AIGC content generation method and system based on content big data monitoring

Abstract

The invention provides an AIGC content generation method and system based on content big data monitoring, and the method comprises the steps: constructing a target text content set from the content behavior big data of a target user, carrying out the text semantic coding representation, and generating a first text semantic vector sequence; then, text content unit sequences and model feature tags corresponding to the text content unit sequences are generated through convergence, and target AIGC attention semantic vector sequences corresponding to the text content unit sequences are generated through an AIGC semantic attention mechanism; and finally, integrating the target AIGC attention semantic vector sequences to generate a second text semantic vector sequence, carrying out AIGC content generation, and outputting an AIGC content generation result, so that accurate content generation can be realized through an AIGC technology by effectively utilizing user content behavior big data, the accuracy and pertinence of content generation are improved, the personalized requirements of users are met, and the user experience is improved. And user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AIGC content generation method and system based on content big data monitoring. Background Art

[0002] In the current era of rapid development of informatization and digitalization, the amount of content data on the Internet has exploded. Faced with such a huge amount of content data, how to effectively mine, analyze and utilize this data has become an urgent problem to be solved. Especially in the field of content creation and distribution, how to generate targeted and attractive content based on user needs and interests is the key to improving user experience and satisfaction.

[0003] Traditional content generation methods mainly rely on manual creation and editing, which is not only inefficient but also difficult to meet large-scale, personalized content needs. With the continuous development of artificial intelligence technology, AIGC (Artificial Intelligence Generated Content) technology has gradually emerged. AIGC technology can use machine learning and natural language processing technologies to automatically generate various types of content, such as text, images, audio, etc., thereby greatly improving the efficiency and scale of content generation.

[0004] However, the existing AIGC technology still has some shortcomings in practical applications. On the one hand, due to the diversity and complexity of content data, how to accurately capture and represent the semantic information of text content is a major challenge facing AIGC technology. On the other hand, how to personalize and optimize the generated content according to user needs and interests is also a problem that existing technologies need to solve. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides an AIGC content generation method based on content big data monitoring, the method comprising:

[0006] Obtain a target text content set formed by multiple text content data from any target user's content behavior big data;

[0007] Performing text semantic encoding representation on the target text content set to generate a first text semantic vector sequence;

[0008] According to the first text semantic vector sequence, the text content data in the target text content set are aggregated to generate a plurality of text content unit sequences and a model feature label corresponding to each of the text content unit sequences, each of the text content unit sequences includes a plurality of the text content data, and the model feature label is a feature label corresponding to an AIGC content mining unit that performs AIGC semantic attention;

[0009] For each of the text content unit sequences, an AIGC content mining unit indicated by a model feature label corresponding to the text content unit sequence performs AIGC semantic attention on the first text semantic vector sequence to generate a target AIGC attention semantic vector sequence corresponding to the text content unit sequence;

[0010] Integrate the target AIGC attention semantic vector sequences corresponding to each of the text content unit sequences to generate a second text semantic vector sequence corresponding to the target text content set;

[0011] Perform AIGC content generation on the second text semantic vector sequence and output the AIGC content generation result.

[0012] On the other hand, an embodiment of the present invention also provides an AIGC content generation system based on content big data monitoring, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0013] Based on the above aspects, the embodiment of the present application constructs a target text content set from the target user's content behavior big data, and performs text semantic encoding representation to generate a first text semantic vector sequence. Subsequently, a text content unit sequence and its corresponding model feature label are generated by convergence, and the target AIGC attention semantic vector sequence corresponding to each text content unit sequence is generated using the AIGC semantic attention mechanism. Finally, these target AIGC attention semantic vector sequences are integrated to generate a second text semantic vector sequence, and AIGC content generation is performed, and the AIGC content generation result is output, thereby effectively utilizing user content behavior big data, realizing accurate content generation through AIGC technology, improving the accuracy and pertinence of content generation, meeting user personalized needs, and improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of the execution flow of the AIGC content generation method based on content big data monitoring provided in an embodiment of the present invention.

[0015] Figure 2It is a schematic diagram of the hardware architecture of the AIGC content generation system based on content big data monitoring provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 It is a flow chart of an AIGC content generation method based on content big data monitoring provided by an embodiment of the present invention. The AIGC content generation method based on content big data monitoring is introduced in detail below.

[0017] Step S110, obtaining a target text content set formed by multiple text content data from the content behavior big data of any target user, performing text semantic coding representation on the target text content set, and generating a first text semantic vector sequence.

[0018] In this embodiment, a large social media platform is taken as an example. This social media platform has a large number of users. Each user performs various interactions on the social media platform, such as posting updates, commenting, liking, etc. These interactive behaviors generate a large amount of content behavior big data.

[0019] The server first identifies a target user. For example, a user named "Xiao Ming" is selected, who is very active on the platform and often posts content about travel, food, and technology.

[0020] Next, we can filter out text content data related to Xiao Ming from the stored content behavior big data. These text content data include his travel notes (such as "I went to the beautiful Bali, where the beaches and sunsets are intoxicating"), food reviews ("The pasta in this restaurant is super delicious, with rich sauce"), and discussions about technology products ("The new smartphone has a very powerful camera function"), etc. These text content data together constitute the target text content collection.

[0021] Then, the server uses a pre-trained word vector model (such as Word2Vec) and a deep learning framework (such as TensorFlow) to perform text semantic encoding. For each word in the target text content set, the word vector model maps it to a low-dimensional vector space. For example, for the word "Bali", it may be mapped to a 100-dimensional vector [0.1, 0.2, -0.3, ...], which can represent the semantic information of "Bali" to a certain extent.

[0022] By sequentially semantically encoding each text content data in the target text content set, the server combines these encoded vectors in the order of the text to form a first text semantic vector sequence. Assuming that Xiao Ming's travel notes, food reviews, and technology discussions are encoded to obtain vector sequences V1, V2, and V3 respectively, then the first text semantic vector sequence is [V1, V2, V3].

[0023] Step S120, based on the first text semantic vector sequence, the text content data in the target text content set are aggregated to generate multiple text content unit sequences and model feature labels corresponding to each of the text content unit sequences, each of the text content unit sequences includes multiple text content data, and the model feature labels are feature labels corresponding to the AIGC content mining units that perform AIGC semantic attention.

[0024] In this embodiment, the server uses a neural network-based attention mechanism to extract semantic elements. For each vector in the first text semantic vector sequence, it is processed by a multi-layer neural network to identify the key semantic elements. For example, in Xiao Ming's travel notes, the vector may contain semantic elements such as location (Bali), scenery (beach, sunset); in food reviews, it contains semantic elements such as food (pasta) and taste (delicious, rich); in technology discussions, it contains semantic elements such as products (smartphones) and functions (taking pictures). After processing, the third text semantic vector sequence is obtained.

[0025] Next, taking the travel notes as an example, the server regards each text semantic vector as a semantic unit and compares it with the predefined semantic concept template library (including various travel-related concepts, such as attractions, activities, etc.). For example, for the vector "beach", it is found through comparison that it belongs to the semantic concept of "attractions", thus obtaining the first target text semantic vector with preliminary semantic element tags.

[0026] Then, based on the similarity and relevance of the marked preliminary semantic elements, the server analyzes the preliminary semantic relationship between the first target text semantic vectors. For example, "beach" and "sunset" may be associated in location, which is a parallel relationship, and generates a second target text semantic vector with preliminary semantic relationship marks.

[0027] Next, each second target text semantic vector is used as a node to construct a semantic relationship network according to the semantic relationship. If there is a relationship between "beach" and "sunset", an edge is established between their corresponding nodes.

[0028] According to the node connectivity and edge weight distribution, the role of each node in the semantic relationship network is analyzed. For example, "Bali" may be a semantic core node because many other semantic elements (such as beach, sunset) are related to it; while some descriptive words (such as "beautiful") may be edge nodes. Then, based on the node role classification results, the preliminary contribution of each text content data to the global semantics is evaluated. For example, the initial contribution of the semantic core node "Bali" is high because it is the core location of the entire travel experience. Next, the semantic path from each node to other nodes is analyzed. For example, the semantic path from "beach" to "Bali" calculates the scope and degree of influence of "beach" on the semantics of "Bali" through this path, and adjusts the preliminary semantic contribution value accordingly to obtain the first semantic contribution value.

[0029] Finally, the adaptation state parameters of each text content data in the corresponding local context are analyzed. For example, in the local context describing food, the word "delicious" has a high degree of adaptation to the overall context. According to this adaptation state parameter, the first semantic contribution value is further adjusted to obtain the second semantic contribution value, that is, the semantic contribution value.

[0030] Thus, the semantic contribution fusion is performed on each semantic contribution value. For example, a threshold is set, and the text content data with a semantic contribution value higher than the threshold is determined as the attention text content data. Assume that in Xiao Ming's content, the semantic contribution values ​​of "Bali", "delicious", "smartphone", etc. are relatively high and are determined as attention text content data.

[0031] In this way, based on these attention text content data, the server generates multiple text content unit sequences. For example, one text content unit sequence can be ["I went to the beautiful Bali, where the beaches and sunsets were intoxicating"], and another can be ["The pasta in this restaurant is super delicious and the sauce is rich"], etc.

[0032] Next, for each text content unit sequence, the server determines the corresponding unit semantic vector sequence in the third text semantic vector sequence. For example, for a text content unit sequence about Bali, find its corresponding vector part in the third text semantic vector sequence.

[0033] The target AIGC content mining distribution is determined based on the unit semantic vector sequence. For example, if this sequence mainly involves tourism content, then the confidence of the AIGC content mining unit related to tourism being called may be higher.

[0034] According to the target AIGC content mining distribution, the model feature label corresponding to the text content unit sequence is determined. For example, for a tourism-related text content unit sequence, its model feature label may point to the tourism-related AIGC content mining unit.

[0035] Step S130, for each of the text content unit sequences, the AIGC content mining unit indicates the model feature label corresponding to the text content unit sequence, performs AIGC semantic attention on the first text semantic vector sequence, and generates a target AIGC attention semantic vector sequence corresponding to the text content unit sequence.

[0036] In this embodiment, the server first performs non-AIGC semantic attention on the first text semantic vector sequence. This may involve the extraction and processing of some general semantic information, and does not involve AIGC-specific mining. For example, for the vectors in the first text semantic vector sequence, some basic semantic compression and filtering operations are performed to remove some noise information and generate a fourth text semantic vector sequence.

