text feature vectors, and obtaining document-level feature vectors from the text feature vectors using a long short-term memory network;

[0012] 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.

[0013] As a preferred solution of the present invention, a multi-level text structure is established for the dataset based on the e-commerce live broadcast product chain as a master-slave relationship, including:

[0014] 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 correlation;

[0015] Search the sorted product nodes in the dataset for description words, generate documents with the description words as word relationship graph, and calculate the weight w of the description words o , the expression is:

[0016]

[0017] Among them, i represents the number of description words, j represents the number of documents, m represents the total number of times the description words appear in the documents, k represents the number of times each description word appears in document j, and n 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;

[0018] 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;

[0019] Merge the description word content and the sub-word connection graph corresponding to the product node in the main word relationship graph to obtain a text structure;

[0020] The text structure is injected with text keywords at corresponding product nodes, the importance of 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.

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

[0022] The dataset is edited according to the product features of the live broadcast products to obtain the product style, the product style is compiled to obtain the style coding features, the style coding features are mapped to the multi-level text structure, and the multi-level text structure is divided into multiple N*N overlapping blocks. The first layer of convolution is used to judge the authenticity of each overlapping block to obtain the true score of each overlapping block;

[0023] According to the true score, 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 are dynamically convolved in the corresponding coding network to obtain hotspot data.

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

[0025] As a preferred solution of the present invention, the text data is subjected to a semantic directional tracking method based on the emotional semantic time sequence to fuse the text emotional features, including:

[0026] Dividing the text data into a plurality of units according to a real-time data stream, and detecting a burst word in each unit, wherein the real-time data stream divides the text data into a plurality of units according to a fixed time length;

[0027] Generate a sentiment score for the burst word using a sentiment dictionary, detect sentiment changes in the text data using Gaussian distribution, and calculate the sentiment score corresponding to the burst word per unit time;

[0028] 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.

[0029] As a preferred solution of the present invention, obtaining dynamic event change features based on the text sentiment features includes:

[0030] Taking the text emotion feature as the original live broadcast information, extracting the text emotion feature using a Chinese text processing library, and calculating the emotion coefficient of the original live broadcast information;

[0031] Determine the forwarding number of the current live content based on the sentiment coefficient, calculate the forwarding frequency characteristics of the original live information in the current state, and calculate the popularity V of the current topic based on the forwarding frequency characteristics. p , whose expression is:

[0032]

[0033] 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;

[0034] According to the popularity of the topic V p The total number of participating users is continuously extracted as the fan attribute feature within T time periods, and the hierarchical clustering algorithm is used to monitor the fan attribute feature in real time to obtain the dynamic event change feature.

[0035] As a preferred solution of the present invention, the dynamic event change features are subjected to feature extraction using a training model to obtain a sentence-level text feature vector, including:

[0036] 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;

[0037] The sentence vector is embedded with description characters using a position embedding vector model, and a sentence-level text feature vector is obtained based on prior knowledge.

[0038] As a preferred solution of the present invention, obtaining a document-level feature vector from the text feature vector through a long short-term memory network includes:

[0039] 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;

[0040] Calculate the weight of the comprehensive text features through the 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 a document-level feature vector.

[0042] As a preferred solution of the present invention, 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:

[0043] 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;

[0044] 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;

[0045] 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;

[0046] The personalized user features are integrated into the document-level feature vector, and the live broadcast content is adjusted in real time.

[0047] A second aspect of the present invention provides a system for generating live e-commerce content based on hotspot fusion, comprising:

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

[0049] The hotspot identification and analysis module performs semantic analysis on the collected data to identify the most popular product categories, brands, or events and assess their popularity level.

[0050] 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 their interests;

[0051] The content generation module generates live broadcast scripts, selects products to be displayed, and designs interactive sessions based on user interests and hot events;

[0052] Feedback adjustment module monitors live broadcast effects in real time and dynamically adjusts live broadcast content to optimize user experience.