[0037] For each text content unit sequence, the fourth text semantic vector sequence is loaded into the AIGC content mining unit indicated by the model feature label corresponding to the text content unit sequence. For example, for a tourism-related text content unit sequence, the corresponding model feature label points to the tourism AIGC content mining unit, and the fourth text semantic vector sequence is loaded into this unit to generate the corresponding first AIGC focus semantic vector sequence. This first AIGC focus semantic vector sequence may be a vector sequence that has been more deeply mined and focused on tourism-related semantics, such as paying more attention to semantic information such as the characteristics of tourist attractions and travel experiences.

[0038] If the first AIGC attention semantic vector sequence is a basic semantic vector sequence for generating a target AIGC attention semantic vector sequence, the server performs non-AIGC semantic attention on the first AIGC attention semantic vector sequence corresponding to each text content unit sequence. This may be a further optimization and adjustment operation, such as removing some less relevant semantic information again to generate a fifth text semantic vector sequence.

[0039] For each text content unit sequence, the fifth text semantic vector sequence is loaded into the AIGC content mining unit indicated by the corresponding model feature label to generate the corresponding second AIGC attention semantic vector sequence. This second AIGC attention semantic vector sequence focuses more accurately on AIGC-related semantic content. For example, in terms of tourism, it may more deeply explore the semantic information such as niche attractions and unique tourism culture of tourist destinations. This second AIGC attention semantic vector sequence is the target AIGC attention semantic vector sequence.

[0040] Step S140, integrating the target AIGC focus semantic vector sequences corresponding to each of the text content unit sequences, and generating a second text semantic vector sequence corresponding to the target text content set.

[0041] In this embodiment, the server integrates the target AIGC attention semantic vector sequences corresponding to each text content unit sequence. For example, there are tourism-related target AIGC attention semantic vector sequences, food-related target AIGC attention semantic vector sequences, and technology-related target AIGC attention semantic vector sequences. The server combines them together according to certain rules (such as sequential connection or weighted summation, etc.) to form a second text semantic vector sequence corresponding to the target text content set. Assuming that the tourism-related target AIGC attention semantic vector sequence is Vt1, the food-related is Vt2, and the technology-related is Vt3, the second text semantic vector sequence Vt=0.4*Vt1+0.3*Vt2+0.3*Vt3 is obtained by weighted summation (such as weights of 0.4, 0.3, and 0.3 respectively).

[0042] Step S150, performing AIGC content generation on the second text semantic vector sequence, and outputting the AIGC content generation result.

[0043] In this embodiment, the server uses a pre-trained AIGC content generation model (such as a generation model based on the Transformer architecture) to process the second text semantic vector sequence. This model has been pre-trained with a large amount of text data and can generate relevant text content based on the input semantic vector sequence. For example, for a second text semantic vector sequence containing tourism, food, and technology semantic information, the model may generate a text content that combines these elements, such as "With the help of technology, we can better explore tourist destinations like Bali, taste the local super delicious specialties, and record these beautiful moments with smartphones." As a result, the server outputs the generated text content as an AIGC content generation result. This result can be used in various application scenarios, such as providing users with personalized content recommendations, automatically generating travel guides, or descriptions of the use scenarios of technology products.

[0044] Based on the above steps, the embodiment of the present application constructs a target text content set from the target user's content behavior big data, and performs text semantic encoding representation to generate a first text semantic vector sequence. Subsequently, a text content unit sequence and its corresponding model feature label are generated by convergence, and the target AIGC attention semantic vector sequence corresponding to each text content unit sequence is generated using the AIGC semantic attention mechanism. Finally, these target AIGC attention semantic vector sequences are integrated to generate a second text semantic vector sequence, and AIGC content generation is performed, and the AIGC content generation result is output, thereby effectively utilizing user content behavior big data, realizing accurate content generation through AIGC technology, improving the accuracy and pertinence of content generation, meeting user personalized needs, and improving user experience.

[0045] In a possible implementation, step S120 includes:

[0046] Step S121, extracting attention semantic elements from the first text semantic vector sequence to generate a third text semantic vector sequence.

[0047] Step S122, estimating the semantic contribution value corresponding to each of the text content data according to the third text semantic vector sequence.

[0048] Step S123, determining at least one attention text content data by performing semantic contribution fusion on each of the semantic contribution value values, wherein the attention text content data is text content data whose corresponding semantic contribution fusion value meets the set conditions.

[0049] Step S124: generating the plurality of text content unit sequences based on the attention text content data.

[0050] In a possible implementation, step S122 includes:

[0051] Step S1221, performing semantic element analysis on each text semantic vector in the third text semantic vector sequence, treating each text semantic vector as a semantic unit, identifying the semantic elements contained in the text semantic vector by comparing with a predefined semantic concept template library, and obtaining a first target text semantic vector with a preliminary semantic element mark corresponding to each text content data, wherein the semantic element is a basic unit for constructing semantic understanding,

[0052] Step S1222, based on the similarity and correlation of the marked preliminary semantic elements, analyze the preliminary semantic relationship between the first target text semantic vectors, and generate a second target text semantic vector corresponding to each text content data marked with the preliminary semantic relationship, wherein the preliminary semantic relationship includes inclusion relationship, parallel relationship, and causal relationship.

[0053] Step S1223, taking each of the second target text semantic vectors as a node, constructing a semantic relationship network according to the semantic relationship between the second target text semantic vectors, wherein if there is a semantic relationship between two second target text semantic vectors, an edge is established between the corresponding nodes.

[0054] Step S1224, analyzing the role of each node in the semantic relationship network according to the node connectivity and edge weight distribution of each node, and generating a node role classification result, wherein the role is a semantic core node, an edge node or an intermediate transition node.

[0055] Step S1225 , evaluating the preliminary contribution of each text content data to the global semantics according to the node role classification result, and generating a preliminary semantic contribution value corresponding to each text content data.

[0056] Step S1226, analyze the semantic path from each node to other nodes, calculate the scope and degree of influence of each node on the semantics of other nodes through the semantic path, and adjust the preliminary semantic contribution value corresponding to each text content data according to the scope and degree of influence of each node on the semantics of other nodes through the semantic path, and generate the first semantic contribution value corresponding to each text content data after the semantic path influence adjustment.

[0057] Step S1227, analyze the adaptation state parameters of each of the text content data in the corresponding local context, and adjust the first semantic contribution value corresponding to each of the text content data after the semantic path influence adjustment according to the adaptation state parameters, generate the second semantic contribution value corresponding to each of the text content data after the local context adaptation adjustment, and output the second semantic contribution value as the semantic contribution numerical value.

[0058] In this embodiment, the first text semantic vector sequence of the target user "Xiao Ming" has been obtained. This sequence contains vector information of his text content data in the areas of travel, food, and technology after being semantically encoded.

[0059] The server uses an attention network to extract semantic elements. This attention network has multiple hidden layers, and the number of neurons in each hidden layer is carefully designed to adapt to the task of semantic element extraction.

[0060] For each text semantic vector in the first text semantic vector sequence, when entering the attention mechanism network, the network will identify the key semantic elements contained therein according to the numerical distribution in the vector and the predefined semantic mapping rules. For example, in the vector sequence corresponding to Xiao Ming's text about travel, "I went to the beautiful Bali, where the beaches and sunsets were intoxicating", the network may identify the semantic elements related to the place name "Bali", including its geographical location, tourism popularity and other related semantic information; for the semantic element "beach", it may identify its natural attributes (such as beach type, surrounding marine environment, etc.); for "sunset", it may identify its visual effects, time of appearance and other semantic elements.

[0061] For the food-related text "This restaurant's pasta is super delicious and the sauce is rich", the attention mechanism network will identify semantic elements such as the food type and production style of "pasta" in the vector, as well as semantic elements such as "delicious" and "rich" that describe the taste.

[0062] In the technology-related text "The new smartphone's camera function is so powerful", semantic elements related to the brand and model of "smartphone" will be identified (if reflected in the vector), as well as semantic elements such as the technical parameters of the "camera function" and imaging effects.

[0063] After the attention network processes all the vectors in the first text semantic vector sequence, the server obtains the third text semantic vector sequence. The vectors in this sequence not only contain the original semantic information, but also highlight the key semantic elements, providing a more targeted basis for the subsequent semantic contribution numerical estimation.

[0064] Next, taking tourism text as an example, each text semantic vector related to tourism in the third text semantic vector sequence is regarded as a semantic unit. These semantic units are compared with the predefined semantic concept template library. For example, the predefined semantic concept template library contains rich concepts such as tourist attractions, tourist activities, and tourist experiences around the world. For the semantic unit "Bali", by comparing it with the template library, it is determined that it is a famous tourist attraction concept, and detailed information such as its region and tourist features may be marked, thereby obtaining the first target text semantic vector marked with preliminary semantic elements.

[0065] For food text, such as the semantic unit "pasta", when compared with the semantic concept template library, it is determined to be a classic Italian food, and its ingredients, cooking methods and other related preliminary semantic elements are marked to form the first target text semantic vector.

[0066] In the scientific and technological text, after comparing the semantic unit "smartphone" with the template library, its preliminary semantic elements such as device type and functional category are marked to obtain the corresponding first target text semantic vector.

[0067] According to the similarity and relevance of the marked preliminary semantic elements, the server begins to analyze the preliminary semantic relationship between the semantic vectors of the first target text. In the tourism text, "Bali" has a geographical inclusion relationship with "beach" and "sunset" because the beach and sunset are the characteristic elements of Bali; and there is a visual parallel relationship between "beach" and "sunset" because they are different scenery that attract tourists to Bali. For the food text, "pasta" has a relationship of describing objects and describing attributes with "delicious" and "rich", and "delicious" and "rich" have a parallel relationship of describing taste. In the science and technology text, "smartphone" and "photographing function" have an inclusion relationship of equipment and function. Based on these analyses, the server generates a second target text semantic vector with preliminary semantic relationship marks for each text content data.

[0068] On this basis, each second target text semantic vector is used as a node, and a semantic relationship network is constructed according to the semantic relationship between them. In the tourism semantic relationship network, "Bali" is an important node, and there are connecting edges with the "beach" and "sunset" nodes. The weight of the edge is determined according to the closeness of the semantic relationship between them. For example, the edge weight between "Bali" and "beach" may be high because the beach is an important part of Bali. In the food semantic relationship network, the "pasta" node is connected to the "delicious" and "rich" nodes, and the different connecting edges reflect the degree of semantic association between them. For the technology semantic relationship network, the "smartphone" node is connected to the "photographing function" node, and the weight of the edge reflects the importance of the function to the device.

[0069] Then, based on the connectivity of each node and the weight distribution of the edges, the server analyzes the role of each node in the semantic relationship network. In the tourism semantic relationship network, "Bali" is classified as a semantic core node because it has high-weight connections with multiple other nodes; modifier nodes such as "beautiful" have low connectivity and small weight and are considered edge nodes; and nodes such as "beach" that connect core nodes and other related nodes are considered intermediate transition nodes. In the food semantic relationship network, "pasta" is a core node, and modifiers such as "super" are edge nodes. In the technology semantic relationship network, "smartphone" is a core node.

[0070] Then, based on the node role classification results, the server evaluates the initial contribution of each text content data to the global semantics. For tourism text, the initial contribution of the semantic core node "Bali" is higher because it is the core of the entire travel experience description; the contribution of the edge node "beautiful" is lower. In the food text, the initial contribution of "pasta" is higher, and the contribution of modifiers is lower. In the technology text, the initial contribution of "smartphone" is higher.

[0071] Then, the server analyzes the semantic path from each node to other nodes. Taking tourism text as an example, the semantic path from "beach" to "Bali" has an enhancing effect on the semantics of "Bali" through this path, because the beach is one of the characteristics of Bali. According to the scope and degree of this influence, the preliminary semantic contribution value of "beach" is adjusted to obtain the first semantic contribution value. For food text, the influence of "delicious" on the overall semantics through the semantic path with "pasta" will adjust its first semantic contribution value. In science and technology text, "camera function" will also adjust its first semantic contribution value through the semantic path with "smartphone".

[0072] Then, the adaptation state parameters of each text content data in the corresponding local context are analyzed. In the travel text, "beach" has a high degree of adaptation in the local context of describing Bali tourism. According to this adaptation state parameter, the first semantic contribution value is further adjusted to obtain the second semantic contribution value, which is the semantic contribution value corresponding to each text content data. In the food text, "delicious" has a high degree of adaptation in the local context of describing pasta, and the final semantic contribution value is obtained after adjustment. For the technology text, "camera function" has a high degree of adaptation in the local context of describing smartphones, and the corresponding semantic contribution value is also obtained.

[0073] Based on the above, the server sets a range of semantic contribution values ​​for each text content data, which is determined according to the specific task and data characteristics. For example, in the range of 0-1, a threshold is set to 0.5.

[0074] The server comprehensively considers the semantic contribution values ​​of all text content data. For example, if the semantic contribution value of "Bali" in the travel text is 0.8, "beach" is 0.6, and "sunset" is 0.7; for the food text, "pasta" is 0.7, and "delicious" is 0.6; for the technology text, "smartphone" is 0.8, and "camera function" is 0.7, etc.

[0075] When these semantic contribution values ​​are merged, weighted summation or simple summation may be used. Assuming that simple summation is used, the semantic contribution values ​​of each text content data are added together and then compared with the set threshold.

[0076] After fusion, if the semantic contribution fusion value of a certain text content data is greater than or equal to the set threshold, then this text content data is determined as attention text content data. In the above example, the semantic contribution fusion values ​​of "Bali", "beach", "sunset", "pasta", "delicious", "smartphone", "photo function", etc. are all greater than 0.5, so they are all determined as attention text content data.

[0077] Next, the server groups the attention text content data to generate multiple text content unit sequences. For example, the tourism-related attention text content data "Bali", "Beach", and "Sunset" are grouped into a text content unit sequence, which specifically describes Xiao Ming's travel experience.

[0078] The food-related words "pasta" and "delicious" are combined into another text content unit sequence to represent food-related content.

[0079] Then, the technology-related "smartphone" and "camera function" are combined into a text content unit sequence to represent technology-related content. In this way, multiple text content unit sequences are generated, each of which focuses on a specific type of content, providing a clear semantic unit for subsequent operations.

[0080] In a possible implementation, step S120 further includes:

[0081] Step S125: for each of the text content unit sequences, determine a corresponding unit semantic vector sequence in the third text semantic vector sequence.

[0082] Step S126, determining a target AIGC content mining distribution according to the unit semantic vector sequence, wherein the target AIGC content mining distribution indicates the confidence level of each AIGC content mining unit being called.

[0083] Step S127, determining the model feature label corresponding to the text content unit sequence according to the target AIGC content mining distribution.

[0084] In this embodiment, the server has generated multiple text content unit sequences, such as tourism-related text content unit sequences including "Bali", "Beach", and "Sunset", food-related text content unit sequences including "Pasta" and "Delicious", and technology-related text content unit sequences including "Smartphone" and "Photographing Function". At the same time, the server also has a third text semantic vector sequence, which is obtained by extracting attention semantic elements from the first text semantic vector sequence and contains rich semantic element information.

[0085] For the unit sequence of tourism-related text content, the server searches and filters in the third text semantic vector sequence. Since the vectors in the third text semantic vector sequence correspond one-to-one to the original text content data, the server selects the corresponding vectors from the third text semantic vector sequence according to the position and semantic coding relationship of the tourism-related text content data ("Bali", "Beach", "Sunset") in the original text. These vectors together constitute the unit semantic vector sequence corresponding to the tourism-related text content unit sequence in the third text semantic vector sequence. For example, the vector corresponding to "Bali" may contain semantic element information such as its geographical location and tourism popularity, the vector corresponding to "Beach" contains information such as the natural attributes of the beach, and the vector corresponding to "Sunset" contains information such as visual effects. Combining these vectors in order will result in a unit semantic vector sequence related to tourism.

[0086] Similarly, for the unit sequence of text content related to food, the server finds the vectors corresponding to "pasta" and "delicious" in the third text semantic vector sequence. These vectors contain semantic element information such as food type and taste description, and are combined to form a unit semantic vector sequence related to food.

[0087] For technology-related text content unit sequences, the server determines the vectors corresponding to "smartphone" and "camera function" from the third text semantic vector sequence. These vectors contain semantic element information such as device type and functional characteristics, thereby obtaining a technology-related unit semantic vector sequence.

[0088] Next, for the tourism-related unit semantic vector sequence, the server first analyzes the semantic features therein. For example, the information contained therein, such as the tourist destination (Bali), attractions (beach, sunset), etc., indicates that this is a semantic vector sequence about tourism geography and tourism experience. The server determines the possibility of different AIGC content mining units being called according to predefined rules and algorithms, which are based on a large amount of tourism-related data and AIGC content mining experience.

[0089] Assume that there are different AIGC content mining units such as tourist attraction recommendation mining unit, tourist culture mining unit, and tourist itinerary planning mining unit. Since the unit semantic vector sequence highlights the tourist destination and tourist attraction information, the confidence of the tourist attraction recommendation mining unit may be higher, for example, it is set to 0.6; the confidence of the tourist culture mining unit may be 0.3, because the current vector sequence reflects less cultural elements; the confidence of the tourist itinerary planning mining unit may be 0.1, because there are not many semantic elements related to itinerary planning.

[0090] For the unit semantic vector sequence related to food, it contains semantic features such as food type (pasta) and taste (delicious). The server uses AIGC content mining units such as food recipe mining unit, food taste analysis mining unit, and food restaurant recommendation mining unit according to predefined food-related rules. Since the vector sequence mainly contains descriptions of the food itself and its taste, the confidence of the food taste analysis mining unit may be higher, such as 0.5; the confidence of the food recipe mining unit may be 0.3; the confidence of the food restaurant recommendation mining unit may be 0.2, because it does not directly involve restaurant information.

[0091] In the sequence of unit semantic vectors related to technology, semantic features such as devices (smartphones) and functions (photographing functions) are included. Assume that there are AIGC content mining units such as technology product function analysis mining unit, technology product evaluation mining unit, and technology product usage scenario mining unit. Since the products and functions are highlighted, the confidence of the technology product function analysis mining unit may be higher, such as 0.7; the confidence of the technology product evaluation mining unit may be 0.2; and the confidence of the technology product usage scenario mining unit may be 0.1.

[0092] Then, the server combines the confidence of each AIGC content mining unit to form the target AIGC content mining distribution. For example, for tourism-related, it is [tourist attraction recommendation mining unit: 0.6, tourism culture mining unit: 0.3, travel itinerary planning mining unit: 0.1]; for food-related, it is [food taste analysis mining unit: 0.5, food recipe mining unit: 0.3, food restaurant recommendation mining unit: 0.2]; for technology-related, it is [technology product function analysis mining unit: 0.7, technology product evaluation mining unit: 0.2, technology product usage scenario mining unit: 0.1].

[0093] Based on the above, for the tourism-related text content unit sequence, since the confidence of the tourist attraction recommendation mining unit in its target AIGC content mining distribution is the highest, the server determines that the model feature label corresponding to the text content unit sequence is the label related to the tourist attraction recommendation mining unit. This label contains identification information related to the tourist attraction recommendation mining unit, such as a specific code or name, so that subsequent operations can accurately identify and call the relevant mining unit.

[0094] For the food-related text content unit sequence, because the food taste analysis mining unit has the highest confidence in the target AIGC content mining distribution, its corresponding model feature label is the label related to the food taste analysis mining unit, and this label also carries the identification information of the food taste analysis mining unit.

[0095] In the sequence of technology-related text content units, since the confidence of the technology product function analysis mining unit is the highest, its corresponding model feature label is the label related to the technology product function analysis mining unit, which contains the identification information of the technology product function analysis mining unit. In this way, each text content unit sequence has a corresponding model feature label, which will guide the subsequent semantic attention and content generation operations based on AIGC.

[0096] In a possible implementation, step S130 includes:

[0097] Step S131: Perform non-AIGC semantic attention on the first text semantic vector sequence to generate a fourth text semantic vector sequence.

[0098] Step S132, for each of the text content unit sequences, load the fourth text semantic vector sequence into the AIGC content mining unit indicated by the model feature label corresponding to the text content unit sequence, and generate a corresponding first AIGC focus semantic vector sequence, wherein the first AIGC focus semantic vector sequence is the target AIGC focus semantic vector sequence, or the first AIGC focus semantic vector sequence is a basic semantic vector sequence used to generate the target AIGC focus semantic vector sequence.