[0053] According to a second aspect of the present invention, a storage medium is provided, wherein the storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the method according to any one of claims 1 to 8 is implemented.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention adopts a dynamic convolution model to dynamically adjust the normalization parameters according to the different styles of live broadcast products to achieve smooth transitions between different styles and predict the user's interest level in the live broadcast product in real time, thereby improving the flexibility and adaptability of the live broadcast generation model;

[0056] The RoBERTa algorithm is used to extract text features and obtain sentence-level text feature vectors, which can obtain the contextual information behind the text in real time. The document-level feature vectors are obtained through the long short-term memory network combined with the attention mechanism. An encoder is built based on the Fastformer model to encode the text vectors. The user features extracted from the personalized live broadcast recommendation model are injected into the pointer generation network decoder, thereby influencing the generated live broadcast content, making the generated live broadcast content in line with the user's interests and improving the live broadcast effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0058] Figure 1 A flow chart of a method for generating live e-commerce content provided by an embodiment of the present invention;

[0059] Figure 2 A system structure block diagram of the live e-commerce content generation method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] like Figure 1 As shown, the present invention provides a method for generating live e-commerce content based on hotspot fusion, comprising the following steps:

[0062] Continuously acquire hot and 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;

[0063] In this embodiment, real-time hot spot information is obtained from multiple online resources and organized into a data set, semantic analysis is performed on the hot spot information, highly relevant goods or services are determined, user attributes and interests are understood, and a multi-level text structure is established with the e-commerce live broadcast product chain as the master-slave relationship, which is used as 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] 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 dynamic event change features;

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

[0067] Extracting the dynamic event change features using a training model to obtain sentence-level text feature vectors, and obtaining document-level feature vectors from the text feature vectors using a long short-term memory network;

[0068] In this embodiment, the dynamic event change features are trained through a training model, and the event heat calculation method of text popularity, content sensitivity, emotional fluctuation value, and user participation is integrated to predict the event heat, obtain a directional document-level feature vector, and obtain an accurate summary of each stage of the event.

[0069] 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.

[0070] In this embodiment, an encoder is constructed based on the Fastformer model to encode the document-level feature vector, and the user features extracted from the personalized recommendation model are injected into the pointer generation network decoder to generate high-hot topics, so that the generated live broadcast content meets the user's interests.

[0071] A multi-level text structure is established for the dataset based on the e-commerce live product chain as the master-slave relationship, including:

[0072] 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 correlation;

[0073] In this embodiment, each type of live broadcast product is used as a node, and the live broadcast content is written in the order of the live broadcast products. The live broadcast products are sorted in order of positive correlation based on the user's popularity of the live broadcast products and the user's emotional characteristics, thereby increasing the user's interest in the live broadcast content.

[0074] Search the sorted product nodes in the dataset for description words, generate documents with the description words as word relationship graph, and calculate the weight w of the description words o , the expression is:

[0075]

[0076] Among them, i represents the number of description words, j represents the number of documents, m represents the total number of times the description words appear in the documents, k represents the number of times each description word appears in document j, and n i,jIndicates 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;

[0077] 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;

[0078] In this embodiment, the TextRank algorithm is used to extract keywords from all descriptive words, a network graph model is constructed by splitting words, the importance of words is calculated based on the similarity between words, and finally the keywords are obtained by sorting the words according to their weights, and the main word relationship graph and the subordinate word connection graph are obtained based on the word order between the keywords.

[0079] Merge the description word content and the sub-word connection graph corresponding to the product node in the main word relationship graph to obtain a text structure;

[0080] The text structure is injected with text keywords at corresponding product nodes, the importance of 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.

[0081] In this embodiment, weights are calculated based on the descriptive words of the users appearing in the live broadcast room, the users' enthusiasm for the live broadcast content is analyzed, and a word relationship graph is constructed based on the enthusiasm to generate a text structure, which makes it easier to grasp the fluctuations in users' emotions.

[0082] The multi-level text structure is subjected to text data fusion using a dynamic convolution model, including:

[0083] The dataset is edited according to the product features of the live broadcast products to obtain the product style, the product style is compiled to obtain the style coding features, the style coding features are mapped to the multi-level text structure, and the multi-level text structure is divided into multiple N*N overlapping blocks. The first layer of convolution is used to judge the authenticity of each overlapping block to obtain the true score of each overlapping block;

[0084] According to the true score, 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 are dynamically convolved in the corresponding coding network to obtain hotspot 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 data set, which can process arbitrary data and improve the flexibility and adaptability of the model.

[0086] The hotspot data is integrated into the multi-level text structure to obtain text data.

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

[0088] The text data is subjected to a semantic directional tracking method based on the emotional semantic time sequence to fuse the text emotional features, including:

[0089] Dividing the text data into a plurality of units according to a real-time data stream, and detecting a burst word in each unit, wherein the real-time data stream divides the text data into a plurality of units according to a fixed time length;

[0090] Generate a sentiment score for the burst word using a sentiment dictionary, detect sentiment changes in the text data using Gaussian distribution, and calculate the sentiment score corresponding to the burst word per unit time;

[0091] 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.