[0099] In this embodiment, the server has determined each text content unit sequence and their corresponding model feature labels, and now needs to process the first text semantic vector sequence. This first text semantic vector sequence contains the vector information of the target user (such as Xiao Ming) in various aspects of travel, food, technology, etc. after semantic encoding.

[0100] Then, the server uses a general semantic compression and information filtering algorithm to perform non-AIGC semantic attention. This algorithm mainly performs preliminary feature extraction and information simplification on the vectors in the first text semantic vector sequence, and does not involve AIGC-specific semantic mining.

[0101] For example, for the tourism-related vectors, the algorithm may abstract the geographic information of the vectors describing the tourist destinations, simplifying some overly detailed geographic location descriptions (such as the longitude and latitude of a specific beach in Bali) into more general tourist destination features, such as only retaining information about tropical beaches. For food-related vectors, some very detailed description vectors about the specific taste of food may be simplified, retaining the general taste category of the food (such as simplifying a vector that describes the taste of a unique sauce on pasta in detail into a vector related to the rich taste category). For technology-related vectors, some complex technical parameter vectors of smartphones may be simplified, retaining only the main functional feature vectors (such as simplifying the numerous optical parameter vectors in the camera function into vectors representing the main functions such as high pixels and night scene mode).

[0102] In this way, the server processes all vectors in the first text semantic vector sequence and generates a fourth text semantic vector sequence. Compared with the first text semantic vector sequence, this sequence is more concise and universal while retaining basic semantic information, preparing for subsequent loading into the AIGC content mining unit.

[0103] Next, for the tourism-related text content unit sequence, the corresponding model feature label indicates a tourist attraction recommendation mining unit. The server loads the fourth text semantic vector sequence into the tourist attraction recommendation mining unit.

[0104] The tourist attraction recommendation mining unit has a specific neural network structure and algorithm. This unit first analyzes the loaded fourth text semantic vector sequence and identifies the semantic information related to the tourist attraction recommendation. For example, the overall characteristics of the tourist destination (Bali), the attractiveness of tourist attractions (beaches, sunsets), etc. are found in the simplified vector.

[0105] Then, based on this semantic information, the mining unit will further perform semantic transformation and feature enhancement on the vector. It may weight the information in the vector according to predefined tourist attraction recommendation rules (such as popularity, tourist evaluation and other factors). For example, the vector features of Bali, a popular tourist destination, are enhanced to increase its influence in the entire vector sequence, because popular destinations are more important in attraction recommendations.

[0106] After being processed by the tourist attraction recommendation mining unit, a first AIGC attention semantic vector sequence related to tourism is generated. If in this scenario, this first AIGC attention semantic vector sequence is the target AIGC attention semantic vector sequence, then it can be used for subsequent operations (such as integration into the final second text semantic vector sequence for AIGC content generation). If it is the basic semantic vector sequence used to generate the target AIGC attention semantic vector sequence, further processing is required.

[0107] In addition, the model feature label corresponding to the text content unit sequence related to food points to the food taste analysis mining unit. The server loads the fourth text semantic vector sequence into the mining unit.

[0108] The food taste analysis mining unit will conduct an in-depth analysis of the food-related semantic information in the vector sequence. For example, for the vector of pasta and its strong taste, the mining unit will use a predefined food taste analysis model, which contains knowledge such as the association between different tastes and the relationship between ingredients. It will further refine the semantic information about the taste, and may analyze the association between the taste "strong" and other related tastes (such as mellow, rich, etc.), and incorporate the analysis results into the vector.

[0109] At the same time, the mining unit will also more accurately locate the taste characteristics of food such as pasta based on the existing food knowledge database. For example, if the database indicates that pasta with a certain sauce will produce a specific taste combination, the mining unit will reflect this association information in the vector. After these processes, the first AIGC attention semantic vector sequence related to food is generated. Similarly, this sequence may be a target AIGC attention semantic vector sequence or a basic semantic vector sequence, depending on the specific processing flow.

[0110] In addition, the model feature label corresponding to the technology-related text content unit sequence is a technology product function analysis mining unit. The server loads the fourth text semantic vector sequence into this mining unit.

[0111] The technology product function analysis mining unit first identifies the semantic information related to the smartphone camera function in the vector sequence. Then, it will analyze the technology product function according to the predefined technology product function analysis rules, which may be based on a large amount of technology product evaluation data and function analysis experience. For example, based on the main functional features of the camera function such as high pixel and night scene mode, the mining unit will analyze the importance of these functions in the entire smartphone function system and adjust the vector accordingly. If high pixel is considered to be a very critical function in the current function analysis, the vector part related to high pixel will be enhanced.

[0112] Through such processing, a first AIGC attention semantic vector sequence related to technology is generated, which may also be a target AIGC attention semantic vector sequence or a basic semantic vector sequence used to generate a target AIGC attention semantic vector sequence.

[0113] In a possible implementation manner, when the first AIGC attention semantic vector sequence is a basic semantic vector sequence used to generate the target AIGC attention semantic vector sequence, step S130 further includes:

[0114] Step S133: Perform non-AIGC semantic attention on the first AIGC attention semantic vector sequences corresponding to each of the text content unit sequences to generate a fifth text semantic vector sequence.

[0115] Step S134, for each of the text content unit sequences, load the fifth text semantic vector sequence into the AIGC content mining unit indicated by the model feature label corresponding to the text content unit sequence, and generate a corresponding second AIGC focus semantic vector sequence, wherein the second AIGC focus semantic vector sequence is the target AIGC focus semantic vector sequence, or the second AIGC focus semantic vector sequence is a basic semantic vector sequence used to generate the target AIGC focus semantic vector sequence.

[0116] In this embodiment, the server has generated first AIGC attention semantic vector sequences for each text content unit sequence (such as travel, food, and technology-related sequences), and these sequences are basic semantic vector sequences for generating target AIGC attention semantic vector sequences. Now, these first AIGC attention semantic vector sequences need to be further processed.

[0117] In detail, for the first AIGC attention semantic vector sequence related to tourism, it contains semantic vectors about tourist destinations (Bali), attractions (beach, sunset), etc. after preliminary processing by the tourist attraction recommendation mining unit. The non-AIGC semantic attention operation performed by the server is intended to further optimize these vectors.

[0118] For example, the server may adjust the information related to the tourist season in the vector. If there is some general information about the peak tourist season in Bali in the first AIGC attention semantic vector sequence, the server will refine it based on more historical tourism data, such as determining which months in the peak season have the most tourists and the most suitable weather, and integrate this information into the vector in a simplified way, and update the original vector about the tourist season.

[0119] At the same time, for the description vectors of tourist attractions, the server may remove some repeated or redundant information. For example, if there is some overlapping information in the description of the beach in the previous processing (such as the description of the sand quality of the beach and the scenery around the beach are partially repeated), the server will optimize these vectors and retain the most core description information. After such processing, the fifth text semantic vector sequence related to tourism is obtained.

[0120] For another example, in the first AIGC attention semantic vector sequence related to food, the semantic vectors about pasta and its taste processed by the food taste analysis and mining unit are included. The non-AIGC semantic attention operation of the server will further refine these vectors.

[0121] For example, if there is some vague or inaccurate information about the relationship between ingredients and taste in the pasta flavor vector, the server will refer to the food knowledge graph to make corrections. If the original vector does not clearly describe the relationship between a certain ingredient and the "rich" taste, the server will readjust the relevant information in the vector based on the general role of the ingredient in pasta making in the food knowledge graph.

[0122] In addition, the server will simplify and optimize the cultural background information of food (if it is involved in the first AIGC focus semantic vector sequence). If there are some complex descriptions of pasta in the local culture of Italy, the server will extract the cultural elements that best reflect the taste and incorporate them into the vector, remove some cultural background information that is not very relevant to taste analysis, and thus generate a fifth text semantic vector sequence related to food.

[0123] For another example, for the first AIGC semantic vector sequence related to technology, which includes semantic vectors about the camera function of smartphones processed by the technology product function analysis and mining unit, the non-AIGC semantic attention operation of the server will optimize these vectors.

[0124] For example, in the vector of the camera function, if there is some comparative information about the image quality in different camera modes, but this information is too complex or contains some inaccurate expressions, the server will refer to the mobile phone evaluation standard database for correction. The complex image quality comparison is simplified into more intuitive evaluation indicators (such as high, medium, and low quality), and this information is accurately updated into the vector.

[0125] At the same time, for the association vector part between the camera function and other mobile phone functions (such as the impact of battery life on long-term photography, etc.), if there is redundant information, the server will clean it up and retain the most critical association information, and then generate a fifth text semantic vector sequence related to technology.

[0126] Next, the model feature label corresponding to the tourism-related text content unit sequence indicates a tourist attraction recommendation mining unit. The server loads the fifth tourism-related text semantic vector sequence into the mining unit.

[0127] The tourist attraction recommendation mining unit will conduct in-depth mining on this vector sequence again. It may adjust the strategy for recommending tourist attractions based on the updated tourist season information in the fifth text semantic vector sequence. For example, if it is found that a particular month in the peak season has the most tourists and the best weather, the mining unit will more prominently recommend tourist attractions (beaches, sunsets, etc.) in Bali in this month.

[0128] For the characteristic vector part of tourist attractions, the mining unit may further refine the recommended content based on the optimized vector information. If the beach was generally recommended before, it may now provide more personalized recommendations based on more accurate beach information (such as sand quality, surrounding facilities, etc.). After such processing, a second AIGC attention semantic vector sequence related to tourism is generated. This sequence may be the target AIGC attention semantic vector sequence in this scenario, which is directly used for subsequent integration operations; or it may still be a basic semantic vector sequence if other processing is required (although no further processing is performed in the current example).

[0129] For another example, the model feature label corresponding to the text content unit sequence related to food points to the food taste analysis mining unit. The server loads the fifth text semantic vector sequence related to food into the mining unit.

[0130] The food taste analysis mining unit reanalyzes the information in the vector. For example, based on the updated relationship vector between pasta flavors and ingredient combinations, the mining unit may discover some new flavor combinations or optimized taste influencing factors. If the revised vector indicates that a new ingredient has a more important impact on the "rich" flavor, the mining unit will strengthen the representation of this relationship in the vector.

[0131] For the vector part of the relationship between food culture background and taste, the mining unit will conduct a deeper taste analysis based on the optimized information. If the previous connection between cultural elements and taste is not clear, now the new vector information can more accurately analyze how cultural elements affect the taste characteristics of pasta. After these processes, a second AIGC attention semantic vector sequence related to food is generated, which may also be the target AIGC attention semantic vector sequence or the basic semantic vector sequence used to generate the target AIGC attention semantic vector sequence.

[0132] For another example, the model feature label corresponding to the technology-related text content unit sequence is a technology product function analysis mining unit. The server loads the fifth technology-related text semantic vector sequence into this mining unit.

[0133] The technology product function analysis mining unit will conduct further analysis based on the updated camera function vector information. For example, based on the simplified camera mode image quality evaluation vector, the mining unit can more accurately evaluate the performance of the camera function in different scenarios. If there was some ambiguity in the image quality evaluation in the night scene mode before, now a more accurate evaluation can be given based on the new vector information.

[0134] For the vector parts related to the camera function and other mobile phone functions, the mining unit will deeply analyze the mutual influence between them based on the optimized information. If the previous analysis of the relationship between battery life and long-term photography was not accurate enough, a more reasonable analysis can now be performed based on the new vector information, and the results can be integrated into the vector to generate a second AIGC attention semantic vector sequence related to technology, which may be the target AIGC attention semantic vector sequence or the basic semantic vector sequence used to generate the target AIGC attention semantic vector sequence.

[0135] In a possible implementation, the method is implemented according to an AIGC content generation model, wherein the AIGC content generation model includes a semantic tag assignment submodel, a basic text semantic encoder, an AIGC semantic attention submodel, and an AIGC content generation submodel, wherein the AIGC semantic attention submodel includes a plurality of AIGC content mining units, and the training steps of the AIGC content generation model include:

[0136] Step A110, obtaining a template text content set formed by multiple template text content data.

[0137] In this embodiment, in order to train the AIGC content generation model, the server obtains template text content data from a pre-built large-scale text database. This database contains a variety of texts, such as travel guides, food reviews, technology product introductions, and other texts in different fields.

[0138] For example, in the field of tourism, the server may select multiple travel notes from different regions (such as Paris, Tokyo, Sydney, etc.), which describe in detail the local attractions, food, culture, etc., such as "In Paris, the Eiffel Tower is a must-see attraction, the surrounding streets are full of artistic atmosphere, and tasting the local croissants is a wonderful experience."

[0139] In the field of food, we collected text content such as menu introductions of various restaurants and food reviews of diners. For example, "The salmon sushi in this sushi restaurant is extremely fresh, the rice grains are full, and it tastes very delicious when paired with special soy sauce and mustard."

[0140] In the field of science and technology, we selected review articles and product manuals of technology products such as mobile phones and computers, such as "The new mobile phone is equipped with a high-performance processor and runs at an extremely fast speed. Its camera uses advanced optical zoom technology and has excellent shooting effects."

[0141] These text contents from different fields together constitute the template text content collection, providing rich materials for subsequent model training.

[0142] Step A120, loading the template text content set into the basic text semantic encoder to generate a template first text semantic vector sequence.

[0143] In this embodiment, the basic text semantic encoder uses deep learning technology, such as a neural network model based on the Transformer architecture. This neural network model has been pre-trained on a large amount of text data and can convert the input text into a semantic vector representation.

[0144] When loading the template text content collection into the basic text semantic encoder, for each word in the travel notes, the encoder converts it into a vector representation based on the pre-trained word vector lookup table, and then processes these vectors through the multi-layer structure of the neural network. For example, for the word "Eiffel Tower", an initial word vector is obtained first, and then after calculations in the network layer, this word vector interacts with other word vectors in the sentence, and finally obtains a vector representation containing richer semantic information.

[0145] The same process is also applied to words such as "Salmon Sushi" in food reviews and "High-Performance Processor" in technology product descriptions. After each text content data in the entire template text content set is processed by the basic text semantic encoder, it is combined in sequence to generate the template first text semantic vector sequence. Each vector in this template first text semantic vector sequence contains the semantic information of the corresponding text content and is the basis for subsequent model operations.

[0146] Step A130, loading the template first text semantic vector sequence into the semantic label assignment sub-model, generating a plurality of template text content unit sequences and a model feature label corresponding to each of the template text content unit sequences.

[0147] In this embodiment, the semantic tag assignment submodel is composed of a sequence aggregation attention submodel and a tag selection submodel. The server loads the template first text semantic vector sequence into the sequence aggregation attention submodel. This submodel first analyzes each vector in the vector sequence and calculates the importance weight of each vector in the entire sequence. For example, in the vector part related to the travel notes, if the vector "Eiffel Tower" is of key significance in describing tourist attractions, then it will be assigned a higher attention weight. By calculating the attention weights of all vectors, the sequence aggregation attention submodel can obtain a re-weighted template third text semantic vector sequence. At the same time, according to the weight of each vector and its relationship with other vectors, the semantic contribution value corresponding to each template text content data can also be determined. For example, for the text content data that describes the Eiffel Tower in detail, due to its high importance, the corresponding semantic contribution value is also large.

[0148] Then, according to the semantic contribution value corresponding to each template text content data, the server determines multiple template text content unit sequences. For example, the text content data related to tourist attractions with higher semantic contribution values ​​(such as descriptions of attractions such as the Eiffel Tower and the Louvre) are combined into a tourism-related template text content unit sequence; the food-related (such as descriptions of salmon sushi and croissants) are combined into a food-related template text content unit sequence; the technology-related (such as descriptions of high-performance processors, optical zoom technology, etc.) are combined into a technology-related template text content unit sequence.

[0149] For each template text content unit sequence, the server determines the corresponding template unit semantic vector sequence in the template third text semantic vector sequence. For example, for a template text content unit sequence related to tourism, find its corresponding vector part in the template third text semantic vector sequence.

[0150] Then, this template unit semantic vector sequence is loaded into the label selection sub-model. The label selection sub-model analyzes the possibility of each AIGC content mining unit being called based on the semantic information in this vector sequence, and generates the template AIGC content mining distribution. For example, for a tourism-related template unit semantic vector sequence, the calling confidence of the tourist attraction recommendation mining unit may be higher, and the calling confidence of the tourism culture mining unit may be lower.

[0151] Finally, according to the template AIGC content mining distribution, the model feature label corresponding to the template text content unit sequence is determined. For example, if the calling confidence of the tourist attraction recommendation mining unit is the highest, then the model feature label corresponding to the tourism-related template text content unit sequence is the label related to the tourist attraction recommendation mining unit.

[0152] Step A140, for each of the template text content unit sequences, perform AIGC semantic attention on the template first text semantic vector sequence according to the AIGC content mining unit corresponding to the corresponding model feature label in the AIGC semantic attention sub-model, and generate a template AIGC attention semantic vector sequence corresponding to the template text content unit sequence.

[0153] In this embodiment, the model feature label corresponding to the tourism-related template text content unit sequence is a label related to the tourist attraction recommendation mining unit. The server performs AIGC semantic attention on the template first text semantic vector sequence according to the tourist attraction recommendation mining unit in the AIGC semantic attention sub-model.

[0154] The tourist attraction recommendation mining unit has a special neural network structure inside to identify semantic information related to tourist attraction recommendations. It searches for vectors related to tourist attractions (such as the Eiffel Tower, the Louvre, etc.) in the first text semantic vector sequence of the template and processes these vectors according to the predefined tourist attraction recommendation rules. For example, if the rule stipulates that popularity and tourist reviews are important factors, the mining unit will enhance the vector part representing the popularity of the attraction and the tourist reviews to highlight the importance of this information in the entire vector sequence.

[0155] After being processed by the tourist attraction recommendation mining unit, a tourism-related template AIGC attention semantic vector sequence is generated. The vectors in this sequence are more focused on the semantic information related to tourist attraction recommendations.

[0156] Step A150, integrating the template AIGC focus semantic vector sequences corresponding to each of the template text content unit sequences, to generate a template second text semantic vector sequence.

[0157] For example, the model feature label corresponding to the food-related template text content unit sequence points to the food taste analysis mining unit. The food taste analysis mining unit in the AIGC semantic attention sub-model will operate on the template first text semantic vector sequence.

[0158] This mining unit will focus on vectors related to the taste of food (such as the freshness of salmon sushi, the texture of croissants, etc.). It may adjust these vectors based on knowledge in the field of food. For example, if the taste of a certain food is closely related to the freshness of a specific ingredient, the mining unit will enhance the part of the vector representing the freshness of the ingredient to highlight its impact on the taste, thereby generating a template AIGC attention semantic vector sequence related to food.

[0159] For the technology-related template text content unit sequence, the corresponding model feature label is the label related to the technology product function analysis mining unit. The technology product function analysis mining unit will perform AIGC semantic attention on the template's first text semantic vector sequence.

[0160] This mining unit will focus on vectors related to the functions of technology products (such as the computing speed of high-performance processors, the zoom capability of optical zoom technology, etc.). It will optimize these vectors according to the performance evaluation criteria of technology products. For example, if the computing speed is a key indicator for measuring processor performance, the mining unit will strengthen the vector part representing the computing speed of the processor, and then generate a technology-related template AIGC attention semantic vector sequence.

[0161] In this embodiment, the server integrates the template AIGC attention semantic vector sequences of different fields such as tourism, food, and technology. Assume that the template AIGC attention semantic vector sequence related to tourism is Vt1, the food-related is Vt2, and the technology-related is Vt3. The server can integrate by weighted summation. For example, according to the importance of different fields in the entire template text content set or the pre-set weights, assuming that the weight of the tourism field is 0.4, the food field is 0.3, and the technology field is 0.3, then the generated template second text semantic vector sequence Vt = 0.4*Vt1+0.3*Vt2+0.3*Vt3. This integrated template second text semantic vector sequence contains semantic information from various fields after AIGC semantic attention.

[0162] Step A160, loading the template second text semantic vector sequence into the AIGC content generation sub-model to generate a template AIGC content generation result.