[0092] In this embodiment, task events are configured for text data, and the frequency of events is tracked regularly. It is set to run once every two hours to detect sudden words in each unit. First, the classifyWords function is written to calculate the sentiment score, and it is determined which sentiment words, negative words and degree adverbs are contained in the text. After using the open function to read all text words line by line in the readlines manner, a dictionary is constructed using the defaultdict function of the collections module. The Keys function is used to determine whether the word exists in the corresponding dictionary. All sentiment words and corresponding sentiment weights are traversed, and the sentiment score is calculated by combining negative words and degree adverbs. The sentiment scores of all texts are added up and the average is taken as the temporal sentiment score. The sentiment semantic temporal sequence is used to construct a word relationship graph in each unit to obtain the text sentiment features.

[0093] Acquiring dynamic event change features based on the text sentiment features includes:

[0094] Taking the text emotion feature as the original live broadcast information, extracting the text emotion feature using a Chinese text processing library, and calculating the emotion coefficient of the original live broadcast information;

[0095] Determine the forwarding number of the current live content based on the sentiment coefficient, calculate the forwarding frequency characteristics of the original live information in the current state, and calculate the popularity V of the current topic based on the forwarding frequency characteristics. p , whose expression is:

[0096]

[0097] 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;

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

[0099] The dynamic event change features are extracted using a training model to obtain sentence-level text feature vectors, including:

[0100] 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;

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

[0102] The sentence vector is embedded with description characters using a position embedding vector model, and a sentence-level text feature vector is obtained based on 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, and the position embedding vector is obtained by modeling the position of the characters in the sentence using relative position encoding. The three vectors are then summed as the input vector of the RoBERTa pre-trained model, and the text feature vector is obtained through the prior knowledge of the RoBERTa pre-trained model.

[0104] Obtaining a document-level feature vector from the text feature vector through a long short-term memory network includes:

[0105] 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;

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

[0107] The text features are sorted according to the weight matrix to construct a document-level feature vector.

[0108] In this embodiment, a long short-term memory network is introduced after the RoBERTa pre-training model layer, and two long short-term memory network models with opposite time sequences are combined together. The text sequence can be encoded from both forward and reverse directions, and forward and backward text features are extracted in the bidirectional long short-term memory network model. Finally, the forward and backward features are fused to obtain a document-level text feature vector.

[0109] The document-level feature vector is constructed by an encoder using 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:

[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 a global document vector;

[0111] 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;

[0112] 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;

[0113] In this embodiment, the text data is encoded using an attention sequence, and multiple self-attention mechanisms can be combined together. The focus of the features learned by each self-attention mechanism may be slightly different, thereby reducing the overfitting phenomenon and learning more comprehensive text feature information.

[0114] The personalized user features are integrated into the document-level feature vector, and the live broadcast content is adjusted in real time.

[0115] In this embodiment, the Fsatformer model is used to model and analyze the document-level feature vector, and output a feature vector containing contextual information. It can obtain more information in combination with the contextual content, greatly improving the ability of the language model to obtain text vector features. At the same time, the user representation information extracted by the personalized recommendation model is injected into the decoding process, so that the generated live broadcast content can meet the behavioral preferences of different users.

[0116] like Figure 2 As shown, the second embodiment: a system for generating live e-commerce content based on hotspot fusion, comprising:

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

[0118] The hotspot identification and analysis module performs semantic analysis on the collected data to identify the most popular product categories, brands, or events and assess their popularity level.

[0119] 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 their interests;

[0120] The content generation module generates live broadcast scripts, selects products to be displayed, and designs interactive sessions based on user interests and hot events;

[0121] Feedback adjustment module monitors live broadcast effects in real time and dynamically adjusts live broadcast content to optimize user experience.

[0122] In this embodiment, the RoBERTa pre-trained model is used to obtain real-time hot spot information from multiple online resources, and semantic analysis is performed on the hot spot information to determine the user's enthusiasm for the live broadcast product and the semantic feature representation of related products. A personalized model is established based on the user's behavior pattern, and a live broadcast plan is formulated. During the live broadcast, the indicator content is adjusted in real time according to the audience interaction.