[0163] In this embodiment, the AIGC content generation sub-model is based on the Transformer architecture or other neural network structures suitable for text generation. It has been pre-trained on a large amount of text data and can generate corresponding text content based on the input semantic vector sequence.

[0164] Then, the template second text semantic vector sequence is loaded into the AIGC content generation sub-model. The model gradually generates text content by predicting the probability distribution of the next word based on the semantic information in the vector sequence. For example, if the template second text semantic vector sequence contains semantic information related to travel, food, and technology, the model may generate a text that combines these elements, such as "When traveling, you can not only enjoy attractions like the Eiffel Tower, but also taste fresh salmon sushi, while using devices with high-performance processors to record beautiful moments." This generated text is the template AIGC content generation result.

[0165] Step A170, determining a training cost parameter, wherein the training cost parameter includes a first error characterizing a characteristic distance between the template AIGC content generation result and the template AIGC content generation result, wherein the template AIGC content generation result is used to characterize a reference AIGC content generation result of the template text content set.

[0166] In this embodiment, the server compares the generated template AIGC content generation result with the template AIGC content generation result used as a reference (which may be manually annotated or an expected result obtained from a high-quality data set). For example, a distance measurement method such as cosine distance or edit distance (depending on the characteristics and requirements of the text content) is used to calculate the characteristic distance between the two, and this distance value is the first error. If the generated result is very similar to the reference result, the first error will be small; conversely, if there is a large difference, the first error will be large.

[0167] Step A180, based on the training cost parameters, optimize the model parameter information of the semantic label allocation submodel, the basic text semantic encoder, the AIGC semantic attention submodel and the AIGC content generation submodel.

[0168] In this embodiment, the server uses an optimization algorithm (such as stochastic gradient descent or its variants) to adjust the model parameters of the semantic label assignment submodel, the basic text semantic encoder, the AIGC semantic attention submodel, and the AIGC content generation submodel according to the calculated training cost parameters (including the first error and possible other error indicators).

[0169] For the basic text semantic encoder, if the first error is found to be large, it may mean that the encoder is not encoding certain words or semantics accurately enough. The server will fine-tune the neural network weights in the encoder based on the results of error back propagation, so that the template text content set can be more accurately converted into a semantic vector sequence in subsequent training iterations.

[0170] For the semantic label assignment sub-model, if the model feature label corresponding to a certain template text content unit sequence is inaccurate, resulting in deviations in AIGC semantic attention and final content generation, the parameters in the sequence aggregation attention sub-model and the label selection sub-model are adjusted according to the training cost parameters to improve the accuracy of label assignment.

[0171] In the AIGC semantic attention sub-model, if it is found that a certain AIGC content mining unit does not process the first text semantic vector sequence of the template well, for example, it does not correctly highlight the key semantic information, then the neural network structure parameters inside the mining unit are adjusted according to the training cost parameters to improve its ability to mine specific semantics.

[0172] For the AIGC content generation sub-model, if the generated template AIGC content generation result is significantly different from the reference result, the server will adjust the parameters in the model, such as adjusting the parameters for predicting the probability distribution of the next word, so that more expected text content can be generated in subsequent training.

[0173] Through multiple such training iterations, the model parameters are continuously adjusted so that the AIGC content generation model can generate more accurate and expected AIGC content when processing various template text content data.

[0174] In a possible implementation, the semantic tag assignment submodel includes a sequence aggregation attention submodel and a tag selection submodel, and step A130 includes:

[0175] Step S131, loading the template first text semantic vector sequence into the sequence aggregation attention sub-model, generating a template third text semantic vector sequence and a semantic contribution value corresponding to each template text content data determined based on the template third text semantic vector sequence.

[0176] Step S132, determining the plurality of template text content unit sequences according to the semantic contribution value corresponding to each of the template text content data.

[0177] Step S133: for each of the template text content unit sequences, determine a corresponding template unit semantic vector sequence in the template third text semantic vector sequence.

[0178] Step S134, loading the template unit semantic vector sequence into the tag selection sub-model to generate template AIGC content mining distribution.

[0179] Step S135, determining the model feature label corresponding to the template text content unit sequence according to the template AIGC content mining distribution.

[0180] In this embodiment, a template first text semantic vector sequence has been obtained through a basic text semantic encoder, and this sequence includes semantic vectors converted from template text content data in multiple fields such as tourism, food, and technology.

[0181] When the server loads the template first text semantic vector sequence into the sequence aggregation attention sub-model, the model first analyzes each vector. Taking the vector corresponding to the text content data in the tourism field as an example, for the vector describing tourist attractions (such as the Eiffel Tower), the model will calculate its attention weight based on its position in the entire sequence, its relationship with other vectors, and predefined rules.

[0182] Assuming that in the first text semantic vector sequence of the template, there are vectors describing the Paris tourism environment (such as street atmosphere) around the vector related to the Eiffel Tower, the sequence aggregation attention sub-model will consider the semantic association between these vectors, because the Eiffel Tower is a landmark attraction in Paris and has a semantic connection with the surrounding environment vectors. According to the closeness of this connection, the model will assign a corresponding attention weight to the Eiffel Tower vector.

[0183] For vectors in the field of food, such as the vector describing salmon sushi, the model will consider its relationship with related vectors representing the freshness of ingredients, matching sauces, etc. If the freshness vector of salmon sushi is closely related to it, this relationship will be taken into account when calculating the attention weight of the salmon sushi vector.

[0184] The same is true in the field of science and technology. For a vector describing a high-performance processor, its attention weight will be determined by considering its relationship with related vectors such as the phone's operating speed and multi-tasking capabilities.

[0185] By calculating the attention weight of each vector, the sequence aggregation attention sub-model re-weights the vectors in the template's first text semantic vector sequence to generate the template's third text semantic vector sequence. In this process, important vectors (such as description vectors of major attractions in the tourism field, description vectors of key features of specialty foods in the food field, and description vectors of core product functions in the technology field) will receive higher weights, making them more prominent in the new template's third text semantic vector sequence.

[0186] Based on the attention weight of each vector and its semantic importance in the entire sequence, the sequence aggregation attention sub-model determines the semantic contribution value corresponding to each template text content data. For example, for the text content data that describes the Eiffel Tower in detail and critically, since its corresponding vector has a higher weight in the template third text semantic vector sequence, its semantic contribution value is relatively large; while for some auxiliary descriptions (such as a small shop briefly mentioned in a Paris travel note), its semantic contribution value is relatively small.

[0187] Next, the server groups the template text content data according to the semantic contribution values ​​based on a preset threshold or grouping strategy.

[0188] For example, in the tourism field, template text content data with high semantic contribution values ​​related to tourist attractions (such as the Eiffel Tower, the Louvre, etc.) and tourism experience (such as tasting local food, watching special performances, etc.) are grouped together to form a tourism-related template text content unit sequence. For example, the detailed description of the Eiffel Tower and the description of the experience of playing around the Eiffel Tower are classified as tourism-related template text content unit sequences because their semantic contribution values ​​meet the tourism-related grouping requirements.

[0189] For another example, in the field of food, template text content data with large semantic contribution values ​​related to food dishes (such as salmon sushi, croissants, etc.) and food flavors (such as freshness, unique taste, etc.) are gathered together to form a food-related template text content unit sequence. For example, the description of the freshness, taste, and production process of salmon sushi is classified into the food-related template text content unit sequence because its semantic contribution value meets the food-related standards.

[0190] For another example, in the field of science and technology, template text content data related to science and technology products (such as high-performance processors, optical zoom technology, etc.) and product functional characteristics (such as operating speed, zoom capability, etc.) whose semantic contribution values ​​meet the requirements are grouped together to form a science and technology-related template text content unit sequence. For example, the description of the performance parameters, functional advantages, etc. of high-performance processors is determined to be a science and technology-related template text content unit sequence because its semantic contribution value meets the science and technology-related grouping standard.

[0191] Next, for the tourism-related template text content unit sequence, the server searches in the template third text semantic vector sequence. Since the vectors in the template third text semantic vector sequence correspond to the original template text content data, the server determines the corresponding vector part in the template third text semantic vector sequence according to the position of the text content data in the tourism-related template text content unit sequence in the original template text content set. For example, the tourism-related template text content unit sequence contains a description of the Eiffel Tower and the surrounding tourism experience, and the server finds the vectors corresponding to these descriptions in the template third text semantic vector sequence, and these vectors together constitute the tourism-related template unit semantic vector sequence.

[0192] For another example, for a food-related template text content unit sequence, the corresponding vector is determined in the template third text semantic vector sequence based on the position of the food-related content in the original template text content set. For example, if the food-related template text content unit sequence contains a description of salmon sushi, the server finds the vector part related to salmon sushi in the template third text semantic vector sequence, and this part of the vector is the food-related template unit semantic vector sequence.

[0193] For another example, in terms of the technology-related template text content unit sequence, the corresponding vector is found in the template third text semantic vector sequence according to the position of the technology-related content in the original template text content set. For example, the technology-related template text content unit sequence contains a description of a high-performance processor, and the server locates the vector related to the high-performance processor in the template third text semantic vector sequence, and these vectors constitute the technology-related template unit semantic vector sequence.

[0194] On this basis, taking the tourism-related template unit semantic vector sequence as an example, when the server loads it into the label selection sub-model, this sub-model has a neural network structure pre-trained based on a large amount of tourism-related data.

[0195] The model will analyze the vectors in the template unit semantic vector sequence and identify the semantic features related to different AIGC content mining units. For example, if the vector contains more semantic information related to tourist attractions, the model will calculate the possibility of calling the tourist attraction recommendation mining unit according to predefined rules; if there is semantic information related to tourism culture, the model will also calculate the possibility of calling the tourism culture mining unit, etc.

[0196] After analyzing the template unit semantic vector sequence, the label selection sub-model calculates the confidence of each AIGC content mining unit being called, and these confidences are combined to form the template AIGC content mining distribution. For example, for the tourism-related template unit semantic vector sequence, the confidence of the tourist attraction recommendation mining unit may be 0.6, the confidence of the tourism culture mining unit is 0.3, and the confidence of the travel itinerary planning mining unit is 0.1. This is the tourism-related template AIGC content mining distribution. The template unit semantic vector sequences related to food and technology will also undergo similar processing to generate their corresponding template AIGC content mining distributions.

[0197] Next, for the tourism-related template text content unit sequence, since the confidence of the tourist attraction recommendation mining unit in its template AIGC content mining distribution is the highest (e.g., 0.6), the server determines that the model feature label corresponding to the template text content unit sequence is a label related to the tourist attraction recommendation mining unit. This label contains identification information related to the tourist attraction recommendation mining unit, so that subsequent operations can accurately identify and call the relevant mining unit.

[0198] For another example, in a sequence of food-related template text content units, if the food taste analysis mining unit has the highest confidence in the template AIGC content mining distribution, then its corresponding model feature label is the label related to the food taste analysis mining unit, and this label carries the identification information of the food taste analysis mining unit.

[0199] For another example, for a sequence of technology-related template text content units, if the confidence of the technology product function analysis mining unit is the highest, its corresponding model feature label is a label related to the technology product function analysis mining unit, which contains the identification information of the technology product function analysis mining unit. In this way, each template text content unit sequence has a corresponding model feature label, which provides guidance for subsequent AIGC semantic attention and content generation operations.

[0200] In a possible implementation manner, the training cost parameter further includes a second error, and the method further includes:

[0201] Step B110, obtaining a target training scale, wherein the target training scale is used to represent a reference training scale of the template text content unit sequence.

[0202] Step B120, performing semantic contribution fusion on the semantic contribution value corresponding to each of the template text content data to generate a semantic contribution fusion result.

[0203] Step B130, determining the second error according to the error between the semantic contribution fusion result and the target training scale.

[0204] In this embodiment, during the training of the AIGC content generation model, the server pre-sets a reference training scale for the template text content unit sequence in each field (such as tourism, food, technology, etc.). This reference training scale is determined based on factors such as the expected learning ability of the model, the diversity of the data, and the final performance requirements.

[0205] For example, for the template text content unit sequence in the tourism field, based on past training experience and the need to cover tourism-related knowledge, the target training scale is set to 1,000 valid semantic units (the semantic unit here can be the semantic representation corresponding to the processed tourism-related text fragment). For the food field, the target training scale is set to 800 semantic units, because the semantic features in the food field may be more concentrated than those in the tourism field, and the main semantic information can be covered without too much text. For the science and technology field, because its technical concepts are more complex and diverse, the target training scale is set to 1,200 semantic units.

[0206] In the tourism field, the server has determined the semantic contribution value for tourism-related template text content data. For example, for the detailed text content data describing the Eiffel Tower, its semantic contribution value is 0.8; for the text content data describing the atmosphere of the small streets around Paris, the semantic contribution value is 0.3, and so on.

[0207] The server merges the semantic contribution values ​​of all tourism-related template text content data. A simple fusion method is to sum them. Assuming that there are 100 template text content data in the tourism field, add up these 100 semantic contribution values, and the sum obtained is the semantic contribution fusion result in the tourism field. For example, the calculated sum is 50.0.

[0208] In the field of food, food-related template text content data also have their own semantic contribution values. For example, the semantic contribution value of text content data describing the freshness of salmon sushi is 0.7, and the semantic contribution value of text content data describing the taste of croissant bread is 0.6.

[0209] Similarly, the semantic contribution values ​​of all food-related template text content data are fused by summing them up. Assuming that there are 80 template text content data in the food field, the calculated semantic contribution fusion result is 45.0.

[0210] In the field of science and technology, for technology-related template text content data, such as text content data describing the computing speed of high-performance processors, the semantic contribution value is 0.9, and text content data describing the zoom capability of optical zoom technology has a semantic contribution value of 0.8, etc.

[0211] The semantic contribution values ​​of all template text content data in the field of science and technology (assuming there are 120 in total) are fused in a summation manner, and the resulting semantic contribution fusion result is 60.0.

[0212] Therefore, the semantic contribution fusion result in the tourism field is 50.0, and the target training scale is 1000 semantic units. To calculate the error between the two, a simple difference calculation can be used here, that is, the second error = |50.0-1000| = 950. This error value reflects the gap between the current semantic contribution numerical fusion result and the expected target training scale. The semantic contribution fusion result in the food field is 45.0, and the target training scale is 800 semantic units. The calculated second error is |45.0-800| = 755. The semantic contribution fusion result in the science and technology field is 60.0, and the target training scale is 1200 semantic units. The calculated second error is |60.0-1200| = 1140.

[0213] In a possible implementation, the training cost parameter further includes a third error, the template AIGC content mining distribution includes a call confidence corresponding to a model feature label corresponding to a template text content unit sequence, and the method further includes:

[0214] Step C110, for each model feature label, determine corresponding weight information, wherein the weight information is used to represent the ratio of the size of the template text content data corresponding to the model feature label to the size of the global template text content data in the template text content set.

[0215] Step C120, determining the model call cost parameter corresponding to the model feature label according to the product of the weight information and the call confidence corresponding to the model feature label.

[0216] Step C130, determining the third error according to the model calling cost parameters corresponding to each of the model feature labels.

[0217] In this embodiment, it is assumed that there are 1000 template text content data in the tourism-related template text content set (the total number assumed here is different from the target training scale mentioned above for the convenience of calculation), of which there are 300 template text content data corresponding to the tourist attraction recommendation mining unit. Then the weight information of the model feature label corresponding to the tourist attraction recommendation mining unit = 300 / 1000 = 0.3. If there are 200 template text content data corresponding to the tourism culture mining unit, its weight information = 200 / 1000 = 0.2. If there are 100 template text content data corresponding to the travel itinerary planning mining unit, its weight information = 100 / 1000 = 0.1.

[0218] Assume that there are 800 template text content data in the food-related template text content set, of which 300 template text content data correspond to the food taste analysis mining unit. The weight information of the model feature label corresponding to the food taste analysis mining unit = 300 / 800 = 0.375. If there are 200 template text content data corresponding to the food recipe mining unit, its weight information = 200 / 800 = 0.25. If there are 100 template text content data corresponding to the food restaurant recommendation mining unit, its weight information = 100 / 800 = 0.125.

[0219] Assume that there are 1200 template text content data in the technology-related template text content set, of which 500 are template text content data corresponding to the technology product function analysis mining unit. The weight information of the model feature label corresponding to the technology product function analysis mining unit = 500 / 1200≈0.417. If there are 300 template text content data corresponding to the technology product evaluation mining unit, its weight information = 300 / 1200 = 0.25. If there are 200 template text content data corresponding to the technology product usage scenario mining unit, its weight information = 200 / 1200≈0.167.

[0220] Next, the calling confidence of the tourist attraction recommendation mining unit is 0.6 (determined earlier when generating the template AIGC content mining distribution), and its weight information is 0.3, so the model calling cost parameter corresponding to the tourist attraction recommendation mining unit = 0.3 × 0.6 = 0.18. The calling confidence of the tourism culture mining unit is 0.3, the weight information is 0.2, and its model calling cost parameter = 0.2 × 0.3 = 0.06. The calling confidence of the travel itinerary planning mining unit is 0.1, the weight information is 0.1, and its model calling cost parameter = 0.1 × 0.1 = 0.01.

[0221] The calling confidence of the food taste analysis mining unit is 0.5, the weight information is 0.375, and its model calling cost parameter = 0.375 × 0.5 = 0.1875. The calling confidence of the food recipe mining unit is 0.3, the weight information is 0.25, and its model calling cost parameter = 0.25 × 0.3 = 0.075. The calling confidence of the food restaurant recommendation mining unit is 0.2, the weight information is 0.125, and its model calling cost parameter = 0.125 × 0.2 = 0.025.

[0222] The call confidence of the technology product function analysis mining unit is 0.7, the weight information is 0.417, and its model call cost parameter = 0.417 × 0.7 ≈ 0.292. The call confidence of the technology product evaluation mining unit is 0.2, the weight information is 0.25, and its model call cost parameter = 0.25 × 0.2 = 0.05. The call confidence of the technology product usage scenario mining unit is 0.1, the weight information is 0.167, and its model call cost parameter = 0.167 × 0.1 = 0.0167.

[0223] One way to determine the third error is to calculate the standard deviation of all model call cost parameters. First calculate the mean: (0.18+0.06+0.01) / 3=0.083. Then calculate the standard deviation, first calculate the square of the difference between each model call cost parameter and the mean:

[0224] (0.18-0.083)2=0.009409

[0225] (0.06-0.083)2=0.000529

[0226] (0.01-0.083)2=0.005329

[0227] Add these squared sums and divide by the number of model feature labels (here 3), and take the square root: Third error = √((0.009409+0.000529+0.005329) / 3)≈0.073.

[0228] In the field of food, the average value is: (0.1875+0.075+0.025) / 3=0.09583.

[0229] Compute the square of the difference:

[0230] (0.1875-0.09583)2=0.008372

[0231] (0.075-0.09583)2=0.000470

[0232] (0.025-0.09583)2=0.004999

[0233] The third error = √((0.008372+0.000470+0.004999) / 3)≈0.065.

[0234] In the field of science and technology, the average value is: (0.292+0.05+0.0167) / 3=0.11957.

[0235] Compute the square of the difference:

[0236] (0.292-0.11957)2=0.029792

[0237] (0.05-0.11957)2=0.004839

[0238] (0.0167-0.11957)2=0.010542

[0239] The third error = √((0.292+0.05+0.0167) / 3)≈0.124.

[0240] These second errors and third errors, together with the first error mentioned earlier (the error that characterizes the characteristic distance between the template AIGC content generation result and the template AIGC content generation result), constitute the training cost parameters. The server can optimize the model parameter information of the semantic label allocation sub-model, the basic text semantic encoder, the AIGC semantic attention sub-model and the AIGC content generation sub-model based on these training cost parameters, thereby improving the performance of the entire AIGC content generation model.

[0241] Figure 2 The hardware structure of the AIGC content generation system 100 based on content big data monitoring for implementing the above-mentioned AIGC content generation method based on content big data monitoring provided by an embodiment of the present invention is shown as follows: Figure 2 As shown, the AIGC content generation system 100 based on content big data monitoring may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0242] In one possible design, the AIGC content generation system 100 based on content big data monitoring can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the AIGC content generation system 100 based on content big data monitoring can be implemented on an AIGC content generation system based on content big data monitoring.

[0243] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the AIGC content generation system 100 based on content big data monitoring uses to execute or use to complete the exemplary method described in the present invention.