[0123] A third embodiment: a storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, the method according to any one of claims 1 to 8 is implemented.

[0124] The present invention adopts a dynamic convolution model to dynamically adjust the normalization parameters according to the different styles of live broadcast products to achieve smooth transitions between different styles and predict the user's interest level in the live broadcast product in real time, thereby improving the flexibility and adaptability of the live broadcast generation model;

[0125] The RoBERTa algorithm is used to extract text features and obtain sentence-level text feature vectors, which can obtain the contextual information behind the text in real time. The document-level feature vectors are obtained through the long short-term memory network combined with the attention mechanism. An encoder is built based on the Fastformer model to encode the text vectors. The user features extracted from the personalized live broadcast recommendation model are injected into the pointer generation network decoder, thereby influencing the generated live broadcast content, making the generated live broadcast content in line with the user's interests and improving the live broadcast effect.

[0126] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection 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 hotspot frequent events to construct a data set; 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 correlation; Search the sorted product nodes in the dataset for description words, generate documents with the description words as word relationship graph, and calculate the weight w of the description words o , the expression is: Among them, i represents the number of description words, j represents the number of documents, m represents the total number of times the description words appear in the documents, k represents the number of times each description word appears in document j, and n 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; Merge the description word content and the sub-word connection graph corresponding to the product node in the main word relationship graph to obtain a text structure; Injecting text keywords into the text structure at corresponding product nodes, calculating the importance of text content by calculating the similarity between words, and dividing the text structure into multiple levels according to the importance of the text content; fusing text data using a dynamic convolution model 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 dynamic event change features; Extracting the dynamic event change features using a training model to obtain sentence-level text feature vectors, and obtaining document-level feature vectors from the text feature vectors using 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. A method for generating live e-commerce content based on hotspot fusion according to claim 1, characterized in that: The multi-level text structure is subjected to text data fusion using a dynamic convolution model, including: The dataset is edited according to the product features of the live broadcast products to obtain the product style, the product style is compiled to obtain the style coding features, the style coding features are mapped to the multi-level text structure, and the multi-level text structure is divided into multiple N*N overlapping blocks. The first layer of convolution is used to judge the authenticity of each overlapping block to obtain the true score of each overlapping block; According to the true score, 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 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.

3. A method for generating live e-commerce content based on hotspot fusion according to claim 2, characterized in that: The text data is subjected to a semantic directional tracking method based on the emotional semantic time sequence to fuse the text emotional features, including: Dividing the text data into a plurality of units according to a real-time data stream, and detecting a burst word in each unit, wherein the real-time data stream divides the text data into a plurality of units according to a fixed time length; Generate sentiment scores for the burst words using a sentiment dictionary, detect sentiment changes in the text data using Gaussian distribution, and calculate the sentiment scores corresponding to the burst words per 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.

4. A method for generating live e-commerce content based on hotspot fusion according to claim 2, characterized in that: Acquiring dynamic event change features based on the text sentiment features includes: Taking the text emotion feature as the original live broadcast information, extracting the text emotion feature using a Chinese text processing library, and calculating the emotion coefficient of the original live broadcast information; Determine the forwarding number of the current live content based on the sentiment coefficient, calculate the forwarding frequency characteristics of the original live information in the current state, and calculate the popularity V of the current topic based on the forwarding 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 as the fan attribute feature within T time periods, and the fan attribute feature is monitored in real time using a hierarchical clustering algorithm to obtain dynamic event change features.

5. A method for generating live e-commerce content based on hotspot fusion according to claim 4, characterized in that: The dynamic event change features are extracted using a training model to obtain sentence-level text feature vectors, 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 sentence-level text feature vector is obtained based on prior knowledge.

6. A method for generating live e-commerce content based on hotspot fusion according to claim 5, 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 the 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.

7. A method for generating live e-commerce content based on hotspot fusion according to claim 6, characterized in that: The document-level feature vector is constructed by an encoder using 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.

8. A system for generating live e-commerce content based on hotspot fusion according to any one of claims 1 to 7, characterized in that: The system comprises: The data collection module 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 to identify the most popular product categories, brands, or events and assess their popularity level. 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 their interests; The content generation module generates live broadcast scripts, selects products to be displayed, and designs interactive sessions based on user interests and hot events; Feedback adjustment module monitors live broadcast effects in real time and dynamically adjusts live broadcast content to optimize user experience.

9. 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 7 is implemented.

Images