[0244] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the AIGC content generation method based on content big data monitoring in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0245] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned AIGC content generation system 100 based on content big data monitoring. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0246] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned AIGC content generation method based on content big data monitoring is implemented.

[0247] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, various features are sometimes combined into one embodiment, drawing or description thereof.

Claims

1. A method for generating AIGC content based on content big data monitoring, characterized in that: The method comprises: Obtain a target text content set formed by multiple text content data from any target user's content behavior big data; Performing text semantic encoding representation on the target text content set to generate a first text semantic vector sequence; According to the first text semantic vector sequence, the text content data in the target text content set are aggregated to generate a plurality of text content unit sequences and a model feature label corresponding to each of the text content unit sequences, each of the text content unit sequences includes a plurality of the text content data, and the model feature label is a feature label corresponding to an AIGC content mining unit that performs AIGC semantic attention; For each of the text content unit sequences, an AIGC content mining unit indicated by a model feature label corresponding to the text content unit sequence performs AIGC semantic attention on the first text semantic vector sequence to generate a target AIGC attention semantic vector sequence corresponding to the text content unit sequence; Integrate the target AIGC attention semantic vector sequences corresponding to each of the text content unit sequences to generate a second text semantic vector sequence corresponding to the target text content set; Perform AIGC content generation on the second text semantic vector sequence and output the AIGC content generation result.

2. The AIGC content generation method based on content big data monitoring according to claim 1 is characterized in that: The method of aggregating text content data in the target text content set according to the first text semantic vector sequence to generate a plurality of text content unit sequences and a model feature label corresponding to each of the text content unit sequences includes: Extracting attention semantic elements from the first text semantic vector sequence to generate a third text semantic vector sequence; estimating a semantic contribution value corresponding to each of the text content data according to the third text semantic vector sequence; By performing semantic contribution fusion on each of the semantic contribution values, at least one attention text content data is determined, where the attention text content data is text content data whose corresponding semantic contribution fusion value meets the set conditions; Based on each of the attention text content data, generate the plurality of text content unit sequences; The step of estimating the semantic contribution value corresponding to each of the text content data according to the third text semantic vector sequence includes: Perform semantic element analysis on each text semantic vector in the third text semantic vector sequence, regard each text semantic vector as a semantic unit, identify the semantic elements contained in the text semantic vector by comparing it with a predefined semantic concept template library, and obtain a first target text semantic vector with preliminary semantic element markings corresponding to each text content data, wherein the semantic element is a basic unit for constructing semantic understanding. Analyze the preliminary semantic relationship between the first target text semantic vectors according to the similarity and relevance of the marked preliminary semantic elements, and generate a second target text semantic vector corresponding to each text content data with preliminary semantic relationship marks, wherein the preliminary semantic relationship includes a containment relationship, a parallel relationship, and a causal relationship; Taking each of the second target text semantic vectors as a node, a semantic relationship network is constructed according to the semantic relationship between the second target text semantic vectors, wherein if there is a semantic relationship between two second target text semantic vectors, an edge is established between the corresponding nodes; Analyze the role of each node in the semantic relationship network according to the node connectivity and edge weight distribution of each node, and generate a node role classification result, wherein the role is a semantic core node, an edge node or an intermediate transition node; Evaluate the preliminary contribution of each text content data to the global semantics according to the node role classification result, and generate a preliminary semantic contribution value corresponding to each text content data; Analyze the semantic paths from each node to other nodes, calculate the scope and degree of influence of each node on the semantics of other nodes through the semantic paths, and adjust the preliminary semantic contribution value corresponding to each text content data according to the scope and degree of influence of each node on the semantics of other nodes through the semantic paths, and generate the first semantic contribution value corresponding to each text content data after the semantic path influence adjustment; Analyze the adaptation state parameters of each of the text content data in the corresponding local context, and adjust the first semantic contribution value corresponding to each of the text content data after the semantic path influence adjustment according to the adaptation state parameters, generate the second semantic contribution value corresponding to each of the text content data after the local context adaptation adjustment, and output the second semantic contribution value as the semantic contribution numerical value.

3. The AIGC content generation method based on content big data monitoring according to claim 2 is characterized in that: The step of aggregating text content data in the target text content set based on the first text semantic vector sequence to generate a plurality of text content unit sequences and a model feature label corresponding to each of the text content unit sequences further includes: For each of the text content unit sequences, determining a corresponding unit semantic vector sequence in the third text semantic vector sequence; Determining a target AIGC content mining distribution according to the unit semantic vector sequence, wherein the target AIGC content mining distribution indicates a confidence level of each AIGC content mining unit being called; According to the target AIGC content mining distribution, the model feature label corresponding to the text content unit sequence is determined.

4. The AIGC content generation method based on content big data monitoring according to any one of claims 1 to 3, characterized in that: The step of performing AIGC semantic attention on the first text semantic vector sequence according to the AIGC content mining unit indicated by the model feature label corresponding to the text content unit sequence for each of the text content unit sequences to generate a target AIGC attention semantic vector sequence corresponding to the text content unit sequence includes: Performing non-AIGC semantic attention on the first text semantic vector sequence to generate a fourth text semantic vector sequence; For each of the text content unit sequences, the fourth text semantic vector sequence is loaded into the AIGC content mining unit indicated by the model feature label corresponding to the text content unit sequence to generate a corresponding first AIGC focus semantic vector sequence, wherein the first AIGC focus semantic vector sequence is the target AIGC focus semantic vector sequence, or the first AIGC focus semantic vector sequence is a basic semantic vector sequence used to generate the target AIGC focus semantic vector sequence.

5. The AIGC content generation method based on content big data monitoring according to claim 4 is characterized in that: In the case where the first AIGC attention semantic vector sequence is a basic semantic vector sequence for generating the target AIGC attention semantic vector sequence, the AIGC content mining unit indicating the model feature label corresponding to the text content unit sequence performs AIGC semantic attention on the first text semantic vector sequence for each of the text content unit sequences to generate the target AIGC attention semantic vector sequence corresponding to the text content unit sequence, further comprising: Performing non-AIGC semantic attention on the first AIGC attention semantic vector sequences respectively corresponding to the text content unit sequences to generate a fifth text semantic vector sequence; For each of the text content unit sequences, the fifth text semantic vector sequence is loaded into the AIGC content mining unit indicated by the model feature label corresponding to the text content unit sequence to generate a corresponding second AIGC focus semantic vector sequence, wherein the second AIGC focus semantic vector sequence is the target AIGC focus semantic vector sequence, or the second AIGC focus semantic vector sequence is a basic semantic vector sequence used to generate the target AIGC focus semantic vector sequence.

6. The AIGC content generation method based on content big data monitoring according to claim 1 is characterized in that: The method is implemented according to the AIGC content generation model, which includes a semantic tag assignment sub-model, a basic text semantic encoder, an AIGC semantic attention sub-model and an AIGC content generation sub-model, wherein the AIGC semantic attention sub-model includes a plurality of AIGC content mining units, and the training steps of the AIGC content generation model include: Obtain a template text content set formed by multiple template text content data; Loading the template text content set into the basic text semantic encoder to generate a template first text semantic vector sequence; Loading the template first text semantic vector sequence into the semantic label assignment submodel to generate a plurality of template text content unit sequences and a model feature label corresponding to each of the template text content unit sequences; For each of the template text content unit sequences, AIGC semantic attention is performed on the template first text semantic vector sequence according to the AIGC content mining unit corresponding to the corresponding model feature label in the AIGC semantic attention sub-model to generate a template AIGC attention semantic vector sequence corresponding to the template text content unit sequence; Integrate the template AIGC attention semantic vector sequences corresponding to each of the template text content unit sequences to generate a template second text semantic vector sequence; Loading the template second text semantic vector sequence into the AIGC content generation sub-model to generate a template AIGC content generation result; Determine a training cost parameter, the training cost parameter including a first error characterizing a characteristic distance between the template AIGC content generation result and a template AIGC content generation result, the template AIGC content generation result being used to characterize a reference AIGC content generation result of the template text content set; According to the training cost parameters, the model parameter information of the semantic label allocation submodel, the basic text semantic encoder, the AIGC semantic attention submodel and the AIGC content generation submodel is optimized.

7. The AIGC content generation method based on content big data monitoring according to claim 6 is characterized in that: The semantic label assignment submodel includes a sequence aggregation attention submodel and a label selection submodel, and the loading of the template first text semantic vector sequence into the semantic label assignment submodel to generate a plurality of template text content unit sequences and a model feature label corresponding to each of the template text content unit sequences includes: Loading the template first text semantic vector sequence into the sequence aggregation attention sub-model, generating a template third text semantic vector sequence and a semantic contribution value corresponding to each template text content data determined according to the template third text semantic vector sequence; Determining the plurality of template text content unit sequences according to the semantic contribution value corresponding to each of the template text content data; For each of the template text content unit sequences, determining a corresponding template unit semantic vector sequence in the template third text semantic vector sequence; Loading the template unit semantic vector sequence into the tag selection sub-model to generate template AIGC content mining distribution; According to the template AIGC content mining distribution, the model feature label corresponding to the template text content unit sequence is determined.

8. The AIGC content generation method based on content big data monitoring according to claim 7 is characterized in that: The training cost parameter also includes a second error, and the method further includes: Acquire a target training scale, where the target training scale is used to represent a reference training scale of the template text content unit sequence; Performing semantic contribution fusion on the semantic contribution value corresponding to each of the template text content data to generate a semantic contribution fusion result; The second error is determined according to an error between the semantic contribution fusion result and the target training scale.

9. The AIGC content generation method based on content big data monitoring according to claim 8 is characterized in that: The training cost parameter also includes a third error, the template AIGC content mining distribution includes a call confidence corresponding to the model feature label corresponding to the template text content unit sequence, and the method further includes: For each model feature label, determining corresponding weight information, the weight information being used to indicate the ratio of the size of the template text content data corresponding to the model feature label to the size of the global template text content data in the template text content set; Determine a model call cost parameter corresponding to the model feature label according to the product of the weight information and the call confidence corresponding to the model feature label; The third error is determined according to the model calling cost parameters corresponding to each of the model feature labels.

10. An AIGC content generation system based on content big data monitoring, characterized in that: The AIGC content generation system based on content big data monitoring includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AIGC content generation method based on content big data monitoring as described in any one of claims 1 to 9 above.

Images